Lucas

Robot action models shouldn't need 256 vision tokens per frame. Pi0.5 spends 400M parameters on SigLIP just to see. We replaced it with a 4.4M encoder that outputs 5 tokens — and action quality barely changes. 91x smaller. 51x fewer tokens. 7.3x faster inference.

Digit’s been clocked in since 2023, streaming continuous operations.

@yacineMTB Sound tied to mosquito flap frequency

What's the best way to detect a mosquito in 3d space?

i can't believe i'm pushing latents over uart lmfao

the only thing on my mind more than my startup is this 11/10 girl that wandered into my life

building a startup is alot more like cookie clicker than I thought

@chesterzelaya theres no option to buy multiple drones on the site

shipping today 🫡 modded drone + nimbus bundles

@janleike putting claude on robots?

Some personal news: I am starting a new research project at Anthropic. Very excited about this! Many things are needed to make AGI go well, and alignment is only one of them. More on this soon…

can someone sell me 50$ worth of anthropic secondaries

we might actually be in the timeline where we crack interpretability before AGI

@PropheticAI @JackSimison sweet

Prophetic is on a mission to transform how humans dream. Meet Jack. One of the first to experience lucid control and awareness in his dreams from the Prophetic Dual. We are excited to bring this capability and more to billions of humans. User interview #2 @JackSimison

My team at @GoodfireAI has been cooking up a new way to do interpretability: decompose a language model’s weights, not its activations. Our decomposition natively handles attention (!) and behaves less like a lookup table and more like a generalizing algorithm. (1/6)

one of the best use cases for cloud robot inference while they aren't very deployed, is for swarm drones as they need a centralized controller anyway

Everything is a skill issue

I can’t believe she would put a 3k gpu on a clothe folding robot

You probably think this is the most boring AI (not generated) video on the internet right now. (And that’s exactly why you need to watch it.) 👇 It’s literally just a robot bending a wire. Over. And over. And over again. If a human had to do this for 8 hours a day, they’d absolutely lose their mind. But this robot? It’s going to do it all day, all night, without a single coffee break, distraction, or complaint. We’re obsessed with robots that can do backflips, parkour, or dance routines. But the real robotics revolution isn't flashy. It looks exactly like this: automating the deeply tedious, mind-numbing, repetitive tasks that quietly hold our supply chains and factories back. Why toil when a machine can do it for you? Humans have better things to do. We have better things to build. Better things to create. Human capital is simply too valuable to be wasted on the mundane. Let the machines do the work. We build autonomous robots for real-world industrial work. Welcome to Factory Intelligence lnkd.in/gX73YBeZ

This is what I woke up to

@bubbleboi great shitpost

You think the AI bottleneck is HBM. You think it’s CoWoS. You think it’s GB200 cables or 800G optics or the Arizona power grid. You are looking at the wrong layer of the stack. The bottleneck is … nicotine. Specifically, the ESSE Change 1mg menthol cigarette made by KT&G (033780.KS) and smoked by Samsung’s HBM packaging engineers in the courtyard of Pyeongtaek Campus P3 between 2:47 AM and 3:04 AM during the third reflow shift. Without this cigarette, Samsung HBM4 yields collapse from 64% to 11%. Nobody knows why. I think I cracked it. I ran a backtest on Samsung’s HBM3 qualification ramp pulling weekly yield disclosures from FSS DART filings, normalizing to a 7-day baseline, and cross-referencing against KT&G’s quarterly shipment manifests at the SKU × distribution-center level. R² of 0.87 on HBM3, with ESSE Change 1mg shipments to the Pyeongtaek-Anseong corridor leading yield prints by 36–40 hours. I assumed I’d messed up the merge. Ran it three more ways. The signal held. So I pulled HBM4E, which Samsung started disclosing weekly in Q1 2026. Same overlay, but this time narrowed to CU branch #4471 which is 340 meters from the P3 fab gate, the only 24-hour store within walking distance of the third-shift smoking area. R² = 0.91, lag of 38 hours, p < 0.001. When I ran SK hynix yields against the same KT&G series, no signal. Samsung yields against Marlboro shipments via Philip Morris Korea, no signal. Every other CU branch within 2km of P3 nothing. Only #4471 prints. This upcoming June, Samsung is qualifying HBM4E for Nvidia Feynman. Samsung supplies ~30% of HBM4E into the ramp or 1.2M Feynman packages, every single one of those HBM stacks is gated by an engineer puffing on an ESSE Change. Samsung HBM4E yield is now the load-bearing dependency for global AI capex, but KT&G is the load-bearing dependency for Samsung. And sadly there is no alternative. Japan Tobacco’s Mevius is described by Samsung engineers as “tasting like a hospital.” Marlboro is what SK hynix smokes and for engineering reasons is structurally incompatible with Samsung’s MR-MUF process. BAT and Imperial have under 4% combined Korean share, and Chinese domestics can’t be imported in volume. Even KT&G’s own ESSE Special Gold was trialed at P3 in March 2025 and caused a 7% yield drop in one week because the engineers said they could taste the paper supplier change and were feeling nauseous. With all that said, look at the financials. ESSE Change runs ~38% gross margin against a 52% group average. KT&G has been underpricing their flagship brand for two decades because nobody realized it isn’t a cigarette, it’s the bottleneck of global AI capex. A 15% price hike is ~80% lift to operating income at this volume, lifting group EBITDA margin from 28% to 34%. On ₩6.1T of revenue, that’s ₩370B of incremental operating profit against a ₩9.8T market cap resulting in a 12% earnings beat from one SKU repricing. And the best part? The engineers literally cannot decline, because there is no substitute. I bet on two hikes in eighteen months which gets you 25% EPS growth without selling a single additional pack. This is what the sell-side is going to figure out in Q3, and it’s why the multiple goes from 9× to 22× before consensus catches up to the revisions. No amount of Arizona engineers chewing 30mg mint Zyns replaces this. KT&G has a lockdown on the engineers who make the world’s HBM. They can name their price.

Using Partiful as a crm blasting tool, wayyyyy higher open rates

Dream mogging with @PropheticAI