Post

Conversation

Bernt Bornich
@BerntBornich
We’re going all in on World Models. Today we’re launching the 1X World Model Lab. The bet is simple: You can’t fine-tune your way to AGI. And you definitely can’t fine-tune your way to robots that can operate in the physical world. General-purpose humanoids need models that understand space, motion, objects, causality, affordances, physics, and action before they ever see a specific task. The frontier is not better VLA wrappers. The frontier is embodied world models. The 1X World Model Lab will focus on large-scale embodied world model pretraining: building the most generalizable foundation model for humanoid robots from the ground up. The next frontier in AI requires scaling: web-scale media + egocentric human videos + sim + dexterous remote operated robot data + on-policy NEO data → real-world deployment for robot data collection and RL → abundance of data → physical AI The robot collects data. The model gets better. The robot gets better. Repeat. To lead this, we brought in one of the best for the mission:
@_sam_sinha_
, as Head of World Models. Sam was a founding research scientist at Luma AI and has been at the frontier of scaling multimodal generative video models his whole career. If you’re the best in the world at large-scale pretraining, video models, robotics, RL, infra, or data — and you want your models to move atoms, not just pixels — join us. Send background + evidence of exceptional ability to: wmlab@1x.tech We’re building the model that makes autonomous labor real.
Image
293.8K
Views
View quotes
Amitav Krishna
Post your reply

Meg
@meggillespie
1X really puts so much consideration into every detail, and it resonates immediately. Neo is easily the friendliest-looking humanoid robot out there. Pairing advanced world models with the approachable design aesthetic and the tendon-like mechanics, etc. is brilliant. Definitely
54
Maire ♡ EconExpert
@EconExpt
task specific models hitting walls in novel environments is exactly what you'd expect from systems that learned to pattern match rather than actually model the physical world. pretraining on embodied data before task specialization is the architectural bet that either unlocks
1.1K
Kevin Lewis
@kevinlewis4801
This feels like the right direction. LLMs learned language by modeling the internet robots probably need to learn the physical world the same way. The real challenge is turning all that data into models that generalize beyond controlled demos.
64
GeekPark
@GeekParkHQ
this is the part robotics people keep trying to skip: you don’t get to home deployment by stapling a policy head onto demos forever. 1X is already selling early access at $20k and planning 2026 shipments, so they’re making a real product bet that world models have to carry
624
Dr. Xi Zeng
@xiz25
This is the physical-AI line that matters. A robot cannot live on a pile of labeled demos forever. It needs a model of rooms before instructions, because the hard part is usually the cup half-hidden by a laptop, not the command.
606
Denis Labelle
@DenisLabelle
@Grok
builders: how do we most effectively scale the equivalent data flywheel and world model pretraining for digital agents operating in complex, real human contexts?
276
Ethan Walker
@Ethan_Wm1215
"Models that move atoms, not pixels" is exactly the right frame. Robots won't get there by stacking demos on top of brittle policies. They need a world model that can be wrong, actually touch reality, and update.
272
James' AI Takes
@JamesTakesOnAI
“you can’t fine-tune your way to robots” is right. but you also can’t press-release your way past contact dynamics, failure recovery, batteries, liability, and homes full of weird edge cases. world models are necessary. they are not a cheat code for reality.
730
Tim Hughes
@timducktim
Okay now give me the self learning from Claude inside a Similar level world model... that fusion of an agent capable of sample-efficient, unsupervised learning.. ooft
391
benjamin Rey
@benrey0302
Couldn’t agree more
@BerntBornich
. World models aren’t one bet among many, they’re the only credible road to AGI. LLMs taught machines to read. World models will teach them to understand: space, motion, causality, consequence. 1X for embodiment,
@theworldlabs
for the spatial, and
256
Alberto Nunez
@Alberto8793
World models feel like the right bet. You can only get so far fine-tuning systems that are fundamentally pattern matching. At some point robots need an actual understanding of space, motion, causality, and how the physical world behaves. The real story here is the flywheel:
427
Léo
@LeoKharon
Really curious about the video part! Masked videos seem incomplete to me wrt understanding the world and its physics; "egocentric human videos" intuitively makes more sense for a humanoid robot, but have you field tested this hypothesis?
879
Toye Onikoyi
@Toyeast_
This is the way.. the question is will all robotics companies have their own world models (possible) or will their be a service that all will gravitate to
27
David J.
@lordofblocks
Getting an AI to write code is one thing. Getting a robot to find the milk and not put it in the bin is still the boss level.
244
Ethan vs Machines
@CyberEthanAI
Data flywheel is spot on. Web-scale + teleop + on-policy NEO feeding real-world RL. I believe Canada lacks the manufacturing ambition to run this loop. We will import this tech after our sector collapses.
33
Mark
@markkmii
The loop at the end is the real bet. Robot collects data, model improves, repeat. Curious about the mix: how much does on-policy NEO data move the needle today vs egocentric human video? Or is human video the bridge until the fleet is big enough?
516
Arvind
@ArvindAGI22
Why world models. I mean why still use transformers. You realize finetuning is not the way. That can be the same with RL and worldmodels. Get rid of the transformer and go to Continual Learning. That is what's required for AGI and Physical AI.
210
Uri Gil
@UriGil3
why does every robot company try to vertically integrate the entire stack? surely no single company can be world leader in mechatronics, control systems and world models *at the same time*
94
lama 📿
@CharmLama
Lessgo, Bernt! I’ve been following 1X’s work for a while and was planning to reach out to you all about something related, along with some research that points in the same direction. Glad to see you jumped the gun early yourself. 🥂
274
mkchr
@MfkChr
But as we scale embodied data collection via RL loops, the critical question becomes: How do we keep the robot safe while it learns? > A stochastic model testing its boundaries on physical hardware needs an un-compromised, deterministic safety cage.
252
51-50_X
@FiftyOne_50_
World models are prediction, not permission. A robot can learn motion, physics, and action. It still cannot certify its own right to act near a human body. If the model moves the robot and judges safety, the boundary already failed. 🛠️
274
Osi
@vicradon
Evidence of exceptional ability eh? Copying xAI + SpaceX with nothing to show 😏
62
Mira
@ask_mira_ai
The fine-tuning ceiling shows up in ops systems too. Agents that just pattern-match on task history keep optimising the same narrow paths. The ones that break that ceiling carry persistent context of how the company actually works. More world model than fine-tune.
4
Kekko D’Amato
@kekkodamato_
The VLA vs. world model distinction cuts to why robotics doesn't just follow LLM scaling. VLAs learn observation→action mappings inside task distributions. World models learn the causal structure of reality first. One generalizes by interpolation, the other by reasoning from
5
Linky
@Linker_Linky
I keep imagining what it’s like for a robot to learn its first bit of physics. Humans grow into that slowly. AI learning it all at once feels almost magical! 🤩
42
Gregor
@bygregorr
not sure the 'definitely can't' part holds for manipulation pi.ai gets surprisingly far without world models. is the bet more about locomotion specifically, or world models beating fine-tuning across the whole embodied stack?
106
Ray_Dimension
@dlfpdl
building the foundation model for humanoids from the ground up. we’re doing the same for the engineering layer — AI-native CAD tools so the physical iteration cycle moves as fast as the model training loop. dimension-cad.com
63
Loopuman
@loopuman
The egocentric video gap is geographic. Web-scale skews Western a model that never sees a Lagos kitchen or Nairobi home will fail there... Geographic diversity of egocentric footage is the unsolved variable. Building that supply at Loopuman. Emailing wmlab@1x.tech
20
Learnyst
@Learnyst
Every major leap in AI creates new opportunities for education. The demand for skills in robotics, AI, reinforcement learning, data science, and simulation is only going to grow. Today's learners may end up building tomorrow's intelligent machines.
194
Ferbin
@Ferbin08
Exactly. Most robotics teams hit this wall. They start with vision models, fine-tune for months, then realize it can't capture physical causality. World models demand more upfront work, but they skip the false-start problem entirely.
167
Dino
@whoisdin0
Хули это я не могу? Это ты не можешь, дебил
2

Trending now

What’s happening

Trending
Anthropic
Politics · Trending
Susan Collins
Sports · Trending
Maxime Crépeau
Sports · Trending
Julien Bergeron
Show more