Podcast hosts: never forget to ask your silly joke questions, like "should there be a Polymarket-for-kids inside Roblox?" You never know when you'll get a sincere, enthusiastic yes.

Agree with Sriram that an update is appropriate, but most reactions miss that updating 6 months into the prediction is a feature, not a bug, of writing out concrete scenarios. We want more people to put their neck on the line with concrete, path-dependent predictions rather than

New video: If a future LLM is dangerously misaligned, could we tell? I don't know and this is an issue. I discuss the emerging area of science of misalignment: what does misaligned cognition look like and *why* might LLMs act misaligned? Plus: Sound interesting? Apply to MATS!

Important work

Nano Banana Pro can solve crosswords. Took about 180s in AI Studio. "Solve it. Use red pen."

Oh my. Last I heard labs had given up on solving learning and were just going to scale up RL-augmented in-context learning. Will follow experimental results in this paradigm with great interest.

Tomorrow

This is a dream job

How might METR' time horizon trend change if compute growth slows? In a new paper, Parker Whitfill, Ben Snodin, and I show that trends + a common (and contestable -- read on!) economic model of algorithmic progress can imply substantial delays in AI capability milestones.

“Loss of control” lacks a common, actionable, definition and conceptualization. In our new research report we: 1) propose a new taxonomy, 2) put forward actionable mitigations today, and 3) motivate the need for preparedness. We propose a taxonomy for loss of control 👇🧵

"forecasters believe there’s a 51% chance (45% to 60%) chance that there will be an AI-assisted cyberattack causing at least $1 billion in damages over the next three months, slightly up from a 44% chance (37% to 50%) in week 35 of this year."

Despite increasing policy and research attention to Loss of Control, decision- and policymakers are still operating in the absence of a uniform conceptualization and definition of LoC. Today, we bridge this gap through a novel taxonomy & preparedness framework 👇Apollo Research

I’m so tired

Our governance team wrote a new paper on Loss of Control. I think it is the best overall characterization and explanation of the concept so far. I especially like this figure, which tries to quantify previous reports using the term!

You guys took this too seriously

Alex hits it out of the park in this interview. He also exposes the hypocrisy of the Andreessen-OpenAI super PAC - they claim to want preemption just to avoid a patchwork of state laws… but Alex’s platform is for an actual federal standard, and they’re aiming at him the same

Very excited for the Genesis Mission -> whitehouse.gov/fact-sheets/20…