Tom Tunguz

A laptop now handles 78% of one user's AI tasks locally, with only complex work routed to the cloud. The system, which classifies tasks as easy or hard and processes them accordingly, achieved 25% mor…

Microsoft added average token usage to a model release card yesterday, introducing a new dual benchmark that measures both performance and cost. The metric shows Microsoft's model achieving a 71.6 SWE…

Tom Tunguz, a writer of 16 years, tested an "AI editorial council" of three different models—Gemini, Claude, and OpenAI Codex—to edit his fourth draft, but the result was a "fingerpaint disaster" as e…