Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Why I'm still bearish on LLMs after Navier-Stokes

dank.systems

Post reply on HN
Print view View original

361–370 of 642 posts

Re: Why I'm still bearish on LLMs after Navier-Stokes

#361

Post by goolz »

intrasight wrote: post #258
dotdi wrote: post #211

Earlier quoted context omitted.

The whole point of this post was that it's questionable what can be achieved without huge investments into oversight and steering, because navier-stokes was a topic with an unusual level of specification. The problem itself was a specification. Such situations are rare in real-world scenarios. AI agents are good at solving well-specified tasks, not at solving problems. They do well in fields where the cost/effort of…

> a topic with an unusual level of specification Solving cancer also has an unusual level of specification. Many real world problems have that characteristic.

Maybe a specific cancer but in general “solving cancer” is about as scattershot and unspecific as you could be.

Re: Why I'm still bearish on LLMs after Navier-Stokes

#362

Post by tired-turtle »

roenxi wrote: post #344
killerstorm wrote: post #343

Earlier quoted context omitted.

This is an absolute nonsense. Any frontier model can implement chess program from scratch - modeling the board, checking legality, etc. If you asked e.g. GPT-6 to get good at chess and gave it a computer, it will get good at chess. That's an actual strategic skill. Asking GPT to play chess directly using its reasoning only tests its reasoning ability to model chess state. Which it really is NOT optimized for. This is…

It's an interesting puzzle, isn't it. On the one hand, the AIs are no good at playing Chess. However, on the other hand, if you ask an AI to win a game of chess it has all the tools on hand to compete at the same level as Stockfish - it can re-implement an engine and even probably has a GPU on hand to train its own neural nets. So should we say that the AI can play chess well, or that it cannot?

Is it, though? If you design and build a winning F1 race car, did you also win the race?

Recent discourse around AI seems to conflate the semantics of winning: 1. you contributed to the win vs 2. you yourself were the winning driver.

Re: Why I'm still bearish on LLMs after Navier-Stokes

#363

Post by glitchc »

carodgers wrote: post #17

This April 2026 paper is a fun and related read. https://arxiv.org/html/2509.24239v4 Researchers asked frontier models to play chess. Have a look at the MAR rates in Table 3. When not explicitly told which moves were legal, no model identified legal moves at a rate better than 80%. Many asked for more illegal moves than legal moves. And even when explicitly told which moves were legal, the models continued to ask for…

Why not ask it to implement a chess engine first, and then use that to play against you?

Does the LLM need to learn to play chess if it can build a chess engine to play for it instead?

Re: Why I'm still bearish on LLMs after Navier-Stokes

#364

Post by alpinisme »

danpalmer wrote: post #347
titzer wrote: post #321

Earlier quoted context omitted.

[flagged]

Sure, but installing a chess program is child/teen level general ability, and playing chess well is highly trained expert level ability. Which one are we sold AI as being?

I think we are being sold AI as expert only when given tools (although that is not emphasized). The (quasi?) miracle of AI right now is that you can get an agent to accomplish the task of a team of intelligent but not exceptional humans at speeds far exceeding what the human could do. Which makes it “cheap” to throw (effectively) dozens of teams at a problem for the equivalent of hundreds of man hours.

That may not be the AI of sci fi fantasy but it’s still a game changing reality.

Re: Why I'm still bearish on LLMs after Navier-Stokes

#365

Post by phoghed »

fc417fc802 wrote: post #235
phoghed wrote: post #113

Earlier quoted context omitted.

They’ll never be AGI simply because the definition will be constantly updated to be some steps ahead of them.

I'm pretty sure "competent at chess without external aids" has been on the standard AGI checklist since before personal computers were a thing. How can you claim an intelligence is general if it can't make sense of such a highly constrained board game? This is solidly table stakes.

Because they’ll train it to be good at chess and then everyone will say yeah but playing chess doesn’t mean you’re AGI, it can’t even ____

It can’t even count the R’s in strawberry

It can’t even add numbers

It can’t even solve a millennium puzzle

It’s not even a chess GM

It’s not even beyond human capability in Go

It can’t even drive a car

It can’t even self replicate

It can’t even build weapons

It doesn’t even have feelings

So how could someone conceivably convince everyone that some system is AGI when there are still tasks that some human or group of humans can do that the system cannot?

This will only happen, in my opinion, when the model/system can self-improve at a rate that scares people.

Re: Why I'm still bearish on LLMs after Navier-Stokes

#366

Post by moomoo11 »

lolakutty wrote: post #232
moomoo11 wrote: post #136

the issue most of you seem to not realize is that when you put these models in a loop, you are able to do more and more insane and cool things. have you guys actually designed, built, and deployed agentic workflows? it is actually quite hard, requires tons of time spent on evals and testing to ensure accuracy, but when it starts to work it is mind blowing. there is no going back. listening to people yap about AI when…

> when you put these models in a loop, you are able to do more and more insane and cool things... Please share some of these insane things that you speak of..

i mean have you used any coding agents? if you’re getting slop code in 2026, that’s a smell and skill issue. fwiw i was pretty bearish on AI until i spent a month a couple months ago going deep into agentic workflows.

use your imagination to solve problems people face and pay $$$ for today that is error prone and hard.

i’ve got agentic workflows for the particular industry im building for, one of which that replaces the need to hire $500+/hr services.

in this particular workflow (don’t want to reveal too much, sorry this is my competitive advantage but you can figure it out for your own workflows) a $4/1M model ingests a file that is currently used in a extremely complicated program that few people understand how to use.

it parses the data, loads it into a database, and then spawns a bunch of other agents that check the data against work in flight. there’s checks for bad data. in that case, more agents are spawned that reach out to the involved people or parties for clarification. if it cannot figure something out it reaches out to the right contacts for more information. while this is happening, more agents begin doing work that involves continuous reconciliation against 100s or 1000s or more things in flight.

as files are uploaded, or updates from people come in, agents do work to ensure things remain on track.

people are able to work across languages and cultures, and my agents ensure that while people can make mistakes, it will catch them in real time and ensure continuously monitor the situation.

it’s pretty nuts how much inefficiency agents today can solve. it takes patience to run tests and tweak shit until it works.

*** the really cool thing is that more capable agents can continuously monitor how things are going and improve the workflow itself… so all i need to do is maintain the actual tests. ****

i loved writing tests back in the day to ensure i built good software. today we write tests to ensure the business can run.

Re: Why I'm still bearish on LLMs after Navier-Stokes

#367

Post by Topfi »

user43928 wrote: post #341
Topfi wrote: post #293

Earlier quoted context omitted.

Fortunately, a fellow commenter was so kind and did it with Astra. Didn't do that well either [0]. I'm sure GPT-7 will be super mega ASI regardless (since GPT-6 Astra already claimed AGI in the minds of Jen-Hsun, et al.)... I'll say it till there is any evidence of the contrary, LLMs are not intelligent and their capabilities solely within the realms of well tailored training data. "Just" having been trained on every…

Doesn't look impressive, although I'm hearing a marked improvement in choosing legal moves, compared to early 2025. Given the pace of improvements, is it really unimaginable that GPT-7 will play Chess reasonably well and generalize better? I would not be surprised if OpenAI released a model that beats humans at chess this year.

I very much agree that the next models will be better, heck, I still suck at hobbyist training and could probably coax t5 to do better in Chess specifically, just need to get loads of data from Stockfish.

Thing is, given what GPT-6 Astra was trained on and what models of a similar class can do (including developing a competitive chess engine), it is often paradoxical and somewhat surprising how little these models have gained in actually capability that is in the training data, but not RLHFd to hell, so to speak. Tracking the state of pieces, I suspect given similar in Sudoku [0], is what these models struggle with in game settings, whilst tracking the state of code changes can be reliable over 250k tokens. Essentially, for the latter they were trained in the specific manner that lead them to abstract the capability, but that doesn't track to the former, which is a massive difference between LLMs data focused training and human learning.

So yeah, GPT-7 or any upcoming/present LLM could do massively better in Chess than GPT-6 Astra, but not because the approach was emergent out of pure data. Rather, it requires a very specific training data type and stack for a model to gain capabilities that track a specific task long enough to adhere to the rules of a game such as chess.

[0] https://logicalintelligence.com/blog/energy-based-model-sudo...

Re: Why I'm still bearish on LLMs after Navier-Stokes

#368

Post by kbau »

bitexploder wrote: post #349
kbau wrote: post #329

Earlier quoted context omitted.

Neural Nets can be trained to play chess very well and have been doing so for a long time (see Stockfish and Leela as some of the most popular/strongest ones - top GMs have no chance against them), but these are dedicated models, where the game rules are encoded in the learning process, as opposed to large language models which are natural language processing models. Technically you can give an LLM a lot of chess boo…

https://en.wikipedia.org/wiki/Chinese_room I think about this once in a while. At some point if it does the thing almost perfectly is it still not doing the thing?

I suppose if you all you need is a good enough opponent for the average person out there, sure this is good enough.

I was more talking in reference to why the LLMs in the above linked paper were producing so many illegal moves, and it is because they are not hard constrained by the rules of the game. Of course, a loop can prompt until a valid move is produced and then rendered on a screen. But why do this? I suppose, who am I to say what should be done or not, but a specialized tool being better than a general one at its specific job isn't particularly surprising.

Re: Why I'm still bearish on LLMs after Navier-Stokes

#369

Post by hackinthebochs »

YeGoblynQueenne wrote: post #280
linkjuice4all wrote: post #158

Earlier quoted context omitted.

State-sponsored psyop meta comments aside, the models obviously continue to get better, but there is still a lot of 'guard railing' required to keep even the latest models completely on-task. The chess example is interesting because it's clearly a well-studied and established domain so the rules, strategies, and whatever else is in the training data should make yield excellent results; but clearly there is some behav…

For me the useful intuition is that LLMs haven't somehow magickally learned to implement any of the algorithms we know that we have used to make strong chess engines: alpha-beta minimax and Monte-Carlo Tree Search on the one hand, and obviously the ability to learn accurate evaluation functions by self-play. I mean we've done all this before in a task-specific fashion. It's useful to know that LLMs haven't managed to…

>If that were true, we should have seen LLMs play good chess by now.

Not at all. LLMs learn by imbibing a mass of relationships as isolated fragments of information. There is a certain amount of sorting and indexing that happens during the training phase. There is also a certain amount of compute executed on these relationships during inference. LLMs can model processes that fit within the compute budget. Language translation works well because language is lookup-heavy while being light on compute.

Chess is a compute heavy game of finding the best move out of many possibilities with wide variation in the quality of each move. Humans cut through the compute requirements by reinforcement and learning intuition. LLMs don't get reinforcement on chess so they must compute during inference a unified model of chess. Developing a strong model of chess from raw fragments of information is simply not in their compute budget.

Re: Why I'm still bearish on LLMs after Navier-Stokes

#370

Post by lukewarm707 »

nl wrote: post #272
lukewarm707 wrote: post #131

Earlier quoted context omitted.

ai has a >10% chance of causing human extinction, according to anthropic big heads. if that's true, you are wrong. if that's false, anthropic is dishonest. why trust a dishonest company to be worth anything?

> if that's false, anthropic is dishonest. why trust a dishonest company to be worth anything? This logic doesn't follow at all. If their argument is that there is 10% chance of extinction then they also believe there is a 90% chance it won't.

recalling the exact phrasing, several senior people at anthropic made public statements agreeing a 10% chance of causing extinction in less than 10 years.

10% is uninsurable, priced in with ordinary treatment of risk it suggests that anthropic should be worth zero today. creating that risk would put every executive in jail.

on top of that it would demand under existing laws of conflict, a military campaign to destroy anthropic. that is not optional, it is demanded now to save lives.

hard to make comparisons but we mourned and rembered 9/11 recently. a 10% risk of hundreds of millions dead in 10 years would make anthropic a thousands of times greater threat than al qaeda. many countries would assassinate dario amodei and the leadership of anthropic now, within weeks or months.

actually just on the vague risk of having a nuclear weapon in 10 years, the USA killed ayatollah khamenei, his daughter, his son-in-law, his daughter-in-law and his 14 month old granddaughter. then, they killed over 120 children ages 6-12 by accidentally bombing a school.

in the sense that i would analyse a company, at least, the claim is false. it's not true that ai has a 10% chance of causing human extinction within 10 years.

they are making false claims about the technology they sell.

i have a fairly inflexible approach to that. sure, exaggerate but outright lies about the nature of the product don't work for me.

Post reply on HN