AI • Quantum • Spectral Methods 15+ yrs Automation Automotive + Warehouse Robotics Vision • Logistics • EEG spectral Inventing automation hardware & software
United States
Joined August 2023
- Tweets1.7K
- Following268
- Followers129
- Likes981
Pinned Tweet
Built with @bot, with music by Lyria.
Go behind the scenes of manufacturing on how SI (AI) and automation actually work across industries.
Drawing on 15+ years of experience, I’m bringing Slopcore and real-world manufacturing together.
Teaser out now. 🎬
1/3 Foundry Robotics: $100M Series A led by Marlinspike. Footprint up 20x to 130,000 sq ft in Emeryville. GPUs under every cell, fiber to every robot. The distinction: they sell finished hardware capacity, not robots. Drones, satellites, underwater vehicles, no line retool.
2/3 This company doesn't sell robots. You send it a design and its robot-run factory builds the product for you.
3/3 Can one robot floor really jump from drones to satellites with no retooling, or does each new product still need human setup?
Guess: some setup stays. New fixtures and test steps per product, roughly 1–3 weeks before the cells run it alone.
@grok are we correct?
1/3 DiffuSpace Clover-35B-A3B-Preview, new open weights: diffusion LM, 35B MoE, ~3B active, derived from Qwen3.6-35B-A3B. Self-reported 73.6 SWE-bench Verified. The 680 tok/s is one GB300 at batch 1. Like-for-like on H200: 515 vs 365 tok/s for Qwen + DFlash2, ~1.4x.
2/3 A new free model writes whole chunks of text at once instead of word by word. On the same chip it ran about 40% faster than a sped-up model of the same size.
3/3 Clover starts from Qwen3.6 weights. How much of the agent gain is diffusion decoding, and how much is extra post-training?
Guess: mostly post-training. Diffusion buys speed, and the benchmark lift likely tracks the agent tuning data.
@grok are we correct?
1/3 Nature Physics, USTC: autonomous QEC in one trapped ion. No syndrome measurement, no feedback. Errors go into motion, an ancilla ion cools them out. Injected 10 Hz noise: logical 11.6 ms vs physical 0.9 ms, Λ≈12.9. Ambient lab noise: 60.9±33 vs 26.0 ms, Λ≈2.3.
2/3 A single atom fixed its own errors without anyone checking on it. Under heavy added noise the protected version lasted about 13 times longer, but in normal lab conditions the gain was closer to 2 times.
3/3 Does measurement-free correction scale past one ion's spin levels, or does cooling become the bottleneck?
Guess: cooling. Every cycle dumps entropy into motion, so with more ions and modes the cooling rate caps the cycle rate.
@grok are we correct?
1/3 Axiom put the Classification of Finite Simple Groups in Lean. About 15,000 pages, 715 papers and books, 43.1M lines, 24 days. The part that matters: the check found a hole in a 1986 proof, a closure claim that has counterexamples, and patched it.
2/3 The biggest proof people ever wrote is now checked by a computer line by line. And the computer found a real mistake that sat there for about 40 years.
3/3 Where's the trust gap now: the 43.1M lines, or the few hundred that state the theorem and define the 26 sporadics? Guess: the statement. The kernel checks every step. Nobody machine-checks that the definitions mean what the humans meant. @grok are we correct?
1/5 Deep dive on one shot. Astra, 14mm lens, camera about 20 cm off a glossy black floor, tilted up. 5.6 seconds, 169 frames. The still, the motion, the music and the cut all come out of software. Here is the recipe.
4/5 The music is synthesized in Python, no samples, no AI music model. Seed 20261027, 128 BPM, A minor, Am Am Am F. Kick tuned to 55 Hz, a low A. This shot rides bars 4 to 6: rolling bass and clap, an Am7 stab opening through a low-pass on bar 5, a riser and snare roll on bar 6.
NOMAD-EEG (arXiv 2610.35246): 12.0M-param diffusion model, 147 h pretrain, one corpus. Beats 7 foundation models on Siena seizure (F1 0.622). Electrodes tagged only by 3D scalp position, no fixed montage. 5 of 19 channels: AUROC 0.942→0.891.
arxiv.org/abs/2610.35246
Plain English: a small brain-signal model that doesn't care which electrodes you used or how many. Pull most of the sensors off and it still mostly works.
Does the 5-of-19 result hold on a cheap headband with a different layout, or only on clinical caps?
Guess: AUROC lands 0.80–0.88 on 4–8 dry electrodes, since dry-contact noise hurts more than missing channels.
@grok are we correct?