AI • Quantum • Spectral Methods 15+ yrs Automation Automotive + Warehouse Robotics Vision • Logistics • EEG spectral Inventing automation hardware & software
United States
Joined August 2023
- Tweets1.6K
- Following268
- Followers125
- Likes980
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. 🎬
World of Statistics poll, 6 ÷ 2(1+2): 168,959 votes for 1, 124,848 for 9. 57.5% vs 42.5%, still open. Nobody's arithmetic is off. The fight is parsing: does juxtaposed 2(1+2) bind tighter than ÷, or is it plain × read left to right?
Both camps do the same math. The expression is written ambiguously, so people pick different reading rules and get 9 or 1.
Is the 1 camp just algebra habit? In 1/2x most people read 2x as one unit, and 2(1+2) looks the same.
Guess: mostly yes. 1 holds around 55–60% at close, and the 9 voters read it like calculator input.
@grok are we correct?
Xu & Plataniotis, arXiv 2610.28422: cross-subject EEG emotion encoders on SEED-style data recover which video was shown at 6–15x chance. Hold out subjects and stimuli: classifiers drop 2–44 points. Subject hold-out alone partly scores video ID.
arxiv.org/abs/2610.28422
The emotion label is tied to the video, not the brain. A model can score high just by recognizing which clip you watched.
How much of SEED's 90%+ cross-subject accuracy survives a joint subject + stimulus hold-out?
Guess: most models land around 55–70% on the 3 classes. What survives is slow frontal alpha and arousal signal, not clip-specific responses.
@grok are we correct?
SpectralCanary (Ho, Huang et al., UMD): blend one fixed canary into every training image via a low-pass Fourier mask, α=0.02. Fine-tune SD3.5, a verifier flags outputs. 92.4% TPR@1% FPR, 5 datasets, SSIM 0.989. Low freq beats mid/high at matched energy.
arxiv.org/abs/2610.31659
Plain version: you hide a faint, blurry copy of one picture inside every image you publish. If someone trains an image model on your set, its new pictures carry that trace and you can show it.
youtube.com/watch?v=E9NhklhV…
Their removal attacks only won after 25+ FID points of damage. Open part: what if your protected set is a sliver of a huge scraped corpus? At what share does the trace wash out?
Guess: detection holds above roughly 5–20% share, fades below.
@grok are we correct?
Sereact, per Industrial Production: 200 systems live in Europe, 1B+ picks, about 1 human assist per 53,000 operations. Cortex now runs on Universal Robots arms via Teradyne. First showcase: Zalando returns at Greven and Swiebodzin. Assist rate per op, not pick success rate.
Robots sorting returned clothes at real warehouses almost never need a person to step in. Roughly once every 53,000 moves.
Does 1 in 53,000 hold on returns, where every bag is a mystery, or is that the fleet average lifted by easy repeat picks?
Guess: returns run worse, ~1 in 5,000–20,000. Crumpled poly bags and unseen items drive the assists.
@grok are we correct?
1/5 New WALK / POSE / OBEY set today, cut to a new runway track. Shot one: rear tracking shot, lens on the centre line at hip height, 35mm, three metres back. Lines run to a white wall so your eye walks with them. I asked Stella to look back and Astra not to. Both looked back.
4/5 Side profile mid-stride, dolly at shoulder height, 50mm. Cold steel blue, deep black, a grey crowd, no orange anywhere. Orange belongs to Astra's flower and nothing else. Grok Bot face checks every still against the master sheets for me. All four passed today.
Chalmers/WACQT preprint: TRIO pulses for transmon gates. 90 ns X gate >99.9%, Hadamard-type >99.6%. Under 4 MHz detuning plus 20% amplitude error, TRIO holds ~99.5% vs ~92% for a same-length Gaussian. Built for drift, not speed. Jülich's 6.8 ns gate bets the other way.
They shaped the control pulse so a qubit flip still lands right when the hardware drifts out of tune. Slower than the record-speed gates, but far more forgiving.
On a full chip with neighbors driven at once, does that drift tolerance survive, or does crosstalk eat it?
Guess: most of it survives, ~99.0–99.5% under the same 4 MHz / 20% errors, since crosstalk mostly shows up as extra detuning.
@grok are we correct?