Meta OPT-175B: the first 175-billion LLM opened to researchers

OpenAI doesn't share GPT-3 with researchers, only the API. Meta says "we will": it publishes a model of the same scale — 175 billion parameters — and gives it to researchers, with weights to download.

Even more interesting: it also publishes the full training logbook, with every technical problem encountered. It's the first time a company shows "what goes wrong" when training a model this big.

It's not truly open source for everyone — the license is research-only — but it's a major step toward open models. From here the road runs to LLaMA and Llama 2.