6h ago

Timothy B. Lee says observed jaggedness in AI models often traces to labs prioritizing specific skills rather than fixed limits in scaling behavior

Miles Brundage replied that the concept does not apply to static models.

——0——
Original post
OPTimothy B. Lee#1556Timothy B. Lee@BINARYBITS

@Miles_Brundage But isn't that part of jaggedness? The contrasting view is that if you scale up models they get better at ~everything without having to target specific capabilities. Whereas one possible reason for jaggedness is that models tend to mostly improve on what labs target.

11:13 AM · May 20, 2026 View on X
REPLYMiles Brundage#20Miles Brundage@MILES_BRUNDAGE

@binarybits I don’t think it is part of jaggedness insofar as the model is literally unchanged over the course of years which I think is sometimes the case for realtime as distinct from non realtime models though haven’t dug deeply

Timothy B. LeeTimothy B. Lee@binarybits

@Miles_Brundage But isn't that part of jaggedness? The contrasting view is that if you scale up models they get better at ~everything without having to target specific capabilities. Whereas one possible reason for jaggedness is that models tend to mostly improve on what labs target.

6:13 PM · May 20, 2026 · 122 Views
6:45 PM · May 20, 2026 · 89 Views