4 ms·
Very interesting. I guess this is the strawberry model that was rumoured. I am a bit surprised that this does not beat GPT-4o for personal writing tasks. My ex
by patapong 2y ago
Very interesting. I guess this is the strawberry model that was rumoured.
I am a bit surprised that this does not beat GPT-4o for personal writing tasks. My expectations would be that a model that is better at one thing is better across the board. But I suppose writing is not a task that generally requires "reasoning steps", and may also be difficult to evaluate objectively.
- markonen 2y agoIn the performance tests they said they used "consensus among 64 samples" and "re-ranking 1000 samples with a learned scoring function" for the best results. If they did something similar for these human evaluations, rather than just use the single sample, you could see how that would be horrible for personal writing.
- janalsncm 2y agoI don’t understand how that is generalizable. I’m not going to be able to train a scoring function for any arbitrary task I need to do. In many cases the problem of ranking is at least as hard as generating a response in the first place.
- afro88 2y agoThe solution of the cipher example problem also strongly hints at this: "there are three r's in strawberry"
- patapong 2y agoConfirmed by the verge: https://www.theverge.com/2024/9/12/24242439/openai-o1-model-reasoning-strawberry-chatgpt https://www.theverge.com/2024/9/12/24242439/openai-o1-model-...
- slashdave 2y ago> My expectations would be that a model that is better at one thing is better across the board. No, it's the opposite. This is simply a function of resources applied during training.
- patapong 2y agoTo some extent I agree, but until now all of the big jumps (GPT2 -> GPT3 -> GPT4) have meant significant improvements across all tasks. This does not seem to be the case here, this model seems to be vastly stronger on certain tasks but not much of an improvement on other tasks. Maybe we will have to wait for GPT5 for that :)
- slashdave 2y ago> (GPT2 -> GPT3 -> GPT4) have meant significant improvements These were all train the same way. It's fairly clear that o1 was not. > Maybe we will have to wait for GPT5 for that :) There will be no GPT5, for the simple reason that scaling has reached a limit and there is no more text data to train on.
- nickfromseattle 2y agoMaybe math is easier to score and do reinforcement learning on because of it's 'solvability' whereas writing requires human judgement to score?