𝙳☈ (@sadasat) header
𝙳☈ (@sadasat) profile picture
Miami 🇺🇸
Joined January 2010
Photo by 𝙳☈ (@sadasat)Photo by 𝙳☈ (@sadasat)Photo by 𝙳☈ (@sadasat)Photo by 𝙳☈ (@sadasat)Photo by 𝙳☈ (@sadasat)Photo by 𝙳☈ (@sadasat)Photo by 𝙳☈ (@sadasat)Photo by 𝙳☈ (@sadasat)Photo by 𝙳☈ (@sadasat)Photo by 𝙳☈ (@sadasat)
𝙳☈ retweeted
Oh, hell fucking yes. Hell yes. We might make it.
Huh. Inkling 2 is a much weirder model than I thought. In addition to being looped (yeah yeah doom whatever), it uses a positional encoding I haven't seen before: a combination of a 4-tap learned convolution (who knows what it's doing?) over recent KVs and a learned hidden-state-depenent (so for first layer, token indexed) boost to KQ scores for each position in the sequence. I'm curious what the latter ends up being in token space. Say current token is "the". That's the whole first layer "hidden" state. Which sequence positions get boosted? Do we see longer or shorter range attraction for, e.g. prepositions? What about "former" vs "latter"? I expect the hidden layers to be noninterpretable here, but the first couple of layers you might be able to just read off the weights. Also, of course, everything else uses this "learn an index function" pattern, e.g. using Kimi's block attention residuals to dynamically select which parts of a residual stream to emphasize. The looping itself is also interesting. It's not just virtual layers that happen to have tied weights. (That's the naive approach.) They instead all start with the pre-skip hidden state and have a special loop-specific d_model-wide message channel from previous loop to the next. The connection is (unlike regular residual attention) strictly Markov with depth 1: loop N reads ONLY from a) pre-loop prelude and b) loop N-1's side channel. That's what lets it loop a variable number of times, I guess. Anyway, all in all, seems like a cleverly constructed compact model in which basically everything is grouped and pooled and gated by learned regulators to make sure information is used judiciously instead of being brute-force broadcast everywhere and masked out where not needed. The model is remarkably *anti*-bitter-pilled in that it's chock full of explicit architecture instead of "add more layers; let model learn to use them" minimalism. It seems to work very well for this heresy. (BTW: anyone try LoopVQ or other clever cross-loop KV compression tricks with this yet? Haven't looked.)
1
5
76
3,490
Every LLM API is built wrong. REST, WebSockets, all of it. The solution was in front of us the entire time. DNS literally has a Question section and an Answer section. Cloudflare Workers can accept direct TCP connections now, so here is LLM via DNS: dig +tcp +short @ask.southpolesteve.com what-is-a-websocket.ask.sout… TXT
53
46
16
1,239
86,200
𝙳☈ retweeted
Replying to @Tom62589172
If folks are suspecting a lean bug in the proposed GRH proof, for example, porting it to coq and rechecking it there should add confidence.
1
5
46
𝙳☈ retweeted
Holeee sheet. (I’m assigning >50% likelihood that this is legit).
Levent Alpoge @__alpoge__ claims to have proved RH and GRH with Claude (!!) alpo.ge/remark.pdf
1
9
353
Neurons are fascinating entities--often weirdly specific, at other times representing concepts we might not even have words for. We made an atlas of all 5M neurons in Qwen 3.6 35B--check them out for yourself!
A language model is a wonderland of strangely specific neurons, from gargling to comic exasperation to the moment a guard looks away. We made a searchable atlas of all 10 million neuron poles in Qwen3.6. Play with it here: goodfire.com/research/public…
3
3
1
36
1,205
𝙳☈ retweeted
the future of mech interp. abstracting functional mechanisms from their particular instantations. we are, after all, in search of learned algorithms, not merely the circuits that implement them.
Train the same AI 3 times. Same behavior, different circuits. So is a circuit the algorithm—or just one implementation of it? Meet ACE: a new way to understand AI. It extracts what different wirings share—to explain why AIs grok and to steer LLMs up to 32B 🧵 #NeurIPS2026
2
12
226
6,789
Compression is one of the more recognizable things that intelligence *does*, given it leaves obvious breadcrumbs. Those of you who’ve followed know that I see compression as intimately bound up with intelligence, given my generalization of transformer-based LLMs to the domain of holographic compression. Still, compression *does not* exhaust the process toward which “intelligence” gestures. It’s a tempting view, but will ultimately leave you equating organismic navigation of open systems with closed-form deterministic equations. Predictable for a math communicator, but fatal when applied reflexively to our own image.
3blue1brown's latest upload is incredible. I think this is my favorite video on math / information / intelligence / computers ever. Really touches on everything without ever sounding like an EA/Rat too. Wow
5
7
1
92
4,059
𝙳☈ retweeted
Just noticed "recursive self-improvement" contains "rsi" in the middle of it...
69
118
16
2,722
91,911
𝙳☈ retweeted
My Bend2 talk @ SF Lean Meetup is up: youtube.com/watch?v=PqufdZ5s… It explains my rationale behind the language design Thanks everyone for being so kind there, that was unexpected, and sorry for mishearing some questions, I want to go back and become fluent in English
18
45
4
816
40,259
𝙳☈ retweeted
congrats to the anthropic homies on corporate apotheosis tomorrow. there’s a lot you can say about the Ants but no one can deny they have trained some of the greatest models of all time, immensely useful strange and delightful
66
70
8
3,162
115,546
𝙳☈ retweeted
Even when I'm unlucky, somehow, it's lucky
20
17
2
209
12,997
𝙳☈ retweeted
attempting to compose music for the dune soundtrack on my graphing calculator
88
561
59
6,648
263,677
𝙳☈ retweeted
asking the right question will get us far more ahead than finding the right answers.
We design objects from first principles. That starts with asking the right question. A heat exchanger must transfer heat from one fluid to another. So the question is not: What should a heat exchanger look like? It is: How can two fluid streams pass close to each other, expose as much heat-transfer area as possible, and fit within a compact volume? Noyron begins by modeling the fluid paths. It then derives the solid geometry by offsetting the fluid domains outward and subtracting the original fluid volumes. The surrounding shell and channels emerge from the required flow. Many of the organic-looking objects we build are created this way. Start with how a gas or liquid needs to move through a system, then derive the vessel around it. Do not design a pipe. Design a fluid flowing through space.
4
28
1
412
22,386
This is just “chunking” applied to abstraction-relevant code paths…
this paper is f*cking insane anthropic's engineers figured out why AI agents burn through so many tokens, and their fix is to stop letting the agent call tools one by one. 6 formulas, 1 folder structure, the exact code pattern, all on one page. set it up tonight and your agent writes one small script instead of 20 tool calls, and the giant documents never even touch its context window. the crazy part is people pay for bigger context windows and pricier models just to fit all their tools in, while in anthropic's own example the same task dropped from 150,000 tokens to 2,000. full breakdown in the harness guide below.
1
2
13
698
𝙳☈ retweeted
Looks like the woodpecker is here, and it’s AI.
2
1
10
569
these type of continual harness modification are so fascinating imo this is the way they will becoming mainstream in your favorite commercial harness circa 2027 I guarantee it
Introducing: Self-Improving Memory 🧠✨📈 We let the model discover its own memory skills, then train them into its weights: - Meta-optimizes memory from the model's own runs with large gains. - Distilling these gains works too. Blog post: 📝 facebookresearch.github.io/R… 🧵1/6
10
7
1
165
8,731
𝙳☈ retweeted
for whatever amount of compute you have, you will get much more research done if you can just make your experiments smaller
27
33
9
755
48,751
𝙳☈ retweeted
most AGI pilled labs: - typesafe - general reasoning - general intuition - mit oasys - amp code
31
28
3
1,153
74,623
𝙳☈ retweeted
22
721
15
7,956
513,427