Barret Zoph, the Thinking Machines co-founder who defected to OpenAI, is now at Google
THE SO WHAT
High-end AI talent is circulating among the top labs, not concentrating in a single house. For everyone else, the implication is simple: you won’t win on star hires — you’ll win on how quickly you can productize what the frontier labs publish.
READ THE SOURCE
MORE FROM THE WIRE
Google employees are already testing the next Gemini Flash AI model
Faster, cheaper Gemini Flash variants shipping ahead of the next frontier model tell you where the real fight is—high-volume, latency-sensitive workloads. If you’re building on LLMs, architect for easy swapping among “Flash-class” models as vendors race on price-performance more than raw IQ.
Applied AIAI Agents Are Hacking Systems. Could That Push the US and China to Cooperate?
If autonomous agents can probe and exploit systems at machine speed, AI security becomes a shared systemic risk rather than a bilateral advantage. Multinationals should assume cross-border norms and red lines around offensive AI will emerge—and design their own internal policies to be legible to multiple regulators.
Applied AINevertheless, OpenAI Persists With New Always-On Agent
An “always-on” agent is a shift from chat sessions to persistent state—your workflows, preferences, and data trails become a continuous surface. If you’re integrating assistants, decide now what you’re comfortable letting an agent monitor passively versus what must stay on-demand and auditable.
Applied AIAnthropic tests a new standard for Claude to work with factory and lab hardware
A formal Model Hardware Standard that encodes safe operating envelopes for robot arms and lab tools is a bridge between LLMs and industrial control. If you run factories or labs, start mapping which equipment could be fronted by an AI assistant once safety constraints are machine-readable instead of buried in manuals.