You've been with the professors
And they've all liked your looks
With great lawyers you have
Discussed lepers and crooks
You've been through all of
F. Scott Fitzgerald's books
You're very well read
It's well known
But something is happening here
And you don't know what it is
Do you, Mister Jones?
Ballard of a Thin Man, Bob Dylan, 1965.
It has been a while since I wrote here :) Now I have a bit of time and a lot of ideas so time to dump one here :)
I'll write (once more) about AI and scientific research. Earlier I talked about Knuth's experience report using Claude (see Knuth's Claude's Cycles bellow). On LinkedIn I also posted a few thoughts on the matter, for example, I proposed a new track for conferences called "AI co-authored" for papers that make heavy use of AI in their production, regardless of its phase: proof-reading, idea generation, idea testing, etc.
Now I want to talk about what has been called auto-science, vibe-math and so on. It has gained traction, probably because users of AI systems are producing proofs in Lean, a proof assistant that is fairly recent, the first version was released in 2013 while other proof assistants had been around for much longer, e.g. Coq (now Rocq) has been around since 1989. With the backing of tech giants like Microsoft and Amazon, Lean nowadays dominate the vibe-coding proving market.
So, what do I understand by auto-science? I call auto-science to the following process: A human (or AI agent) proposes a theoretical problem (CS or math) to an AI agent. The agent works the problem and produces a proof or counter-example. To me, for this process to be strictly called auto-science the human must trust the agent's output to be right (hence the necessity that it must produce a proof or a counter-example). If the human needs to review the agent's output then I would call it AI-assisted-science. It is clear that in both cases (auto-science and AI-assisted-science) the AI agent lifts the greatest weight. Now, let's get into it shall we?
The academic context
If you publish papers probably you work at a university, perhaps as a professor (in any kind of flavor) a post-doc or a PhD student; I do not consider master and bachelor students because many of them do not get paid for publishing. Publishing papers shows you are at the top of the chain. It also means you depend on publishing for the greatest part. This is no secret, it is common that a post-doc or professorship position requiring something like strong publication record. Or if you are looking for a PhD you may see something like our group has a solid track record at top venues like... The publishing dependency is on two dimensions, the first is the publication output itself and the second the impact it has on the community.
It is not enough to publish, your papers must have readers, similar to online social media content creators: the more views and reactions with their content, the more they get paid. In a social media post, tracking reactions and views is conceptually trivial, you just need a couple of atomic counters. For papers that's a bit different, it doesn't just matter that your paper is cited it also matters who cite it; still we can draw a parallel with online social media, it is not the same to be reposted by a random person than by a celebrity.
The average paper builds upon existing knowledge and does so following the broad structure: we study problem X, also studied by [A,B,C..] with the lenses developed by [Z]. Ideally every claim in a paper must be able to be traced back following the references (up to honest paper availability). This check is 'relatively easy' I have done it since my bachelor days and I was not a particularly bright student, so, most students should be able to do it reasonably well.
For the most part, for an academic, their publications is their reputation and that has a direct impact on employment security and growth opportunities.
The tech context
Thanks to Claude now I can work on a small domain-specific language for CGR (see Undusting CGR in the Random CS section if you want to learn what CGR is). That was always the main goal of CGR, to have a language resembling SETL but specialized for computational geometry. However, I have always been super lazy to write parsers and lexers, those two just fill me with stress instead of joy, so I avoid them for the most part. With Claude I was able to obtain a parser and a lexer for a prototype DSL in Objective-C; that is pretty cool. I gave Claude a few examples, this is one of them:
Now, take any book or any paper on AI that you want and implement what you find in there. If you would like to start from scratch, you could start by implementing a neural network. Soon enough you will realize that you now need two inputs: the network weights and the prompt. If you have no network weights, you just have a few lines of code that will give you either constant answers to all your prompts or random answers every time. So, you don't only need the problem (prompt) but also you require to provide a computing context (network weights). The problems are easy to come by, the computing context not so much.
It starts with a training data that is transformed to vectors of numbers and those vectors are used to define the network weights. Without the training data, it is almost sure that you have an useless program. So, let's assume you have training data that is successfully translated into network weights that yield a good performing network. This means that the network can reasonably generalize (provide good approximations for the problem context).
With this in mind, let me propose the Programming Poet hypothesis: If a network is solely trained on poems the probability that it writes correct computer programs for any algorithmic problem in any programming language is 0.
If the Programming Poet hypothesis is true, then to obtain good results in any field, the network must be trained on material relevant to the field (no, it doesn't count that you regard your favorite proof as a beautiful poem!). Then, let us assume that we have access to a network trained on all the CS (math) papers published from 1980 to date and call it The Oracle. Why did I choose 1980? well, because if you started to publish when you were around 20 years old and your first paper was published in 1980, then by now your employment problems should be non-existent and your reputation well-established, so you don't need to produce novel results (please note that I talk about the need not the ability).
The experiment
Now the fun part, the experiment. Let us assume that all academics have access to The Oracle all can question it and read its answers. At this point the playing field is leveled: all academics have access to the same network; what will set them apart is the questions they ask, the prompts. If an academic asks a prompt the network was trained on, the answer must be close to the ground-truth. If an academic asks a prompt the network was not trained on, the answer is expected to be somewhat reasonable. If The Oracle can generalize, then it is reasonable to assume it will produce new results.
To check The Oracle's work, we must rely on the academic's factual knowledge: for the academic the problem could be novel, maybe unaware that someone else has already solved it or considered it, without any means to track The Oracle's work, from the academic's point of view, the answer is novel, The Oracle is producing novel results.
There are famous problems everyone 'knows' or know they exist, for those problems there is much training data: different approaches, sub-results, related problems, etc. Again an academic prompting these problems and getting a good result could reasonably conclude the network is producing new results, again the answer is novel.
But is it? if the academic prompted a problem already in the training data, by definition the problem is not novel. If the academic prompted a problem community regarded as open and gets a good result then it must be checked for 'novelty'. If the solution is actually novel, using insight no one has proposed before or the solution is mix of already present sub-solutions that no one had put together. By academic standards mixing already present sub-solutions is considered novel but not as novel as having an insight no one else has proposed. Again, judging the output of The Oracle falls back to the academic's factual knowledge base. Put into my bachelor's thesis adviser words: "I do not know everything that is being published" and he is regarded as a leading researcher in Computational Geometry.
Current models under-perform at tracing back their outputs to the training data, we have seen a lot of problems with models fabricating references: papers that do not exist. This makes the evaluation of novelty a real headache because you can not track the roots of the solution. The fallback option: If the proof checks then call it solved and move on.
As good as your last paper
If an academic's worth is in their reputation that comes from not only published papers but their impact, what will happen if we accept The Oracle's outputs and we can't trace back (without any doubts) it's reasoning? Well, the big winners are the ones that get the result first and publish it first, because their names are the ones that will be on the paper and headlines saying they solved, with AI, a long-standing problem. All the giant's shoulders upon which that solution stands will not be fully known, because we can't trace back The Oracle's reasoning. Furthermore, that paper will be part of the training set of The Oracle and live as a standalone paper as long as someone remembers it, hence, academics' reputation will be as good as their last paper.
We have already seen signs of this, it seems now that you broadcast your Lean proof to claim the prize, then you can write the paper. It is a race of prompts, a race of problems.
I suspect soon there will be databases of solved problems that you can automatically cross-check to see if someone else has already produced a Lean proof. Lean has already moved in this direction with their ambition to "Formalize CS".
We already have seen a massive surge of submitted papers to top conferences and the pressure on peer-review is building up, if even before this 'revolution' peer-review was under pressure, these days it seems it is on the brink of breaking.
Perhaps escaping this race by a lateral move could explain why recently I started to see colleagues calling for producing 'disposable proofs'. A 'disposable proof' is defined as a proof that no one cares for but that along with other 'disposable proofs' are generalized in a 'general theory'. The idea is to find 'the shape' of the underlying mathematics. So, say good-bye to your favorite equation in physics just because there is something called 'general theory of relativity'. I can't help but to see this as a profound lack of respect to the field, rounding up results just to claim you have 'the global' view and then dispose them. Like if Kepler burned Tycho Brahe's observations after he derived his laws of planetary motion. Again, it seems to be about claiming the result as soon as possible, and how the foundational data came to be? who cares, it grows on (the ashes of) trees!
The AI Purge
By now, we are familiar with headlines like "The end of programming", "The AI revolution", and many others full of optimism. In his essay "The Fate of Empires" Sir John Glubb writes: "... decadence is the disintegration of a system, not of its individual members. The habits of the members of the community have been corrupted by the enjoyment of too much money and too much power for too long a period. The result has been, in the framework of their national life, to make them selfish and idle. A community of selfish and idle people declines, internal quarrels develop in the division of its dwindling wealth, and pessimism follows..." Are we heading towards a new dawn in science or to its decline, towards the AI purge?
I believe the underlying goal of colleagues lobbying for 'disposable auto-science' is to bump the complexity by gathering enough data current models would struggle to interpret and interpret it in order to stay relevant. In the present context, as a problem, we can not deny it is relevant and novel, moreover, it is a generalization of what experts in the field are able to do (and do extremely well): literature review. Any good literature review paper classifies papers in categories discussing similarities and differences, spot open and partially addressed problems, basically is step 0 when you want to do research at any serious level. What sets apart your literature review when you are a student and one published in a journal is the scope and the depth. It is expected that an established expert provides a deeper and richer insight than a beginner. This seems to be a safe bet, you use your knowledge and skills to classify many data points (results, proofs) into a coherent model and then publish that model; of course you need the money first to create those data points and that's where the university's/project's budget comes into play, this however, isn't new to an academic. This approach is a much, much safer bet than sitting down and try to come up with a problem the model can not solve *and* that appeals to your peers. By construction, the problems you find in your 'disposable proofs' review, are open with respect to that model and hence relevant to the current standards. Someone would fill some holes and others will be regarded as 'irrelevant' following typical 'novelty/interesting' scientific canons.
But the interesting insight is that the drive of the proposal is to stay relevant by spotting and plugging holes of a technology that would not evolve beyond a hard threshold if we accept the Programming Poet hypothesis. In short, we are heading towards a "selfish and idle" community. I'm not saying these problems and analyses will be easy or trivial, surely some will be quite challenging, but if the Programming Poet hypothesis is true, then we are heading to a stale; there is of course, the possibility that in the meantime a breakthrough is made and we face new smarter models and we throw away all the advances made so far or that the generalization of the observed data gives way to a new insight that seed the current models and open new "frontiers". The stale state may not be reached in our generation but in the coming ones, it appears that we are not that interested in the future generations or in the pursuit of knowledge as a human endeavor because we are focused on staying relevant with respect to artificial oracles regardless of the cost.
That being said, it's hard to argue against sealing the holes in current models, beyond the usual arguments of resource consumption and accountability.
If only...
I love reading papers, I think there is something magical on how science is communicated. I love reading them cover to cover, not only the results and references. Many times reading a paper has made me read a lot more just to verify the claims :) That being said, I wouldn't mind reading papers written by The Oracle if and only if all the derived results could be tracked down to its roots. That is the only way we can confidently judge the novelty of a paper. If The Oracle could able to trace all results back to the training set can you imagine the well of knowledge we could have access to? I mean the well is there already in raw data and you should be able to dive in it with Google but it lacks of the 'necessary' structure.
A paper's worth does not lay solely on its result but in its process. If we can achieve effective, reliable explainable AI, then we would see an increase in human-made scientific production because we would be exposed to new thinking patterns that we could quickly incorporate to our own, if we are flexible enough, that is.
If we can achieve effective, reliable explainable AI we could finally unmask AI for what it is: a biased linear combination of previous data points or marvel at its beautiful ability to create truly novel knowledge using just linear equations.
So, yes, right now I think we are heading closer to a new dawn in science after we experience and survive the AI purge; we shall stay optimist as a species we have survived at least one, well documented, dark age.
Postscript After I finished writing this entry as a draft, I head there was a drama around Open AI's Navier-Stokes counter-example. Apparently a pair of researchers were using Open AI's public models and made some progress. Before publishing their results one wrote to Open AI, it seems, before Open AI started to work on the problem. In this kind of cases it is hard to establish a trustworthy timeline so, it is hard to say with confidence what actually happened. However, it is a real challenge to the imagination that a company that built its welth around data would not mine its user's data.
A more detailed account of a similar ocurrence, this one involving a couple of papers with humans in-the-loop was given by Suvrit Sra on LinkedIn. Sra and his colleague wrote a paper that later was allegedly plagiarized by another paper a little more than 15 days after. The second paper was AI co-authored and according to Sra disregards their contributions and plagiarize their work. Once more, without any means to track down AI efforts back to its training dataset it is impossible to prove, without any doubt, who is to blame. I have met my fair share of authors that had barely skimmed through the papers they cite.
They'll stone ya when you're at the breakfast table
They'll stone ya when you are young and able
They'll stone ya when you're tryin' to make a buck
They'll stone ya and then they'll say, "good luck"
Tell ya what, I would not feel so all alone
Everybody must get stoned
Rainy Day Women #12 & 35, Bob Dylan, 1966.
And they've all liked your looks
With great lawyers you have
Discussed lepers and crooks
You've been through all of
F. Scott Fitzgerald's books
You're very well read
It's well known
But something is happening here
And you don't know what it is
Do you, Mister Jones?
Ballard of a Thin Man, Bob Dylan, 1965.
It has been a while since I wrote here :) Now I have a bit of time and a lot of ideas so time to dump one here :)
I'll write (once more) about AI and scientific research. Earlier I talked about Knuth's experience report using Claude (see Knuth's Claude's Cycles bellow). On LinkedIn I also posted a few thoughts on the matter, for example, I proposed a new track for conferences called "AI co-authored" for papers that make heavy use of AI in their production, regardless of its phase: proof-reading, idea generation, idea testing, etc.
Now I want to talk about what has been called auto-science, vibe-math and so on. It has gained traction, probably because users of AI systems are producing proofs in Lean, a proof assistant that is fairly recent, the first version was released in 2013 while other proof assistants had been around for much longer, e.g. Coq (now Rocq) has been around since 1989. With the backing of tech giants like Microsoft and Amazon, Lean nowadays dominate the vibe-coding proving market.
So, what do I understand by auto-science? I call auto-science to the following process: A human (or AI agent) proposes a theoretical problem (CS or math) to an AI agent. The agent works the problem and produces a proof or counter-example. To me, for this process to be strictly called auto-science the human must trust the agent's output to be right (hence the necessity that it must produce a proof or a counter-example). If the human needs to review the agent's output then I would call it AI-assisted-science. It is clear that in both cases (auto-science and AI-assisted-science) the AI agent lifts the greatest weight. Now, let's get into it shall we?
The academic context
If you publish papers probably you work at a university, perhaps as a professor (in any kind of flavor) a post-doc or a PhD student; I do not consider master and bachelor students because many of them do not get paid for publishing. Publishing papers shows you are at the top of the chain. It also means you depend on publishing for the greatest part. This is no secret, it is common that a post-doc or professorship position requiring something like strong publication record. Or if you are looking for a PhD you may see something like our group has a solid track record at top venues like... The publishing dependency is on two dimensions, the first is the publication output itself and the second the impact it has on the community.
It is not enough to publish, your papers must have readers, similar to online social media content creators: the more views and reactions with their content, the more they get paid. In a social media post, tracking reactions and views is conceptually trivial, you just need a couple of atomic counters. For papers that's a bit different, it doesn't just matter that your paper is cited it also matters who cite it; still we can draw a parallel with online social media, it is not the same to be reposted by a random person than by a celebrity.
The average paper builds upon existing knowledge and does so following the broad structure: we study problem X, also studied by [A,B,C..] with the lenses developed by [Z]. Ideally every claim in a paper must be able to be traced back following the references (up to honest paper availability). This check is 'relatively easy' I have done it since my bachelor days and I was not a particularly bright student, so, most students should be able to do it reasonably well.
For the most part, for an academic, their publications is their reputation and that has a direct impact on employment security and growth opportunities.
The tech context
Thanks to Claude now I can work on a small domain-specific language for CGR (see Undusting CGR in the Random CS section if you want to learn what CGR is). That was always the main goal of CGR, to have a language resembling SETL but specialized for computational geometry. However, I have always been super lazy to write parsers and lexers, those two just fill me with stress instead of joy, so I avoid them for the most part. With Claude I was able to obtain a parser and a lexer for a prototype DSL in Objective-C; that is pretty cool. I gave Claude a few examples, this is one of them:
The convex_hull function calculates the convex hull of the vertex set of the geometic graph g. The idea is that a com(trans)piler takes this program and produces an Objective-C program linked to CGR's library and then the Objective-C compiler produce an executable program. If the program is run and an input graph is given, then the program calculates the convex hull of the graph.List<Vertex> convex_hull:(graph) g { List<Vertex> CH <- []; Vertex P0 <- g.lowestLeftmostVertex; List<Vertex> V <- g.sortRadiallyWRT(P0); for Vertex v in V: while |CH| > 1 && g.turn(CH[1],CH[0],v) = g.LEFT_TURN: CH.pop() CH.push(v); return CH; }
Now, take any book or any paper on AI that you want and implement what you find in there. If you would like to start from scratch, you could start by implementing a neural network. Soon enough you will realize that you now need two inputs: the network weights and the prompt. If you have no network weights, you just have a few lines of code that will give you either constant answers to all your prompts or random answers every time. So, you don't only need the problem (prompt) but also you require to provide a computing context (network weights). The problems are easy to come by, the computing context not so much.
It starts with a training data that is transformed to vectors of numbers and those vectors are used to define the network weights. Without the training data, it is almost sure that you have an useless program. So, let's assume you have training data that is successfully translated into network weights that yield a good performing network. This means that the network can reasonably generalize (provide good approximations for the problem context).
With this in mind, let me propose the Programming Poet hypothesis: If a network is solely trained on poems the probability that it writes correct computer programs for any algorithmic problem in any programming language is 0.
If the Programming Poet hypothesis is true, then to obtain good results in any field, the network must be trained on material relevant to the field (no, it doesn't count that you regard your favorite proof as a beautiful poem!). Then, let us assume that we have access to a network trained on all the CS (math) papers published from 1980 to date and call it The Oracle. Why did I choose 1980? well, because if you started to publish when you were around 20 years old and your first paper was published in 1980, then by now your employment problems should be non-existent and your reputation well-established, so you don't need to produce novel results (please note that I talk about the need not the ability).
The experiment
Now the fun part, the experiment. Let us assume that all academics have access to The Oracle all can question it and read its answers. At this point the playing field is leveled: all academics have access to the same network; what will set them apart is the questions they ask, the prompts. If an academic asks a prompt the network was trained on, the answer must be close to the ground-truth. If an academic asks a prompt the network was not trained on, the answer is expected to be somewhat reasonable. If The Oracle can generalize, then it is reasonable to assume it will produce new results.
To check The Oracle's work, we must rely on the academic's factual knowledge: for the academic the problem could be novel, maybe unaware that someone else has already solved it or considered it, without any means to track The Oracle's work, from the academic's point of view, the answer is novel, The Oracle is producing novel results.
There are famous problems everyone 'knows' or know they exist, for those problems there is much training data: different approaches, sub-results, related problems, etc. Again an academic prompting these problems and getting a good result could reasonably conclude the network is producing new results, again the answer is novel.
But is it? if the academic prompted a problem already in the training data, by definition the problem is not novel. If the academic prompted a problem community regarded as open and gets a good result then it must be checked for 'novelty'. If the solution is actually novel, using insight no one has proposed before or the solution is mix of already present sub-solutions that no one had put together. By academic standards mixing already present sub-solutions is considered novel but not as novel as having an insight no one else has proposed. Again, judging the output of The Oracle falls back to the academic's factual knowledge base. Put into my bachelor's thesis adviser words: "I do not know everything that is being published" and he is regarded as a leading researcher in Computational Geometry.
Current models under-perform at tracing back their outputs to the training data, we have seen a lot of problems with models fabricating references: papers that do not exist. This makes the evaluation of novelty a real headache because you can not track the roots of the solution. The fallback option: If the proof checks then call it solved and move on.
As good as your last paper
If an academic's worth is in their reputation that comes from not only published papers but their impact, what will happen if we accept The Oracle's outputs and we can't trace back (without any doubts) it's reasoning? Well, the big winners are the ones that get the result first and publish it first, because their names are the ones that will be on the paper and headlines saying they solved, with AI, a long-standing problem. All the giant's shoulders upon which that solution stands will not be fully known, because we can't trace back The Oracle's reasoning. Furthermore, that paper will be part of the training set of The Oracle and live as a standalone paper as long as someone remembers it, hence, academics' reputation will be as good as their last paper.
We have already seen signs of this, it seems now that you broadcast your Lean proof to claim the prize, then you can write the paper. It is a race of prompts, a race of problems.
I suspect soon there will be databases of solved problems that you can automatically cross-check to see if someone else has already produced a Lean proof. Lean has already moved in this direction with their ambition to "Formalize CS".
We already have seen a massive surge of submitted papers to top conferences and the pressure on peer-review is building up, if even before this 'revolution' peer-review was under pressure, these days it seems it is on the brink of breaking.
Perhaps escaping this race by a lateral move could explain why recently I started to see colleagues calling for producing 'disposable proofs'. A 'disposable proof' is defined as a proof that no one cares for but that along with other 'disposable proofs' are generalized in a 'general theory'. The idea is to find 'the shape' of the underlying mathematics. So, say good-bye to your favorite equation in physics just because there is something called 'general theory of relativity'. I can't help but to see this as a profound lack of respect to the field, rounding up results just to claim you have 'the global' view and then dispose them. Like if Kepler burned Tycho Brahe's observations after he derived his laws of planetary motion. Again, it seems to be about claiming the result as soon as possible, and how the foundational data came to be? who cares, it grows on (the ashes of) trees!
The AI Purge
By now, we are familiar with headlines like "The end of programming", "The AI revolution", and many others full of optimism. In his essay "The Fate of Empires" Sir John Glubb writes: "... decadence is the disintegration of a system, not of its individual members. The habits of the members of the community have been corrupted by the enjoyment of too much money and too much power for too long a period. The result has been, in the framework of their national life, to make them selfish and idle. A community of selfish and idle people declines, internal quarrels develop in the division of its dwindling wealth, and pessimism follows..." Are we heading towards a new dawn in science or to its decline, towards the AI purge?
I believe the underlying goal of colleagues lobbying for 'disposable auto-science' is to bump the complexity by gathering enough data current models would struggle to interpret and interpret it in order to stay relevant. In the present context, as a problem, we can not deny it is relevant and novel, moreover, it is a generalization of what experts in the field are able to do (and do extremely well): literature review. Any good literature review paper classifies papers in categories discussing similarities and differences, spot open and partially addressed problems, basically is step 0 when you want to do research at any serious level. What sets apart your literature review when you are a student and one published in a journal is the scope and the depth. It is expected that an established expert provides a deeper and richer insight than a beginner. This seems to be a safe bet, you use your knowledge and skills to classify many data points (results, proofs) into a coherent model and then publish that model; of course you need the money first to create those data points and that's where the university's/project's budget comes into play, this however, isn't new to an academic. This approach is a much, much safer bet than sitting down and try to come up with a problem the model can not solve *and* that appeals to your peers. By construction, the problems you find in your 'disposable proofs' review, are open with respect to that model and hence relevant to the current standards. Someone would fill some holes and others will be regarded as 'irrelevant' following typical 'novelty/interesting' scientific canons.
But the interesting insight is that the drive of the proposal is to stay relevant by spotting and plugging holes of a technology that would not evolve beyond a hard threshold if we accept the Programming Poet hypothesis. In short, we are heading towards a "selfish and idle" community. I'm not saying these problems and analyses will be easy or trivial, surely some will be quite challenging, but if the Programming Poet hypothesis is true, then we are heading to a stale; there is of course, the possibility that in the meantime a breakthrough is made and we face new smarter models and we throw away all the advances made so far or that the generalization of the observed data gives way to a new insight that seed the current models and open new "frontiers". The stale state may not be reached in our generation but in the coming ones, it appears that we are not that interested in the future generations or in the pursuit of knowledge as a human endeavor because we are focused on staying relevant with respect to artificial oracles regardless of the cost.
That being said, it's hard to argue against sealing the holes in current models, beyond the usual arguments of resource consumption and accountability.
If only...
I love reading papers, I think there is something magical on how science is communicated. I love reading them cover to cover, not only the results and references. Many times reading a paper has made me read a lot more just to verify the claims :) That being said, I wouldn't mind reading papers written by The Oracle if and only if all the derived results could be tracked down to its roots. That is the only way we can confidently judge the novelty of a paper. If The Oracle could able to trace all results back to the training set can you imagine the well of knowledge we could have access to? I mean the well is there already in raw data and you should be able to dive in it with Google but it lacks of the 'necessary' structure.
A paper's worth does not lay solely on its result but in its process. If we can achieve effective, reliable explainable AI, then we would see an increase in human-made scientific production because we would be exposed to new thinking patterns that we could quickly incorporate to our own, if we are flexible enough, that is.
If we can achieve effective, reliable explainable AI we could finally unmask AI for what it is: a biased linear combination of previous data points or marvel at its beautiful ability to create truly novel knowledge using just linear equations.
So, yes, right now I think we are heading closer to a new dawn in science after we experience and survive the AI purge; we shall stay optimist as a species we have survived at least one, well documented, dark age.
Postscript After I finished writing this entry as a draft, I head there was a drama around Open AI's Navier-Stokes counter-example. Apparently a pair of researchers were using Open AI's public models and made some progress. Before publishing their results one wrote to Open AI, it seems, before Open AI started to work on the problem. In this kind of cases it is hard to establish a trustworthy timeline so, it is hard to say with confidence what actually happened. However, it is a real challenge to the imagination that a company that built its welth around data would not mine its user's data.
A more detailed account of a similar ocurrence, this one involving a couple of papers with humans in-the-loop was given by Suvrit Sra on LinkedIn. Sra and his colleague wrote a paper that later was allegedly plagiarized by another paper a little more than 15 days after. The second paper was AI co-authored and according to Sra disregards their contributions and plagiarize their work. Once more, without any means to track down AI efforts back to its training dataset it is impossible to prove, without any doubt, who is to blame. I have met my fair share of authors that had barely skimmed through the papers they cite.
They'll stone ya when you're at the breakfast table
They'll stone ya when you are young and able
They'll stone ya when you're tryin' to make a buck
They'll stone ya and then they'll say, "good luck"
Tell ya what, I would not feel so all alone
Everybody must get stoned
Rainy Day Women #12 & 35, Bob Dylan, 1966.
When I was working on Programming Without Problems V (PwP V), for some reason
that I can't remember I came to the conclusion that I would need a
kind of simple machinery to derive "symbolic" proofs.
Without further due I started to work on it, but at the same time I
keep working on PwP V and came up with a suitable version that didn't
required the mentioned machinery so it never made it to PwP V.
I named it "Elementary School Proof Assistant" (ESPA) because it is meant to
help with truly elementary proofs.
At the same time I think it is fun, it kind of remind me of my first
proofs when I was in high school.
ESPA is composed of two main modules, an "induction" and a symbolic modules both written in Prolog (SICStus Prolog). The "induction" module is quoted because it is not a real induction as in the mathematical sense. In mathematical induction, the induction step helps you to assume that if your proposition holds for a given natural, then it holds for the successor of that natural. I remember to imagine that maybe somewhere there was a monastery in which monks were testing each proposition with each possible natural. Like imagine the famous n(n+1)/2 for the sum of the first n natural numbers to be checked one value of n at the time. The Peano construction of the naturals defines a natural number as the successor of a natural number of 0 (or any other initial thing like origin). So, for example, the natural number three is defined as suc(suc(suc(0))). You can define addition and multiplication using the Peano construction, Chapter 3 of Aaron Stump's book "Verified Functional Programming in Agda" address this within Agda while the Volume 1 "Logical Foundations" of the famous "Software Foundations" series shows it within Rocq. Peano construction is the sensitive and natural implementation of the natural numbers but I didn't want to build something that formal but also not having something inconsistent. The answer was to implement a "ranged" induction. It basically means to "prove" a proposition by testing it against all the numbers on that range. In testing lingo it similar to property-based testing. My version of ranged induction does not boil down to SICStus arithmetic, it implements its own version of successor and predecessor based on SICStus increment and decrement operations (it is a nitpick but it is just to have a bit more of control). This model is really weak but that is what I was looking for, because I wanted to "prove" really basic stuff.
When you use Peano numbers, you must prove commutativity and neutrality, but I took an approach closer to abstract algebra and say: "I have this operation + that has a neutral and is commutative". I'm known to be a bit paranoid when it comes to simple things because I tend to make mistakes (I have many many stories about middle-school algebra in which I flipped signs, etc). So, I didn't want to introduce an axiom into my system that would render it inconsistent, therefore, I created the "induction" module so it could "property test" simple facts. The facts I'm talking about are: reflexivity (a = a), symmetry (a = b iff b = a), addition commutativity (a + b = b + a), addition neutral (a + 0 = a), and so on. Ranged induction is really weak and for example can not test addition associativity (a + (b + c) = (a + b) + c) so that has to be a 'true' axiom, i.e. introduced into the system as true. An axiom is induced into the system iff the weak "induction" is unable to prove it. This has the nice property that if you wish to insert an axiom that can be proved by the weak "induction", the system will fail.
The next module, the symbolic module is a bit more interesting because it is a rewrite module. The symbolic module is where you use what you have proved, either symbolically or 'inductively' to prove new propositions. For example, we can prove the simple proposition: 0 + a = a. This is the "code" using ESPA predicates:
The first line says we want to prove a theorem (yeah I know, it's too much ;)) the name of the theorem "lneu"(left-neutral), the list of assumptions, a natural number "a" and the theorem itself: 0 + a = a.
The second line says to unify the assumptions (a,b) in the theorem "add_conm" with the first assumption and the constant 0. The theorem add_conm is defined as: iprove(add_conm,[inumber(a),inumber(b)], a+b = b+a). Which means it is proved by induction (iprove).
The third line uses the theorem add_rneut defined as: iprove(add_rneut,[inumber(a)],a+0 = a).
The 4th line rewrites the ls of the second step with the right side of the 3rd step.
The 5th line uses symmetry to get the actual form of the theorem.
When this small program is executed, SICStus answers:
As last example, the following extract shows the definition of a group's operations (see e.g. Rotman, J., "An Introduction to the Theory of Groups", 4th Ed., 2nd printing, Springer, 1999) and a proof of the left cancelation law:
And how ESPA responds:
Appendix
To be complete, I wrote in Agda and in Rocq the proof for addition commutativity without using that 0 is neutral on the right side. But first I'll write the 'paper' double induction proof.
Paper proof
Def. (Natural number)
A natural number is either zero or the successor of a natural number n', written as suc n'.
Def. (Addition)
Let x,y be two natural numbers x + y is defined as: y if x is zero, suc (x+y) otherwise.
Theo. Let x, y be two natural numbers, then x + y = y + x.
By induction on x and y.
Base (x): x = zero. To prove: zero + y = y + zero.
Base (y): y = zero To prove: zero + zero = zero + zero.
Obvious.
I.H. (y): zero + y' = y' + zero To prove: zero + (suc y') = (suc y') + zero.
zero + suc y' = suc y' [Left side, + definition]
(suc y') + zero = suc (y' + zero) [Right side, + definition]
suc (y' + zero) = suc (zero + y') = suc y' [I.H. (y), + definition]
I.H. (x): x' + y = y + x' To prove: suc x' + y = y + suc x'.
suc x' + y = suc (x' + y) = suc (y + x') = y + suc x' [ + definition, I.H. (x), sucr]
sucr is an auxiliary lemma that states the successor of addition: suc(x+y) is equal to the addition of x and the successor of y (suc(x+y) = x + suc y).
I wrote this proof in Agda and Rocq, the results are in the files Agda, Rocq. For the Agda version I used "Programming Language Foundations in Agda" by Philip Wadler, Wen Kokke and Jeremy G. Siek.
What I like about the Agda version is that you can write chains of equations, so it feels closer to the paper version, but to me, induction (using cong) does not feel natural; I also love the use of Unicode characters in Emacs. On the other hand, the induction in Rocq's version feels more natural, as you can name the induction variable and the induction hypothesis. Using simpl takes away some of the equations but it doesn't feel weird and in Emacs you can always see the goal and the context.
ESPA is composed of two main modules, an "induction" and a symbolic modules both written in Prolog (SICStus Prolog). The "induction" module is quoted because it is not a real induction as in the mathematical sense. In mathematical induction, the induction step helps you to assume that if your proposition holds for a given natural, then it holds for the successor of that natural. I remember to imagine that maybe somewhere there was a monastery in which monks were testing each proposition with each possible natural. Like imagine the famous n(n+1)/2 for the sum of the first n natural numbers to be checked one value of n at the time. The Peano construction of the naturals defines a natural number as the successor of a natural number of 0 (or any other initial thing like origin). So, for example, the natural number three is defined as suc(suc(suc(0))). You can define addition and multiplication using the Peano construction, Chapter 3 of Aaron Stump's book "Verified Functional Programming in Agda" address this within Agda while the Volume 1 "Logical Foundations" of the famous "Software Foundations" series shows it within Rocq. Peano construction is the sensitive and natural implementation of the natural numbers but I didn't want to build something that formal but also not having something inconsistent. The answer was to implement a "ranged" induction. It basically means to "prove" a proposition by testing it against all the numbers on that range. In testing lingo it similar to property-based testing. My version of ranged induction does not boil down to SICStus arithmetic, it implements its own version of successor and predecessor based on SICStus increment and decrement operations (it is a nitpick but it is just to have a bit more of control). This model is really weak but that is what I was looking for, because I wanted to "prove" really basic stuff.
When you use Peano numbers, you must prove commutativity and neutrality, but I took an approach closer to abstract algebra and say: "I have this operation + that has a neutral and is commutative". I'm known to be a bit paranoid when it comes to simple things because I tend to make mistakes (I have many many stories about middle-school algebra in which I flipped signs, etc). So, I didn't want to introduce an axiom into my system that would render it inconsistent, therefore, I created the "induction" module so it could "property test" simple facts. The facts I'm talking about are: reflexivity (a = a), symmetry (a = b iff b = a), addition commutativity (a + b = b + a), addition neutral (a + 0 = a), and so on. Ranged induction is really weak and for example can not test addition associativity (a + (b + c) = (a + b) + c) so that has to be a 'true' axiom, i.e. introduced into the system as true. An axiom is induced into the system iff the weak "induction" is unable to prove it. This has the nice property that if you wish to insert an axiom that can be proved by the weak "induction", the system will fail.
The next module, the symbolic module is a bit more interesting because it is a rewrite module. The symbolic module is where you use what you have proved, either symbolically or 'inductively' to prove new propositions. For example, we can prove the simple proposition: 0 + a = a. This is the "code" using ESPA predicates:
I know this looks messy it is not meant to be 'friendly' at this stage :D:- siprove(theorem,lneu,[inumber(a)],0+a = a), siprove_intro(lneu,[1-a,cte(0)-b],add_conm), siprove_intro(lneu,[1-a],add_rneut), siprove_rewrite(lneu,2,ls,3,rs), siprove_intro(lneu,[1-a,expr(0+a)-b],sym).
The first line says we want to prove a theorem (yeah I know, it's too much ;)) the name of the theorem "lneu"(left-neutral), the list of assumptions, a natural number "a" and the theorem itself: 0 + a = a.
The second line says to unify the assumptions (a,b) in the theorem "add_conm" with the first assumption and the constant 0. The theorem add_conm is defined as: iprove(add_conm,[inumber(a),inumber(b)], a+b = b+a). Which means it is proved by induction (iprove).
The third line uses the theorem add_rneut defined as: iprove(add_rneut,[inumber(a)],a+0 = a).
The 4th line rewrites the ls of the second step with the right side of the 3rd step.
The 5th line uses symmetry to get the actual form of the theorem.
When this small program is executed, SICStus answers:
ESPA uses commutativity of + and a theorem stating 0 is a neutral on the right side. This is different to what you find in "Logical Foundations" and in Aaron Stump's book, they first prove 0 is a neutral on the left, then it is a neutral on the right and finally that addition is commutative. The goal I had in mind for ESPA was to be only symbolic manipulation, kind in the spirit of Programming Without Problems; express properties of algebraic structures by rewriting, such that when a concrete example fullfills the axioms, all the properties could be used by rewriting.Theorem: lneu proved. Theorem Name: lneu Theorem Goal: 0+a=a ========== Proof Context ========== Asm --- Assumption The --- Therefore/Conclusion Der --- Derived Asm 1. a The 2. a+0=0+a add_conm-[1-a,cte(0)-b] The 3. a+0=a add_rneut-[1-a] Rew 4. a=0+a 3-2 The 5. 0+a=a sym-[1-a,expr(0+a)-b]
As last example, the following extract shows the definition of a group's operations (see e.g. Rotman, J., "An Introduction to the Theory of Groups", 4th Ed., 2nd printing, Springer, 1999) and a proof of the left cancelation law:
:- iprove(refl,[inumber(a)],a=a). :- iprove(sym,[inumber(a),inumber(b),a=b],b=a). :- iprove(ladd,[inumber(a),inumber(b),inumber(c),b=c], a+b = a+c). :- iprove(radd,[inumber(a),inumber(b),inumber(c),b=c], b+a = c+a). :- iprove(assoc, [inumber(a),inumber(b),inumber(c)],a+(b+c)=(a+b)+c,axiom). :- iprove(rneutral, [inumber(a)], a + 0 = a). :- iprove(lneutral, [inumber(a)], 0 + a = a). :- iprove(rinverse, [inumber(a)], a + -a = 0). :- iprove(linverse, [inumber(a)], -a + a = 0). :- siprove(theorem,lcancel, [inumber(a),inumber(b), inumber(c), expr(a + b = a + c)], b = c), siprove_intro(lcancel,[expr(-a)-a,expr(a+b)-b,expr(a+c)-c], ladd), siprove_intro(lcancel,[expr(-a)-a,1-b,2-c], assoc), siprove_intro(lcancel,[expr(-a)-a,1-b,3-c], assoc), siprove_intro(lcancel,[1-a],linverse), siprove_intro(lcancel,[2-a,expr(-a+a)-b,cte(0)-c],radd), siprove_intro(lcancel,[3-a,expr(-a+a)-b,cte(0)-c],radd), siprove_intro(lcancel,[2-a],lneutral), siprove_intro(lcancel,[3-a],lneutral), siprove_rewrite(lcancel, 9-rs, 11-rs), siprove_rewrite(lcancel, 10-rs, 12-rs), siprove_rewrite(lcancel, 5-ls, 6-rs), siprove_rewrite(lcancel, 15-rs, 7-rs), siprove_rewrite(lcancel, 16-ls, 13-rs), siprove_rewrite(lcancel, 17-rs, 14-rs).
And how ESPA responds:
Theorem: lcancel proved. Theorem Name: lcancel Theorem Goal: b=c ========== Proof Context ========== Asm --- Assumption The --- Therefore/Conclusion Der --- Derived Asm 1. a Asm 2. b Asm 3. c Asm 4. a+b=a+c The 5. -a+(a+b)= -a+(a+c) ladd-[expr(-a)-a,expr(a+b)-b,expr(a+c)-c] The 6. -a+(a+b)= -a+a+b assoc-[expr(-a)-a,1-b,2-c] The 7. -a+(a+c)= -a+a+c assoc-[expr(-a)-a,1-b,3-c] The 8. -a+a=0 linverse-[1-a] The 9. -a+a+b=0+b radd-[2-a,expr(-a+a)-b,cte(0)-c] The 10. -a+a+c=0+c radd-[3-a,expr(-a+a)-b,cte(0)-c] The 11. 0+b=b lneutral-[2-a] The 12. 0+c=c lneutral-[3-a] Rew 13. -a+a+b=b 11-9 Rew 14. -a+a+c=c 12-10 Rew 15. -a+a+b= -a+(a+c) 6-5 Rew 16. -a+a+b= -a+a+c 7-15 Rew 17. b= -a+a+c 13-16 Rew 18. b=c 14-17
Appendix
To be complete, I wrote in Agda and in Rocq the proof for addition commutativity without using that 0 is neutral on the right side. But first I'll write the 'paper' double induction proof.
Paper proof
Def. (Natural number)
A natural number is either zero or the successor of a natural number n', written as suc n'.
Def. (Addition)
Let x,y be two natural numbers x + y is defined as: y if x is zero, suc (x+y) otherwise.
Theo. Let x, y be two natural numbers, then x + y = y + x.
By induction on x and y.
Base (x): x = zero. To prove: zero + y = y + zero.
Base (y): y = zero To prove: zero + zero = zero + zero.
Obvious.
I.H. (y): zero + y' = y' + zero To prove: zero + (suc y') = (suc y') + zero.
zero + suc y' = suc y' [Left side, + definition]
(suc y') + zero = suc (y' + zero) [Right side, + definition]
suc (y' + zero) = suc (zero + y') = suc y' [I.H. (y), + definition]
I.H. (x): x' + y = y + x' To prove: suc x' + y = y + suc x'.
suc x' + y = suc (x' + y) = suc (y + x') = y + suc x' [ + definition, I.H. (x), sucr]
sucr is an auxiliary lemma that states the successor of addition: suc(x+y) is equal to the addition of x and the successor of y (suc(x+y) = x + suc y).
I wrote this proof in Agda and Rocq, the results are in the files Agda, Rocq. For the Agda version I used "Programming Language Foundations in Agda" by Philip Wadler, Wen Kokke and Jeremy G. Siek.
What I like about the Agda version is that you can write chains of equations, so it feels closer to the paper version, but to me, induction (using cong) does not feel natural; I also love the use of Unicode characters in Emacs. On the other hand, the induction in Rocq's version feels more natural, as you can name the induction variable and the induction hypothesis. Using simpl takes away some of the equations but it doesn't feel weird and in Emacs you can always see the goal and the context.
If you have searched for a software developer job, you may have noticed
positions advertised as software engineer. And upon closer inspection
both roles seem to have the main core responsibilities. So, it seems
that both are synonymous, kind of the first phrase in The Silmarillion:
"There was Eru, the One, who in Arda is called Ilúvatar".
If we search in Wikipedia we would find that Software Engineering is "... focused on designing, developing, testing and maintaining software applications. It involves applying engineering principles and computer programming expertise to develop software systems that meet user needs." If we then click on "developing" it would take us to the Software development page which is described as: "... is the process of designing, creating, testing and maintaining software applications to meet specific user needs or business objectives. The process is more encompassing than programming, writing code, because it includes conceiving the goal, evaluating feasibility, analyzing requirements, design, testing and release. The process is part of software engineering which also includes organizational management, project management, configuration management and other aspects."
Both are striking similar in their definitions, both deal with design, creation, testing and maintenance of software it is only when you continue to read that you may see a hint of their differences, if any, specially in the software development entry: "The process is part of software engineering which also includes organizational management, project management, configuration management and other aspects". It turns out that the software engineering entry does not mention any of these aspects as part of its "workload".
It is impossible not to draw a parallel with other engineering disciplines like civil engineering. Yes, building software is different than building a building but, they are not that different. Again the Wikipedia says that civil engineering "... deals with the design, construction and maintenance of the physically and naturally built environment..." If we click on "construction" it is defined as "... is the process of delivering buildings, infrastructure, industrial facilities, and associated activities through to the end of their life. It typically starts with planning, financing and design that continues until the asset is built and ready for use."
Here the separation is clearer, and it should be as civil engineers are always present at construction time but they do not lay the bricks, that is the job of the bricklayer. The engineers make sure the walls are being built according to the specification by overseeing the construction process. Of course they must be familiar with the properties of the build materials so the end product has the properties of the designed product, but they do not make sure of this by building the thing themselves. A bricklayer must know how to read a schematic (specification) and knows the best way to deal with concrete and steel rods, how to make sure a wall is straight and so on.
In today's software construction, we do not have that clear separation. There are no roles specific for programming nor roles specific for designing. Software engineer/developer positions do not mention, for example, knowledge of formal specification languages.
Unless critical software is being constructed, formal methods are neglected. This hinder software creation evolution as a discipline. Recently Dave Plummer wrote on Twitter (now X :D ) that in his 40 years of programming, his first takeaway is that "90% of the job is debugging and fixing, not creating new code." 90% of a person's productive energy is a lot, and it is devoted to fixing errors. That doesn't seem right, I mean, from an employer/investor point of view, I don't think it would look like a sustainable enterprise if your top engineers/developers fix whatever others got 'wrong' and introduce errors themselves that would take other engineers 90% of their time to fix. If instead of fixing errors, those top engineers would focus on designing and proving software properties and the next layer of engineers focus on refinement of those formal designs and the next layer in further refinement and so on until executable code is obtained, we would observe a massive jump in software correctness.
Ian Craig books: "Formal Models of Operating System Kernels" and "Formal Refinement for Operating System Kernels" showed that specifying and refining code of micro-kernels is possible and feasible (in modern times). Before, in 1985 and 1990, Hartmut Ehrig and Bernd Mahr wrote Fundamentals of Algebraic Specification 1 and 2. The first book is 41 years old and I can not help to wonder how would an average developer react to "composition of interface specifications". Surely a CS educated developer could understand and find gain in algebraic specifications, but it is clear that the actual practitioner's field of software creation has paid little attention to formal models, otherwise platforms like GitHub and StackOverflow would be flooded with different kinds of specifications next to code.
Nix/NixOS is a mainstream software designed to provide reproducible environments. It is a step towards correct software. Endless are the stories paraphrased as "It runs on my computer", with Nix, those stories may cease to exist. It is similar to OCaml's opam package manager in which you can have independent environments and be able to install specific versions of packages.
With the advent of AI code generation it seems we are taking one step forward and 3/4 of a step backwards because software specification is made using natural language. Getting truly reproducible results, that is, the same code for the same prompt seems to be a challenge. However, I think it can yield some positive results if an AI is capable of interpreting a basic specification in natural language and synthesize code in a given language. I think an interesting exercise would be to see how to write AI understandable natural language specifications of some examples in Ian Craig's books and have the AI to synthesize code in different languages that would be equivalent to correct refinements.
However, I'm not sure if I'm ready to burn a small forest for that.
If we search in Wikipedia we would find that Software Engineering is "... focused on designing, developing, testing and maintaining software applications. It involves applying engineering principles and computer programming expertise to develop software systems that meet user needs." If we then click on "developing" it would take us to the Software development page which is described as: "... is the process of designing, creating, testing and maintaining software applications to meet specific user needs or business objectives. The process is more encompassing than programming, writing code, because it includes conceiving the goal, evaluating feasibility, analyzing requirements, design, testing and release. The process is part of software engineering which also includes organizational management, project management, configuration management and other aspects."
Both are striking similar in their definitions, both deal with design, creation, testing and maintenance of software it is only when you continue to read that you may see a hint of their differences, if any, specially in the software development entry: "The process is part of software engineering which also includes organizational management, project management, configuration management and other aspects". It turns out that the software engineering entry does not mention any of these aspects as part of its "workload".
It is impossible not to draw a parallel with other engineering disciplines like civil engineering. Yes, building software is different than building a building but, they are not that different. Again the Wikipedia says that civil engineering "... deals with the design, construction and maintenance of the physically and naturally built environment..." If we click on "construction" it is defined as "... is the process of delivering buildings, infrastructure, industrial facilities, and associated activities through to the end of their life. It typically starts with planning, financing and design that continues until the asset is built and ready for use."
Here the separation is clearer, and it should be as civil engineers are always present at construction time but they do not lay the bricks, that is the job of the bricklayer. The engineers make sure the walls are being built according to the specification by overseeing the construction process. Of course they must be familiar with the properties of the build materials so the end product has the properties of the designed product, but they do not make sure of this by building the thing themselves. A bricklayer must know how to read a schematic (specification) and knows the best way to deal with concrete and steel rods, how to make sure a wall is straight and so on.
In today's software construction, we do not have that clear separation. There are no roles specific for programming nor roles specific for designing. Software engineer/developer positions do not mention, for example, knowledge of formal specification languages.
Unless critical software is being constructed, formal methods are neglected. This hinder software creation evolution as a discipline. Recently Dave Plummer wrote on Twitter (now X :D ) that in his 40 years of programming, his first takeaway is that "90% of the job is debugging and fixing, not creating new code." 90% of a person's productive energy is a lot, and it is devoted to fixing errors. That doesn't seem right, I mean, from an employer/investor point of view, I don't think it would look like a sustainable enterprise if your top engineers/developers fix whatever others got 'wrong' and introduce errors themselves that would take other engineers 90% of their time to fix. If instead of fixing errors, those top engineers would focus on designing and proving software properties and the next layer of engineers focus on refinement of those formal designs and the next layer in further refinement and so on until executable code is obtained, we would observe a massive jump in software correctness.
Ian Craig books: "Formal Models of Operating System Kernels" and "Formal Refinement for Operating System Kernels" showed that specifying and refining code of micro-kernels is possible and feasible (in modern times). Before, in 1985 and 1990, Hartmut Ehrig and Bernd Mahr wrote Fundamentals of Algebraic Specification 1 and 2. The first book is 41 years old and I can not help to wonder how would an average developer react to "composition of interface specifications". Surely a CS educated developer could understand and find gain in algebraic specifications, but it is clear that the actual practitioner's field of software creation has paid little attention to formal models, otherwise platforms like GitHub and StackOverflow would be flooded with different kinds of specifications next to code.
Nix/NixOS is a mainstream software designed to provide reproducible environments. It is a step towards correct software. Endless are the stories paraphrased as "It runs on my computer", with Nix, those stories may cease to exist. It is similar to OCaml's opam package manager in which you can have independent environments and be able to install specific versions of packages.
With the advent of AI code generation it seems we are taking one step forward and 3/4 of a step backwards because software specification is made using natural language. Getting truly reproducible results, that is, the same code for the same prompt seems to be a challenge. However, I think it can yield some positive results if an AI is capable of interpreting a basic specification in natural language and synthesize code in a given language. I think an interesting exercise would be to see how to write AI understandable natural language specifications of some examples in Ian Craig's books and have the AI to synthesize code in different languages that would be equivalent to correct refinements.
However, I'm not sure if I'm ready to burn a small forest for that.
When I was growing up I had the chance to decide between playing console
games or PC games.
My parents used to give me an allowance that I dedicated to games and
computing books (when I grew older 13+).
I sided for PC games just because they were (and had the potential to be)
more complex than console games.
I came to this conclusion not because I had many games (I didn't)
but it was natural if you look at the input mechanisms:
The Super Nintendo had a control with 11 "playing" buttons vs the
potential 100+ keyboard keys.
I started with flight simulators, I like them so much that I saved many months to be able to play SU-27 Flanker and EF2000 "properly" when I could finally afford a Thrustmaster joystick. But I also love First-Person Shooters (FPS). Like many I started with Doom, Wolfenstein, and one of my all-time favorite series: Quake. But later, I met Rainbow Six and it was AWESOME!
I liked that you could shoot guns but also plan ahead when and how you do it; it was very refreshing; it combined the shooting action with a bit more of "tactical" thinking. The genre is actually called "Tactical shooter" and some reviewers and players focus on "realism" when judging such games. The next step towards "realism", is called "Milsim" (ARMA series) and finally "War Games" used by people dedicated to war.
I have never shot a gun in my life, nor I have been shot at, so I have no idea what makes a FPS game more "realistic" than other. I mean, I know that you can't jump and fire a "sniper rifle" the way you do in Counter Strike, to me Counter Strike is closer to Quake 3 than it is to the first Call of Duty (I mean the PC first version). However, I do know that bullets do not flight in a straight line, there is something called gravity that curves the trajectory and other factors that are "simulation-friendly" (there is some interesting discussion on simulations in philosophy of computing).
So, I like when a game incorporates simulation-friendly elements but to some extend. For example, I don't like much to walk my character on an ARMA mission for hrs just to get killed; I do not have a lot of time to devote :) Rainbow Six fits that bill, you can spend a considerable amount of time planning your mission and seeing parts of it work and others don't, honing your plan and preparation; just so you finish the mission with all your soldiers alive (and preferably healthy) in 3-4 minutes. I think that is a good 'simulation' on itself, you spend a lot of time studying, training, etc. for a short test.
So, if you are not familiar with the Rainbow Six 1-3 series, you plan the actions of 3 teams (up to 8 soldiers) to execute a mission. Most of them are to kill enemies and rescue hostages. All soldiers can be player-controllable at any moment; if they are not controlled by you, they are controlled by the game (AI). So, the focus of the game is "to kill the bad guys and rescue the innocent" that's it. The planning allows you to arrange the 3 teams to your liking to execute a mission, one team may enter a room from one side, another from another door and the last team is making sure no one enters the building. The idea (I think) is teamwork. This was released in 1998.
Ready or Not came as a glass of clear water, it is sold as conceptually close to the SWAT series (which I also loved but mostly online). In it, you control a character and 'command' 4 characters, I say command because you can't take control of them just direct them and let the AI work. Ready or Not core idea is to "arrest-injure-kill suspects and rescue innocents". The missions vary but in one, you are tossed to 'secure' a Mail dispatch warehouse with (I think) 20 suspects, that is: 5 suspects per character.
When it came out it was advertised as developed with the help of real SWAT officers to get a nice amount of 'realism' (again with the word). I'm not expert, but I don't think 5 people (a leader and 4 team members) would be what any police force would send to arrest 20 'heavily armed suspects', but probably (maybeee) what a specialized army branch would send to kill 20 regular soldiers.
It feels Ready or Not is conceptually closer to a water-down version of Rainbow Six, perhaps Rainbow Six Lockdown, but with goals of SWAT 4 (kill the least amount of suspects). Ready or Not places a lot of focus on weapons: ammunition type, barrel attachments, sights, etc. and zero focus on planning. You command the other members 'live' during the game, and even that part feels a bit clumsy, the contextual menu for 'tazing' a suspect while it is still 'a threat' takes to 1) click the middle button while pointing to the suspect 2) Press 3 (Deploy) 3) Press 1 (Tazer). All while keep shouting at the suspect, keep your finger ready on the left button in case the suspect raises the weapon and keep an eye on the rest of the surroundings; oh by the way, each time you shout, your weapon sways so you may loose the contextual menu in the middle of the action. It is not surprising that 99.9% of PvE (the only type of multiplayer available) matches I have played, my digital colleagues shoot to kill.
I think there is a lot of potential in Ready or Not, even if they make a 180 turn and focus it on killing rather than arresting. But it also must run on consoles, and well, consoles have limited input. I do not know why speech recognition has not been pushed, particularly in consoles, it appeared in Rainbow Six Black Arrow (I think, I'm not sure). This would help to create more complex games.
The other side is the AI, in comparison to Rainbow Six, it is horrible. I agree that Rainbow Six 3 soldiers sometimes are either super good or super bad, but they are 'enjoyable'. It is frustrating to see one of your police officers just pass-by a suspect without doing anything just to get shot on the back. This shows that a lot of work has been lost, most likely due to IP (Intellectual Property). So, how am I supposed to arrest 20 suspects with noob cops? :D
Rainbow Six 3 Athena Sword and Ready or Not are separated by 22 years, yes Ready or Not looks gorgeous even on my 12 years old computer, but if I have to choose I would play a lot more a more complex, equally looking Rainbow Six 3 version today, and I have played 617 hrs on Ready or Not. Now that computer games are trending (there are even persons called "pro-players") I would have expected Tactical Shooters to be more Tactical more immersive, but I believe if you are a small game studio and had to redo everything from the ground up, because we prefer to die with our secrets than see others thrive and build upon our successes, it doesn't matter if you are in the late 90's or in 2026. It makes no difference having a 1.3 GHz computer or a 4.20 GHz, 512 MB of RAM or 8 GB or a video card with 128 MB or 3 GB. Having faster, bigger hardware does not mean you will get your path finding algorithms correct the first years. In many gameplay aspects Rainbow Six 3, released in 1998 is far superior to Ready or Not.
I started with flight simulators, I like them so much that I saved many months to be able to play SU-27 Flanker and EF2000 "properly" when I could finally afford a Thrustmaster joystick. But I also love First-Person Shooters (FPS). Like many I started with Doom, Wolfenstein, and one of my all-time favorite series: Quake. But later, I met Rainbow Six and it was AWESOME!
I liked that you could shoot guns but also plan ahead when and how you do it; it was very refreshing; it combined the shooting action with a bit more of "tactical" thinking. The genre is actually called "Tactical shooter" and some reviewers and players focus on "realism" when judging such games. The next step towards "realism", is called "Milsim" (ARMA series) and finally "War Games" used by people dedicated to war.
I have never shot a gun in my life, nor I have been shot at, so I have no idea what makes a FPS game more "realistic" than other. I mean, I know that you can't jump and fire a "sniper rifle" the way you do in Counter Strike, to me Counter Strike is closer to Quake 3 than it is to the first Call of Duty (I mean the PC first version). However, I do know that bullets do not flight in a straight line, there is something called gravity that curves the trajectory and other factors that are "simulation-friendly" (there is some interesting discussion on simulations in philosophy of computing).
So, I like when a game incorporates simulation-friendly elements but to some extend. For example, I don't like much to walk my character on an ARMA mission for hrs just to get killed; I do not have a lot of time to devote :) Rainbow Six fits that bill, you can spend a considerable amount of time planning your mission and seeing parts of it work and others don't, honing your plan and preparation; just so you finish the mission with all your soldiers alive (and preferably healthy) in 3-4 minutes. I think that is a good 'simulation' on itself, you spend a lot of time studying, training, etc. for a short test.
So, if you are not familiar with the Rainbow Six 1-3 series, you plan the actions of 3 teams (up to 8 soldiers) to execute a mission. Most of them are to kill enemies and rescue hostages. All soldiers can be player-controllable at any moment; if they are not controlled by you, they are controlled by the game (AI). So, the focus of the game is "to kill the bad guys and rescue the innocent" that's it. The planning allows you to arrange the 3 teams to your liking to execute a mission, one team may enter a room from one side, another from another door and the last team is making sure no one enters the building. The idea (I think) is teamwork. This was released in 1998.
Ready or Not came as a glass of clear water, it is sold as conceptually close to the SWAT series (which I also loved but mostly online). In it, you control a character and 'command' 4 characters, I say command because you can't take control of them just direct them and let the AI work. Ready or Not core idea is to "arrest-injure-kill suspects and rescue innocents". The missions vary but in one, you are tossed to 'secure' a Mail dispatch warehouse with (I think) 20 suspects, that is: 5 suspects per character.
When it came out it was advertised as developed with the help of real SWAT officers to get a nice amount of 'realism' (again with the word). I'm not expert, but I don't think 5 people (a leader and 4 team members) would be what any police force would send to arrest 20 'heavily armed suspects', but probably (maybeee) what a specialized army branch would send to kill 20 regular soldiers.
It feels Ready or Not is conceptually closer to a water-down version of Rainbow Six, perhaps Rainbow Six Lockdown, but with goals of SWAT 4 (kill the least amount of suspects). Ready or Not places a lot of focus on weapons: ammunition type, barrel attachments, sights, etc. and zero focus on planning. You command the other members 'live' during the game, and even that part feels a bit clumsy, the contextual menu for 'tazing' a suspect while it is still 'a threat' takes to 1) click the middle button while pointing to the suspect 2) Press 3 (Deploy) 3) Press 1 (Tazer). All while keep shouting at the suspect, keep your finger ready on the left button in case the suspect raises the weapon and keep an eye on the rest of the surroundings; oh by the way, each time you shout, your weapon sways so you may loose the contextual menu in the middle of the action. It is not surprising that 99.9% of PvE (the only type of multiplayer available) matches I have played, my digital colleagues shoot to kill.
I think there is a lot of potential in Ready or Not, even if they make a 180 turn and focus it on killing rather than arresting. But it also must run on consoles, and well, consoles have limited input. I do not know why speech recognition has not been pushed, particularly in consoles, it appeared in Rainbow Six Black Arrow (I think, I'm not sure). This would help to create more complex games.
The other side is the AI, in comparison to Rainbow Six, it is horrible. I agree that Rainbow Six 3 soldiers sometimes are either super good or super bad, but they are 'enjoyable'. It is frustrating to see one of your police officers just pass-by a suspect without doing anything just to get shot on the back. This shows that a lot of work has been lost, most likely due to IP (Intellectual Property). So, how am I supposed to arrest 20 suspects with noob cops? :D
Rainbow Six 3 Athena Sword and Ready or Not are separated by 22 years, yes Ready or Not looks gorgeous even on my 12 years old computer, but if I have to choose I would play a lot more a more complex, equally looking Rainbow Six 3 version today, and I have played 617 hrs on Ready or Not. Now that computer games are trending (there are even persons called "pro-players") I would have expected Tactical Shooters to be more Tactical more immersive, but I believe if you are a small game studio and had to redo everything from the ground up, because we prefer to die with our secrets than see others thrive and build upon our successes, it doesn't matter if you are in the late 90's or in 2026. It makes no difference having a 1.3 GHz computer or a 4.20 GHz, 512 MB of RAM or 8 GB or a video card with 128 MB or 3 GB. Having faster, bigger hardware does not mean you will get your path finding algorithms correct the first years. In many gameplay aspects Rainbow Six 3, released in 1998 is far superior to Ready or Not.
Among all the noise in AI, sometimes you hear that beautiful harmonic
tone in the background.
I believe Knuth's diligence, attention to detail and academic honestly are truly hard to find. We can peek at it in his latest note: Claude's Cycles
In it, Knuth gives an account of his friend Filip Stappers' journey with Claude Opus 4.6 to prove a problem that will appear in the upcoming "The Art of Computer Programming" book.
What I think is the most valuable about this note is that Knuth describes the process from first attempt to proof; which must be something similar to what is done when AI is applied to problem-solving.
On the actual problem, I think it is interesting that Claude could solve the problem (roughly in an hr) and at the same time was unable to remember it was told to log progress carefully. It sounds strange.
"Not all open problems on a paper are worthy of a paper..." (see Appendix for context on how I came to this)
When you have finite time to work, you want to make the most of it and a ranking of problems is natural. If something sounds simple or evident, you can say: "This must be true/false" and move along. That leaves a lot of "open problems" that are formally open but not really, a small number of people work on them.
AI noise makers constantly use the phrase "Open problem" as something the community has been actively working on but not getting it. They evoke the Hollywood science movies in which people work around the clock on a single problem until they get it.
At the same time, they don't give any worthy details, like a transcript of the entire session, or a careful account of events; at most you get a prompt of 800 lines or something. You basically get a "it compiled" line and move on.
While this is true in formal verification, I think it is not enough: procedure is as important as the result.
There are many dimensions you can use to analyze Knuth's paper.
Dimension 1: Effectiveness
The problem was solved quite fast (leaving aside the human intervention) however, how can a system solve an arguably difficult problem and yet forget a simple instruction as keep a detailed progress log? It makes little sense unless...
It had been trained with closely related literature. A paper or series of papers, blog posts, forum posts or loose files on the Internet unknown to Knuth but not to the crawlers. Knuth himself acknowledges of a reference related to the problem but unknown to him.
Open problems are holes in the knowledge fabric. If you work on an open problem, you are weaving knowledge to cover that hole. Sometimes you are able to cover the whole hole or you make some partial progress, as Knuth reports Claude did, and someone else covers the rest. Other times progress is made to make the hole smaller. You develop something that clarifies the problem, restricts it or solve special cases but it remains open whether the general case can be addressed in the same way. This dynamic gets more complicated when progress in an area can be translated to progress in another.
So, from where did Claude got its ideas? it is unknown as probably they were not logged.
But this also shows an exciting behavior of AI models: they mix things. In this case the mixing worked and it would be wonderful to know exactly "how open" the problem really is (was)? If there was training data that was really close to solve it then the mixer of things didn't do much work, but if there was not much training data, then the mixer did more work and that would be interesting.
Dimension 2: Credit
It is plausible this anecdote will be in the upcoming book, and the name Claude's Cycles is already catching on.
We will not know (most probably) from where Claude got its inspiration. For all we know it could be a small paper somewhere made by someone published in who knows where that was hard to find in Claude's structure but it was there and it is reachable, however, who wrote that? was it a close enough paper? we may never know. On the other hand, the result is now known to be authored by Claude.
This is a tangible example of giving credit to the mixer and not the ingredients. In science people live from reputation, "the one who solved ..." we have constructed it that way. Now Claude gets credit from a Computing pillar and that is not small feat.
Should it get credit for it?
Science has also its fair share of drama, people stealing other people's work, often a more famous stealing from a lesser known. Are we seeing a normalization of this stealing?
Dimension 3: Filler/Filter
If an open problem is small it stands to reason that ML approaches could be used to close them. If it is unable to do so, then perhaps it could be looked at.
The role of ML approaches would be that of filling the gaps or filter the interesting problems from the "already solvable by existing means"
In this role the process of solving a problem would be more important than the solution itself because it could help to develop ML transparency and ultimately trust.
The noise makers are already saying one of the most prominent figures in CS is approving "vibe coding".
Could not be further from the truth, this is one problem for the next book, if you have seen a Knuth's book, in particular a volume of "The Art of Computing Programming", where this problem will appear, you will see there are TONS of problems, and this is one of them.
Does that make it simple?
Heck no!
But it doesn't justify nor implies that we should invest 90% of our eggs into this basket!
--------------- Appendix ---------------
I think I learned (quite) late in my career the deep social ties underlying academia.
A long time ago I wanted to do my PhD with certain a professor at a certain university across the Atlantic. I learned there was an opening and sent my application and he said: "that looks good". His webpage stated something like: if you want to work with me, surprise me. How do you surprise a professor you want to work with? Easy, you pick one of his latest papers, look at the open problems section, pick one and solve it (kudos if they wrote on the paper that they tried and couldn't). It took me about 3 weeks or so but I managed to solve one of two problems, and I was ready to present it once I get the call.
The call never came.
After a couple of months I decided to turn my proof into a paper. The paper with the open problem was published on a Tier-1 conference so that's what I aimed for. Obviously it was rejected as it was poorly written, so I tried again (and again.. and again) each time the rejections were getting thinner, up to 5 reviewers would read the paper. The rejections were never about the math, no correction was made on that side, so, I got it right since day one; they were about the degree of novelty of the problem, basically 'is this hard enough?'
I fondly remember one reviewer on my second attempt: "This is an elegant proof and an important result that definitely deserves a spot on the program". I gave up after 3 or so attempts (each time adding more results) when the final decision was phrased as: "Not all open problems on a paper are worthy of a paper on a top-tier conference. We suggest the author to publish it on a lower tier venue" That was completely counter-intuitive to me (it still is :) ).
I did listen and when I submitted it to an international but still local (Latin America) big conference, it was rejected because "the problem has already been solved" and I don't remember the exact wording but the reviewers implied that I was stealing the result! The problem was not at all solved, it was a draft on arXiv with a much more complicated proof than mine but with all the authors of the original paper.
A few years later, I found the original authors (plus one) finally published the complicated proof in a journal. I mean really complicated.
That's ok, no hard feelings :)
However, the important bit that stuck with me: "Not all open problems on a paper are worthy of a paper..."
I believe Knuth's diligence, attention to detail and academic honestly are truly hard to find. We can peek at it in his latest note: Claude's Cycles
In it, Knuth gives an account of his friend Filip Stappers' journey with Claude Opus 4.6 to prove a problem that will appear in the upcoming "The Art of Computer Programming" book.
What I think is the most valuable about this note is that Knuth describes the process from first attempt to proof; which must be something similar to what is done when AI is applied to problem-solving.
On the actual problem, I think it is interesting that Claude could solve the problem (roughly in an hr) and at the same time was unable to remember it was told to log progress carefully. It sounds strange.
"Not all open problems on a paper are worthy of a paper..." (see Appendix for context on how I came to this)
When you have finite time to work, you want to make the most of it and a ranking of problems is natural. If something sounds simple or evident, you can say: "This must be true/false" and move along. That leaves a lot of "open problems" that are formally open but not really, a small number of people work on them.
AI noise makers constantly use the phrase "Open problem" as something the community has been actively working on but not getting it. They evoke the Hollywood science movies in which people work around the clock on a single problem until they get it.
At the same time, they don't give any worthy details, like a transcript of the entire session, or a careful account of events; at most you get a prompt of 800 lines or something. You basically get a "it compiled" line and move on.
While this is true in formal verification, I think it is not enough: procedure is as important as the result.
There are many dimensions you can use to analyze Knuth's paper.
Dimension 1: Effectiveness
The problem was solved quite fast (leaving aside the human intervention) however, how can a system solve an arguably difficult problem and yet forget a simple instruction as keep a detailed progress log? It makes little sense unless...
It had been trained with closely related literature. A paper or series of papers, blog posts, forum posts or loose files on the Internet unknown to Knuth but not to the crawlers. Knuth himself acknowledges of a reference related to the problem but unknown to him.
Open problems are holes in the knowledge fabric. If you work on an open problem, you are weaving knowledge to cover that hole. Sometimes you are able to cover the whole hole or you make some partial progress, as Knuth reports Claude did, and someone else covers the rest. Other times progress is made to make the hole smaller. You develop something that clarifies the problem, restricts it or solve special cases but it remains open whether the general case can be addressed in the same way. This dynamic gets more complicated when progress in an area can be translated to progress in another.
So, from where did Claude got its ideas? it is unknown as probably they were not logged.
But this also shows an exciting behavior of AI models: they mix things. In this case the mixing worked and it would be wonderful to know exactly "how open" the problem really is (was)? If there was training data that was really close to solve it then the mixer of things didn't do much work, but if there was not much training data, then the mixer did more work and that would be interesting.
Dimension 2: Credit
It is plausible this anecdote will be in the upcoming book, and the name Claude's Cycles is already catching on.
We will not know (most probably) from where Claude got its inspiration. For all we know it could be a small paper somewhere made by someone published in who knows where that was hard to find in Claude's structure but it was there and it is reachable, however, who wrote that? was it a close enough paper? we may never know. On the other hand, the result is now known to be authored by Claude.
This is a tangible example of giving credit to the mixer and not the ingredients. In science people live from reputation, "the one who solved ..." we have constructed it that way. Now Claude gets credit from a Computing pillar and that is not small feat.
Should it get credit for it?
Science has also its fair share of drama, people stealing other people's work, often a more famous stealing from a lesser known. Are we seeing a normalization of this stealing?
Dimension 3: Filler/Filter
If an open problem is small it stands to reason that ML approaches could be used to close them. If it is unable to do so, then perhaps it could be looked at.
The role of ML approaches would be that of filling the gaps or filter the interesting problems from the "already solvable by existing means"
In this role the process of solving a problem would be more important than the solution itself because it could help to develop ML transparency and ultimately trust.
The noise makers are already saying one of the most prominent figures in CS is approving "vibe coding".
Could not be further from the truth, this is one problem for the next book, if you have seen a Knuth's book, in particular a volume of "The Art of Computing Programming", where this problem will appear, you will see there are TONS of problems, and this is one of them.
Does that make it simple?
Heck no!
But it doesn't justify nor implies that we should invest 90% of our eggs into this basket!
--------------- Appendix ---------------
I think I learned (quite) late in my career the deep social ties underlying academia.
A long time ago I wanted to do my PhD with certain a professor at a certain university across the Atlantic. I learned there was an opening and sent my application and he said: "that looks good". His webpage stated something like: if you want to work with me, surprise me. How do you surprise a professor you want to work with? Easy, you pick one of his latest papers, look at the open problems section, pick one and solve it (kudos if they wrote on the paper that they tried and couldn't). It took me about 3 weeks or so but I managed to solve one of two problems, and I was ready to present it once I get the call.
The call never came.
After a couple of months I decided to turn my proof into a paper. The paper with the open problem was published on a Tier-1 conference so that's what I aimed for. Obviously it was rejected as it was poorly written, so I tried again (and again.. and again) each time the rejections were getting thinner, up to 5 reviewers would read the paper. The rejections were never about the math, no correction was made on that side, so, I got it right since day one; they were about the degree of novelty of the problem, basically 'is this hard enough?'
I fondly remember one reviewer on my second attempt: "This is an elegant proof and an important result that definitely deserves a spot on the program". I gave up after 3 or so attempts (each time adding more results) when the final decision was phrased as: "Not all open problems on a paper are worthy of a paper on a top-tier conference. We suggest the author to publish it on a lower tier venue" That was completely counter-intuitive to me (it still is :) ).
I did listen and when I submitted it to an international but still local (Latin America) big conference, it was rejected because "the problem has already been solved" and I don't remember the exact wording but the reviewers implied that I was stealing the result! The problem was not at all solved, it was a draft on arXiv with a much more complicated proof than mine but with all the authors of the original paper.
A few years later, I found the original authors (plus one) finally published the complicated proof in a journal. I mean really complicated.
That's ok, no hard feelings :)
However, the important bit that stuck with me: "Not all open problems on a paper are worthy of a paper..."
There is a great deal of evidence that seems to indicate the creation
of computer programs is difficult, from research studies on
programming teaching and learning to general public perception surveys.
But why is this? and from the perspective of the general public
why should we care?. In here I'll briefly address both questions.
Why is it difficult?
We do not know exactly why programming is difficult, to the best of our knowledge after teaching it to the students to the best of our abilities, students fail tests in sufficient numbers allowing us to conclude it is a difficult task.
We could argue it is not the discipline but how it is taught, however, for the sake of the argument, let us assume it is not. Moreover, there exists an ample repertoire of techniques that had been devised and tested to teach programming.
While we do not know precisely why programming is difficult, we do know that it involves certain competences, and perhaps one of the most mentioned is the ability to solve problems, also known as problem solving.
Solving problems is the core of programming, currently if we do not have a problem to solve, there is little motivation to program a computer. But we solve problems on a daily basis. Solving problems with the computer requires another ability precision. When we express a solution to a problem as a computer program we are being precise, each instruction help us to arrive to a solution, to track this progress each instruction must fulfill two conditions: the first, it must be a valid instruction in a programming language and secondly the inputs-outputs relationship must have a meaning in our problem domain. So, in few words creating computer programs means solving problems precisely.
The precision requirement is what brings programming closer to exact sciences like mathematics than to any other disciplines, like, literature or law. Even within mathematics and theoretical computer science, we can consider programs to be (much) more precise than a mathematical proof (in here I'm not considering computer-checked proofs, I mean it in the more traditional way). A mathematical or theoretical computer science paper can make a small omission or even a small mistake and still be understood by the reader. If a computer program is missing a part, it simply won't work, if it does, then you can remove that part altogether and we can conclude the solution was not completely precise (which is fine), if a computer program has a small mistake, say, start counting from 1 instead of 0, it can have disastrous consequences.
On the other hand, this kind of precision has implied a great deal of repetition. Various tasks are solved in humanly similar ways but computationally different ones. The creation of different programming languages with higher expressiveness has mitigated this to some extend, for example, the verbosity required to read an integer from the keyboard is different in assembly language than in Python, but the problem still persists. The more we perform related tasks the more we search for ways to abstract them and hence, simplify them.
I think the need for precision and the implication of detail and repetition contributes significantly in making programming a challenging and time consuming intellectual endeavor, and hence, difficult.
Why should we care?
As general public, why should we care about programming at all? let alone whether it is difficult or not? Advocates of Computational Thinking argue that Computational Thinking is required to function in modern society because computers are omnipresent. My counter-argument is: shoes are omnipresent and we are not required to understand how a shoe is made or repaired. However, shoes fulfill only one task while computer can fulfill many. Every time we use a computer we are programming it. The simple action of dialing a phone number on our cellphones trigger a series of program actions that accumulated allow us to call a number. It can be regarded as simple programming.
Since computers are omnipresent, we are constantly programming, following a series of steps to accomplish a goal, for example, if we want to connect to a WiFi network we first select the network and then type in the password, not the other way around, even though there is no computational limitation to do it in reverse. As general users, we are limited to the programs (applications) installed in our computers, and if we remove them, we are left with an expensive paper weight. For example, let's say a lawyer specialized in international commerce laws would like to keep track of new trading laws and amendments in different countries. Without a program to do so, or any programming capacity, the lawyer's only option is to manually check the new laws constantly. Despite the fact that this problem could be addressed with the simple solution of matching keywords, that solution is simply out of reach.
So, there is a need/demand for new programs, but there are not enough programmers to satisfy that demand, and as we already discussed, programs are not exactly fast to create. This makes learning to program an appealing proposition, however, as we have already pointed out, it is not an easy undertaking.
AI marketers are positioning it to fill this gap arguing that English, or any other natural language for that matter, will be the only programming language that you will need. I think Jensen Huang (NVIDIA CEO) said it at the London Tech Week (I'm not sure, I tried to find the video of the interview but I didn't find it fully nor a reliable source transcript). So far I have seen the discussion on this focusing on two simplistic extremes, on the one hand, programming is seeing as cumbersome syntax, on the other hand, programming is regarded as a highly specialized thinking process. Both are examples of the excluded-middle they are wrong and correct at the same time. Take our example of the lawyer, the program the lawyer would require is not at all a highly specialized thinking process, it is actually on the simple lane with the current available libraries, it is technically challenging, finding the correct libraries, gluing them together and so on, but it is not at all a intellectual endeavor. On the other hand, people that argues programming is cumbersome syntax conveniently forget the vast majority of simplified programming language available, mostly specialized to teach programming.
In the small clips that I saw, Jensen Huang selected his words carefully: "In the future, you'll tell the computer what you want" I like this quote because with my Introductory Programming students we always discuss what problems a computer can solve and how we tell it what we want. We always tell the computer to do something, the thing is, that we find it difficult to tell the computer what we want in a way that it can compute it. This quote was meant to be understood that telling means writing it in a natural language, in this case English. Huang also elaborates on how to achieve this communication with the computer thru English using AI and explains this is what Prompt Engineering is all about.
Let me make a detour here. The programming language of the future should not be only English, presumably the talk was addressed to native English speakers (if it was given at the London Tech Week), so it makes sense that you want to pitch a 'democratization' of programming without being 'overly democratic'. Nowadays we have computational translators, it should be, once more, an intellectually simple task to write an interface from your favorite natural language into English and finally into code generation. We may argue that translators are not very good, sometimes they translate literally instead of the intention, but programming texts are not made of cliches or literary beautified they are technical and as such easier to translate; I'm just saying. -End of detour-
The critics that I find more 'grounded' argue that English is not precise enough for that task and hence this would fail. We are not sure of this, but we have evidence of the shortcomings of natural languages when we consider precision.
Arguably, long before humans started to think about mathematics and logic, they had laws, and laws are expressed in natural language. Laws are one of our first attempts to reach precision with natural languages and so far, we have not reached a satisfactory point. We still have loop-holes, judges and lawyers constantly argue about 'the spirit/intent of the law'. However, laws and its discipline are logical (or should be). Having evidence A,B,C we can conclude D, classic Aristotelian Logic. It is not surprising that one of the first attempts of Expert Systems were to interpret law. However, back then, it was evident that while laws were logical, they also had a big influence of 'intent', ambiguity; big enough to render Expert Systems ineffective. So, we have massive evidence that our communication using natural language is imprecise. But, if we can accept a mathematics or theoretical computer science paper with small omissions or even tiny errors, as correct, couldn't be possible to create correct programs from an imprecise natural language?
Let's make a though experiment. An old-fashioned programmer writes in a programmers forum a solution to sort numbers in an array and describe it as: "The following function accepts a pointer to an array of numbers and the dimension of the array. It sorts the numbers in-place and returns a pointer to the array". The code is published and people consult it when they google "sort numbers in array", "sort array", "sort numeric array" and so on. Then come students, let's say high school students, that as part of an assignment they need to sort numbers in an array. They have their own forum, an online community. One of them, is fond of how movie-depicted gangsters talk and wants to exercise his skills and write: "yo man, I'm in deep need of some code to throw me my numbers from smallest to biggest and baddest bro! ASAP dude I need to pop it" One of the members, that is familiar with the code the old-fashioned programmer wrote replies: "Here you go" and attaches a link to the program and the original poster (aka OP) replies "Spot on dude! Thanks I owe you big!". Then another student that is fond of poetry might write: "Dear fellow members: To complete my task I require to ask How can I task My computer to rank The numbers I have." And once again he is pointed to the code and the brewing poet positively acknowledge the pointer. And like this many interactions happen simultaneously with all kinds of speech.
It is not brainier that an AI trained on this data will make all kinds of relations of different speeches to the sorting numbers code. So, when this AI is asked for a code to sort numbers it will be able to deliver the 'correct' code, even when our examples never used the word 'sorting' nor 'array'. But what happens if this code was and implementation of 'bubble sort' then the generated code will be 'highly inefficient'. Of course, AI advocates are hoping the dataset is diverse enough to be able to 'search', thru Prompt Engineering, for an efficient solution. If you don't know what a computationally efficient solution is or how it looks like, how can you direct the AI?
Yes, English allowed you to find A solution but not THE solution. If you want THE solution, you need THE RIGHT code-set, and THE RIGHT natural language (i.e. technical language) dataset. But to use this you need to know the correct technical language which the general public does not posses. I do not doubt that big tech companies use this kind of technology, perhaps trained solely on specialized data, e.g. Computer Science and Mathematics books and papers and code repositories, to generate and modify code; it is an tool for experts trained on experts generated data. I do not doubt that as a niche user case it should render adequate results, for example, if you want to factor a piece of a function that does X into a new function it is definitely simpler to write to a bot "Factor the functionality X to a new function funx" and the bot do the program analysis required to deduce the inputs and outputs (if the are not stated as pre/post-conditions in comments) and then produce the function.
The language is precise, but to reach that precision you need to know how to program, so, it is a chicken-egg type of problem.
We can argue that the general public would be happy with A solution instead of THE solution, hence voiding my argument. This definitely may help with sells, as slow solutions on old hardware are painful and getting the newer phone, the newest computer might accelerate that and hence increase our waste and our energy requirements.
The fact that we are having this discussion of 'programming democratization' shows the underlying problem, we, as computing experts, had been unable to raise the level of abstraction of our daily tools. I think this is an exciting endeavor, have we reached the abstraction limit? or can we get some inspiration from these developments and how they came to be to make a breakthrough and devise a simpler, more efficient and ultimately greener solution?
Or, in a few years time, will we be stuck in discussions about the intent of the prompt?
Why is it difficult?
We do not know exactly why programming is difficult, to the best of our knowledge after teaching it to the students to the best of our abilities, students fail tests in sufficient numbers allowing us to conclude it is a difficult task.
We could argue it is not the discipline but how it is taught, however, for the sake of the argument, let us assume it is not. Moreover, there exists an ample repertoire of techniques that had been devised and tested to teach programming.
While we do not know precisely why programming is difficult, we do know that it involves certain competences, and perhaps one of the most mentioned is the ability to solve problems, also known as problem solving.
Solving problems is the core of programming, currently if we do not have a problem to solve, there is little motivation to program a computer. But we solve problems on a daily basis. Solving problems with the computer requires another ability precision. When we express a solution to a problem as a computer program we are being precise, each instruction help us to arrive to a solution, to track this progress each instruction must fulfill two conditions: the first, it must be a valid instruction in a programming language and secondly the inputs-outputs relationship must have a meaning in our problem domain. So, in few words creating computer programs means solving problems precisely.
The precision requirement is what brings programming closer to exact sciences like mathematics than to any other disciplines, like, literature or law. Even within mathematics and theoretical computer science, we can consider programs to be (much) more precise than a mathematical proof (in here I'm not considering computer-checked proofs, I mean it in the more traditional way). A mathematical or theoretical computer science paper can make a small omission or even a small mistake and still be understood by the reader. If a computer program is missing a part, it simply won't work, if it does, then you can remove that part altogether and we can conclude the solution was not completely precise (which is fine), if a computer program has a small mistake, say, start counting from 1 instead of 0, it can have disastrous consequences.
On the other hand, this kind of precision has implied a great deal of repetition. Various tasks are solved in humanly similar ways but computationally different ones. The creation of different programming languages with higher expressiveness has mitigated this to some extend, for example, the verbosity required to read an integer from the keyboard is different in assembly language than in Python, but the problem still persists. The more we perform related tasks the more we search for ways to abstract them and hence, simplify them.
I think the need for precision and the implication of detail and repetition contributes significantly in making programming a challenging and time consuming intellectual endeavor, and hence, difficult.
Why should we care?
As general public, why should we care about programming at all? let alone whether it is difficult or not? Advocates of Computational Thinking argue that Computational Thinking is required to function in modern society because computers are omnipresent. My counter-argument is: shoes are omnipresent and we are not required to understand how a shoe is made or repaired. However, shoes fulfill only one task while computer can fulfill many. Every time we use a computer we are programming it. The simple action of dialing a phone number on our cellphones trigger a series of program actions that accumulated allow us to call a number. It can be regarded as simple programming.
Since computers are omnipresent, we are constantly programming, following a series of steps to accomplish a goal, for example, if we want to connect to a WiFi network we first select the network and then type in the password, not the other way around, even though there is no computational limitation to do it in reverse. As general users, we are limited to the programs (applications) installed in our computers, and if we remove them, we are left with an expensive paper weight. For example, let's say a lawyer specialized in international commerce laws would like to keep track of new trading laws and amendments in different countries. Without a program to do so, or any programming capacity, the lawyer's only option is to manually check the new laws constantly. Despite the fact that this problem could be addressed with the simple solution of matching keywords, that solution is simply out of reach.
So, there is a need/demand for new programs, but there are not enough programmers to satisfy that demand, and as we already discussed, programs are not exactly fast to create. This makes learning to program an appealing proposition, however, as we have already pointed out, it is not an easy undertaking.
AI marketers are positioning it to fill this gap arguing that English, or any other natural language for that matter, will be the only programming language that you will need. I think Jensen Huang (NVIDIA CEO) said it at the London Tech Week (I'm not sure, I tried to find the video of the interview but I didn't find it fully nor a reliable source transcript). So far I have seen the discussion on this focusing on two simplistic extremes, on the one hand, programming is seeing as cumbersome syntax, on the other hand, programming is regarded as a highly specialized thinking process. Both are examples of the excluded-middle they are wrong and correct at the same time. Take our example of the lawyer, the program the lawyer would require is not at all a highly specialized thinking process, it is actually on the simple lane with the current available libraries, it is technically challenging, finding the correct libraries, gluing them together and so on, but it is not at all a intellectual endeavor. On the other hand, people that argues programming is cumbersome syntax conveniently forget the vast majority of simplified programming language available, mostly specialized to teach programming.
In the small clips that I saw, Jensen Huang selected his words carefully: "In the future, you'll tell the computer what you want" I like this quote because with my Introductory Programming students we always discuss what problems a computer can solve and how we tell it what we want. We always tell the computer to do something, the thing is, that we find it difficult to tell the computer what we want in a way that it can compute it. This quote was meant to be understood that telling means writing it in a natural language, in this case English. Huang also elaborates on how to achieve this communication with the computer thru English using AI and explains this is what Prompt Engineering is all about.
Let me make a detour here. The programming language of the future should not be only English, presumably the talk was addressed to native English speakers (if it was given at the London Tech Week), so it makes sense that you want to pitch a 'democratization' of programming without being 'overly democratic'. Nowadays we have computational translators, it should be, once more, an intellectually simple task to write an interface from your favorite natural language into English and finally into code generation. We may argue that translators are not very good, sometimes they translate literally instead of the intention, but programming texts are not made of cliches or literary beautified they are technical and as such easier to translate; I'm just saying. -End of detour-
The critics that I find more 'grounded' argue that English is not precise enough for that task and hence this would fail. We are not sure of this, but we have evidence of the shortcomings of natural languages when we consider precision.
Arguably, long before humans started to think about mathematics and logic, they had laws, and laws are expressed in natural language. Laws are one of our first attempts to reach precision with natural languages and so far, we have not reached a satisfactory point. We still have loop-holes, judges and lawyers constantly argue about 'the spirit/intent of the law'. However, laws and its discipline are logical (or should be). Having evidence A,B,C we can conclude D, classic Aristotelian Logic. It is not surprising that one of the first attempts of Expert Systems were to interpret law. However, back then, it was evident that while laws were logical, they also had a big influence of 'intent', ambiguity; big enough to render Expert Systems ineffective. So, we have massive evidence that our communication using natural language is imprecise. But, if we can accept a mathematics or theoretical computer science paper with small omissions or even tiny errors, as correct, couldn't be possible to create correct programs from an imprecise natural language?
Let's make a though experiment. An old-fashioned programmer writes in a programmers forum a solution to sort numbers in an array and describe it as: "The following function accepts a pointer to an array of numbers and the dimension of the array. It sorts the numbers in-place and returns a pointer to the array". The code is published and people consult it when they google "sort numbers in array", "sort array", "sort numeric array" and so on. Then come students, let's say high school students, that as part of an assignment they need to sort numbers in an array. They have their own forum, an online community. One of them, is fond of how movie-depicted gangsters talk and wants to exercise his skills and write: "yo man, I'm in deep need of some code to throw me my numbers from smallest to biggest and baddest bro! ASAP dude I need to pop it" One of the members, that is familiar with the code the old-fashioned programmer wrote replies: "Here you go" and attaches a link to the program and the original poster (aka OP) replies "Spot on dude! Thanks I owe you big!". Then another student that is fond of poetry might write: "Dear fellow members: To complete my task I require to ask How can I task My computer to rank The numbers I have." And once again he is pointed to the code and the brewing poet positively acknowledge the pointer. And like this many interactions happen simultaneously with all kinds of speech.
It is not brainier that an AI trained on this data will make all kinds of relations of different speeches to the sorting numbers code. So, when this AI is asked for a code to sort numbers it will be able to deliver the 'correct' code, even when our examples never used the word 'sorting' nor 'array'. But what happens if this code was and implementation of 'bubble sort' then the generated code will be 'highly inefficient'. Of course, AI advocates are hoping the dataset is diverse enough to be able to 'search', thru Prompt Engineering, for an efficient solution. If you don't know what a computationally efficient solution is or how it looks like, how can you direct the AI?
Yes, English allowed you to find A solution but not THE solution. If you want THE solution, you need THE RIGHT code-set, and THE RIGHT natural language (i.e. technical language) dataset. But to use this you need to know the correct technical language which the general public does not posses. I do not doubt that big tech companies use this kind of technology, perhaps trained solely on specialized data, e.g. Computer Science and Mathematics books and papers and code repositories, to generate and modify code; it is an tool for experts trained on experts generated data. I do not doubt that as a niche user case it should render adequate results, for example, if you want to factor a piece of a function that does X into a new function it is definitely simpler to write to a bot "Factor the functionality X to a new function funx" and the bot do the program analysis required to deduce the inputs and outputs (if the are not stated as pre/post-conditions in comments) and then produce the function.
The language is precise, but to reach that precision you need to know how to program, so, it is a chicken-egg type of problem.
We can argue that the general public would be happy with A solution instead of THE solution, hence voiding my argument. This definitely may help with sells, as slow solutions on old hardware are painful and getting the newer phone, the newest computer might accelerate that and hence increase our waste and our energy requirements.
The fact that we are having this discussion of 'programming democratization' shows the underlying problem, we, as computing experts, had been unable to raise the level of abstraction of our daily tools. I think this is an exciting endeavor, have we reached the abstraction limit? or can we get some inspiration from these developments and how they came to be to make a breakthrough and devise a simpler, more efficient and ultimately greener solution?
Or, in a few years time, will we be stuck in discussions about the intent of the prompt?
It is not bad to say "I told you so".
Sometimes I'm surprised by how experts react to developments that were bound to happen. Frankly, I feel the reactions are signs of self-censorship that harm everyone.
I'm referring to the latest AI 'misuse' trend, the abuse of women by deep-fakes requested to Grok. To see why this was foreseeable, let's run a quick mental exercise.
Imagine you have 10 objects: Object1, Object2, and so on to Object10. Each object has the 'pleasantness' property. ObjectK.pleasantness is a number that reflects how pleasant ObjectK is. We can rewrite ObjectK.pleasantness by its proportion with respect to all the objects, i.e., ObjectK.pleasantness = ObjectK.pleasantness / Tp. Where TP = Object1.pleasantness + ... + Object10.pleasantness
Now the pleasantnesses (what a word) make a probability function, call it Pp.
Throw all the objects into an opaque urn and draw objects from the urn following Pp, that is, if ObjectI is more pleasant than ObjectL, then it is more likely to be drawn from the urn.
Once you draw an object, you log which object you drew and then return it to the urn. You keep going until you are tired.
Perhaps your pleasantness perception does not agree with Pp. For example, Pp might say Object7 is the most pleasant, but to you, Object2 is the most pleasant.
It is quite easy to go over your log and discard the Objects that you don't like or that you find less pleasant than others. This could be tiresome, especially if you disagree with Pp. If Object7 is the most likely and Object2 the least likely, you should expect to discard many more objects than the ones you keep.
On the other hand, if you somewhat agree with Pp, you might be happy with any draw.
Moreover, on your log, you could keep a record of the (partial) sum of the objects you have seen, instead of the full list.
For example, if your original log showed the 4th draw as Object7, Object5, Object2, it will now show Object7+Object5+Object2, assuming objects can be added.
Let's complicate this experiment a bit more. Imagine the objects are colors, and you throw "magical" 1-unit colors into the Main urn. You also have a second opaque urn that we will call "Mixing urn". Each time you draw a magical 1-unit color from the Main urn, you empty its contents into the Mixing urn and then return the empty unit into the Main urn. Since the unit is magical, it will refill with its color once it is inside the Main urn.
You cannot peek inside the Mixing urn, and you cannot see which color you draw from the Main urn. Colors are drawn following Pp.
The Mixing urn will give you a 1-unit color, the result of mixing all the colors you drew. Once you mix two colors, you can not separate them, so by only seeing the resulting color, you can not determine the proportions of each color.
Let's say you know the 10 base colors, you are allowed to create as many mixed colors as you wish, but you do not know Pp. Would you be able to make "good" guesses about Pp? Probably, let’s say Red is one of the most liked colors, if all the mixed colors seem to have a 'red touch' (whatever that means), you could conclude that Red must have a high probability. On the other hand, if all the colors seem to be bright, you could conclude that black or dark blues (assuming they are base colors) must have low probabilities.
Probably now you are devising how well you can approximate Pp by playing with the base colors.
For example, you can take all the subsets of 2 and create mixed colors, see what happens, then all subsets of 3, and so on.
Final complication. Imagine you add a new dimension: "intensity" to each base color. How many? To keep it 'real', let's use the RGBA model and consider 256 different intensities. Now we have 2560 'different' colors and Pp is adjusted to consider the intensity. For example, light blue could be more pleasant than dark blue, however, both are 'blue'. If we repeat the previous experiment, deducing the proportions from the mixed color seems to become a bit more difficult problem; in essence, it is the same problem, but with bigger numbers.
Now, what does this mental exercise have to do with Grok and corporate AI in general? Well, our mental exercise is an oversimplification of how dominant AI models work: Each time you give the model a prompt (a series of draws in our example), the model gives you a response (a magical mixed color). We do not know the precise characteristics of the training data (base colors' pleasantness), but we may know some of the base colors (academic, formal, humorous, sexist, racist). Moreover, our own pleasantness concept may not agree at all with the underlying pleasantness of the training set. But just as in our mental exercise, we could draw some conclusions about the pleasantness of the categories in the training set by examining the responses to various prompts. However, it is impossible to filter out categories (colors) from the response (mixed color).
The signs were/are all over, or, as I said at the beginning, "I told you so".
Remember a famous Time article about OpenAI hiring Kenyans to make ChatGPT 3 'less toxic' from 2 or 3 years ago? If you create and create and create mixed colors over and over and over again, and all seem to have a Red component, what can you conclude?
If you have enough 'problems' with what your AI generates, big enough to devise a way to make it less toxic, is your training set valuable? I'll give you a sec to think about this one from a social and corporate perspective.
Perhaps you are thinking what I'm thinking, we can fix it, we just did it in our colors mental exercise, just remove the base colors we don't want. Yes, we can do that, but there are 2 things to consider. 1) Read this post backwards. If we consider colors and intensities, it is difficult to correctly classify them: is this Red, or Rose, or Light Red?
2) If you remove that which has a higher probability of being liked, what does it mean to your average user, to your business?
As biases in AI models came to light for the general public, experts' debates often focused on the implications of using these tools for society. How it reinforced stereotypes, how it was detrimental for minorities, e.g., AI-approved bank loans. However, let us not forget that, with our "Internet data trained" AI models, we are being forced to look at ourselves in a mirror that, many times, enhances a reflection we have conveniently decided to bury; a reflection that scares us.
The Internet didn't build itself; we built it.
If we want better AI models, we need better training data.
To get better massive training data, we need a better Internet content.
To get a better Internet content, we need a healthier society.
Sometimes I'm surprised by how experts react to developments that were bound to happen. Frankly, I feel the reactions are signs of self-censorship that harm everyone.
I'm referring to the latest AI 'misuse' trend, the abuse of women by deep-fakes requested to Grok. To see why this was foreseeable, let's run a quick mental exercise.
Imagine you have 10 objects: Object1, Object2, and so on to Object10. Each object has the 'pleasantness' property. ObjectK.pleasantness is a number that reflects how pleasant ObjectK is. We can rewrite ObjectK.pleasantness by its proportion with respect to all the objects, i.e., ObjectK.pleasantness = ObjectK.pleasantness / Tp. Where TP = Object1.pleasantness + ... + Object10.pleasantness
Now the pleasantnesses (what a word) make a probability function, call it Pp.
Throw all the objects into an opaque urn and draw objects from the urn following Pp, that is, if ObjectI is more pleasant than ObjectL, then it is more likely to be drawn from the urn.
Once you draw an object, you log which object you drew and then return it to the urn. You keep going until you are tired.
Perhaps your pleasantness perception does not agree with Pp. For example, Pp might say Object7 is the most pleasant, but to you, Object2 is the most pleasant.
It is quite easy to go over your log and discard the Objects that you don't like or that you find less pleasant than others. This could be tiresome, especially if you disagree with Pp. If Object7 is the most likely and Object2 the least likely, you should expect to discard many more objects than the ones you keep.
On the other hand, if you somewhat agree with Pp, you might be happy with any draw.
Moreover, on your log, you could keep a record of the (partial) sum of the objects you have seen, instead of the full list.
For example, if your original log showed the 4th draw as Object7, Object5, Object2, it will now show Object7+Object5+Object2, assuming objects can be added.
Let's complicate this experiment a bit more. Imagine the objects are colors, and you throw "magical" 1-unit colors into the Main urn. You also have a second opaque urn that we will call "Mixing urn". Each time you draw a magical 1-unit color from the Main urn, you empty its contents into the Mixing urn and then return the empty unit into the Main urn. Since the unit is magical, it will refill with its color once it is inside the Main urn.
You cannot peek inside the Mixing urn, and you cannot see which color you draw from the Main urn. Colors are drawn following Pp.
The Mixing urn will give you a 1-unit color, the result of mixing all the colors you drew. Once you mix two colors, you can not separate them, so by only seeing the resulting color, you can not determine the proportions of each color.
Let's say you know the 10 base colors, you are allowed to create as many mixed colors as you wish, but you do not know Pp. Would you be able to make "good" guesses about Pp? Probably, let’s say Red is one of the most liked colors, if all the mixed colors seem to have a 'red touch' (whatever that means), you could conclude that Red must have a high probability. On the other hand, if all the colors seem to be bright, you could conclude that black or dark blues (assuming they are base colors) must have low probabilities.
Probably now you are devising how well you can approximate Pp by playing with the base colors.
For example, you can take all the subsets of 2 and create mixed colors, see what happens, then all subsets of 3, and so on.
Final complication. Imagine you add a new dimension: "intensity" to each base color. How many? To keep it 'real', let's use the RGBA model and consider 256 different intensities. Now we have 2560 'different' colors and Pp is adjusted to consider the intensity. For example, light blue could be more pleasant than dark blue, however, both are 'blue'. If we repeat the previous experiment, deducing the proportions from the mixed color seems to become a bit more difficult problem; in essence, it is the same problem, but with bigger numbers.
Now, what does this mental exercise have to do with Grok and corporate AI in general? Well, our mental exercise is an oversimplification of how dominant AI models work: Each time you give the model a prompt (a series of draws in our example), the model gives you a response (a magical mixed color). We do not know the precise characteristics of the training data (base colors' pleasantness), but we may know some of the base colors (academic, formal, humorous, sexist, racist). Moreover, our own pleasantness concept may not agree at all with the underlying pleasantness of the training set. But just as in our mental exercise, we could draw some conclusions about the pleasantness of the categories in the training set by examining the responses to various prompts. However, it is impossible to filter out categories (colors) from the response (mixed color).
The signs were/are all over, or, as I said at the beginning, "I told you so".
Remember a famous Time article about OpenAI hiring Kenyans to make ChatGPT 3 'less toxic' from 2 or 3 years ago? If you create and create and create mixed colors over and over and over again, and all seem to have a Red component, what can you conclude?
If you have enough 'problems' with what your AI generates, big enough to devise a way to make it less toxic, is your training set valuable? I'll give you a sec to think about this one from a social and corporate perspective.
Perhaps you are thinking what I'm thinking, we can fix it, we just did it in our colors mental exercise, just remove the base colors we don't want. Yes, we can do that, but there are 2 things to consider. 1) Read this post backwards. If we consider colors and intensities, it is difficult to correctly classify them: is this Red, or Rose, or Light Red?
2) If you remove that which has a higher probability of being liked, what does it mean to your average user, to your business?
As biases in AI models came to light for the general public, experts' debates often focused on the implications of using these tools for society. How it reinforced stereotypes, how it was detrimental for minorities, e.g., AI-approved bank loans. However, let us not forget that, with our "Internet data trained" AI models, we are being forced to look at ourselves in a mirror that, many times, enhances a reflection we have conveniently decided to bury; a reflection that scares us.
The Internet didn't build itself; we built it.
If we want better AI models, we need better training data.
To get better massive training data, we need a better Internet content.
To get a better Internet content, we need a healthier society.
It is January 3rd 2026, so far I had 2 topics in the queue that I
wanted to write about, but today it made 3 topics. What an eventful
start of the year!
ACM Digital Library Goes "Open Access"
When I was a bachelor's student, one of my professors said, "If you publish in the ACM, you are somebody in computing." That statement stuck with me. I became a member when I was a bachelor's student. The main reason to be a member was not the conference benefits but the ability to read any paper published in the Digital Library (DL). If you are a member, you can download any paper. They also had agreements with book publishers, giving you more access to computing literature. That was a lot of resources for about 50 USD.
A few years back the "publishers crisis" hit and hit hard. In a nutshell, a publisher's job is to publish papers while ensuring access to every published paper. In the old days (pre-computing dominance), this was mostly manual; in particular, a physical library must have existed somewhere to ensure access. As a publisher, you must also ensure your publications meet a certain level of "quality". To meet this goal, publishers hired editors who decide what gets published and how a prospective paper should be modified to meet the quality desired. Editors rely on reviewers to perform this task, so the basic workflow is writers -> editor -> reviewers -> editor -> writers and, so on.
The same overall structure is adopted for conferences. If a conference is backed by a publisher, e.g., Springer, Elsevier, IEEE, ACM, etc., authors have a certain "quality" assurance, but mostly a certain liveness assurance; their paper would not be lost in 10 or 15 years. In the era of computing dominance, authors and readers get an extra assurance: instant availability. Before computers and the Internet, if you were an academic and wanted to read a paper published at a conference held abroad, you had to file a petition to your university's library, which in turn filed a petition (by physical mail) to the library of the hosting university. The hosting university staff would search their records, make a copy of the requested paper, and send it back (again by physical mail) to the requesting library, which would finally deliver it to the academic; this process, more or less, was followed for each and every paper you required a copy of. Nowadays, you go to the publisher's Webpage, search for the paper, and click on 'PDF' and seconds (at most) later, you hold a copy of the published paper.
Publishers then act as middlemen: readers rely on them to access papers, and writers rely on them to publish papers. And so, it is not difficult to see the power-position publishers have, and they make the most of it: access fees. This model contradicts the popular belief that science should be "free" (as in freedom); mostly performed by scientists working at (public) universities, the results should be for the benefit of all (mankind). Access fees, also known as 'paywalls', intuitively contradict this essence. However, the mere existence of scientific publishers is evidence of how the main actors of scientific creation neglected the proper keeping of their legacy. By "main actors", I encompass "all": from university management down to writers. In theory, it is evident that we do not require publishers, if all (and by all I mean ALL, not just the ones we like) universities of the world shared a common knowledge library. If by any chance this seems like a strange idea, think about blockchain and cryptocurrencies :) A library owned by all, and cared for by all.
Going back to the main topic. I became aware of a petition at iptetitions.com. If you can not access it, you can read the petition here. Written by Terence Patrick Kelly, a well-known computer scientist. The ACM steering committee decided to make metrics, 'advanced search', and comments available only to members; previously they were available to all. You can read about this on the current ACM Premium description here. In brief, that means if you are a "Basic Edition" user and you visit the page of a paper, you can no longer write a comment, and (perhaps) more importantly, you can not see how many times the paper has been downloaded and cited.
The petition is a call to take that back.
It has been some time since I downloaded papers by searching the DL directly. Like most people (or so I thought), I use Google Scholar and download the papers from there. I mostly use a publisher's website to download the paper's BibTeX entry. I learned the hard way that Google Scholar is not the best source for this kind of information. However, I very little 'action' in the comments section when I visit the DL page, so, I wouldn't mind having it entirely removed.
I regard the citation download metrics more like 'gossip' than something truly meaningful. But I have noticed them, and I admit they have sometimes biased my logic: "Oh, this must be good if it was downloaded 1K times last week". It is kind of when you see a post on your feed (now called timeline) with a lot of likes. Likes would be times downloaded, citations, perhaps comments? I could also live without them.
AI-generated article summaries are another Premium feature. I could definitely live without them. I mean, is the ACM's ruling body aware that papers have a section called Abstract/Résumé?
Bulk citation exports and PDF downloads. That could be a downside for me. I have folders with relevant papers in PDF that I can consult at any time, so if I cannot download PDFs, that could be a problem for me.
Advanced search is also a miss for me. I mostly search for papers published within a certain date range; without it, I'm shown all papers published, and well, those are a lot. I think this is detrimental to the spirit of fast availability; you are unable to search for papers published in a given venue or within a given time frame. The ACM is truly hurting itself because it cannot provide accurate access to its Open Access papers; any computer scientist would laugh at how difficult it is to access the information within its main society.
On the last two points alone, I should be signing the petition, but I did not. The reason is: I agree with the goals, but I do not agree with the reasons given in the petition. It reads like an influencer-written complaint: if people can not see my likes and posts, they can't know I'm an influencer.
"Author recognition is the social contract that binds writers into a sustainable community: Authors fill the ACM Digital Library by donating articles, and the DL rewards authors by enabling readers and prospective employers to find and appraise authors' contributions. The new paywall voids this social contract, shortchanging authors and removing much of the incentive that formerly motivated them."
I'm probably very naive, but I don't write to get some metrics that anyone can see. I do understand, on the other hand, that you may want to evaluate a prospective employee by checking those metrics. However, in that case, I expect the employer to have greater access to the world's leading computing association than the Basic one. Hence, you would not have any problem at all.
This statement also raises an interesting question: How many authors are affiliated with institutions that DO NOT have an institutional ACM DL access? I would expect the number to be small.
"Volunteer labor is the foundation of ACM's existence. The new paywall attempts to monetize volunteers' contributions, which were intended for the benefit of the public."
Just recently, the ACM moved to Open Access only after intensive pressure to do so. Unlike other Editors, the ACM is a "society"; hence, at its core, it should not be a business. Therefore, going Open Access was the right move; should statistics be "open" as well?
The short answer is yes, just because if you collect data from "free" raw data, the results should also be "free". Not because you lose visibility to the "outside world".
The move is clearly detrimental if you judge by numbers.
After reading this petition, I must say that hiding stats would be a good idea; it may force readers to actually read the paper and decide for themselves, instead of "going with the flow". If you are an employer and read the paper, you get to decide whether it is a good paper, rather than relying on "it has X citations, so it must be good". Especially after we have seen how these metrics have been abused.
Kongens nytårstale
While you live in a new place, you get to pick some of the traditions of the place you are living in.
In my case, it was the Queen (now King) of Denmark's New Year's Eve speech. I learned that it is tradition to guess the topics the King would address in his speech. This year, an obvious choice would be Greenland (Grønland in Danish).
The King visited Greenland during the summer, but there is also the constant threat from the USA to get it "One way or the other". So it was obvious that a clear message should be sent. But it seems the message was not exactly what was expected: a brief mention of Greenland and the danger lurking... unaddressed.
At first, I thought that my limited Danish prevented me from capturing more, but then I saw a couple of news reports at DR highlighting this.
The speech of the Prime Minister Mette Frederiksen did not seem to add anything new, but it was not as tepid as the one from the King (to my view of course). The policy of silence is worrisome. Specially when your "closest ally" (nærmeste allierede in danish) is the one that just put a stunt (see next).
Make America Great Again
I have always wondered what Donald Trump actually means by "Great Again". To what greatness is he referring? I'm not a Trump expert, but to the best of my knowledge, he hasn't actually defined a time frame for the greatness he would like to transition the USA back to. Perhaps he grew up with his grandfather's stories about Andrew Carnegie and John D. Rockefeller, and that is the time frame he would like his country to go back to. Or perhaps he would like to go back to the early 50's or 60's, I don't know.
It actually doesn't really matter what he means; it is the end, and "the end justifies the means", and the actual focus should be on "the means" as they are today's end.
It is not a secret that the República Bolivariana de Venezuela, commonly known as Venezuela, is an oil-rich country. And oil is craved by Trump's USA.
Trump's logic is plain: we need oil, Venezuela has it, we try to seize power by proxy (i.e. we back-up a Venezuelan citizen to seize power), we fail, we declare a Venezuelan cartel a terrorist organization, we blow-up some boats near their shores, we attack a sovereign country and kidnap its president, and declare ourselves "managers of the country". Does this sound remotely familiar to you?
Not so long ago, almost four years ago, a country attacked another, and a bunch of countries cried to the heavens. It was unacceptable, unimaginable that a country wanted to annex another, the drums of war could be heard, money donations, bullets donations were made, and are still being made.
In this case? Keir Starmer would like to "clarify the facts"; Ursula von der Leyen: "We stand by the people of Venezuela and support a peaceful and democratic transition. Any solution must respect international law and the UN Charter". As we say in Mexico, "lo caido caido" or tough luck, better luck next time. It is laughable, but following the playbook to the letter.
What happened and is planned has a very common name: theft. At his press conference, Trump didn't specify how long the occupation will last, but a good guess can already be made: as long as it takes to bleed out the country of its oil reserves, not a minute longer, not a minute shorter. It is not like Trump is in the business of trying to annex countries or part of them to the 50 stars flag, oh, with the obvious exception, of course.
I kind of expected more from Trump. I would have respected him more if he had come out a year earlier and said something like "I want Venezuela's oil, and we are going to get it one way or the other. It is vital for our interests." And then come a year later and say: "See? I told you we were going to get it." That straight, honest, and confrontational attitude I could respect, and I would entertain the idea that he might be able to "Make America Great Again". Instead, he beats around the bush, hiding behind the narco-whatever rhetoric. I think I created that expectation after seeing him behave with Ukraine's president, stating plainly that they would squeeze the best deal possible out of him. Or seeing him threaten an ally, who buys his combat planes, with getting a piece of their territory. I think it is easy to threaten those who are already under your thumb.
Don't get me wrong, I'm not downplaying the determination and courage shown by the USA president. It takes both for a 79-year-old man, apparently not at the peak of his health, to sit down for 3+ hours and follow through a TV show, which is how he described how he felt watching the feed.
My grandmother had trouble following an action movie plot when she indulged me and watched it with me, and action movies are bite-sized. I imagine real-time military feeds are much more complicated, and to be able to follow them, regulating your emotions and your bodily responses must require paramount determination and courage.
Voices within the USA spoke against the attack, however, I wonder how long it will take to silence the internal opposition. Perhaps when the projected financial gains are shared, the internal disagreement will reconsider its position. Who knows, perhaps it is a big enough gain to put the Epstein's files to rest (for a while at least).
I'm not an expert on Venezuela, nor on Trump. I do not know how much the recent 20 years of occupation taught the USA about bleeding out natural resources.
I do think that, among the top administration that approved the strike, the logic was as simple as in a Vin Diesel movie: wander around cartel territory to arrest the kingpin, and everyone will be cool about it, embracing the deliverance. Reality, however, is a little bit more complicated than a movie.
It remains to be seen how much more complicated it will be in this case.
The day the Ukraine war began, my wife and I were talking about it. I conjectured that Ukraine would end like Germany after WWII. I'm not a consummated politician nor a political science professor, but I wasn't born yesterday either. If I recall correctly, the colorful character Henry Kissinger stated something like the USA (or EU) should negotiate with Russia. Of course, he got crucified. It turns out that is exactly what is happening. Now, I will use my "I wasn't born yesterday" powers to conjecture that Trump really, really wants the Venezuelan succession government to work, because it is in his best interest to minimize the troops' presence in Venezuela. Not because he is afraid, but because he has 3 years left as president, he can not afford a mishap that would cost time. He has his sights set on other countries like Colombia, Nicaragua, Brazil, Cuba, and, of course, Mexico; as well as his strong interest in adding a new star to the USA flag. Therefore, he requires his army to be intact. He is betting/hoping that no country would put up a fight. He wants his legacy to be the one that brought stability to the USA's influence sphere and, surprise, surprise, "Make America Great Again". However, we might be witnessing the last attempt of a dying empire.
ACM Digital Library Goes "Open Access"
When I was a bachelor's student, one of my professors said, "If you publish in the ACM, you are somebody in computing." That statement stuck with me. I became a member when I was a bachelor's student. The main reason to be a member was not the conference benefits but the ability to read any paper published in the Digital Library (DL). If you are a member, you can download any paper. They also had agreements with book publishers, giving you more access to computing literature. That was a lot of resources for about 50 USD.
A few years back the "publishers crisis" hit and hit hard. In a nutshell, a publisher's job is to publish papers while ensuring access to every published paper. In the old days (pre-computing dominance), this was mostly manual; in particular, a physical library must have existed somewhere to ensure access. As a publisher, you must also ensure your publications meet a certain level of "quality". To meet this goal, publishers hired editors who decide what gets published and how a prospective paper should be modified to meet the quality desired. Editors rely on reviewers to perform this task, so the basic workflow is writers -> editor -> reviewers -> editor -> writers and, so on.
The same overall structure is adopted for conferences. If a conference is backed by a publisher, e.g., Springer, Elsevier, IEEE, ACM, etc., authors have a certain "quality" assurance, but mostly a certain liveness assurance; their paper would not be lost in 10 or 15 years. In the era of computing dominance, authors and readers get an extra assurance: instant availability. Before computers and the Internet, if you were an academic and wanted to read a paper published at a conference held abroad, you had to file a petition to your university's library, which in turn filed a petition (by physical mail) to the library of the hosting university. The hosting university staff would search their records, make a copy of the requested paper, and send it back (again by physical mail) to the requesting library, which would finally deliver it to the academic; this process, more or less, was followed for each and every paper you required a copy of. Nowadays, you go to the publisher's Webpage, search for the paper, and click on 'PDF' and seconds (at most) later, you hold a copy of the published paper.
Publishers then act as middlemen: readers rely on them to access papers, and writers rely on them to publish papers. And so, it is not difficult to see the power-position publishers have, and they make the most of it: access fees. This model contradicts the popular belief that science should be "free" (as in freedom); mostly performed by scientists working at (public) universities, the results should be for the benefit of all (mankind). Access fees, also known as 'paywalls', intuitively contradict this essence. However, the mere existence of scientific publishers is evidence of how the main actors of scientific creation neglected the proper keeping of their legacy. By "main actors", I encompass "all": from university management down to writers. In theory, it is evident that we do not require publishers, if all (and by all I mean ALL, not just the ones we like) universities of the world shared a common knowledge library. If by any chance this seems like a strange idea, think about blockchain and cryptocurrencies :) A library owned by all, and cared for by all.
Going back to the main topic. I became aware of a petition at iptetitions.com. If you can not access it, you can read the petition here. Written by Terence Patrick Kelly, a well-known computer scientist. The ACM steering committee decided to make metrics, 'advanced search', and comments available only to members; previously they were available to all. You can read about this on the current ACM Premium description here. In brief, that means if you are a "Basic Edition" user and you visit the page of a paper, you can no longer write a comment, and (perhaps) more importantly, you can not see how many times the paper has been downloaded and cited.
The petition is a call to take that back.
It has been some time since I downloaded papers by searching the DL directly. Like most people (or so I thought), I use Google Scholar and download the papers from there. I mostly use a publisher's website to download the paper's BibTeX entry. I learned the hard way that Google Scholar is not the best source for this kind of information. However, I very little 'action' in the comments section when I visit the DL page, so, I wouldn't mind having it entirely removed.
I regard the citation download metrics more like 'gossip' than something truly meaningful. But I have noticed them, and I admit they have sometimes biased my logic: "Oh, this must be good if it was downloaded 1K times last week". It is kind of when you see a post on your feed (now called timeline) with a lot of likes. Likes would be times downloaded, citations, perhaps comments? I could also live without them.
AI-generated article summaries are another Premium feature. I could definitely live without them. I mean, is the ACM's ruling body aware that papers have a section called Abstract/Résumé?
Bulk citation exports and PDF downloads. That could be a downside for me. I have folders with relevant papers in PDF that I can consult at any time, so if I cannot download PDFs, that could be a problem for me.
Advanced search is also a miss for me. I mostly search for papers published within a certain date range; without it, I'm shown all papers published, and well, those are a lot. I think this is detrimental to the spirit of fast availability; you are unable to search for papers published in a given venue or within a given time frame. The ACM is truly hurting itself because it cannot provide accurate access to its Open Access papers; any computer scientist would laugh at how difficult it is to access the information within its main society.
On the last two points alone, I should be signing the petition, but I did not. The reason is: I agree with the goals, but I do not agree with the reasons given in the petition. It reads like an influencer-written complaint: if people can not see my likes and posts, they can't know I'm an influencer.
"Author recognition is the social contract that binds writers into a sustainable community: Authors fill the ACM Digital Library by donating articles, and the DL rewards authors by enabling readers and prospective employers to find and appraise authors' contributions. The new paywall voids this social contract, shortchanging authors and removing much of the incentive that formerly motivated them."
I'm probably very naive, but I don't write to get some metrics that anyone can see. I do understand, on the other hand, that you may want to evaluate a prospective employee by checking those metrics. However, in that case, I expect the employer to have greater access to the world's leading computing association than the Basic one. Hence, you would not have any problem at all.
This statement also raises an interesting question: How many authors are affiliated with institutions that DO NOT have an institutional ACM DL access? I would expect the number to be small.
"Volunteer labor is the foundation of ACM's existence. The new paywall attempts to monetize volunteers' contributions, which were intended for the benefit of the public."
Just recently, the ACM moved to Open Access only after intensive pressure to do so. Unlike other Editors, the ACM is a "society"; hence, at its core, it should not be a business. Therefore, going Open Access was the right move; should statistics be "open" as well?
The short answer is yes, just because if you collect data from "free" raw data, the results should also be "free". Not because you lose visibility to the "outside world".
The move is clearly detrimental if you judge by numbers.
After reading this petition, I must say that hiding stats would be a good idea; it may force readers to actually read the paper and decide for themselves, instead of "going with the flow". If you are an employer and read the paper, you get to decide whether it is a good paper, rather than relying on "it has X citations, so it must be good". Especially after we have seen how these metrics have been abused.
Kongens nytårstale
While you live in a new place, you get to pick some of the traditions of the place you are living in.
In my case, it was the Queen (now King) of Denmark's New Year's Eve speech. I learned that it is tradition to guess the topics the King would address in his speech. This year, an obvious choice would be Greenland (Grønland in Danish).
The King visited Greenland during the summer, but there is also the constant threat from the USA to get it "One way or the other". So it was obvious that a clear message should be sent. But it seems the message was not exactly what was expected: a brief mention of Greenland and the danger lurking... unaddressed.
At first, I thought that my limited Danish prevented me from capturing more, but then I saw a couple of news reports at DR highlighting this.
The speech of the Prime Minister Mette Frederiksen did not seem to add anything new, but it was not as tepid as the one from the King (to my view of course). The policy of silence is worrisome. Specially when your "closest ally" (nærmeste allierede in danish) is the one that just put a stunt (see next).
Make America Great Again
I have always wondered what Donald Trump actually means by "Great Again". To what greatness is he referring? I'm not a Trump expert, but to the best of my knowledge, he hasn't actually defined a time frame for the greatness he would like to transition the USA back to. Perhaps he grew up with his grandfather's stories about Andrew Carnegie and John D. Rockefeller, and that is the time frame he would like his country to go back to. Or perhaps he would like to go back to the early 50's or 60's, I don't know.
It actually doesn't really matter what he means; it is the end, and "the end justifies the means", and the actual focus should be on "the means" as they are today's end.
It is not a secret that the República Bolivariana de Venezuela, commonly known as Venezuela, is an oil-rich country. And oil is craved by Trump's USA.
Trump's logic is plain: we need oil, Venezuela has it, we try to seize power by proxy (i.e. we back-up a Venezuelan citizen to seize power), we fail, we declare a Venezuelan cartel a terrorist organization, we blow-up some boats near their shores, we attack a sovereign country and kidnap its president, and declare ourselves "managers of the country". Does this sound remotely familiar to you?
Not so long ago, almost four years ago, a country attacked another, and a bunch of countries cried to the heavens. It was unacceptable, unimaginable that a country wanted to annex another, the drums of war could be heard, money donations, bullets donations were made, and are still being made.
In this case? Keir Starmer would like to "clarify the facts"; Ursula von der Leyen: "We stand by the people of Venezuela and support a peaceful and democratic transition. Any solution must respect international law and the UN Charter". As we say in Mexico, "lo caido caido" or tough luck, better luck next time. It is laughable, but following the playbook to the letter.
What happened and is planned has a very common name: theft. At his press conference, Trump didn't specify how long the occupation will last, but a good guess can already be made: as long as it takes to bleed out the country of its oil reserves, not a minute longer, not a minute shorter. It is not like Trump is in the business of trying to annex countries or part of them to the 50 stars flag, oh, with the obvious exception, of course.
I kind of expected more from Trump. I would have respected him more if he had come out a year earlier and said something like "I want Venezuela's oil, and we are going to get it one way or the other. It is vital for our interests." And then come a year later and say: "See? I told you we were going to get it." That straight, honest, and confrontational attitude I could respect, and I would entertain the idea that he might be able to "Make America Great Again". Instead, he beats around the bush, hiding behind the narco-whatever rhetoric. I think I created that expectation after seeing him behave with Ukraine's president, stating plainly that they would squeeze the best deal possible out of him. Or seeing him threaten an ally, who buys his combat planes, with getting a piece of their territory. I think it is easy to threaten those who are already under your thumb.
Don't get me wrong, I'm not downplaying the determination and courage shown by the USA president. It takes both for a 79-year-old man, apparently not at the peak of his health, to sit down for 3+ hours and follow through a TV show, which is how he described how he felt watching the feed.
My grandmother had trouble following an action movie plot when she indulged me and watched it with me, and action movies are bite-sized. I imagine real-time military feeds are much more complicated, and to be able to follow them, regulating your emotions and your bodily responses must require paramount determination and courage.
Voices within the USA spoke against the attack, however, I wonder how long it will take to silence the internal opposition. Perhaps when the projected financial gains are shared, the internal disagreement will reconsider its position. Who knows, perhaps it is a big enough gain to put the Epstein's files to rest (for a while at least).
I'm not an expert on Venezuela, nor on Trump. I do not know how much the recent 20 years of occupation taught the USA about bleeding out natural resources.
I do think that, among the top administration that approved the strike, the logic was as simple as in a Vin Diesel movie: wander around cartel territory to arrest the kingpin, and everyone will be cool about it, embracing the deliverance. Reality, however, is a little bit more complicated than a movie.
It remains to be seen how much more complicated it will be in this case.
The day the Ukraine war began, my wife and I were talking about it. I conjectured that Ukraine would end like Germany after WWII. I'm not a consummated politician nor a political science professor, but I wasn't born yesterday either. If I recall correctly, the colorful character Henry Kissinger stated something like the USA (or EU) should negotiate with Russia. Of course, he got crucified. It turns out that is exactly what is happening. Now, I will use my "I wasn't born yesterday" powers to conjecture that Trump really, really wants the Venezuelan succession government to work, because it is in his best interest to minimize the troops' presence in Venezuela. Not because he is afraid, but because he has 3 years left as president, he can not afford a mishap that would cost time. He has his sights set on other countries like Colombia, Nicaragua, Brazil, Cuba, and, of course, Mexico; as well as his strong interest in adding a new star to the USA flag. Therefore, he requires his army to be intact. He is betting/hoping that no country would put up a fight. He wants his legacy to be the one that brought stability to the USA's influence sphere and, surprise, surprise, "Make America Great Again". However, we might be witnessing the last attempt of a dying empire.
A couple of days ago (give or take) my social media feed showed a post
that goes like this: Here is a real task from my job. I have a 100GB binary
file. Produced daily. I can't grep it. But I can decode it.
However, I can't store the decoded version either. It's too big.
How do I efficiently query it?
Decoding piped to grep takes 2 minutes. I want 2 seconds.
When I saw this post, my initial reaction was to question the approach. Unless the binary file is highly encoded, for example, encrypted, it seems inefficient to decode it to plain text and then search with grep; especially within a company that reported a revenue in the billions last year. While this method may be understandable for those with limited resources, one would expect a much more optimized solution in a large enterprise setting.
The engagement generated by the post led the author to disclose more details, for example, that this file was generated by a microservice. And that it had a record-type structure with various record types. (1) We all know what microservices use to talk to each other, right? (2) 100GB files seem to defeat the micro purpose, but that is relative. Also, someone suggested to grep the binary file, but the author seemed to not be aware of such a tool, and the reply was something along the lines of wanting to avoid translating the grep queries into binary grep ones. This seems reasonable if the queries are complex, which is kind of expected.
My suspicions were confirmed when the author disclosed that the decoded file is a JSON file worth at least 1 TB. Then I thought "Who greps such a big file when there are a lot of efficient tools that parse JSON?, again, you must be kidding, right? I mean, you are aware that there are a bunch of smart folks who create software specialized in storing and querying data, right?" But again, probably you don't want to add an extra layer of complexity, so you prefer to stick to big file transactions, or perhaps (and more likely) a high number of small transactions.
The cherry of the cake was when the author posted a follow-up post: You just gotta love how many people jump with solutions without clarifying requirement. This is a common mistake I see while interviewing people at *****. People propose solutions without trying to clarify constraints, or input data patterns, or edge cases.
So, this might as well have been an interview question and an interview behavior. I think it is troublesome. If you are given a problem to solve as part of an interview for a job about solving problems, the first natural reaction is to attempt to solve the problem with the information you are given. Moreover, if the person doing the interview is your prospective immediate supervisor, then it is reasonable to assume you are given all the information necessary or at least available at the time to solve the task. I think most people try to be polite and won't tell the interviewer, "Who the f..k does that? Are you incompetent or just plain lazy?" even if they think about it.
If you are given a half-baked problem by the person who is going to supervise your work, then you, as a prospective employee, can easily make the conclusion "this guy is plain incompetent, he wants me to boost his production by more than 10x without having a complete idea of the problem he wants to solve." If you want to assess the candidate's ability to correctly frame a problem, there are many more meaningful ways to do so.
Interviews go both ways, just as the interviewer concludes the candidate makes a mistake by not asking more questions about the problem, the candidate (even if they ask questions) can easily conclude that the interviewer is lazy, incompetent, or withholds information on purpose to other colleagues.
When I saw this post, my initial reaction was to question the approach. Unless the binary file is highly encoded, for example, encrypted, it seems inefficient to decode it to plain text and then search with grep; especially within a company that reported a revenue in the billions last year. While this method may be understandable for those with limited resources, one would expect a much more optimized solution in a large enterprise setting.
The engagement generated by the post led the author to disclose more details, for example, that this file was generated by a microservice. And that it had a record-type structure with various record types. (1) We all know what microservices use to talk to each other, right? (2) 100GB files seem to defeat the micro purpose, but that is relative. Also, someone suggested to grep the binary file, but the author seemed to not be aware of such a tool, and the reply was something along the lines of wanting to avoid translating the grep queries into binary grep ones. This seems reasonable if the queries are complex, which is kind of expected.
My suspicions were confirmed when the author disclosed that the decoded file is a JSON file worth at least 1 TB. Then I thought "Who greps such a big file when there are a lot of efficient tools that parse JSON?, again, you must be kidding, right? I mean, you are aware that there are a bunch of smart folks who create software specialized in storing and querying data, right?" But again, probably you don't want to add an extra layer of complexity, so you prefer to stick to big file transactions, or perhaps (and more likely) a high number of small transactions.
The cherry of the cake was when the author posted a follow-up post: You just gotta love how many people jump with solutions without clarifying requirement. This is a common mistake I see while interviewing people at *****. People propose solutions without trying to clarify constraints, or input data patterns, or edge cases.
So, this might as well have been an interview question and an interview behavior. I think it is troublesome. If you are given a problem to solve as part of an interview for a job about solving problems, the first natural reaction is to attempt to solve the problem with the information you are given. Moreover, if the person doing the interview is your prospective immediate supervisor, then it is reasonable to assume you are given all the information necessary or at least available at the time to solve the task. I think most people try to be polite and won't tell the interviewer, "Who the f..k does that? Are you incompetent or just plain lazy?" even if they think about it.
If you are given a half-baked problem by the person who is going to supervise your work, then you, as a prospective employee, can easily make the conclusion "this guy is plain incompetent, he wants me to boost his production by more than 10x without having a complete idea of the problem he wants to solve." If you want to assess the candidate's ability to correctly frame a problem, there are many more meaningful ways to do so.
Interviews go both ways, just as the interviewer concludes the candidate makes a mistake by not asking more questions about the problem, the candidate (even if they ask questions) can easily conclude that the interviewer is lazy, incompetent, or withholds information on purpose to other colleagues.
UPDATE:
After writing this I reached out to Marcus Pilgaard to thank him for posting the scores at musescore and showed him the result. He didn't just spot the mistakes and pointed them out, he was so kind enough to offer the solution! That's super cool.
Marcus handed me a revised score for the melody adjusted to the kind of sounds I was generating and he also provided me with a couple of MIDI files for a keyboard. I was able to revise the whole melody. It was nice to hear what I generated along with a trusted baseline.
So here are the results:
Søren Banjomus Revised
I mixed Marcus MIDI melody and what I generated and here is the result
I also tweaked the sound shaper to produce quieter sounds. The revised version is here
Marcus: Tusind tak, du er den bedste!
Original:
A while ago I wrote a white noise generator to avoid going to an app. I wanted to be able to hear it and tune some elements on the fly. Since the generator is written in Python and use Simpleaudio to construct the wav file, I consulted the documentation to figure out how to play raw data.
There I found an example using numpy. The numpy example had some appealing syntax so I decided to create a small program to generate some notes. My "wife" had the idea to play the Imperial March from Star Wars on it and from there I decided to make it play one of my favorites Christmas songs: Søren Banjomus.
Luckily I found a score written by Marcus Pilgaard HERE which was wonderful for my attempts :)
These are the .wav results: Imperial March, Søren Banjomus
And finally this is the code: Player the configuration settings (i.e. filename and playback) are the last two lines
I think it is not bad for a couple of days work considering I had to recall how to read scores, interpret them, and so on :D
After writing this I reached out to Marcus Pilgaard to thank him for posting the scores at musescore and showed him the result. He didn't just spot the mistakes and pointed them out, he was so kind enough to offer the solution! That's super cool.
Marcus handed me a revised score for the melody adjusted to the kind of sounds I was generating and he also provided me with a couple of MIDI files for a keyboard. I was able to revise the whole melody. It was nice to hear what I generated along with a trusted baseline.
So here are the results:
Søren Banjomus Revised
I mixed Marcus MIDI melody and what I generated and here is the result
I also tweaked the sound shaper to produce quieter sounds. The revised version is here
Marcus: Tusind tak, du er den bedste!
Original:
A while ago I wrote a white noise generator to avoid going to an app. I wanted to be able to hear it and tune some elements on the fly. Since the generator is written in Python and use Simpleaudio to construct the wav file, I consulted the documentation to figure out how to play raw data.
There I found an example using numpy. The numpy example had some appealing syntax so I decided to create a small program to generate some notes. My "wife" had the idea to play the Imperial March from Star Wars on it and from there I decided to make it play one of my favorites Christmas songs: Søren Banjomus.
Luckily I found a score written by Marcus Pilgaard HERE which was wonderful for my attempts :)
These are the .wav results: Imperial March, Søren Banjomus
And finally this is the code: Player the configuration settings (i.e. filename and playback) are the last two lines
I think it is not bad for a couple of days work considering I had to recall how to read scores, interpret them, and so on :D
In her CACM 2006 landmark paper "Computational Thinking" Jeannette
M. Wing coined the term Computational Thinking.
Wing concludes: Professors of computer science should
teach a course called "Ways to Think Like a Computer Scientist" to
college freshmen, making it available to non-majors, not just to
computer science majors.
If we recognize that all professions are equally important, then the
first years (not semester nor semesters) of college would be a
plethora of courses "Ways to Think Like a ...".
Which, actually, seem to be more suitable to be taught at
High-School level and should give a more informed landscape of
professional development possibilities.
But why is necessary to have the concept Computational Thinking? According to Wing, "Computational thinking is a grand vision to guide computer science educators, researchers and practitioners as we act to change society's image of the field". As this paper resonated within the Computing Education community, research projects and papers found a conductive thread to spread the "Computational Thinking" concept and how to promote it in pre-college learners, although it remains challenging to define exactly what you should observe to conclude that students have reached "Computational Thinking".
Long before Wing's paper, Leone Minna Gold Burton in 1984 published "Mathematical Thinking: The Struggle for Meaning" in JRME (Journal for Research in Mathematics Education) in which she argues that "mathematical thinking is not thinking about the subject matter of mathematics but a style of thinking that is a function of particular operations, processes, and dynamics recognizably mathematical". In essence this is similar to Wing's definition: "Computational thinking involves solving problems, designing systems, and understanding human behavior, by drawing on the concepts fundamental to computer science". While Wing acknowledges Computational Thinking is related to Mathematical Thinking, she states this relation is due to "like all sciences, its formal foundations rest on mathematics". Indeed Computer Science is a branch of Mathematics, we study problems solvable by algorithms, that is, well-defined (hopefully unambiguous) processes. There should be little doubt that computational problems are more restricted than Mathematical problems, and hence, Computational Thinking should be a restricted, specialized version of Mathematical Thinking.
Wing claims Computational Thinking is a "Fundamental skill", "something every human being must know to function in modern society". As far as I know, Wing does not justify this requirement, other authors state this is because computers are "everywhere". Wing does mention: "Ubiquitous computing was yesterday's dream that became today's reality" but she doesn't relate this reality with the "Fundamental skill" characteristic of Computational Thinking.
Arguably most people own a watch: wristwatch, wall-clock, alarm clock, etc. It is almost impossible to imagine going on your daily modern routine without a tool that marks the time down to the second. Clocks are a marvel of human ingenuity and nowadays 'smart-watches' even check your oxygen saturation. And yet, how to think like a clock designer/builder/tinker is not a course nor a "Fundamental skill". Shall we teach "Car engineering Thinking" just because in your country cars are everywhere? Shall we teach "Railroad design" to European students because you can reach many small towns by train? Or shall we teach "How to think like a librarian" just because a vast majority of persons in the "modern society" own a book? The difference is clear: a computer does much more than mark the time, it doesn't take you from A to B but it can show you how to get from A to B the quickest, it is not a book, but it can contain millions of books in just a tiny fraction of the space.
Is this enough to justify Computational Thinking?
It is futile to cement Computational Thinking on the "Computer" object because problems solvable by well-defined processes transcend computers and had been, in fact, a constant presence throughout the history of mankind: you can't cook an animal you haven't hunted yet. Computers, at the present moment, are instrumental just because they are our current predictable model in which we can test our theories.
But why is necessary to have the concept Computational Thinking? According to Wing, "Computational thinking is a grand vision to guide computer science educators, researchers and practitioners as we act to change society's image of the field". As this paper resonated within the Computing Education community, research projects and papers found a conductive thread to spread the "Computational Thinking" concept and how to promote it in pre-college learners, although it remains challenging to define exactly what you should observe to conclude that students have reached "Computational Thinking".
Long before Wing's paper, Leone Minna Gold Burton in 1984 published "Mathematical Thinking: The Struggle for Meaning" in JRME (Journal for Research in Mathematics Education) in which she argues that "mathematical thinking is not thinking about the subject matter of mathematics but a style of thinking that is a function of particular operations, processes, and dynamics recognizably mathematical". In essence this is similar to Wing's definition: "Computational thinking involves solving problems, designing systems, and understanding human behavior, by drawing on the concepts fundamental to computer science". While Wing acknowledges Computational Thinking is related to Mathematical Thinking, she states this relation is due to "like all sciences, its formal foundations rest on mathematics". Indeed Computer Science is a branch of Mathematics, we study problems solvable by algorithms, that is, well-defined (hopefully unambiguous) processes. There should be little doubt that computational problems are more restricted than Mathematical problems, and hence, Computational Thinking should be a restricted, specialized version of Mathematical Thinking.
Wing claims Computational Thinking is a "Fundamental skill", "something every human being must know to function in modern society". As far as I know, Wing does not justify this requirement, other authors state this is because computers are "everywhere". Wing does mention: "Ubiquitous computing was yesterday's dream that became today's reality" but she doesn't relate this reality with the "Fundamental skill" characteristic of Computational Thinking.
Arguably most people own a watch: wristwatch, wall-clock, alarm clock, etc. It is almost impossible to imagine going on your daily modern routine without a tool that marks the time down to the second. Clocks are a marvel of human ingenuity and nowadays 'smart-watches' even check your oxygen saturation. And yet, how to think like a clock designer/builder/tinker is not a course nor a "Fundamental skill". Shall we teach "Car engineering Thinking" just because in your country cars are everywhere? Shall we teach "Railroad design" to European students because you can reach many small towns by train? Or shall we teach "How to think like a librarian" just because a vast majority of persons in the "modern society" own a book? The difference is clear: a computer does much more than mark the time, it doesn't take you from A to B but it can show you how to get from A to B the quickest, it is not a book, but it can contain millions of books in just a tiny fraction of the space.
Is this enough to justify Computational Thinking?
It is futile to cement Computational Thinking on the "Computer" object because problems solvable by well-defined processes transcend computers and had been, in fact, a constant presence throughout the history of mankind: you can't cook an animal you haven't hunted yet. Computers, at the present moment, are instrumental just because they are our current predictable model in which we can test our theories.
I have heard and read the statement "Programming isn't Math" too
many times, but this one is new, it doesn't just tell you what
programming "isn't" but what "is".
Today a tweet with a video claiming this popped up and I
watched it.
The video dwells on different kinds of "ambiguities" that can happen on C++ programs with examples of "English counterparts". For some reason omitted the fact that Mathematics has ambiguities too, called paradoxes, one of the most famous is "Russell's paradox".
Like all the times I have seen "Programming isn't Math", the video failed to explain this point at any serious level, it follows the "conter-example fallacy"; showing counter examples that curiously fall in the claimed category and conclude the subject at hand must belong to the claimed category. I feel the expression "Programming isn't Math" is meant to point out that having a formal mathematical education is not a pre-requisite to write programs. It is obvious that you can create programs without any formal Mathematical background just as you can build a brick house without any civil engineering knowledge. Imagine you can get a building manual with recipes to build any possible two-story house that explains how you should arrange all the elements. Then, clearly you don't need any civil engineering knowledge because every possible case is solved in the manual, how that manual came to be is beyond this discussion, the point is the manual exists. The existence of the manual can be considered as the culmination of civil, structural and material engineering research into two-story houses. The interesting, open question is: if we accumulate all the programs ever written, can we write any possible program by taking only parts of the existent programs?
I usually do not check "the credentials" of authors because I feel those bias us, but since some of the people I follow, follow that account, then I was curious. I was "surprised" to find the profile indicates the author to be a researcher at the company that created the most successful search engine.
I can not help to be left with this taste of marketing paving: if to create programs you don't need "math" just to be able to have an idea of how to "write proper English" then you are all set now the last step is to use this super intuitive AI tool to help you...
The video dwells on different kinds of "ambiguities" that can happen on C++ programs with examples of "English counterparts". For some reason omitted the fact that Mathematics has ambiguities too, called paradoxes, one of the most famous is "Russell's paradox".
Like all the times I have seen "Programming isn't Math", the video failed to explain this point at any serious level, it follows the "conter-example fallacy"; showing counter examples that curiously fall in the claimed category and conclude the subject at hand must belong to the claimed category. I feel the expression "Programming isn't Math" is meant to point out that having a formal mathematical education is not a pre-requisite to write programs. It is obvious that you can create programs without any formal Mathematical background just as you can build a brick house without any civil engineering knowledge. Imagine you can get a building manual with recipes to build any possible two-story house that explains how you should arrange all the elements. Then, clearly you don't need any civil engineering knowledge because every possible case is solved in the manual, how that manual came to be is beyond this discussion, the point is the manual exists. The existence of the manual can be considered as the culmination of civil, structural and material engineering research into two-story houses. The interesting, open question is: if we accumulate all the programs ever written, can we write any possible program by taking only parts of the existent programs?
I usually do not check "the credentials" of authors because I feel those bias us, but since some of the people I follow, follow that account, then I was curious. I was "surprised" to find the profile indicates the author to be a researcher at the company that created the most successful search engine.
I can not help to be left with this taste of marketing paving: if to create programs you don't need "math" just to be able to have an idea of how to "write proper English" then you are all set now the last step is to use this super intuitive AI tool to help you...
The first time I heard about Roskilde University (Roskilde Universitet) was when I was a bacherlor student: the PhD advisor of one of my bachelor advisors worked there on program analysis of logic programs, the topic I was working on at that time. Little I knew what a special place Denmark and Roskilde would have in my wife's heart and mine. How our breath would be taken away standing in the shadow of Roskilde Domkirke, our weekly trips and special trips to the Vikingeskibsmuseet to see the ships pulled out of the water for the winter while discussing how nice it would be to live in Roskilde and be part of the maintenance crew. Little I knew how we would wait for summer to buy Carolina Reaper chilly peppers at the weekend market and how patient the sweet couple would be with my less-than-acceptable danish or how nice was the 'Honey man' from whom we would buy our Roskilde Ring honey, how much we would cherish our wood crafts we bought at the local market or seeing all the cherished belongings people would sell on weekends or our 'tours' to find and photograph all Kilder. I wouldn't even imagine telling to my office mate at CMU over lunch (during a research stay) about how much I would have loved to have an academic position at RUC while staying at one of the most prestigious computer science universities. I can not help, but feel disappointed and somewhat heartbroken about the proposal to close RUC.
What I write here is not based on exhaustive research, as I said, my Duolingo fueled danish is not what it should be, so there is a lot that I do not understand; it is based on perspective and feeling. It seems some doubtful decisions were made regarding the acceptance of 300 students from Bangladesh in 2024 into a masters program in business that had financial consequences. The consequence for RUC is that it seems it 'profited' millions of danish krone. Perhaps if it were just that, then it would be kind of ok. However, politicians seem to suspect the students and accompanying spouses true purpose is to seek jobs; a discourse often heard worldwide nowadays. Suspicion alone is not a fact; suspicion draws attention and requires investigation, research and debate. If wrong doing with intent is found, then corrective actions must be taken, if necessary fire the university's entire ruling body, but closing a regional university has no sense at all!; specially when a higher-education decentralization strategy is in the works. It seems to me the goal of RUC shutdown is to send a message not an actual step in the right direction to fix the root causes. A message that will have far more negative consequences than 'letting' other universities known whatever there is for them to know.
I feel for the RCU and Roskilde communities but I also feel for the foreigner students that may be prevented from experiencing the warmness and kindness my wife and I had the opportunity to be part of.
I got news regarding a position I applied to, it was rejected. What is particularly frustrating about this situation, is that you had to write a proposal and attach it to the application. It took me several days of work to write the application, reading books, papers, and so on to frame the proposal within the scope of the job advertisement. In the end, I got a template email basically saying "thank you for your interest". From my perspective, that kind of response does not help with your self-grow as a prospective employee. There is no feedback, nothing, just a boolean answer. Of course, I reached out in the hopes of getting a more elaborated feedback, after all we are humans and not AI models using backpropagation. I got an email wrote by a human but still template, however, what was interesting is that only short-listed applicants will get feedback on their application. This seems reasonable if you get a significant number of applicants. That being said, if complete feedback is only given to short-listed applicants, then why not require the proposal only to short-listed applicants? It seems to be a proportional reasonable request. It would be the equivalent of programming tests some companies perform as part of their hiring process.
I still have another issue tho. This is the 3rd position of this type that I have applied and the second that uses a hiring platform. The first position was sent by email, this was a long time ago. In that case I got an interview and then the employer reached out saying they picked the other candidate because he had more experience with Y which I totally get; there was a well-stated reason behind my rejection, of the n boxes we had to tick I checked only n-1. In the second case I got accepted. But in this case there was nothing. I find unsettling that in an era where there are platforms that help to bring talent in, it is still a black hole: information gets sucked in and nothing gets out. There is no theoretical/technical limitation that prevents employers from providing feedback even if you only send your CV; someone, somewhere had to say 'no because ...' and if you don't wish to type there are speech synthesizers or check boxes. I think even LinkedIn provides you with some metrics, if you pay them of course.
I feel there is a culture of opacity which obviously seem reasonable to some employers, perhaps feedback would give a glimpse into the thinking process behind selection which might be exploitable or legal-abiding? Bottom-line if we accept employers have the right to reject candidates for any reason, for example, if you have your head tilted to the right slightly on your profile picture, then nowadays, it is a piece of cake to write a program that desk-reject such applications and write a template email: "We thank you for your interest unfortunately your head was slightly tilted...". Then it is up to you as a job seeker to take it or leave it; unlike AI model training, we do have a choice to incorporate feedback or not.
I have many reusable paper sheets. Sometimes I print on one side and the other is clean, others I buy extra paper sheets that end up not being what I needed and sometimes friends gift me some extra clean or reusable paper sheets. I love handwriting, specially when I use fountain pens, there is something I like about them. My handwriting is terrible; I often tell my students that by the end of the semester they would master two skills, the academic goals of the assignment and decoding hieroglyphs :). I often craft my own notebooks and blocks. Don't get me wrong, I also have my fair share of Moleskine, Fabriano and 365 notebooks. Lately I'm using I block I made a long time ago, perhaps 7 or so years ago with some block sheets that were too thin for me at the time. The sheets are even older, perhaps 12 years? I don't even remember. But I bought a couple of new fountain pens, a Platinum Preppy Extra-Fine and a Pilot kaküno also Extra-Fine and both are great!. If the pen doesn't come with a converter, it is not a problem, I have a syringe, ink bottles and an affinity to mix inks :)
I have been spending a lot of time preparing Programming Without Problems there were many ups and downs and corrections and so on. At some point it was no fun to see so much HTML code together. So I wrote a small Python (go figure :) ) script to help me strip all HTML code and just focus on my writing. It seems to fit the bill so far so, I'm happy. It also helps me to manage my references automatically, so that is nice; if BibTeX is an interstellar ship, my code is the Wright's Brothers sketch :D.
When I was in high-school at UNAM's CCH-Sur I used to spend quite sometime at the library. It is a beautiful place, not only because the building is gorgeous but also because its mathematics collection is amazing. I spent hours there reading and browsing books and of course, napping.
I got the same feelings on the early days of Google. After clicking on the 3rd page or so :)
With AI (ChatGPT, etc.) it is like entering a library in which all books are blacked out and there is this annoying librarian asking what you want and giving you a paragraph of a black book.
When you know exactly what you are looking for, this annoying librarian could be useful. I have spent half a day looking for a reference of some quote I have buried deep in my brain. However, when you look for guidance, the annoying librarian is the least help you should seek.
I understand the need for empirical evidence on a claim. What I can't grasp is the reaction of surprise when the evidence supports common sense, like today on my social media feed regarding a paper showing accumulation of "Cognitive Debt" when using AI for creative writing.
AI will accelerate the success of those that 1) Know exactly what thing they are looking for, 2) it is not easy to find or tedious to achieve and 3) it is common enough in the training set.
AI will accelerate the creative process destruction of those that are novice to a subject and are seeking guidance.
Of course, I would not be surprise that a model happens to be advertised as fine-tuned to provide guidance.
Sometime ago I found that a "white noise" broadcast on YouTube helped me to concentrate while listing to it on my headphones. However, I feel it is too much of a waste to just listen to "random noise". So, I wrote a small Python program to generate some "white noise". We all know it is not trully random, but it is close enough :) And, since I had control over what I created, I decided to mix in some effects.
The first Passing Rain is white noise with the same sinusoidal volume control on both ears.
The second Passing Rain Adjusted is white noise with the same sinusoidal volume control on both ears, but before adjusting the volume, a linear interpolation with 35 points is performed between pairs of "sufficiently off" points. Effectively removing some randomness (actually a lot :D).
The third Alternating Rain is white noise with a sinusoidal volume control starting at for the left channel and another starting at for the right channel.
The fourth and fifth are more if you are into regular beats.
The fourth Slow Beat is a sine wave each second using only the upper part of the intensity spectrum. I'm using 16 bits for each number but only using the [0,32767] range.
The fifth Slow Motor are 15 sine waves per second, again using only the upper part of the intensity specturm.
Watch for your volume if you are using headphones!
I asked Gemini to generate a code to solve the 24 arithmetical puzzle in Ada. I haven't made progress checking the correctness of the generated code. I was planning to cross check it with implementations on other languages in RosettaCode, but it has been down for the best part of the week. What got my eye was the "tests" Gemini generated, here is an example:
It is difficult to believe this kind of code generation is taken seriously. Of course I chose this language and this problem because I don't expect to exist a lot of Ada implementations of 24. At RosettaCode there was none. And if you are using AI to generate code that has been done and redone for 1000 times, then, isn't really a cheaper way?
I watched a GitHub Copilot tutorial video about basics. At some point Copilot is asked what error handling functionality should be added to some code. The first one to be suggested was an ArgumentNullException and I immediately remembered back when I was an undergrad. The Java compiler used to refuse to compile your code if you failed to catch a NullPointerException that some method you called may throw. I had this professor that used to say: "Just compile the damn thing, I KNOW what I'm DOING".
I came across this post on the once
great, then not so great before they sold it and now it is just a
zombie. If I recall correctly it was one of the main aims of
C++: to not impose a certain mindset. But for some reason they lobby
a lot for modern C++. I think one of the reasons is that you need
certain standard guidelines in order to be able to work quickly
on different teams without the (re)learning curve. Some changes are
justified to help the careless developer. I think it is one of
the worst reasons to change a language, it is like saying: We are
going to introduce these changes into (insert your favorite natural
language here) because some people fail to speak it "properly".