🚀 Introducing svmLLM and evmLLM 🚀 Now you can run Solana and Ethereum smart contract bytecode directly inside LLM transformer weights, checked against reference VMs. Threads, repos, explanations and sample code below.👇

there’s a perps platform on Solana that did over $100M in volume this month. It’s called @ImperialPerps no QT from Toly and Matt has no idea it exists I guess? or he’s just acting like he doesn’t? Either way, his performative ignorance is telling

I am for an AI pause, meaning a pause to training frontier models. I think the AI models we have now are amazing. They save an incredible amount of human labor and can help drive science and technology forward and accelerate them enormously. I reckon it would take a decade to even disseminate current AI capabilities through society and make full use of them. Driving the frontier forward from here will massively increase the risks. If we value human life, we should try to keep things at timescale that humans can react. The risks of things happening that most of us find abhorrent will increase a lot otherwise. This will mean not getting some of the benefits of AI, but on balance, I think the risks are now getting too high. Practical ways to implement this pause are: - USA to ban training of new frontier models in a suitable way - USA and China making a treaty and encouraging more countries to join - Enforcement can be via limits on hardware or large data centers, probably both I don't think that this pause should be forever. The goal is to slow down the frontier, let society catch up, and make it more resilient (better resistance to cyber and biological threats, individuals managing more of their life with AI , democratic resilience). In a few years time, we can be in a better place to take the next steps. The goal is not to stop AI and its benefits forever, it's preserving and improving human life through it, and that's only possible by slowing it down to human timescales for now.

1/4 Meet Qwenputer: Qwen3.5-4B with a full RISC-V CPU inside its weights, using the model's own Gated DeltaNet layers as RAM. It runs C and MicroPython in its forward passes, no external runtime. On DeepMind math: base 55–90%, base + Python tool 95–100%, Qwenputer 97.5–100%.

US can’t have a war and AI bubble popping at the same time First the war will end then they will backstop the big labs (OpenAI raising less than expected and Anthropic debt rating isn’t good enough)

Solana has no single document that sets out its protocol rules. They are spread across years of improvement proposals (SIMDs) and the code of its validator clients. @anza_xyz's Brennan Watt (@bw_solana) has proposed one living Solana protocol spec instead: each change becomes a pull request against it, merged only when the change goes live on Mainnet. He has already posted an 18-chapter draft. It is a proposal for feedback; nothing has been adopted.