Anthropic and other AI LLM providers have censorship, so companies that need unrestricted AI are going to Chinese opensource LLMs. They'll start with the core model, use their own data, and then make their own model.
Anthropic in particular, because it is doing mandatory surveillance. They are retaining every prompt and output (and context) for 30 days. Anthropic's LLM does this to build a profile on each user and decide what kind of capabilities to lock or unlock for him. They were lowering people's access, after the people asked about chip design, LLMs, and also simple things like routine biology questions and cryptography, but they weren't even telling the user they were doing this. Sacks says this feels anticompetitive. They were even rewriting your prompt in the background and give you a 'nerfed' answer, and still charge you. 'They will start to pick which corporations they want to benefit,' added Chamath.
Meanwhile, Anthropic's leadership was making public statements that it wants the government to approve (by creating a new government agency for the purpose) or not all LLMs. Perhaps they want some way to stop opensource models.
A gigawatt now costs $100b, said Chamath, while commenting on the perhaps need to build datacenters just for opensource to have a place. When Chamath started his datacenter proposal, it cost $4 or 5b.
They seem very non-censorship, but they wouldn't be/won't be speaking the same after the first event no one wants to happen except some extremely stupid or malignant person or group, which there is. It's just too bad the government can't control itself, because if it could limit its higher abilities to just real terrorism and harm, it would be supportable, but it can't, it just goes further and further into anything at all, really, and just calls it 'terrorism' or sometimes doesn't even, and goes after non-harmful actions too.
AI censorship might be what unites the governments of the world, if some event causes people to ask for it.