[redox99]

Google (and others) may not have released model weights, but they've published papers, which is ultimately what makes the field advance. OpenAI not only did not publish any GPT4 paper, they haven't even said how many parameters it has.

[mach1ne]

Indeed Google came up with Transformers and decided to gift the model to humanity. By broad strokes it was luck that OpenAI chose the seemingly right path of AI.

Closest competitor DeepMind played games, which is intuitively closer to what humans do, but its relevance given aspects of deep learning is questionable.

[teruakohatu]

> DeepMind played games, which is intuitively closer to what humans do, but its relevance given aspects of deep learning is questionable.

Reinforcement Learning is part of what OpenAI is doing. I don't think Google went down the wrong path. If anything they should have run down the path they were on.

[buildbot]

Then what is this? 99 pages of bullshit? https://arxiv.org/pdf/2303.08774.pdf

[redox99]

> Given both the competitive landscape and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar.

It's 99 pages of marketing material

[buildbot]

Personally I disagree, there are lot of interesting tidbits in this paper. More than marketing would need at least.

[whatshisface]

What good bits did you find? (I'm not sure how fruitful the "OpenAI is a Microsoft department" debate is given that they are almost one and everybody knows it, but I am curious if anyone has found anything good in those many pages.)

[buildbot]

I think the most interesting thing is the their ability to predict performance from loss and on a wide range of tasks using a much smaller model - this lets them fine tune their architecture and hypers, then run a single large training run to get full scale gpt4 - from the paper it sounds like they only trained the large model once, then did a Reinforcement learning with human feedback finetune.

Disclaimer - I work at Microsoft, in AI, and have no internal knowledge about gpt4.

[tempusalaria]

This isn’t that interesting imo. This is the basic outcome of the scaling laws from Kaplan, Chinchilla papers pushed to a larger final model delta.

They likely did extensive small model building on the gpt-4 architecture to establish hyperparameter scaling laws and then did a predicted build in exactly the same way chinchilla did.

[thatsadude]

Google published papers but has anybody be able to replicate their results?

[redox99]

Yes? Attention is all you need and Alpha Zero are the first that come to mind, but there are thousands.