If you couldnt provide over 10k hours egocentric data monthly.
Stop jumping into this biz.
73
Armitage retweeted
How much runway does Physical AI have?
If LLM scaling laws are any indication, a lot.
The largest publicly announced physical AI dataset today is ~1M hours (Dyna-2). Map that onto LLM training tokens and it lands right around GPT-2.5. Past the "cool demo" stage, nowhere near "this changes everything."
Getting to GPT-5-equivalent takes something like 2B hours. That's ~2,000× from where the field is today.
The pace is real though. Publicly announced datasets went from ~10K hours to 1M in under two years. Even if you assume a more conservative 10× a year from here, GPT-5-equivalent shows up around 2030.
Internet = zero-trust environment from now on.
Introducing Griffin, the first model to pass the video Turing test.
48% of people who talked to it live thought it was a real human. Previous systems have had a pass rate <3%. It is #1 on NVIDIA's benchmark for full-duplex AI video.
It’s the first Human Interaction Model (HIM).
Readers added context they thought people might want to know
The 48% figure and "video Turing test" claim are from Tavus's own study of 54 one-minute calls, not independently verified or using a standard protocol. Griffin-Lite leads NVIDIA's VideoFDB benchmark on their public leaderboard.
cellcog.ai/blog/tavus-gri…
research.nvidia.com/labs/amri/proj…
tech-ish.com/2026/10/02/tav…
73
Give me your crack dealer bro
No wonder OpenAI delayed their IPO plans. Secondary shares are being offered at an implied $475 bn valuation, down roughly 44% from the $852 bn primary round earlier this year, that makes that $1 tn IPO look near impossible right now.
Hearing Anthropic is being offered around a $700 bn valuation on the secondary market, yet they’re still trying to pitch their IPO between $1 tn and $2 tn. Good luck with that.
We all know how flaky secondary markets can be, so make of it what you will.
256
Armitage retweeted
7:
The community started pushing the idea that this was NVIDIA's version of the old Fartcoin story
You know the joke;
Humans fart
Machines…
BURST 😂
It's basically trying to create an AI-era cousin to the Fartcoin narrative
If this kind of app could be an effective way for collecting data.
Mecka guys could be billionaires.
We're now over 43,000 weekly active users collecting data to train Helix, our AI model for F.03 robots
This data collection project, Index, is our answer to the data problem: the largest useful robot training dataset in the world
Extrapolate out, and this is the pretraining needed for large scale robot generalization
The effort started by purchasing data from vendors but this data was scarce and really poor quality - there was simply no way to make this work correctly at scale. So we did it ourselves. It was quite a massive effort that has now resulted in a Figure-owned data pipeline that is unmatched
1
259
Egocentric data market is really weird.
Many companies are building their own hws and trying to dump on data vendors.
1
95
Seeing a hype wave around GEN-1.5, and rightfully so. Lots of respect to Pete & Andy for executing so well. The secret is in the naturally repetitive motions in human-collected data. There're 2 main sources for such repetitions:
(1) Symmetric patterns. Sorting, tidying, and assembling almost never finish in one motion. Open any assembly manual from IKEA, and you find most objects symmetrical. You drive one bolt, then its twin, then the next pair. Every {bolt A, bolt B} pair is a natural continuation in context, and the second instance is a free training signal that imitates the first ("prompt").
(2) Recovery. Humans drop things all the time, but we pick them up so fast, we don’t even notice. That reflex to fix is half of our physical competence. The key insight is to keep the failed first half instead of trimming it away. If the model consumes the full arc, fumble, catch, continue, then recovery shows up organically at test time. It's funny that in-context improvement results from *NOT* over-sanitizing your data.
The other critical ingredient is UMI. I've been saying for a while that teleop will not last, and GEN-1.5 is driving the final nail in the coffin. UMI is essentially a human wearing the robot gripper to collect data directly (human → data). Teleop inserts a layer of separation: human → VR/skeletal device → robot → data, which bleeds out all the human "physical intuition". The subtle sleight of hand we perform constantly with objects, the micro-adjustments, the feel of a part snapping into place, is nearly impossible to capture when you can't feel the environment directly.
Once you have enough data, many behaviors can actually be zero-shot. For example, you don't even need finetuning to pick up a novel object. The model "just knows" what to do given a similar scene in the training distribution. Whether in-context learning truly works or not also depends on how far away the test is from training. Currently, the demos are still a bit too simple to conclude.
I'm cautiously optimistic. Still, it's a great day in robotics.
Looks really interesting but i don know why 2 fingers will work.
We build humanoid with hands for complicated tasks.
Introducing GEN-1.5, a one-shot learner.
It can learn new tasks in a few seconds. Show it what to do, and it generalizes.
This capability emerged from pretraining on physical data at scale, as a step towards our mission of building general intelligence for the physical world.
74
Armitage retweeted
35% of OTS deals in egocentric data might involve fraud.
Vendors send a perfect sample. The final dataset is either full of recordings you already own or recordings already open-sourced on Hugging Face.
Sellers can Rename, Trim, Crop, Re-encode. Checksums change. The content doesn’t.
We just open-sourced Overlap. It indexes any OTS data and compares it against your inventory & public HuggingFace datasets.
Ask sellers to share indexes first.
Detect duplicates without ever seeing the real content.
No GPUs needed. A normal MacBook can index 5k hrs data in an hour.
Let’s change the industry: share index files instead of just 5mins samples. Stop paying for free or repeat data.
Github - github.com/World-Archive/ove…
GUI - overlap.worldarchive.co
Their hw is really good, recommend if you need MC egocentric data.
SN/06 is live.
Meet RoboCap: our egocentric platform trusted by top labs — now available to you.
Preorder $499 ($649 w/ RoboWrist) to access the highest Bolts multiplier + cash rewards.
Limited supply (encrypted for use on BitRobot Network only) → shop.bitrobot.ai
1
132
Armitage retweeted
The race for the AI application layer is not only about UI, workflows, or GTM... it is a fight for the intelligence layer itself.
Larping at the finest. In community call they said their goal is 1m-2m$ arr in next 6 months.
Now they said they have 150k hours egocentric data 😂😂
Over the past month, we have continued pushing forward across growth, data production, commercialization, and ecosystem expansion.
On the growth side, Axis Hub has now surpassed 133K total registered users, up from 80K+ about a month ago. Within this user base, more than 40,000 contributors have passed Axis’s quality-contributor filters. Among this qualified-contributor cohort, we currently have 13,480 MAU and 9,409 WAU, representing a 69.8% WAU/MAU ratio, which demonstrates strong recurring participation.
Our weekly active users are contributing at a very high frequency, averaging around 9–10 valid data trajectories per user per day. We are also seeing a group of emerging “human experts” within the contributor network who are able to use our post-training features to correct model failure states and help models continuously improve and iterate.
On the data side, Axis has now generated more than 3.4Mn valid simulation data trajectories, up from around 2Mn one month ago, representing approximately 9,500 hours of precise trajectory data. In addition, we have captured more than 150,000 cumulative hours of ego-centric real-world data, and 100 hours of loco-manipulation data.
84
If sam is still there, OA is still OA.
It brings me great warmth to announce that OpenAI is collapsing.
This is 100% happening bc Sam is such a terrible person.
OpenAI’s 2026 departures so far:
• Scott Gray - longtime GPU systems engineer
• Denise Dresser - Chief Revenue Officer
• Brad Lightcap - former COO
• Fidji Simo - CEO of Applications
• Kevin Weil - Head of OpenAI for Science
• Bill Peebles - Head of Sora
• Srinivas Narayanan - CTO of B2B Applications
• Kate Rouch - Chief Marketing Officer
• Barret Zoph - Enterprise Sales lead
• Johannes Heidecke - Head of Safety Systems
• Joshua Achiam - Chief Futurist
• Chloé Bakalar - Head of Ethics
• Caitlin Kalinowski - Head of Hardware and Robotics
108

















