Seth M

building superintelligence with thermodynamic chips.

i dream of exploring exoplanets.

[Seth M @_sethmorton]

Why what you attend to can't be static:

https://t.co/4R5A5ljyah

Transformers can't change what they attend to after training. Backprop is too global and too destructive for continual learning. The brain doesn't work this way. 1/2

[Seth M @_sethmorton]

@GaoShanghua the two fixes that i think would be great: a falsification layer before promotion (anti-collapse metric + mandatory ablation) and a quality-diversity archive instead of one global champion

[Seth M @_sethmorton]

@GaoShanghua after having used it, this still feels like hill-climbing in a multi-team costume: there's one global champion, one-comment critique gate, serial gpu etc. also - no falsification layer, so it tends to game metrics

[Seth M @_sethmorton]

@bpjzy @alexn5264 , @suganton thoughts?

[Seth M @_sethmorton]

@beffjezos @CIMCAI @stephen_wolfram this is incredibly interesting

[Seth M @_sethmorton]

@internetvin based

[Seth M @_sethmorton]

@roland_graser this is a cool benchmark

[Seth M @_sethmorton]

@iocapon true

[Seth M @_sethmorton]

@beffjezos you should check out @Primer - i really respect the mission & founders of this company. I had a fantastic convo with Robert (their head of engineering) a bit ago

[Seth M @_sethmorton]

@arthur_hyper88 is this real?

[Seth M @_sethmorton]

@poetengineer__ how did you make this visual? it's beautiful

[Seth M @_sethmorton]

@iocapon true

[Seth M @_sethmorton]

@iocapon @interfere_ so exciting!

[Seth M @_sethmorton]

@bercankilic your graphics are awesome

[Seth M @_sethmorton]

@beffjezos i've been super impressed with 5.5 - definitely feels like I can trust it more than Opus

[Seth M @_sethmorton]

@beffjezos i first heard of atomic labs and Sam all the way in Munich from @moritzthuening , the work they're doing is amazing and Sam is incredibly cracked - unfortunate but makes sense that they're moving to Texas

[Seth M @_sethmorton]

@beffjezos more investment is needed in AI for biology - we need better tooling such that bio companies can increase their internal rate of return.

[Seth M @_sethmorton]

@punit_arani i think this is really focused on directed execution toward pre-formed goals - but the real distinguishing variable is something earlier: your capacity to attend to things outside of distribution in the first place

[Seth M @_sethmorton]

@nilscmr check out this blog https://t.co/ibfvFM8SnT

[Seth M @_sethmorton]

@punit_arani true

[Seth M @_sethmorton]

@lukas_bongartz @misovalko reality -> we observe messy data π -> We try decompositions μ -> we score them by preferring the most simple explanatory decompositions and the
best scoring is our understanding of the world