5 Comments
User's avatar
Teemu Maatta's avatar

Great post Nathan, I always like to read your take as super interesting.

I think open source movement is great and required, but we need to see the potential risks of releasing these models. I would prefer a large open source foundation taking charge of training data licensing and model training. Then, this would be licensed to multiple commercial players. I think some current players in training “open source” models - represent their specific interests. Thus the current open source models have lacked (a) quality like BLOOM, Galactica (b) commercial licensing terms (OPT, LLaMA) (c) risk/threat management: are any open source models producing data regards their misuse at scale?

For example practically nobody talks about the needs to have alternatives as well for OpenAI’s Moderation endpoint, which I think will be included to most commercial products released to end-clients by the enterprises.

Expand full comment
Nathan Lambert's avatar

Yeah I agree. You may like some of the post I released today too (touches on this https://robotic.substack.com/p/rlhf-battle-lines-2023).

I think about this safety issue almost daily. Not sure the right solutions now, but it's front of mind.

Expand full comment
Cubeloid's avatar

If they attach a crypto currency for micropayments on prompts would be profitable

Expand full comment
Vinay's avatar

Hi Nathan, Your articles helped me a lot in understanding recent trends. Thank you for that!.

Have a basic question: How do models like OpenAI Codex achieve syntactically correct code every time they generate code? I can understand that they can generate code but I am not able to understand how it can generate correct syntax everytime

Expand full comment
Nathan Lambert's avatar

I don't think they necessarily do, but I think GitHub Codex probably added a rule-set added to the human feedback. So when doing the reward labels, this syntax feedback is added.

We are thinking about doing a project that is similar and adds this to unit tests too.

Expand full comment