3 Comments
User's avatar
Seixas's avatar

$698B of global VC through Q3 is a big number. Do you know how much of it is concentrated in a handful of AI mega-rounds versus spread across the broader market?

Selim Yoruk's avatar

Six agent hours per human hour makes me wonder how much of the human hour is spent reviewing. If the output grows faster than our ability to check it, the bottleneck just moves to whoever has to sign off.

Crosscurrents's avatar

The Terminal-Bench review stood out. Scoring defects in 45.5% of tasks is a big number when labs and investors lean on those leaderboards to justify valuations and funding rounds. Alongside OpenAI calling SWE-bench Verified unreliable back in February, it suggests some of the progress being marketed may partly reflect flaws in the tests themselves. That reaches beyond startups, since policymakers often reach for the same benchmarks when comparing American and Chinese models. The financing item adds another layer, with CoreWeave bonds still yielding 15% and Firmus pulling its IPO. If the scorecards are shaky and the debt keeps getting more expensive, the next few quarters could sort durable AI companies from the ones riding headline numbers.