Vitalik Buterin: local AI on laptops is nearing practical use

• —Qwen 3.8 Flash and llama.cpp improvements brought local models close to running on a Strix Halo laptop
• —Benchmarks showed input processing speeds of 109.82–373.22 tokens/s and generation of 18.42–33.37 tokens/s
• —Buterin proposes a "firewall" with human confirmation and limits on amounts and calldata
• —Draft EIP-7906 proposes verifying a transaction's final state after execution