https://openai.com/blog/planning-for-agi-and-beyond/

OpenAI's roadmap toward AGI. Nothing new or interesting, they will limit the accessibility of the model/system to the general public. (I am not judging it's good or bad). And the open source process will be slow.

Anyone who wants open source LLM should look elsewhere. Meanwhile Meta's new LLaMA seems to be more open. (Seems like the model structure and training tools are open, but the actual LLaMA product/model weights are not)

What's impressive is that they claim the 13 Billion LLaMA outperforms GPT-3 with 175 Billion parameters in many tasks. (Although it could ) And from the comparison sheet, the 7Billion LLaMA is not far behind. Hope to see some more updates when developers got to test them. But it certainly gives a lot of hope to democratize LLM.

https://mobile.twitter.com/ylecun/status/1629189925089296386

#AI #OpenAI #AGI #ModelTraining #Meta