Worth separating the headline from the download: release 1 is 10,124 hours live, with the 100K staged. But the direction is the story. Three separate 100K-hour egocentric corpora are now free for commercial use.
Generic first-person video just became a public utility, and the value moves to whatever cannot be crowdsourced.
Today we're open-sourcing HubEgo-100K: the most diverse egocentric dataset on 🤗 @huggingface.
100,000 hours of skilled tasks from 100,000+ different Hub contributors.
Free for research and commercial use 🧵
39
Where does physical-AI data come from?
I’m mapping collection services, capture hardware and robotics operations, with the public sources attached.
57 organizations. Locations mapped across 18 countries.
Explore the Radar: archivesignal.ca/radar/
2
3
51
This is awesome. The clearest explanation yet of where the value in company data actually sits: not in the logs, in the corrections your experts made and the graded work on top of them.
If you run a business and someone has emailed you about licensing your data, read this before you answer.
1
118
The man had every GPU cluster on Earth modelling protein folds and frontier reasoning, and the civilization-critical failure mode turned out to be the oldest one in the dataset.
Some things don't scale, and some things don't need to.
BREAKING: Nvidia, $NVDA, CEO Jensen Huang has resigned as CEO, President, and member of the Board of Directors, effective immediately.
56
Who’s collecting the data behind physical AI?
We put 47 organizations across 15 countries on one map: data collectors, capture hardware, robotics teams and public demand signals.
Every entry links to its evidence.
Explore: archivesignal.ca/radar/
1
18
Files got abundant. Proof got scarce.
So I built The Proof Room: one fictional workshop archive, opened for inspection. A fault, an intervention that passed its check, and the symptom that came back anyway. Source, failures and gaps all left in.
No email gate. Take it apart: archivesignal.ca/proof-room/
1
39
Judging by the offers coming in, a lot of bad businesses have sold their data. The good ones are profitable, busy, and not interested in a contingent payday.
Surely the labs want some success data sprinkled in, no?
1
1
102
RLWRLD launched DexBench yesterday with 15 dexterous robot tasks across actual industrial categories, including assembly, logistics and hospitality.
Connector insertion and screw fastening are much more interesting tests than another robot waving at a conference. At least we can start asking what the machine can reliably accomplish.
15
Sunday thought: don’t sell ten years of harvest for the first bid at the farm gate.
Your company’s history took a decade to grow. A buyer’s deadline doesn’t make their price right.
Store it. Secure it. Keep your rights.
You can’t wait for better crop prices after signing away the silo.
46
The internet is downloaded. Simulators can rehearse a world, but they can’t go live in one.
The next important datasets won’t be found online.
They’ll be captured from real work, in real places, with real outcomes attached.
Downloading the whole internet and running RL simulators is not enough for intelligence. Clever new real world datasets are needed. The science of AI is far from finished. IMHO.
39
Physical Intelligence is reportedly in talks at $20B, NVIDIA gives GR00T away free, and Tesla is building Optimus in Fremont.
Inside, two of the three brains are built almost the same way. Here's how all three think, and the free lectures that teach it.
The data market did not slow down. It got flooded. Six months of archives arrived in one month, the big buyers narrowed their focus, and the silence everyone is hearing is digestion, not death.
But archives are stock, not flow.
Once this is digested there is no next six months behind it, the buyer count is higher than ever, and most of the willing sellers already locked themselves into exclusives. The scarce side gets scarcer.
archivesignal.ca/insights/th…
30
Affiliate style networks are going to lose this industry.
What's being sold here isn't a widget. It's a person's life's work. Twenty years of records. The thing they built instead of weekends. It's sensitive, it's personal, and for a lot of owners it's their retirement.
And the industry's answer is a funnel. An intake form. A screening call with an AI broker. A fit quiz. Thresholds.
Nobody wants to explain their life's work to a funnel.
Nobody wants to feel like a sheep herded into a pen, tagged and numbered.
I can't even send my own referrers through these systems. These are people I know, sending me owners they know. I'm not feeding a friend's life's work into a form so it can be scored by someone who will never learn his name.
This work is white glove or it doesn't happen. One human who picks up the phone. Who learns the business. Who takes the time. Who treats the owner like the asset is theirs, because it is.
The labs think the scarce thing is data. It isn't. It's an owner who trusts someone enough to say yes.
The best archives never fill out a form.
39
The AI industry's PACs spent tens of millions on the midterms and barely said the word AI in their own ads. Too toxic. I spend the first five minutes of every business call doing what their money couldn't: getting one owner past the fear to the actual economics.
Education is the job now."
25
The AI data market isn't ending. It's graduating.
The frontier labs aren't done buying data. They're done buying more of the categories they've already ingested. What comes next is narrower, deeper, outcome-linked, and the terms will matter more than the first price.
archivesignal.ca/insights/th…
73
The frontier labs aren't done buying data. They're done buying your category.
Behind them, a second wave of labs has to acquire everything the first wave already ingested just to catch up.
The market isn't ending. It's graduating.
1
1
111
A startup called Dream raised $17m after training AI on records covering 100m+ square feet of permitted buildings.
Pretty interesting use of municipal archives. I'm even more curious about the engineering revisions, failed inspections and field changes that never made it into the final approved drawings.
35













