Top all-time · #2681
Adam Karvonen
Personal blog · adamkarvonen.github.ioaiprogramming
Adam Karvonen's personal blog at adamkarvonen.github.io covers ai and programming. Since 2024, 23 different Hacker News readers have linked it 30 times, most often “Chess-GPT's Internal World Model”. It is still writing: the last post was on 2025-11-28, and you can follow it by feed.
Adam Karvonen’s personal website
Most-linked posts?
- Chess-GPT's Internal World Model14 people · 19 mentions · first linked 2024-01-06
- Frontier AI Models Still Fail at Basic Physical Tasks5 people · 5 mentions · first linked 2025-04-14
- An Intuitive Explanation of Sparse Autoencoders for LLM Interpretability3 people · 3 mentions · first linked 2024-06-25
What Hacker News says?
None of your examples refute the direct evidence of internal world model building which has been demonstrated (for example: https://adamkarvonen.github.io/machine_learning/2024/01/03/c... ). Instead you have retreated to qualia like "well" and "sucks hard". > hallucinating Literally every human memory. They may seem tangible to you, but they're all in your head. The result of neurons behaving in ways which have dire…
Comment linking /machine_learning/2024/01/03/chess-world-models.html
on “Andrej Karpathy – It will take a decade to work through the issues with agents”Submitted Chess-GPT's Internal World Model as a story · 120 points
HN discussion1000% this. I would only add this has been demonstrated explicitly with chess: https://adamkarvonen.github.io/machine_learning/2024/01/03/c...
Comment linking /machine_learning/2024/01/03/chess-world-models.html
on “Andrej Karpathy – It will take a decade to work through the issues with agents”Submitted Manipulating Chess-GPT's World Model as a story · 119 points
HN discussion> LLMs can't think. They are generating tokens one at a time Huh? They are generating tokens one at a time - sure that's true. But who's shown that predicting tokens one at a time precludes thinking? It's been shown that the models plan ahead, i.e. think more than just one token forward. [1] How do you explain the world models that have been detected in LLMs? E.g. OthelloGPT [2] is just given sequences of games to t…
Comment linking /machine_learning/2024/01/03/chess-world-models.html
on “Thoughts on thinking”Submitted An Intuitive Explanation of Sparse Autoencoders for LLM Interpretability as a story · 110 points
HN discussionrelated : Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task https://arxiv.org/abs/2210.13382 Chess-GPT's Internal World Model https://adamkarvonen.github.io/machine_learning/2024/01/03/c... discussed here https://news.ycombinator.com/item?id=38893456
Comment linking /machine_learning/2024/01/03/chess-world-models.html
on “Something weird is happening with LLMs and chess”Submitted An Intuitive Explanation of Sparse Autoencoders for LLM Mech Interpretability as a story · 4 points
HN discussionChess is not stateless. En Passant requires last move and castling rights requires nearly all previous moves. https://adamkarvonen.github.io/machine_learning/2024/01/03/c...
Comment linking /machine_learning/2024/01/03/chess-world-models.html
on “Something weird is happening with LLMs and chess”Submitted Frontier AI Models Still Fail at Basic Physical Tasks as a story · 3 points
HN discussionHey - Thanks for sharing! Will take a closer look later but if you are hanging around now, it might be worth asking this now. I read this blog post recently: https://adamkarvonen.github.io/machine_learning/2024/06/11/s... And the author talks about challenges with evaluating SAEs. I wonder how you tackled that and where to look inside your repo for understanding the your approach around that if possible. Thanks agai…
Comment linking /machine_learning/2024/06/11/sae-intuitions.html
on “Show HN: Llama 3.2 Interpretability with Sparse Autoencoders”Submitted Frontier AI Models Still Fail at Basic Physical Tasks: A Manufacturing Case Stu as a story · 3 points
HN discussion>We don't build our world model fresh with every sentence someone says to us Neither do LLMs. The state for the current text perhaps. Definitely not the entire world model(s) which is learnt from the training process and stored in its weights. >They quite literally have no ability to have a "world model" You keep repeating this so let's get one thing straight. You're wrong. You're just wrong. I'm not trying to convi…
Comment linking /machine_learning/2024/01/03/chess-world-models.html
on “Hallucination is inevitable: An innate limitation of large language models”Submitted Frontier AI Models Fail at Basic Physical Tasks: A Manufacturing Case Study as a story · 2 points
HN discussion>We don’t know if it’s using implicit search either. Sure >it’s also possible it has just memorized the evaluations from 10M games and is performing some function of the similarity of the input to those previously seen. That's not possible. The possible set of moves in chess is incredibly large and it is incredibly easy to play a game that has diverged from training. a model that has just memorized all evaluations w…
Comment linking /machine_learning/2024/01/03/chess-world-models.html
on “Grandmaster-Level Chess Without Search”Submitted Frontier AI Models Still Fail at Basic Physical Tasks as a story · 2 points
HN discussionhttps://adamkarvonen.github.io/machine_learning/2024/01/03/c... would be a better link here.
Comment linking /machine_learning/2024/01/03/chess-world-models.html
on “Chess-GPT's Internal World Model”Submitted Chess-GPT's Internal World Model – Adam Karvonen as a story · 2 points
HN discussionAs described in the OP's blog post https://adamkarvonen.github.io/machine_learning/2024/01/03/c... - one of the incredible things here is that the standard GPT architecture, trained from scratch from PGN strings alone, can intuit the rules of the game from those examples, without any notion of the rules of chess or even that it is playing a game. Leela, by contrast, requires a specialized structure of iterative tree…
Comment linking /machine_learning/2024/01/03/chess-world-models.html
on “Chess-GPT's Internal World Model”Submitted LLM inference is nearly deterministic. We use this to audit providers as a story · 1 points
HN discussionThe previous blog post [0] mentioned they were using stockfish with 0.1 seconds per move. [0] https://adamkarvonen.github.io/machine_learning/2024/01/03/c...
Comment linking /machine_learning/2024/01/03/chess-world-models.html
on “Manipulating Chess-GPT's World Model”I'm waiting for a build so I went fully down this trail. XD A1) Yes it can be done with a 4090. (2a). A2) 2 days. (4d/4e). B) Up to you, author did both and settled on 50M which they got to 1300 ELO. (note: they also did subsequent work with the same model to increase perf without further training) (2a). C) Mu: there's no limit as you can page in/out (3a, 3c), and its trivial to store in both RAM and VRAM. 30B fits…
Comment linking /machine_learning/2024/01/03/chess-world-models.html
on “Manipulating Chess-GPT's World Model”More details in the previous blog post: https://adamkarvonen.github.io/machine_learning/2024/01/03/c... > A 50 million parameter GPT trained on 5 million games of chess learns to play at ~1300 Elo in one day on 4 RTX 3090 GPUs. And from the paper: https://arxiv.org/abs/2403.15498 > The 25M parameter model took 72 hours to train on one RTX 3090 GPU. The 50M parameter model took 38 hours to train on four RTX 3090 GPUs…
Comment linking /machine_learning/2024/01/03/chess-world-models.html
on “Manipulating Chess-GPT's World Model”>I really think chess is just a terrible example though. You're really asking a lot of it but I'm honestly shocked it can do what it can. It seems to know some opening books, but falls down immediately. Which really makes sense because you'll find a lot of reading material on specific openings, but the problem space of the game is just too big to find texts about any given game state. Maybe if you have it reason abo…
Comment linking /machine_learning/2024/01/03/chess-world-models.html
on “Is GPT-4 a good data analyst? (2023)”>Similarly, it had no idea about the quality (and sometimes legality) of the chess moves it generated. LLMs can play chess just fine. This isn't some inability. You just tested a model that couldn't. GPT-3.5-instruct-turbo plays at about 1800 Elo and makes legal moves consistently well past openings. https://adamkarvonen.github.io/machine_learning/2024/01/03/c... >Your snake site is probably a good example. ChatGPT…
Comment linking /machine_learning/2024/01/03/chess-world-models.html
on “Is GPT-4 a good data analyst? (2023)”>they don't appear to do the obvious thing which is display the reconstructed "internal" board state This is literally figure 4 This also re-constructs the board state of a chess-playing LLM https://adamkarvonen.github.io/machine_learning/2024/01/03/c...
Comment linking /machine_learning/2024/01/03/chess-world-models.html
on “Do Large Language Models learn world models or just surface statistics? (2023)”Since you can train an LLM to play chess from scratch, I would not be surprised if you could also train one to play Set. I might experiment with it tomorrow. https://adamkarvonen.github.io/machine_learning/2024/01/03/c...
Comment linking /machine_learning/2024/01/03/chess-world-models.html
on “Are LLMs able to play the card game Set?”Also a nice article[0] featured on here[1] recently. [0] https://adamkarvonen.github.io/machine_learning/2024/01/03/c... [1] https://news.ycombinator.com/item?id=38893456
Comment linking /machine_learning/2024/01/03/chess-world-models.html
on “Neural Chess”CAD and machining are different fields, true, but I see a lot of the same flaws that Adam Karvonen highlighted in his essay on LLM-aided machining a few months ago: https://adamkarvonen.github.io/machine_learning/2025/04/13/l... Do any people with familiarity on what's under the hood know if the latent space produced by most transformer paradigms is only capable of natively simulating 1-d reasoning and has to kludge…
Comment linking /machine_learning/2025/04/13/llm-manufacturing-eval.html
on “Show HN: AI in SolidWorks”
