Yeah I got that. It seems like we’re between a rock (needing compute that scales with growing high-entropy neural-behavioral data) and a hard place (only having one proven path to achieve scale but that path blinds you along the way, unless you only care about prediction).

I meant in context of hw lottery and funding push for large-scale neural fdn models. Are we confident frontier transformers are “good enough” to embrace complexity of brain data in the way @tyrellturing.bsky.social wants? What if SOTA models are missing something critical for understanding brains?

Could either of you explain the value of benchmark-based “SOTA” architectures if current base models get 0% on ARC-AGI-2?

Do we care at all about generalizability viz. novel problem-solving, which human brains can do but Claude 3.7 just falls down? Language is not enough.

bsky.app/profile/fcho...

I think I’ll dig in further and perhaps reach out via email. Very cool paper!

I agree the two peaks reflect distinct functions (esp. b/c no second beta dip in 3c) but concern is you are looking at theta power vs. peak accel when Kropff was theta freq vs. instant accel. Plus, I assume homing rotation was ballistic so peak accel indexes swing amplitude (and thus correction).

Beautiful support for oscillatory interference mechs of theta-phase reset. Curious though if that second peak in Fig 3b could be better explained as theta-power speed-tuning during stationary rotation, e.g., Supp. Fig 3a for rat head-scanning in my 2014 paper. dx.doi.org/10.1038/nn.3...

Congrats! I know that isn’t easy.