New releases

Most AI safety conversations centre on alignment: ensuring AI systems share our values and goals. But despite progress, we’re unlikely to know we’ve solved the problem before the arrival of human-level and superhuman systems in as little as three years.

So some are developing a backup plan to safely deploy models we fear are actively scheming to harm us — so-called “AI control.” While this may sound mad, given the reluctance of AI companies to delay deploying anything they train, not developing such techniques is probably even crazier.

Today’s guest — Buck Shlegeris, CEO of Redwood Research — has spent the last few years developing control mechanisms, and for human-level systems they’re more plausible than you might think. He argues that given companies’ unwillingness to incur large costs for security, accepting the possibility of misalignment and designing robust safeguards might be one of our best remaining options.

Buck asks us to picture a scenario where, in the relatively near future, AI companies are employing 100,000 AI systems running 16 times faster than humans to automate AI research itself. These systems would need dangerous permissions: the ability to run experiments, access model weights, and push code changes. As a result, a misaligned system could attempt to hack the data centre, exfiltrate weights, or sabotage research. In such a world, misalignment among these AIs could be very dangerous.

But in the absence of a method for reliably aligning frontier AIs, Buck argues for implementing practical safeguards to prevent catastrophic outcomes. His team has been developing and testing a range of straightforward, cheap techniques to detect and prevent risky behaviour by AIs — such as auditing AI actions with dumber but trusted models, replacing suspicious actions, and asking questions repeatedly to catch randomised attempts at deception.

Most importantly, these methods are designed to be cheap and shovel-ready. AI control focuses on harm reduction using practical techniques — techniques that don’t require new, fundamental breakthroughs before companies could reasonably implement them, and that don’t ask us to forgo the benefits of deploying AI.

As Buck puts it:

Five years ago I thought of misalignment risk from AIs as a really hard problem that you’d need some really galaxy-brained fundamental insights to resolve. Whereas now, to me the situation feels a lot more like we just really know a list of 40 things where, if you did them — none of which seem that hard — you’d probably be able to not have very much of your problem.

Of course, even if Buck is right, we still need to do those 40 things — which he points out we’re not on track for. And AI control agendas have their limitations: they aren’t likely to work once AI systems are much more capable than humans, since greatly superhuman AIs can probably work around whatever limitations we impose.

Still, AI control agendas seem to be gaining traction within AI safety. Buck and host Rob Wiblin discuss all of the above, plus:

• Why he’s more worried about AI hacking its own data centre than escaping
• What to do about “chronic harm,” where AI systems subtly underperform or sabotage important work like alignment research
• Why he might want to use a model he thought could be conspiring against him
• Why he would feel safer if he caught an AI attempting to escape
• Why many control techniques would be relatively inexpensive
• How to use an untrusted model to monitor another untrusted model
• What the minimum viable intervention in a “lazy” AI company might look like
• How even small teams of safety-focused staff within AI labs could matter
• The moral considerations around controlling potentially conscious AI systems, and whether it’s justified

This episode was originally recorded on February 21, 2025.

Video: Simon Monsour and Luke Monsour
Audio engineering: Ben Cordell, Milo McGuire, and Dominic Armstrong
Transcriptions and web: Katy Moore