> A large language model is trained to predict the next word of a piece of text. … simply asking a model to “think step by step” makes its answers better.
I thought both these statements were no longer held to be broadly true? At least in big lab or frontier models.
Next token prediction is still how pre-training works, but that's a relatively small part of the whole training run nowadays. Most of the time is spent in post-training reinforcement learning scenarios, where models are scored based on an evaluation of their ability to complete a task, not whether any specific token correctly predicts a known next token.
Generally, anyone talking about next token prediction being some limiting factor of LLM capability is a good reason to ignore them, as they haven't updated their understanding of LLMs since 2023 or so.
All that happens in reinforcement learning and everything from pre-training to generation, including the last token, at the end is about adjusting the probability of the next token given the sequence.
Anyone who thinks reinforcement learning changes how the generation works hasn’t worked on inference engines and you can dismiss them as inexperienced and misguided.
I think there's a question of semantics here. RL is way more expensive in terms of compute, as you have to generate a significant amount of output before you can judge whether it's good or bad. So compute is one angle, but another is how much impact RL has on the weights vs pre-training
The method could not be performed on the resource because the server is unable to store the representation needed to successfully complete the request. There is insufficient free space left in your storage allocation.
Additionally, a 507 Insufficient Storage error was encountered while trying to use an ErrorDocument to handle the request."
As a Prolog enthusiast and someone who had followed Newell & Simon's SOAR work etc., this was intriguing. There are some respositories that seem related:
https://github.com/pcoz/operational-representation-of-system...
But, it is disappointing that the summary does seem to be AI-generated slop and I can't find code reproducing the claims as described.
I'm pretty sure it's a different project. It's not that inconceivable that the same author has done more than one project in a related area, especially if the one you found on github is obviously different to the one described here.
It's really a nitpick but putting a big, obvious AI generated slop photo at the top of your post is a signal for bad quality content. I would recommend to avoid doing it.
I wouldn't even mind it if it weren't so blatantly lazy (lazier). It's always that off-putting overly busy default look. Like, you couldn't even be bothered to ask for some refinement.
I can't see where you can try this out, so at the moment it's a lot of claims and no real evidence. Also the text feels like it was written by an LLM.
If it's really such a small model why not provide a web interface where we can test it for ourselves?
This feels like high risk of being crackpottery.
And all illustrations have the AI look (they're not AI images, but AI-generated SVG or Powerpoint or something)
It was definitely written by an LLM, the entire article is full of “it’s not X, it’s Y” overconfident truisms
Yep:
> The relation between them always has to be supplied, and supplying it is what the system’s logic is.
100% sure this is Claude's recent writing style at work
The associated paper is (I will say generously) not forthcoming: https://zenodo.org/records/23526702
I can also find no working model. I expect this may be disregarded for now.
> A large language model is trained to predict the next word of a piece of text. … simply asking a model to “think step by step” makes its answers better.
I thought both these statements were no longer held to be broadly true? At least in big lab or frontier models.
Next token prediction is still how pre-training works, but that's a relatively small part of the whole training run nowadays. Most of the time is spent in post-training reinforcement learning scenarios, where models are scored based on an evaluation of their ability to complete a task, not whether any specific token correctly predicts a known next token.
Generally, anyone talking about next token prediction being some limiting factor of LLM capability is a good reason to ignore them, as they haven't updated their understanding of LLMs since 2023 or so.
All that happens in reinforcement learning and everything from pre-training to generation, including the last token, at the end is about adjusting the probability of the next token given the sequence.
Anyone who thinks reinforcement learning changes how the generation works hasn’t worked on inference engines and you can dismiss them as inexperienced and misguided.
I think you're conflating prediction and training. The models are auto-regressive so they always predict the next token.
> Next token prediction is still how pre-training works, but that's a relatively small part of the whole training run nowadays
Only if you call 90% small.
No-one is claiming <50%.
It's only about 10% of total compute capacity. https://x.com/EugeneNg/status/2099315982959616369
I think there's a question of semantics here. RL is way more expensive in terms of compute, as you have to generate a significant amount of output before you can judge whether it's good or bad. So compute is one angle, but another is how much impact RL has on the weights vs pre-training
I cannot see how cost is relevent. What we're discussing is impact.
Only the meaning of "next" and "word" as well as "next word" has become more complicated, otherwise it still holds true as a general mechanism.
Never seen that one before:
"Insufficient Storage
The method could not be performed on the resource because the server is unable to store the representation needed to successfully complete the request. There is insufficient free space left in your storage allocation.
Additionally, a 507 Insufficient Storage error was encountered while trying to use an ErrorDocument to handle the request."
This is what you get when you employ a stochastic parrot to deliver a 507.
When I read the title, I thought this was a joke and the article would just say: "Think by yourself".
Although that's arguably way superior to AI-style reasoning.
As a Prolog enthusiast and someone who had followed Newell & Simon's SOAR work etc., this was intriguing. There are some respositories that seem related: https://github.com/pcoz/operational-representation-of-system... But, it is disappointing that the summary does seem to be AI-generated slop and I can't find code reproducing the claims as described.
> without a LLM https://github.com/pcoz/decisionator
2. Obtain an OpenAI API Key
I'm pretty sure that's a different project
Its the only repo by the author(Edward Chalk) that focuses on decision making models? Its his github account linked from the front page.
I'm pretty sure it's a different project. It's not that inconceivable that the same author has done more than one project in a related area, especially if the one you found on github is obviously different to the one described here.
Chicanery! As others have stated, this is total AI slop. How do these things get upvoted?
I want to believe… but..
Ironically it seems to have been made by LLM ...?
This might have been written by someone suffering from AI psychosis.
It's really a nitpick but putting a big, obvious AI generated slop photo at the top of your post is a signal for bad quality content. I would recommend to avoid doing it.
And using AI's favourite phrase, a deep dive, is also offputting.
(I mean, is he grading the dive?)
I wouldn't even mind it if it weren't so blatantly lazy (lazier). It's always that off-putting overly busy default look. Like, you couldn't even be bothered to ask for some refinement.