pull down to refresh
To the end that it doesn't really matter if concurrently the US and Chinese govt capture all the models out there and put access controls in place, because you can already have something running locally across 4 Mac Studios that outperforms (in capabilities and precision, not speed) Grok-4.20.
And with 48-64 H200s (and if you tune hard, 64 H100s) you can finetune every model out there, including massive ones like Kimi K2.6 and GLM 5.1, meaning you can evolve it sovereignly and no one will know until (and if) you put it out there. All the huggingface tooling is ultimately built on open source python libs, git LFS and some well described APIs which you can replicate single-shot with a 3rd grade sovereign coding LLM, and enables you to pop up whatever you did anywhere at any time. It's not just that resistance can be theoretically possible, you can do it today.
So it doesn't really matter what any government does at this point. If you have a few TBs of disk you can just keep a rolling archive of bleeding edge open weights and tooling until someone shuts it down. And honestly, it makes no sense to gatekeep or shut it down because all that will do is further embolden the boldest among us to defy the overlords. It's out there. It cannot be stopped.
Bottom line, you cannot be denied access today, and you cannot be denied access in the future. The US corpos can deny you as a paying customer but that is easy to overcome for anyone not broke and technically capable.
Relevant topics in different areas that apply in the same way (off the top of my head):
That's the direction I thought you were going.
One of the other things I wonder about is what the rate of diminishing marginal returns is. I imagine AI for personal use doesn't need to be nearly as powerful to still deliver almost the full value.
That's a really good question.
If people really just want a personal assistant ala clawbot then you can run a 5 month old model like GLM-4_7-Flash-REAP-23B on a mac mini with 32-64 GB unified memory easily and it will read, summarize and reply to all your emails, post on SN for ya so we have something to downzap, and manage your calendar for you. Even participate in that non-existing agent economy that the entirety of bitcoin twitter + nostr tries to convince the world is real. The model overall performs worse than Deepseek R1 (from early 2025) for knowledge/chat tasks, but you can run it on low-end hardware and it is instructed in executing tasks.
I have ran this exact thing for a month and a bit until I got dissatisfied with the entire *claw architecture and simply shut it off. Also I had nothing for it to do because (a) it is a retarded vision in the first place and (b) since I'm not giving it my pgp keys or access to signal, it can't really read anything worth reading anyway. For the normie, when combined with search integration, is probably all they'll ever need. But, as Apple is just going to put this on everyone's phone through Siri, normies don't need that mac mini at all.
I think that if normies stay as enslaved to the prefabricated crap they doomscroll every day and do not empower themselves, there will be nothing but UBI a-la The Expanse for them in the future. Maybe with 5% of humanity we can do all the breakthroughs and good riddance on the 95% though, but they'll drown in compliance hell and I don't really wish that to the worst of my enemies. I'd rather have everyone at least have the opportunity to do great things.
Most roads seem to lead towards UBI but it could be relatively unimpactful, like most minimum wage levels, if the economy is allowed to function because almost everyone will be producing far more than it.
To what end?