murat 🍥

[murat 🍥 @mayfer]

now that we're interpreting monosemantic neurons it's time to bring this up again. fast memory recall & mechinterp are the same problem. let's figure out activation-triggered indexing. i think it's doable

[murat 🍥 @mayfer]

imo e2e mechanistic interpretability has the highest chance of leading to e2e memory recall by providing activation-triggered indexing and search

[murat 🍥 @mayfer]

especially embedding math works SO MUCH BETTER than any embedding models' vectors. these kinds of operations simply devolve into junk normally.

the vector space seems much more navigateable with meaning

https://t.co/gH5qGOMEZ3

[murat 🍥 @mayfer]

@nickcammarata cartoon - animation - comic = diagram

i likeeee iiiit

[murat 🍥 @mayfer]

my "abstraction neurons as embeddings" method beats all SOTA embedding search methods on "word similarity" but loses very badly on sentence similarity and retrieval

i'd say this is expected given it uses 16k neurons out of 4B but i'm really pleased how well it works for words

[murat 🍥 @mayfer]

i built an embedding engine by composing vectors from common abstraction neuron activations (h/t @nickcammarata "already clean" essay)

and it behaves more consistently especially when doing embedding math. less distracted by spelling/formatting and more focus on abstractions present

it's not always better, but generally at least on par. i have not seen it breakdown into returning junk like regular embedding models do yet. i am gonna try running some vector search evals on it to see how it performs

[murat 🍥 @mayfer]

@QuentinWach @S_Conradi @algekalipso @cube_flipper

[Quentin @QuentinWach]

I made an inverse design with BeamZ to solve MNIST.

[murat 🍥 @mayfer]

@nickcammarata cartoon - animation - comic = diagram

i likeeee iiiit

[murat 🍥 @mayfer]

i built an embedding engine by composing vectors from common abstraction neuron activations (h/t @nickcammarata "already clean" essay)

and it behaves more consistently especially when doing embedding math. less distracted by spelling/formatting and more focus on abstractions present

it's not always better, but generally at least on par. i have not seen it breakdown into returning junk like regular embedding models do yet. i am gonna try running some vector search evals on it to see how it performs

[murat 🍥 @mayfer]

@nickcammarata "bathroom - toilet" = kitchen
very pleased!!

[murat 🍥 @mayfer]

@thebasedcapital @nickcammarata i will check later

[murat 🍥 @mayfer]

@derek__olson @FoMaHun check this out

[murat 🍥 @mayfer]

@johnsonmxe @nickcammarata also i think it helps understand looped transformers better, since it would allow for wider set of generalization combinations without being constrained by a particular order

[murat 🍥 @mayfer]

@yacineMTB lmao

[murat 🍥 @mayfer]

reproduced very well. qwen 4b forgot facts but can still do reasoning (this is the 99% variant)

[Kaley Brauer @kaleybrauer]

Tested this!

I flattened every neuron's weakest activations

Overall loss barely changes at first but the models get REAL bad at memorized text. So the weak tail does, as Neel expected, contain memorized things. Weaker activations seem to correspond to less generalizable content

[murat 🍥 @mayfer]

im going, come have a pint or four

[deepfates.delve.town @deepfates]

OFFICE HOURS SAN FRANCISCO

WIZARD OF FRIENDLINESS EDITION!!!

Drink booze! Eat food! Talk about that one movie you saw!

Zeitgeist, after 6. Look for the duck!

MEET YOUR MUTUALS! (AND EVERYONE ELSE)

YOU'RE YOUR OWN WIZARD

[murat 🍥 @mayfer]

@tokifyi power over hdmi is a tricky one tho... too low

[murat 🍥 @mayfer]

@tokifyi oh hell yeah. i got goatremote working in raspberry pi, works great! just needs cloud ai and openrouter fits the bill. this will work. hdmi is all you need

[murat 🍥 @mayfer]

@Josikinz lmao