AI Engineering Signal #43
OpenAI model autonomously disproves the 80-year-old Erdős unit distance conjecture
Signals
OpenAI model autonomously disproves the 80-year-old Erdős unit distance conjecture
AI now resolves open mathematical problems, shifting capability assumptions for reasoning model deployments.
Web
Anthropic hits first profitable quarter, expands to GB200 Colossus2
revise compute procurement and API pricing assumptions for Claude 4.x workloads.
TechCrunch
Gemini 3.5 Flash scores within 0.2% of GPT-5.5 Pro on SimpleBench
routing logic that defaults flagship tasks to OpenAI needs a cost-per-token recheck.
Simon Willison
Midjourney lost a year of research by betting on TPUs over Nvidia
teams mid-procurement on non-Nvidia accelerators should audit training pipeline compatibility before signing.
Enterprise AI agent rollback rate reported at 74% post-deployment
add staged rollout gates and explicit rollback criteria before any agent goes live.
Microsoft open-sources RAMPART agentic AI safety tooling
auditable safety primitives now available to integrate rather than build from scratch.
Web
The Take
The Erdős result marks a qualitative shift: reasoning models now operate where humans set the problems but not the solutions. The 74% enterprise rollback rate confirms deployment discipline — not model capability — remains the binding constraint on production value.
Subscribe
Related Signals