pull down to refresh

27 sats \ 2 replies \ @optimism OP 28 Jul \ parent \ on: Open Source and America's AI Action Plan AI
Models themselves are a red-herring in all of this
I've felt that for a longer time. Even if there are triggers baked into weights, we can remove those from the open models simply by reinforcing those weights inversely on top - that's kind of what is done to get uncensored models (decensored would be a better name, imho.)
All the panic early this year when Deepseek released their open source model was 99% openai & co panicking that their proprietary crap got beat again and 1℅ people saying that it ought to be impossible due to export constraints. I can't help but think that that was the intended effect. Nothing burger though: only the retarded plan to prevent development through exporting only nerved chips backfired. That's the real lesson.
only as secure as the flow of energy, chips, and communications networks in the real world.
Exactly. Which is why the less hops between you and your LLM, the better, and the less dependency on 3rd parties, the better. So running an open model that you've validated yourself (maybe can use the eval stack from #1052545 for that), on your own hardware, with your own tooling... is useful no matter whether you're a sovereign state or individual.
write
preview
100 sats \ 1 reply \ @justin_shocknet 28 Jul
Deepseek
That seemed to me very coordinated to pin-prick the model bubble, ever since the focus (on CNBC et al) has been on hardware/energy more than the models
your own hardware
From the nation-state perspective absolutely, but individuals have no recourse at this stage and i'm not sure that will change... even "your" hardware is permissioned technology from the natsec supply chain. There's no meaningful opensource hardware and I suspect that's kept that way intentionally.
reply
17 sats \ 0 replies \ @optimism OP 28 Jul
That seemed to me very coordinated to pin-prick the model bubble
Yep. And that's not a loss I mourn. The bubble was ridiculous. Focusing on infra is time and money spent much better.
There's no meaningful opensource hardware and I suspect that's kept that way intentionally.
I'm still hopeful for RISC-V implementations especially now that generic compute makes space for optimized compute, which does fit. I fear it won't go much past the hobbyist horizon but if I don't use it myself, I'm not part of the solution. So perhaps I should get back into that.
reply