I'm not exactly sure how you would go about grading mathematical proficiency. I went through calculus two and discrete mathematics, but I'm sure that there are things I have forgotten now even though I would be considered familiar with most leading edge AI technology. If I'm being honest, I'm not sure I could pass the final exams I took to get my CS degree right now.
“It is the first step in sociological wisdom, to recognize that the major advances in civilization are processes which all but wreck the societies in which they occur:—like unto an arrow in the hand of a child. The art of free society consists first in the maintenance of the symbolic code; and secondly in fearlessness of revision, to secure that the code serves those purposes which satisfy an enlightened reason. Those societies which cannot combine reverence to their symbols with freedom of revision, must ultimately decay either from anarchy, or from the slow atrophy of a life stifled by useless shadows.” A. N. Whitehead
AIs should be forced to show their work. Every tool they use, every program they generate and run in the background, and every logical inference. They should be forced to produce Lean or Rocq proofs or execution traces for all the computation they use. For facts, they should be able to produce sources. For any abstract reasoning, they should be able to break it down into explainable steps.
Then, on top of that, they should be able to explain any of that, at any level of detail, whether talking to an expert or a layperson.
AIs aren't people. I can demand whatever I want from them.
The entire premise of building machines that can provide intelligence is that they can be arbitraryly rigorous and do far more work then is reasonable or practical to demand of a human.
That's an implementation preference, and you can have all of that today if you write your own harness. Vibe coding a cat meme generator doesn't need a Lean proof.
> AIs should be forced to show their work. Every tool they use, every program they generate and run in the background, and every logical inference.
And they should be 100% deterministic: same input (including a random seed / "temperature" if one wants), same output.
On a not totally dissimilar note I like the projects saying: "AI PRs aren't accepted but if you do something worthwhile, give us your prompts instead".
Yeah how many of us know how to build an ICE engine or shoes of any meaningful quality
We set upon end of human craftsmanship decades ago
Math is probably the easiest to reclaim given its right in front our faces going about daily life. The syntax of math is not that important; real world quantification the syntax is meant to represent will still exist. Our biochemistry implicitly operates on senses of enough food and water, etc.
Such measures are so embedded in the daily routines we live an intuition will always exist
No one is born knowing how to make a computer as we know them today. A cup half filled is obvious
Arguably, the loss of specialized knowledge is what is holding countries like the US back in domains such as shipbuilding[1], rare-earth refining, and semiconductors.
Losing that knowledge puts the losing community in a bad position, and those communities who maintain this information somehow will end up being able to preserve their ability to reason and improve things. Though I don’t expect our current crop of US leaders and business owners to realize that.
“The product of mathematics is clarity and understanding. Not theorems, by themselves. [Their importance is not just in their specific statements], but their role in challenging our understanding, presenting challenges that led to mathematical developments that increased our understanding.
The world does not suffer from an oversupply of clarity and understanding (to put it mildly)… In short, mathematics only exists in a living community of mathematicians that spreads understanding and breaths life into ideas both old and new. The real satisfaction from mathematics is in learning from others and sharing with others. All of us have clear understanding of a few things and murky concepts of many more. There is no way to run out of ideas in need of clarification. The question of who is the first person to ever set foot on some square meter of land is really secondary. Revolutionary change does matter, but revolutions are few, and they are not self-sustaining --- they depend very heavily on the community of mathematicians.”
Curious to see if that can map to what's happening in the software industry/community.
> The product of software engineering (or computer science) is clarity and understanding. Not programs, by themselves. Their importance is not just in their specific statements (lines of code in a specific language), but their role in challenging our understanding, presenting challenges that led to computational (?) developments that increased our understanding.
> ..In short, software only exists in a living community of developers that spreads understanding and breaths life into ideas both old and new. The real satisfaction from computers is in learning from others and sharing with others.
That seems to work. What about other industries that are currently being consumed by automation and "AI"? Like writing or the arts.
> The product of software engineering (or computer science) is clarity and understanding.
Hard disagree. Software engineering is about satisfying requirements while optimizing costs. Computer science is the part where clarity and understanding are the goal.
> The product of software engineering (or computer science) is clarity and understanding. Not programs, by themselves.
The industry demonstrates that in practice, typically, it doesn’t agree with this. “Get the program working” is goal 1, and everything after that is optional, depending on the environment.
> it’s that we may stop producing people who know enough to notice when AI is confidently wrong.
The running-joke is that a LinkedIn-lunatic AI booster, with a Nano Banana-generated profile-pic, will immediately slide into the chat to tell you that that this is already a solved problem: just spin up another agent to do the work to verify the first agent. Token-cost-be-damned. And we laugh and downvote them to oblivion and carry on with our day.
But today I had some exposure to a SotA agentic team coding loop thingie which had been running almost hands-off for a few weeks on a (pretty serious) Win32+Direct3D-to-Emscripten+WebGL porting project - and I'm genuinely spooked at how well it all works; I mention this example because all the agents' processes involved a decently rigorous verification step: any time any agent confidently asserts something then it has to provide associated evidence, such as a unit test report, or build artefact, or external citation, and the system will spawn a new agent (perhaps using a different backing LLM) to verify the claim. I know a unit-test pass/fail isn't quite the same thing as, say, a medical AI agent confidently wrong about me having/not-having terminal spleen cancer, but the capability for a team-of-agents to be self-checking is definitely there.
----
Also, the past 3 years of AI/LLM/etc developments have taught me to never cling to any shortcoming or weakness they have because plenty of them do seem to have been solved or mitigated, either directly or indirectly.
How is the token-cost-be-damned part in the latter example?
I do find that both porting and translation projects have a much higher signal given the ease of mapping to tokens, when there is a proven working source to refer to - the source itself provides the validation. In a new project, you don’t have that validation.
> How is the token-cost-be-damned part in the latter example?
It's there, but...
1. The project owner figured out a way to minimize token usage for agent claim verification tasks.
2. Verification agents used older and much cheaper models, including local models for the most trivial things.
3. They could afford it anyway; but I think it's an inevitability that the token-cost for a task will approach some limit for some quality threshold - concurrent with the dollar-cost-per-token shrinking over time as better hardware comes out.
> In a new project, you don’t have that validation.
I'm still trying to understand that part of the project's history, actually. Obviously the HTML5+WebGL+Emscripten+Etc entrypoint was a "new" project; one of the first things they did was build their own means of verification, I just don't know how that part worked-out in practice (besides the agents dogpiling in on TODO.md).
It does happen. But this is why the “ralph loop” works: over dozens (or hundreds) of iterations, eventually every regression, implicit assumption, fake passing test, is noticed and fixed by another agent. The code slowly but continuously converges to a better state. I’m surprised myself, but haven’t seen it fall into chaos or degradation so far.
I haven't seen a code base which hasn't fallen into degradation from that loop.
From personal experience, any "correction" leads to an LLM writing more code, or stripping apart interfaces and mixing logic poorly, or adding some hidden control flow just to catch that edge case rather than the general case for that category of input.
I wouldn't accept that junk from any engineer in my team.
My experience with good developers has been:
- Create something a bit ugly which works
- As more use cases come in, refactor code and clean up old gnarly bits
LLMs tend to work architecture astronaut style mixed with a junior for implementation. Build theoretically good interfaces but then implement dependencies across all of them. And when changes happen, keep feeding that cycle.
But doesn't things like that happen with people? Take academy where some professor implicitly states "facts" or military or software development where some highly praised senior always has his last word...
The one I saw that curled my toes was an LLM that would take the input of a text field and turn it into an entry into a workplace health and safety log, and attribute the results to a human.
It only needs to be wrong once to cause a legal shitstorm. You also wont be getting a chance to merge in a change to roll in a second agent for consistency checks from prison.
> But today I had some exposure to a SotA agentic team coding loop thingie which had been running almost hands-off for a few weeks on a (pretty serious) Win32+Direct3D-to-Emscripten+WebGL porting project - and I'm genuinely spooked at how well it all works; I mention this example because all the agents' processes involved a decently rigorous verification step: any time any agent confidently asserts something then it has to provide associated evidence, such as a unit test report, or build artefact, or external citation, and the system will spawn a new agent (perhaps using a different backing LLM) to verify the claim. I know a unit-test pass/fail isn't quite the same thing as, say, a medical AI agent confidently wrong about me having/not-having terminal spleen cancer, but the capability for a team-of-agents to be self-checking is definitely there.
It taking "a few weeks" of wall-clock runtime suggests to me that it's making a lot of mistakes that keep needing correction, you're seeing the output after it fixes them. FWIW, my experience with them includes just enough fragile self-made unit tests to be a problem for entirely hands-off work, despite what it gets right.
However, on the topic of time rather than quality: Last summer I decided to convert an old game from C++/SDL2 to JS/WebGL, and was doing that with manual copying each function one at a time to ChatGPT and then code reviewing each line by hand. The wall-clock speedup from LLMs is much much faster than me working like this, so much so that it would have been done in a week if it had managed without making mistakes.
I might go back to that. While I was enjoying the process of code archeology (the code being a mix of many platforms starting with 68k era classic MacOS software rendering and resource forks upwards), I had to focus on other things for a bit in the meanwhile, which is of course the actual benefit of code factories.
Don't worry. A lot of people were captivated by computers and programming before it ever became profitable. A lot of them will continue to be, even after it does not pay well..
I see many people reacting with fear towards IA and many of them do not feel the same level of danger in other places which are clear to me (and I think that what they fear most is the unknown, as it has to be bad, for sure). I like the quote of the other message from Whitehead !!
It sounds ‘the singularity is near’ not so much because AI is reaching escape velocity forward, but because we’re systematically pushing humans back beyond the threshold where computers are legible?
You both seem to be using a different definition of "singularity" from the one I'm familiar with. I've always understood it to mean a rapid feedback loop in which AI creates successive, increasingly capable generations of AI outside human control, rather than simply a level of technological advancement that would be incomprehensible to someone
The singularity is when an AI can improve itself, which among the results include humans no longer being able to understand its code and exponential improvements so we can't catch up either.
It requires a level of self-driven intent LLMs aren't capable of on their own.
The reason the term "singularity" was used when that term was coined was in analogy to a black hole, which has a point on its radius where the gravitational pull is strong enough to stop light from escaping, thereby making it impossible to see beyond that horizon
This is important to the framing of the idea: "the technological singularity" refers to a point in the future past which further developments can't be meaningfully foreseen or maybe even understood. This is often associated with particular paths people think will cause this, but not defined by those. As such, the prior comments are making the claim that you can get there not just by accelerating technological progress, but also by reducing the ability of everyone to comprehend what's happening, which I agree is unfortunately a plausible outcome in the current world
To expand on this, I prefer calling this "the event horizon" rather than "the singularity", as that makes it clear (I hope) that I mean "can't predict" rather than "goes to infinity".
With this framing, my prediction for that horizon has been "around 2030" for over a decade now, as several different currently-exponential growth trends start giving results with weird implications around that year.
(LLMs look like they're ahead of schedule for this, but consider their power requirements and limitations on AI in robotics that mean they're not quite good enough on the "G" part of "AGI" to affect non-desk jobs).
I think the differentiation going forward is going to be if you're a person that understands things or if you're a person that delegates your understanding. I don't think there's going to be a lot of economic or social value to prompting an AI to do things you don't understand. First, anyone can do it, so why should someone give you money or credibility for that? Second, as fewer and fewer people bother to learn hard things, the scarcity of that knowledge will increase its value.
I guess even if I end up being wrong about that, I'd rather enrich my life and grow as a person by continuing to learn and do hard things rather than become an annoying cheerleader dependent on unreliable tools.
Extrapolating to other fields, we can say this development is the worst what one can do to mankind and society: pushing for AI but at the same time decimating the human capital that helped produce and understand it. What is the point of producing countless mathematical proofs or code that one cannot verify nor understand? It’s like having a machine to produce endless things that no one needs or will use. Just because the machine can bring profits for the (initial) investors?! That is the rottenest form of capitalism imaginable.
I strongly suspect most people already live in a world, that is not reasonable to them.
If you observe most policies to come from authority, you might not feel the loss of your own capabilities as much. Most people who are subjected to rules of law, bureaucracy, politics, or religion, might see these rules as arbitrary.
If you think of the way math gets taught in a this-is-right-everything-else-is-wrong way, it will evoke the same feeling of arbitrary authority.
Playing around and exploring math is usually only sometimes taught to children from academic/STEM families. So why would someone mourn the loss of capability for mathematical reasoning in people, if they only experienced it as following something incomprehensible anyway.
54 comments
[ 4.4 ms ] story [ 58.5 ms ] threadThen, on top of that, they should be able to explain any of that, at any level of detail, whether talking to an expert or a layperson.
The entire premise of building machines that can provide intelligence is that they can be arbitraryly rigorous and do far more work then is reasonable or practical to demand of a human.
And they should be 100% deterministic: same input (including a random seed / "temperature" if one wants), same output.
On a not totally dissimilar note I like the projects saying: "AI PRs aren't accepted but if you do something worthwhile, give us your prompts instead".
We set upon end of human craftsmanship decades ago
Math is probably the easiest to reclaim given its right in front our faces going about daily life. The syntax of math is not that important; real world quantification the syntax is meant to represent will still exist. Our biochemistry implicitly operates on senses of enough food and water, etc.
Such measures are so embedded in the daily routines we live an intuition will always exist
No one is born knowing how to make a computer as we know them today. A cup half filled is obvious
Losing that knowledge puts the losing community in a bad position, and those communities who maintain this information somehow will end up being able to preserve their ability to reason and improve things. Though I don’t expect our current crop of US leaders and business owners to realize that.
[1] https://www.gao.gov/products/gao-25-108136
https://mathoverflow.net/questions/43690/whats-a-mathematici...
“The product of mathematics is clarity and understanding. Not theorems, by themselves. [Their importance is not just in their specific statements], but their role in challenging our understanding, presenting challenges that led to mathematical developments that increased our understanding.
The world does not suffer from an oversupply of clarity and understanding (to put it mildly)… In short, mathematics only exists in a living community of mathematicians that spreads understanding and breaths life into ideas both old and new. The real satisfaction from mathematics is in learning from others and sharing with others. All of us have clear understanding of a few things and murky concepts of many more. There is no way to run out of ideas in need of clarification. The question of who is the first person to ever set foot on some square meter of land is really secondary. Revolutionary change does matter, but revolutions are few, and they are not self-sustaining --- they depend very heavily on the community of mathematicians.”
> The product of software engineering (or computer science) is clarity and understanding. Not programs, by themselves. Their importance is not just in their specific statements (lines of code in a specific language), but their role in challenging our understanding, presenting challenges that led to computational (?) developments that increased our understanding.
> ..In short, software only exists in a living community of developers that spreads understanding and breaths life into ideas both old and new. The real satisfaction from computers is in learning from others and sharing with others.
That seems to work. What about other industries that are currently being consumed by automation and "AI"? Like writing or the arts.
Hard disagree. Software engineering is about satisfying requirements while optimizing costs. Computer science is the part where clarity and understanding are the goal.
The industry demonstrates that in practice, typically, it doesn’t agree with this. “Get the program working” is goal 1, and everything after that is optional, depending on the environment.
The running-joke is that a LinkedIn-lunatic AI booster, with a Nano Banana-generated profile-pic, will immediately slide into the chat to tell you that that this is already a solved problem: just spin up another agent to do the work to verify the first agent. Token-cost-be-damned. And we laugh and downvote them to oblivion and carry on with our day.
But today I had some exposure to a SotA agentic team coding loop thingie which had been running almost hands-off for a few weeks on a (pretty serious) Win32+Direct3D-to-Emscripten+WebGL porting project - and I'm genuinely spooked at how well it all works; I mention this example because all the agents' processes involved a decently rigorous verification step: any time any agent confidently asserts something then it has to provide associated evidence, such as a unit test report, or build artefact, or external citation, and the system will spawn a new agent (perhaps using a different backing LLM) to verify the claim. I know a unit-test pass/fail isn't quite the same thing as, say, a medical AI agent confidently wrong about me having/not-having terminal spleen cancer, but the capability for a team-of-agents to be self-checking is definitely there.
----
Also, the past 3 years of AI/LLM/etc developments have taught me to never cling to any shortcoming or weakness they have because plenty of them do seem to have been solved or mitigated, either directly or indirectly.
I do find that both porting and translation projects have a much higher signal given the ease of mapping to tokens, when there is a proven working source to refer to - the source itself provides the validation. In a new project, you don’t have that validation.
It's there, but...
1. The project owner figured out a way to minimize token usage for agent claim verification tasks.
2. Verification agents used older and much cheaper models, including local models for the most trivial things.
3. They could afford it anyway; but I think it's an inevitability that the token-cost for a task will approach some limit for some quality threshold - concurrent with the dollar-cost-per-token shrinking over time as better hardware comes out.
> In a new project, you don’t have that validation.
I'm still trying to understand that part of the project's history, actually. Obviously the HTML5+WebGL+Emscripten+Etc entrypoint was a "new" project; one of the first things they did was build their own means of verification, I just don't know how that part worked-out in practice (besides the agents dogpiling in on TODO.md).
And this is enforced by... another LLM? Seems like it would work alright until something is asserted implicitly and not categorized as an assertion.
From personal experience, any "correction" leads to an LLM writing more code, or stripping apart interfaces and mixing logic poorly, or adding some hidden control flow just to catch that edge case rather than the general case for that category of input.
My experience with good developers has been:
- Create something a bit ugly which works
- As more use cases come in, refactor code and clean up old gnarly bits
LLMs tend to work architecture astronaut style mixed with a junior for implementation. Build theoretically good interfaces but then implement dependencies across all of them. And when changes happen, keep feeding that cycle.
https://en.wikipedia.org/wiki/Space_Shuttle_Challenger_disas...
https://en.wikipedia.org/wiki/Bay_of_Pigs_Invasion
It only needs to be wrong once to cause a legal shitstorm. You also wont be getting a chance to merge in a change to roll in a second agent for consistency checks from prison.
It taking "a few weeks" of wall-clock runtime suggests to me that it's making a lot of mistakes that keep needing correction, you're seeing the output after it fixes them. FWIW, my experience with them includes just enough fragile self-made unit tests to be a problem for entirely hands-off work, despite what it gets right.
However, on the topic of time rather than quality: Last summer I decided to convert an old game from C++/SDL2 to JS/WebGL, and was doing that with manual copying each function one at a time to ChatGPT and then code reviewing each line by hand. The wall-clock speedup from LLMs is much much faster than me working like this, so much so that it would have been done in a week if it had managed without making mistakes.
I might go back to that. While I was enjoying the process of code archeology (the code being a mix of many platforms starting with 68k era classic MacOS software rendering and resource forks upwards), I had to focus on other things for a bit in the meanwhile, which is of course the actual benefit of code factories.
Let's enjoy the ride. It might be last one!
It requires a level of self-driven intent LLMs aren't capable of on their own.
This is important to the framing of the idea: "the technological singularity" refers to a point in the future past which further developments can't be meaningfully foreseen or maybe even understood. This is often associated with particular paths people think will cause this, but not defined by those. As such, the prior comments are making the claim that you can get there not just by accelerating technological progress, but also by reducing the ability of everyone to comprehend what's happening, which I agree is unfortunately a plausible outcome in the current world
To expand on this, I prefer calling this "the event horizon" rather than "the singularity", as that makes it clear (I hope) that I mean "can't predict" rather than "goes to infinity".
With this framing, my prediction for that horizon has been "around 2030" for over a decade now, as several different currently-exponential growth trends start giving results with weird implications around that year.
(LLMs look like they're ahead of schedule for this, but consider their power requirements and limitations on AI in robotics that mean they're not quite good enough on the "G" part of "AGI" to affect non-desk jobs).
In other words, the mathematicians want more funding.
Ancaps are in for a major cope and seethe when they are forced to admit that the communists won the cold war.
I guess even if I end up being wrong about that, I'd rather enrich my life and grow as a person by continuing to learn and do hard things rather than become an annoying cheerleader dependent on unreliable tools.
If you observe most policies to come from authority, you might not feel the loss of your own capabilities as much. Most people who are subjected to rules of law, bureaucracy, politics, or religion, might see these rules as arbitrary.
If you think of the way math gets taught in a this-is-right-everything-else-is-wrong way, it will evoke the same feeling of arbitrary authority. Playing around and exploring math is usually only sometimes taught to children from academic/STEM families. So why would someone mourn the loss of capability for mathematical reasoning in people, if they only experienced it as following something incomprehensible anyway.