Tags
GPT-6 Astra caught a lab plate labeled with the wrong antibiotic, which GPT-5.6 Sol missed. OpenAI still says its new wet-lab scores don’t yet mean reliable lab judgment. Who reads your raw experiment data first right now, a model or a person?
Anthropicanalysis
In Slack polls, Anthropic staff rated their AI uplift at 12x. METR, after eleven weeks with the lab’s internal data, puts its capability acceleration at around 2x. Would your own AI speedup estimate survive an outside audit?
A C++ team wires a clangd language server into Claude Code to cut tokens. In a ByteAsk Research preprint, the bill went up 11% while fixes held at 270 against 275. Is your coding agent finding C++ symbols with grep or a language server right now?
948 working root exploits for the Linux kernel came out of one OpenAI pipeline run in October, while GPT-5.6 Sol managed 166 with the same setup back in July. Would you rescan old code with every new model or trust the earlier audit?
