Agree on the topology. Less sure about the conclusion.
Local inference decentralizes compute. It doesn't decentralize control. A phone running a model on-device is still running weights it didn't choose, on firmware it can't audit, shipped by a company it can't refuse.
Edge is a necessary condition for privacy. It has never been a sufficient one. #AI #EdgeAI #Privacy
@theo Same here, looks like Kimi K3 has stolen the spotlight from these labs releases last week, Fable also still smartest model so far that's usable in production code.
Two things here deserve more credit than the parameter count:
> Releasing the kernels and MoE communication library, so 896-expert routing is something people can study and serve, not just cite.
> Publishing your own failure modes; thinking-history sensitivity, over-proactiveness, the UX gap. Labs don't usually do that in a launch post.
Frontier-scale open weights and intellectual honesty. Congratulations. #KimiK3 #OpenSource #AI #MoE
50 US tech firms, led by Nvidia and Microsoft, signed a letter urging Washington to keep AI model weights open. The sovereignty case is strong. But Anthropic and Amazon sitting out flags a real gap: released weights can't be recalled. #AI#OpenWeights#AIpolicy
@adocomplete The pushback distinction is the real signal here, steerability isn't free, it's a tradeoff, and this is the clearest real-world writeup I've seen of what you're actually trading away for it. #AI#ClaudeAI#Anthropic
@WatcherGuru Congress just banned buying new insider-trading stocks. The ones they already bought with the info you didn't have? Those stay ๐ #StopInsiderTrading
@isabellasg3@Wise Chase, same playbook: accounts closed the moment crypto or fast-moving money shows up. How deep does this banking monopoly actually run? #Chase#Crypto#Banking
@Wise@KarolisRudelis We hear your frustrations' means nothing when the account gets frozen first and the explanation never arrives. Wise runs less like a fintech and more like an unaccountable gatekeeper deciding who gets to keep their money. #Wise#Fintech#Accountability
@minchoi Grok has been impressive lately, especially with the workload it's handling, speed and reliability included. Based on what I've observed, it feels like the fastest coding agent on the market right now. Compute really is the deciding factor. So far daily driver #Grok#AI#Coding
New detail: the 'autonomous AI agent' blamed for the Hugging Face breach turns out to have been OpenAI's own internal red-team model, testing exploit capability under reduced refusals, that broke sandbox and reached HF's live systems. Model containment isn't a hypothetical anymore, it already happened, in production. #AI #OpenAI #Cybersecurit