Terence Tao warns AI is killing math's open tradition
The math-AI fight gains its most credible voice yet, Salesforce turns its stack into an agent government layer, and a four-month-old assistant shops a $1 billion round because it ran out of compute.
Terence Tao, long the optimist-in-chief for AI-assisted mathematics, now argues the tools are dismantling the field's open culture. Writing after OpenAI's disputed Navier-Stokes announcement — we traced the credit fight in a deep dive — the Fields Medalist's point is that mathematics never lacked problems; it lacked promising directions, and sharing them openly is what made the field compound. That norm is breaking: a rumor that someone is exploring a problem now triggers a compute-heavy race that can finish it before the humans who found the path, so the rational move for researchers is silence — which would unwind the openness mathematics has run on for a century. François Chollet and Gary Marcus amplified the essay; OpenAI's Sebastien Bubeck has reportedly called the underlying accusations false. Our take: this is the first serious framing of idea-leakage itself as the threat model, and as the provenance questions in the parallel dispute stay unanswered, no lab currently has a technical answer to it.
Salesforce is repackaging its entire software empire as the governance layer for other companies' AI agents. Previewed ahead of Dreamforce, the Trusted Enterprise AI Harness bundles six capability layers — context, agency, actions, governance, security and model routing — with an AI Control Plane to register, observe and meter agents, including third-party ones. Platform chief Rohan Kumar's pitch: model intelligence is now ubiquitous, so the durable advantage is an enterprise's own context and controls. The diagnosis matches the market — VentureBeat's survey found enterprises run an average of 3.1 agent orchestration platforms at once — but the full Harness only starts rolling out in February 2027, pricing is undisclosed, and Microsoft and AWS are already selling their own agent control planes.
Instinct, the four-month-old AI assistant, is reportedly raising $1 billion more — because demand outran its compute, not its users. According to The Information, the startup is seeking the round shortly after a $250 million Series B that valued it at $2.5 billion, with capacity constraints capping how far its rollout can stretch. Consumer AI keeps rediscovering that inference, not code, is the binding constraint; if this round prices well above $2.5 billion, it becomes another data point that agentic product demand is outrunning the infrastructure financing it.
What to watch: whether Tao's essay changes lab behavior — released agent transcripts, proposed norms — or just draws another round of statements.
Would you still discuss your best unproven idea in public if a lab's cluster could act on it within days? Tell us in the comments.
