“So they’re in the wild now. You may not be legally able to build a commercial product on them, but the genie is out of the bottle.”
Facebook gated LLaMA as a research preview, and someone opened a pull request on its own repo linking a BitTorrent download. llama.cpp’s 4-bit quantization shrinks the 13B model, the one Facebook claims is competitive with GPT-3, to under 8GB on a laptop. OpenAI’s safety story depended on controlling access through its API. That thin layer of defence is gone, and the model makes things up as easily as it states facts.