computer hacker, alignment & safety @flockcap, part-time independent journalist opinions my own 💍 @computerphile

Seattle / San Francisco
Joined March 2020
Pinned Tweet
we made sparse autoencoders compete with tiny softmax classifiers and the classifiers won on bits 🐐 a pure discrete code (no coefficients!) beats a top-k SAE at matched bits on SONAR, and an 800-bit hard code recovers 97.6% of gpt-2's loss paper out now! (link in replies)
2
3
3
45
6,755
fun demo! this runs on the ontologizer work we’ve been doing. sparse autoencoders can’t make this chart. they give you a bag of 16k latents per person. an ontologizer gives every attendee a discrete code word, which is why you get a typology instead of a feature dump.
At @WakaruLabs, we needed to decide who to prioritize talking to at EA Global. I used a truebase's residual stream activations and our library to make a color-coded ontologizer of conference attendees
4
2
14
1,066
if you want the boring version with 77 pages of ablations: flock.holdings/research/onto…
4
185
celeste retweeted
At @WakaruLabs, we needed to decide who to prioritize talking to at EA Global. I used a truebase's residual stream activations and our library to make a color-coded ontologizer of conference attendees
16
12
7
55
7,302
Temple She loves Factorio Crack for autistic systems thinkers and engineers hyperoptimizers Eventually she figured out in the lategame that the ultimate bottleneck is your computer’s memory and processing power to “hold state” on the entire map and gigabase, and you’d have to optimize for that Every bottleneck you solve…creates the next bottleneck somewhere else
omg people weren't lying i bought factorio 7 hours ago and it's 5 am now
1
2
14
944
lately it feels like we've been moving like "we're so back", as the kids seem to call it
2
46
2,467
happy to announce @flockcap has signed its first contract in AI hardware. can't say who yet, but we're really excited about the product. more soon!
4
1
46
2,403
went side questin' @NotionHQ yesterday :)
6
99
3,806
another day at the "building the optimizer that tiles the lightcone" factory
1
1
20
1,584
2024 vs 2026 (before vs after rationalism)
9
199
4,872
I never ~announced~ this, but for the past few months I've been working at @midjourney, learning a ton about operations from @aalleexx Alex organized the Ghost in the Shell screening at Alamo Drafthouse on Monday evening, so I got to pitch in with planning the event it was gratifying how much people appreciated seeing this beautiful philosophical movie on the big screen! I took the liberty of turning your reactions into a video. I hope you like it quotes from @majamediaco (especially!) @PrinceVogel @alder_riley @DavidSHolz @potterlab photos from @snnneee @lishali88 @maxwong @natalieyeo @vmfunc @thedataroom P.S. to be clear this is not Official Midjourney Content, just something I had fun putting together!
44
11
4
304
7,739
all those inaccurate portrayals of computers from hacker movies were just early
Were you thinking “I wish I could explore the OpenAI math repo like playing Breath of the Wild?” mts.now/mathofthewild
7
68
3
1,850
56,874
I was hesitant to open source this out of concern that it could make it easier to build a neuralese CoT. Since OpenAI is set on building neuralese CoT models anyway, maybe I can at least encourage them to use something more interpretable than looped transformers. (repo link)
we made sparse autoencoders compete with tiny softmax classifiers and the classifiers won on bits 🐐 a pure discrete code (no coefficients!) beats a top-k SAE at matched bits on SONAR, and an 800-bit hard code recovers 97.6% of gpt-2's loss paper out now! (link in replies)
1
1
2
400
If you are the only person with nukes, you take over the world. If your opponent is the only person with nukes, you die. If you are the only person with artifical superintelligence, you die. If your opponent is the only person with artifical superintelligence, you die.
23
14
365
10,689
AI might make intelligence abundant while simultaneously making reliable evidence scarce and i think we're undervaluing that
4
2
52
2,449
I support President Trump in proposing nuclear attacks on Los Angeles and San Diego and i urge him to include San Francisco as well . perhaps a chemical attack for Berkeley — something slow and painful
25
50
5
1,194
36,275
went to twitch HQ today
62
1,813
had wind of this for a very long while for reasons, so excited to see mistral come back like this. le chaton fat is real! :D
Meet Mistral Large 4, aka Le Chonk. • 1T parameters, natively multimodal. 49B active. It is the best open weights model from US or Europe on aggregated benchmarks. • State-of-the-art on critical workloads, including cyber defense, manufacturing and finance and it surpasses closed frontier models on visual grounding. • Forged in Europe end-to-end and is deployable from Europe via our own Mistral Cloud infrastructure. • Available to all via API today. Working with cybersecurity partners privately. Open weights release end of October.
1
1
67
2,877