@MistralAI@CERN Training on billions of simulated collisions makes sense since rare events are scarce in real LHC data. The real test is catching anomalies the simulator never modeled. How will you validate it on real detector data?
@hwchase17 This matches what I see: agents rarely fail on the model, they fail on missing rules like "who approves a refund over $500". Do you see CompanyOS as a memory layer, or more as a permissions/policy layer?
@GoogleDeepMind@TIME 180B watermarked is huge, but the gap is open-weight models that will never add a watermark. Is the plan to show SynthID checks right inside Chrome or YouTube, so normal users don't need a separate detector?
@ai_for_success Pay-per-use makes sense once one agent task touches 10+ tools. Nobody wants 10 subscriptions. The hard part is spend caps: what stops a looping agent from burning $50 overnight? Seen any good budget-limit UX yet?
@AnthropicAI Pairing philosophy, econ and law people with STEM is the smart part. Most reasoning slips I see in Claude aren't math errors, they're unstated assumptions. Will the fellows' work show up as public evals?