I think you're super wrong about the local model issue and that's a huge risk for companies like OpenAI.

Apple products as an example have an excellent architecture for local AI. Extremely high-bandwidth RAM.

If you run an OSS model like gpt-oss on a Mac with 32GB of RAM it's already very similar to a cloud experience.

i dont have the hardware to run or try them, but from the huggingfaces discussion forums, gpt-oss seems to be pretty hard censored. I would not consider it as being a viable self-hosted LLM except for the very narrowest of domains (like coding for example).

▲

dangus 4 days ago | parent [-]

I'm not sure where censorship comes in with this discussion, it seems like cloud models are censored as well? And local models are frequently created that are abliterated? Correct me if I'm wrong or misunderstanding you.

Either way, it's just an example model, plenty of others to choose from. The fact of the matter is that the base model MacBook Air currently comes with about half as much RAM as you need for a really really decent LLM model. The integrated graphics are fast/efficient and the RAM is fast. The AMD Ryzen platform is similarly well-suited.

(Apple actually tells you how much storage their local model takes up in the settings > general > storage if you're curious)

We can imagine that by 2030 your base model Grandma computer on sale in stores will have at least 32GB of high-bandwidth RAM to handle local AI workflows.

	▲	chii 4 days ago \| parent [-]
		which is why i made the claim that hardware "problem" will be solved in the near future (i don't consider it solved right now, because even the apple hardware is too expensive and insufficient imho), but the more difficult problem of model availability is much, much harder to solve.