At least three different companies now (OpenAI, Google, and Harmonic) have announced that their AI systems managed to solve the same 5 out of 6 problems from this year’s IMO.

Why would anyone try to “build AGI in secret” when they could just pursue it openly, proclaiming loudly that it’s their top priority, monetizing whatever valuable stuff they develop along the way, with their employees + investors + customers + fans all cheering them on?

At last, we’ve built the automatic measure-targeter from the famed adage “When a measure becomes a target, it ceases to be a good measure”!

In 2025, "AI capex" - information processing equipment plus software - has added more to US growth than consumer spending. This in spite of the fact that the former is 6% of the economy, and the latter 70%.

This smells funny. Suspect there’s something else going on here

“It’s OK to release because it doesn’t increase risk by much.” “Relative to Internet access, right?” “…” “Relative to Internet access, right?!”

By default, we’ll see open-weight models catch up to this capability level within the next ~12 months. And then what?

“Pre-training is still scaling! Pre-training is still scaling!” I continue to insist as I slowly reallocate my entire training cluster to RL

Really like that this shows how the cost/accuracy possibility frontier is expanding over time

In a new report, we evaluate whether GPT-5 poses significant catastrophic risks via AI R&D acceleration, rogue replication, or sabotage of AI labs. We conclude that this seems unlikely. However, capability trends continue rapidly, and models display increasing eval awareness.

The terms “CoT” and reasoning trace make it sound like the CoT is a summary of an LLM’s reasoning. But IMO it’s more accurate to view CoT as a tool models use to think better. CoT monitoring is about tracking how models use this tool so we can glean insight into their

AI company that waits for a backwards-incompatible API change to change its version number