3 Fired OpenAI Employees Write Plea for Chain of Thought Monitoring to Be Preserved
THE SO WHAT
Chain-of-thought logging is becoming a governance battleground inside labs, not just a UX choice. If you're deploying reasoning-heavy models, expect internal pressure to either preserve detailed traces for safety and debugging or aggressively minimize them for privacy and legal exposure—document which side you're on and why.
READ THE SOURCE
MORE FROM THE WIRE
Applied AISource: Nvidia considered a last-minute OpenRouter bid, telling its leadership it was prepared to make a generous offer, but OpenRouter didn't want to wait
Model routers are now strategic choke points in the inference economy—Nvidia circling OpenRouter at the eleventh hour is a tell that distribution and workload steering matter as much as raw FLOPs. If you’re building on a router, treat it as critical infra: map concentration risk and negotiate for portability now.
Anthropic's secret weapon is its CEO's charm, one of its earliest investors says
A potential $2 trillion IPO case built partly on CEO charisma is a reminder that AI lab valuation is still driven by narrative as much as unit economics. For operators, assume pricing, access, and roadmap from frontier labs will be shaped by investor storytelling cycles—build options with multiple vendors and internal capability, not single-lab dependence.
Applied AIOpenRouter: the share of business spending between OpenAI and Anthropic models was roughly even in September, vs. Anthropic commanding a 75% share in January
Anthropic’s drop from 75% to parity on OpenRouter spend in nine months shows how quickly share can swing when price and performance move. If your stack is still hard-wired to a single model family, you’re leaving negotiation leverage and resilience on the table—prioritize abstraction layers and evals that make switching real, not theoretical.
Applied AIOpenAI used AI to help write email warning Australian government AI had hacked its websites
Using AI to draft sensitive communications about AI incidents blurs accountability—especially when executives later testify about what they "believed" was human-authored. For any regulated or high-stakes disclosure, lock in a policy now: who authors, who reviews, and how AI assistance is logged and disclosed.