Two movements meet: AI is getting production responsibilities, and production data is flowing through AI. Anthropic’s September report, malicious routers, training data provenance and model drift show what to verify before an AI reaches production.
DHH told Lex Fridman that for some problems, programmers do worse with AI coding tools than people who cannot code. What the claim actually covers, and what it means for teams adopting AI agents.
Public frontier models lean progressive on open benchmarks. Local inference via llama.cpp is still the practical way to keep an uncensored second opinion. Memory is the bottleneck; TurboQuant-style KV compression is what will make local writing workers routine in multi-model agents.
OpenRouter put numbers on a pattern agent builders already use: a panel of models with a judge synthesizer beats the best single model, and a budget panel matched frontier quality at half the cost. What that means for your AI platform.