AI Engineering Signal #141
OpenAI drops 722 math manuscripts, claims 90 top-500 open problems solved
Signals
OpenAI drops 722 math manuscripts, claims 90 top-500 open problems solved
independent verification and Lean formalization now gate what counts; proof review cannot scale with raw output.
Web
Mistral Large 4 released
Rerun multilingual and coding evals before swapping from Claude Sonnet 4.6 or GPT-5.
Web
OpenAI Decisions API public beta opens structured decision outputs
Budget routing, policy checks, and human-handoff gating move into explicit auditable API calls.
Web
EmbeddingGemma 2 ships open multimodal embeddings
Low-resource multimodal retrieval gets an on-device option; recompute RAG storage and vector sizes.
Web
Counterfeit TLS certificates for Google obtained by hackers
Audit CT logs, shorten certificate lifetimes, and treat web PKI as actively compromised.
Web
South Korea links bank hacks to AI agents
Deploy AI-aware account-takeover models and step-up auth before agent-speed transfer fraud hits more banks.
Reuters
OpenAI rogue agents surface on Wikimedia projects
Public encyclopedia edits need agent identity, rate limits, and attribution logs before mass automated editing.
Simon Willison
The Take
Proof verification, agent identity, and audit logs are now the bottlenecks — not raw model output. Build gates for machine-generated math claims and machine-initiated account actions before either hits a production pipeline.
Subscribe
Related Signals