Examples of success for people in frontier AI:
Emergent Misalignment coauthor Daniel used to get burnout lasting weeks or months; not since our eight calls 9 months ago. He’s also gained ~5h/day after losing interest in distractions.
Founding engineer’s burnout lifted in weeks: he was planning to leave; instead he started shipping again. He also shared feedback he’d been withholding for half a year; the founders acted on it immediately.
Frontier lab researcher went from avoiding conflict with respected colleagues to challenging them. 18mo later he still does this, and even runs his own team.
These were bounties: payment forbidden until lasting results. People dm’d me to set bounties, I chose which to hunt, they chose if/when to pay. I follow up 6, 12, 18+ months later to track results.
Year 0
In the first year, my research focused on what can be done with ~1–4 afternoons. Because I was learning how to screen for who I could help, the sessions were basically the screening (though I also tried to screen for people who’d done years of therapy, coaching, and meditation already so I knew my work was creating counterfactual change). Of 8 AI people who did 1–4 sessions, 6 chose to pay bounties and 2 didn’t.
It was very experimental. All payment was forbidden until their genuine satisfaction with results.
Year 1
In my second year, I wondered what could be done with more time and commitment. For this reason, the incentive structure became half upfront, half on results, 8 weeks.
last person on the table paused due to external crisis
My super boring conclusion from my work with both AI and non-AI people is it’s significantly easier for me to help people who are more successful. People who were less successful had worse external circumstances (most were non-AI people)—eg, sleep, financial, housing issues—and had trouble showing up to meetings and completing the agreed work, so results suffered accordingly. I’m also finding I’m weirdly disincentivized from figuring out if it’s possible to work with people in less than ideal circumstances because it looks like ‘my’ failure.
Also re the success thing, there’s something where people who are successful necessarily already have learning solved, so they’re mainly bottlenecked on unlearning. Meanwhile I don’t help with ~skill-formation/discipline, which is maybe closer to what less successful people need to become successful.
In my second year, I also learned how to track results and share stories better:
Year 2
Now in my third year, I’m trying to unblock positive ASI work I like.
Because many of the people who have internal bottlenecks blocking their work don’t realize it, I’ve started offering internal bottleneck evals so they can see this for themselves. The eval also teaches prerequisite skills and serves as comprehensive screening for both parties.
The incentive alignment for unlearning has also evolved to more closely align with impact. Roughly: refundable effort bond + payment forbidden until their genuine satisfaction similar to Paul Christiano’s Certificates of Impact.
Finding & removing bottlenecks for people helping ASI go well
Is there something you’re trying to do that keeps getting stuck? e.g., write something, run something, start something, coordinate, …
Impressive, usually I’m skeptical of these sorts of interventions but kudos on tracking and publicizing results. Amy write ups on what your process involves?
Which types of people can’t seem to benefit?