skinner_

Okay, I'm not sure about the original one, but here is the prompt of a successful reproduction:

https://aaronlou.com/jacobian_counterexample_prompt.pdf

Obviously it is not random, but it's very generic. No mention of search space or how to reduce it.

The prompt for the Jacobian conjecture was obviously not random. the search space is too big to just try all the combinations of 3 variable polynomials.

Maybe the prompt contained a part like this: "the search space is too big to just try all the combinations of 3 variable polynomials, so be clever about it". Or maybe this part was omitted from the prompt, because modern LLMs are smart enough to figure this out without us having to mention it.

You can skirt around not reasoning in research math because so much of it is just extremely tedious symbolic manipulation.

LOL

You just know nothing about math and are happy to parrot bullshit AI salesmen are selling you.

Not the parent poster here. I do know things about math. I wrote a few papers related to the unit distance problem (https://arxiv.org/abs/2311.10069, https://arxiv.org/abs/2406.15317) and spent quite some time trying to solve it. I had no chance of coming up with the proof that the spicy autocomplete came up with. Dumb benchmark, sure.

My use case is wildly different from CSS, HTML etc. I use numerical algorithms to solve problems in pure mathematics. AI models are now better than me and my colleagues at writing code, and we are pretty good in the first place. The catch is that we do not ask AI to write whole applications, we ask it to implement building blocks like "For finite X subset R^d, find all pairs (Y, Z), Y subset X, Z subset X such that Y and Z are congruent".

Also, if Claude had regurgitated a known solution, it would have come up with it in the first exploration round, not the 31st, as it actually did.

I think the nuanced take on Joel's rant is this: it was good advice for 26 years. It became slightly less good advice a few months ago. This is a good time to warn overenthuastic people that it’s still good advice in 2026, and to start a discussion about which of its assumptions remain to be true in 2027 and later.

Then I think you’ll like our project which aims to find the missing link between transformers and swarm simulations:

https://github.com/danielvarga/transformer-as-swarm

Basically a boid simulation where a swarm of birds can collectively solve MNIST. The goal is not some new SOTA architecture, it is to find the right trade-off where the system already exhibits complex emergent behavior while the swarming rules are still simple.

It is currently abandoned due to a serious lack of free time (*), but I would consider collaborating with anyone willing to put in some effort.

(*) In my defense, I’m not slacking meanwhile: https://arxiv.org/abs/2510.26543 https://arxiv.org/abs/2510.16522 https://www.youtube.com/watch?v=U5p3VEOWza8

https://www.astralcodexten.com/p/in-search-of-ai-psychosis is very relevant, but the main reason I’m posting it here is that, unlike this paper, it takes the opportunity to build the cleverest pun out of the same ingredients:

Folie A Deux Ex Machina

I interpreted it loosely, as "be aware of the possibility, and stop looking at it at the first signs of issues".

100% frontpage-worthy! Frankly I was already bored with all those pelicans, and a bit worried that the labs are overfitting on pelicans specifically. This test clearly demonstrates that they are not.

That's very cool, but it's not an apples to apples comparison. The reasoning model learned how to do long multiplication. (Either from the internet, or from generated examples of long multiplication that were used to sharpen its reasoning skills. In principle, it might have invented it on its own during RL, but no, I don't think so.)

In this paper, the task is to learn how to multiply, strictly from AxB=C examples, with 4-digit numbers. Their vanilla transformer can't learn it, but the one with (their variant of) chain-of-thought can. These are transformers that have never encountered written text, and are too small to understand any of it anyway.

If being probabilistic prevented learning deterministic functions, transformers couldn’t learn addition either. But they can, so that can't be the reason.

But which contributes more, they ask? Who gives a shit, really?

Funding agencies? Should they prioritize established researchers or newcomers? Should they support many smaller grant proposals or fewer large ones?

My uninformed and perhaps overly charitable interpretation: he warned them they were going to be steamrolled, they built their product anyway, and now OpenAI is buying them because (1) OpenAI doesn't want the negative publicity of steamrolling them all, and (2) OpenAI has the money and is a bit too lazy to build a clone.

Amazing! I looked into your ADAM claim, and it checks out. Thanks! Now I'm curious. I you have the time, could you please follow up with the 'etc...'?

You dismiss parent's example test because it's in the training data. I assume you also dismiss the Sally-Ann test, for the same reason. Could you please suggest a brand new test not in the training data?

FWIW, I tried to confuse 4o using the now-standard trick of changing the test to make it pattern-match and overthink it. It wasn't confused at all:

https://chatgpt.com/share/67b4c522-57d4-8003-93df-07fb49061e...

When you build a new model, there is a spectrum of how you use the old model: 1. taking the weights, 2. training on the logits, 3. training on model output, 4. training from scratch. We don't know how much advantage #3 gives. It might be the case that with enough output from the old model, it is almost as useful as taking the weights.

Normally it would answer with a number and an explanation. This one just asks it to skip the explanation so that string comparison can be used to evaluate it.

In your opinion, why did they choose the open source way instead of doing it in a military bunker? (Metaphorical not literal bunker.)

Greater totality of experiences than having read the whole internet? Obviously they are very different kind of experiences, but a greater totality? I'm not so sure.

Here is what we know: The Pile web scrape is 800GB. 20 years of human experience at 1kB/sec is 600GB. Maybe 1kB/sec is bad estimate. Maybe sensory input is more valuable than written text. You can convince me. But next challenge, some 10^15 seconds of currently existing youtube video, that's 2 million years of audiovisual experience, or 10^9GB at the same 1kB/sec.

Sure, these are standard problems, I’ve said so myself. My point is that my productivity is multiplied by ChatGPT, even if it can only solve standard problems. This is because, although I work on highly non-standard problems (see https://arxiv.org/abs/2311.10069 for an example), I can break them down into smaller, standard components, which ChatGPT can solve in seconds. I never ask ChatGPT "where's the race condition" kind of questions.

I work on very complex problems. Some of my solutions have small, standard substeps that now I can reliably outsource to ChatGPT. Here are a few just from last week:

- write cvxpy code to find the chromatic number of a graph, and an optimal coloring, given its adjecency matrix.

- given an adjecency matrix write numpy code that enumerates all triangle-free vertex subsets.

- please port this old code from tensorflow to pytorch: ...

- in pytorch, i'd like to code a tensor network defining a 3-tensor of shape (d, d, d). my tensor consists of first projecting all three of its d-dimensional inputs to a k-dimensional vector, typically k=d/10, and then applying a (k, k, k) 3-tensor to contract these to a single number.

All were solved by ChatGPT on the first try.

Okay, but is this similar to how humans bill you?

Wow, I wasn't aware. @jart, do you have any harsh comments on the actual reneging of an actual promise? @antirez, do you have any kind of comments?

in such a way that could be construed as reneging on your promise

Nothing could be construed in such a way, because such a reneging is not possible.

Here is my prediction, FWIW: the hard part of the problem has already been solved, in the following technical sense: there is a few 1000 lines program that has not been invented yet, but it will be invented soon, that loads a current LLM model, runs fast on current hardware, and you will deem it to be an AGI. In other words, the conditional Kolmogorov complexity of undisputable AGI given the Llama weights is only a few 1000 bytes. We are at the pre-AlphaGo, post Clark-Storkey stage of reasoning. That's my guess, anyway.

Interesting, as a very frequent HN reader, I do not associate this pattern with Python zealots at all. Rust, Swift, Go, more like.

Bingo. That was not obvious to me before your one-word reply, and now, in retrospect, it is.

And I say this as a person who usually just rolls his eyes seeing the typical cynical HN hot takes. I believe Altman voices his deepest convictions here. The smarter thing to be cynical about is, what is his goal with voicing them now, and in this very format? As others have observed, this is the first time he is not even paying lip service to the question of existential risk.

Neither the author of the GPT from scratch post, nor eclectic29 who recommended it above did ever promise that the post is about building LLMs from the ground up. That was the original post.

The GPT from scratch post explains, from the ground up, ground being numpy, what calculations take place inside a GPT model.