You really think that an Othello board could represent the world? Talk about being delusional. This is pathetic. The subheadline: “OTHELLO-GPT USUALLY PREDICTS LEGAL MOVE” says it all - it cannot have a representative internal world model, because in that case the “USUALLY” would go away, as anyone who has played othello for longer than 10 minutes will tell you - and we are talking about a game with very simple rules here, the complexity does not even reach chess - and people expect these things to spit out coherent and valid solutions from target selection in wars to the color their bedroom walls should be painted in.
Edit: I really have to add here the following excerpt (percentages are the rate of invalid moves as the top choice):
The error rate is 0.01% for Othello-GPT trained on the synthetic dataset and 5.17% for Othello-GPT trained on the championship dataset. For comparison, the untrained Othello-GPT has an error rate of 93.29%.
So even if you just feed them real world data without completely synthesizing it, 1 in 20 moves are invalid in a very simple (rule-side) game. If you do not give them a whole lot of played games as a basis, it fails in over 93% of cases (not at making a GOOD move, just at making any valid move!), which is probably very close to a random guess for a move. Something that has even an inkling of an internal world view would be able to adapt after being told that their move is invalid. Also, an internal world view would allow for simply dropping the rulebook onto ChatGPT as a basis, as it works with humans that can imagine.
You really think that an Othello board could represent the world?
No, you seem to be misunderstanding what a “world model” is. That research was on a toy model a fraction of a fraction of the size of modern models and exclusively trained on board notation moves. And yet it modeled the geometry of a board and positions of pieces and ‘my’ vs ‘opponent’ status of those pieces (there’s been multiple follow-ups and replications looking deeper at what encoded).
The question it answered is that transformers don’t simply learn surface statistics from the data but build world models from the data they are trained on.
So when you take a model orders of magnitude larger and train them on collective human culture orders (of orders) of magnitude larger than the board moves in the Othello-GPT paper, what do you think happens? No world modeling at all? Well there’s been a number of interpretability papers since 2023 finding that modem LLMs have functional models of all sorts of things, even including tracking emotions for themselves vs others.
Also, you are assuming that a trained world model is a lossless identical version, which it isn’t and nobody claimed. In fact, arguably the best way to identify AI generated videos these days is by over-expressed emotions of characters. Why? Because the training data is over indexed on influencer exaggerated faces which game clicks so video and image world models of human emotions are miscalibrated.
Transformers have a world model, but often it’s of a kind of parallel imagined/extrapolated world that’s not the real one. Another good example of this is how most models are certain that real world news after their cutoff isn’t true because extrapolating from their internal world model (from the training data before their cutoff) those events seem improbable. The real world doesn’t exactly match the internal model and that drift grows the more they need to extrapolate.
If you want to continue to discuss (I’m happy to help explain and educate you), I’d recommend reading up a bit on what “world model” means in relation to the tech.
You really think that an Othello board could represent the world? Talk about being delusional. This is pathetic. The subheadline: “OTHELLO-GPT USUALLY PREDICTS LEGAL MOVE” says it all - it cannot have a representative internal world model, because in that case the “USUALLY” would go away, as anyone who has played othello for longer than 10 minutes will tell you - and we are talking about a game with very simple rules here, the complexity does not even reach chess - and people expect these things to spit out coherent and valid solutions from target selection in wars to the color their bedroom walls should be painted in.
Edit: I really have to add here the following excerpt (percentages are the rate of invalid moves as the top choice):
So even if you just feed them real world data without completely synthesizing it, 1 in 20 moves are invalid in a very simple (rule-side) game. If you do not give them a whole lot of played games as a basis, it fails in over 93% of cases (not at making a GOOD move, just at making any valid move!), which is probably very close to a random guess for a move. Something that has even an inkling of an internal world view would be able to adapt after being told that their move is invalid. Also, an internal world view would allow for simply dropping the rulebook onto ChatGPT as a basis, as it works with humans that can imagine.
No, you seem to be misunderstanding what a “world model” is. That research was on a toy model a fraction of a fraction of the size of modern models and exclusively trained on board notation moves. And yet it modeled the geometry of a board and positions of pieces and ‘my’ vs ‘opponent’ status of those pieces (there’s been multiple follow-ups and replications looking deeper at what encoded).
The question it answered is that transformers don’t simply learn surface statistics from the data but build world models from the data they are trained on.
So when you take a model orders of magnitude larger and train them on collective human culture orders (of orders) of magnitude larger than the board moves in the Othello-GPT paper, what do you think happens? No world modeling at all? Well there’s been a number of interpretability papers since 2023 finding that modem LLMs have functional models of all sorts of things, even including tracking emotions for themselves vs others.
Also, you are assuming that a trained world model is a lossless identical version, which it isn’t and nobody claimed. In fact, arguably the best way to identify AI generated videos these days is by over-expressed emotions of characters. Why? Because the training data is over indexed on influencer exaggerated faces which game clicks so video and image world models of human emotions are miscalibrated.
Transformers have a world model, but often it’s of a kind of parallel imagined/extrapolated world that’s not the real one. Another good example of this is how most models are certain that real world news after their cutoff isn’t true because extrapolating from their internal world model (from the training data before their cutoff) those events seem improbable. The real world doesn’t exactly match the internal model and that drift grows the more they need to extrapolate.
If you want to continue to discuss (I’m happy to help explain and educate you), I’d recommend reading up a bit on what “world model” means in relation to the tech.