Anthropic just open-sourced FlashProve, a provable inference system using cryptographic zero-knowledge proofs to verify an AI response actually came from the claimed model. Verifiers do not need weights or GPUs, checking proofs on a standard CPU in seconds, though generating proofs still takes 72 to 94 times more GPU compute.
Anthropic just open-sourced FlashProve, a cryptographic system using zkSNARKs to prove an AI response genuinely came from a specific model without revealing its private weights. Running on a single H100 GPU, it generates proofs for a 400B-parameter Llama 4 Maverick at roughly 110 tokens per second, which can be verified on a standard CPU in seconds. While proof generation is 70–95x slower than standard inference, it creates a practical defense against rogue model substitutions and tampering.
LMSYS just dropped over 2,000 battle logs for Llama-4 on Arena and took aim at Meta for transparency issues. The evaluation platform revealed that Meta failed to disclose its experimental Maverick model was heavily tuned to game human preference prompts, prompting an immediate overhaul of leaderboard policies.
Meta’s head of GenAI just denied rumors that Llama 4 Maverick and Scout were trained on test sets to game benchmark scores. The speculation flared over the weekend after users noticed performance gaps across platforms. Al-Dahle blamed the mixed quality on early rollout issues across cloud providers, promising bug fixes over the coming days.
Meta just unveiled Llama 4 Scout and Maverick, marking its official pivot to open-weight, natively multimodal mixture-of-experts architectures. Both models operate on 17B active parameters, with Scout pushing context windows to a massive 10 million tokens, while an upcoming 2T-parameter teacher model, Behemoth, trains in the background.
