Herbie Bradley Frames RSI as Economy-Wide Data Loop
Cambridge researcher argues data scarcity across expert domains will constrain AI progress.
AI researcher Herbie Bradley, a Cambridge PhD student formerly at EleutherAI/CarperAI, appeared on MTSlive and described recursive self-improvement as an economy-wide process limited by scarce high-quality training data from fields such as finance and law. Replies on the platform, including from AI safety researcher Neil Chowdhury, amplified the framing. The discussion remains confined to social media posts with no further confirmed developments or institutional responses.
Combined views
3K
2 posts, first seen 1h ago
Reactions from ranked influencers
2 posts
this was my favorite part of going on @MTSlive RSI as an ecosystem-wide process is an underrated lens
AI researcher Herbie Bradley reframes recursive self-improvement as a loop that closes across the economy, not inside one lab: "My current sense is this is actually just going to be very data bottlenecked. Suppose you get the AI to be as good as Ilya at AI research. You still need to get that AI to train itself to be as good as a senior investment banker at Goldman Sachs, or a senior lawyer, and these are just very, very data bottlenecked right now." "If you speak to any of the RL alignment companies and any of the data vendors, it is just this massive effort of collectively figuring out what is missing, iterating, trying to buy data. In a lot of cases, the enterprises didn't want to sell the data." "I have this vague concept of RSI as the economy-wide improvement loop, where RSI is made up of the data vendors, the forward deployed engineers, the AI labs, in this closed loop." "But it's only closed at an economy-wide level, rather than this sort of Yudkowskian vision of the AI lab improving within itself." @herbiebradley