Shamoon Siddiqui

This is Genie 4, our most realistic and interactive world model yet. 🌐 Explore dynamic environments in real time to see how the scene responds, sounds, and how the people in it move, react, and speak.

When we started @SereactAI , people told us returns were impossible to automate. This is Cortex handling @Zalando returns at a @cevalogistics fulfillment center. It’s live and in production, every shift. Proud of the team that made this real.

One question I get from business owners all the time: How to turn everyday work into robotics training data? Now, I finally write an article about that. The best way to "catch up" on the physical AI transformation is to embrace it.

LLMs are magical alien technologies we’ve managed to conjure up from a distillation of the internet. Even if progress completely stalls, we have years worth of value to harness from what we already have. In other words, these lego bricks still haven’t been put together in the most obvious combinations.

1. Humans will not stop creating text, images, and music. We will always do this for joy and self-expression. 2. AI will create most content in the future, and it will displace many humans doing so, just as machines did for physical products. The result will be an abundance of affordable, high-quality text, images, etc. for everyone to use and enjoy. 3. You might be sad about this. I won't tell you that you are wrong, because I don't police people's feelings. Regardless, this transformation will be, on net, a very good thing. I understand if it's emotionally hard to process, though.

The AI boom has created insatiable demand for data centers. Nobody knows how long it will last. Everything that goes into data centers, from cement, to the labor that builds them, to the electricity that runs them, to the chips that provide their intelligence, is overpriced by a large factor. Gas turbines are on track to triple in price from 2019 levels. An electrician wiring a data center makes about double what electricians make everywhere else. Nvidia sells its chips at a 75% gross margin, as if it were a software company. Companies making these investments are drawing straight lines into the future.

Thanks to @Nasdaq for celebrating our 100M ARR milestone in Times Square.

Announcing the beta release of the Pi API. We’ve partnered with companies across diverse industries to adapt our models to their robots and applications, from harvesting strawberries to assembling servers. Now we’re opening access to more teams.

150 GW of new production brought online over 20 years, or 7.5 GW per year. Most of it would be wind power. Offshore turbines currently generate about 25 MW, so they would have re-industrialized Glasgow's dockyards to produce about one per day. This is about the same rate that the Kaiser shipyards produced finished liberty ships, which are substantially more complex. One turbine per day doesn't sound like all that much, but do it for a decade and Britain doubles its power production capacity. Or stay poor forever.

sat in a room full of robotics founders in SF tonight. everyone agrees VLAs continuously fail, so real deployments rely on reasoning models and vision. most demos can be vibe coded on GPT 6.1 Sol every company has the same progress. only the ones in the media look ahead.

Most robot policies treat time implicitly, using sequence position as a proxy for physical time. But real robot systems rarely run on a fixed clock. Sensors run at different rates, observations arrive with delays, and actions take time to compute and execute. Successful control depends not only on what the robot observes but also on the timing of observations and actions. We introduce 🕒 Chronos Policy, a diffusion policy that explicitly conditions on observation capture times and requested action execution times. We use time-aware positional encodings to make attention depend on elapsed physical time. Because timestamps let the policy interpret each sequence's timing, we can randomize sensing and action timing during training without making those variations ambiguous. This exposes the policy to a wider range of timing conditions, improving robustness to timing changes at deployment.

A new kind of company is emerging in robotics: the deployment company. Broadly speaking, a deployment company builds on a frontier model instead of training its own. Instead, it builds/assembles its own hardware, fine-tunes the model on vertical-specific data, and owns all the hard work of deployment: commissioning, ops, support, and workflows. Until recently, this wasn't an option. If you wanted to deploy a robot that did real manipulation work, you had to build the autonomy yourself. This meant a big robotics team and years of data collection before you could deliver a reliable robot. Now, labs like @Physical_int and @GeneralistAI are changing this pattern. So far, Pi has announced three deployment partners publicly: -> @Weaverobotics builds robots that do laundry and take care of the home -> Actor Labs brings Physical AI to heavy machinery like excavators and loaders, starting with construction. -> @Ultraroboticsco robots pack e-commerce orders in warehouses Notably, each of these companies uses different embodiments in different verticals but can still benefit from the same base model. Generalist hasn't announced any partnerships publicly, but today has released the waitlist for Gen 1.5 (and I've heard whispers of deployment partnerships there as well). The economics of the partnerships aren't public yet, but the setup is mutually beneficial. This lets the labs focus 100% of their effort on developing better models and avoid the demanding deployment schlep. On the flip side, this lets startups benefit from the billions of dollars and world-class researchers needed to train these models. This is incredible for startups. With SOTA models available to build on, teams need less capital to get started, reach their first deployment faster, and spend their time obsessing over customers’ problems, workflows, and the long tail of details that create real value.

Gemini 4 Argon is a step change and will have massive impact on robotics. We have published a developer blog to show various ways in which Argon accelerates robotics. developers.googleblog.com/ge…

Someone is going to be the Hugging Face of robotics. Probably the OpenRouter too. It won’t be a robot company. And almost nobody is building it yet. Yesterday I stood in a London robot gym. Three arms, three makers: FANUC, ABB, Universal Robots.Three controllers. Three SDKs. Three ways of saying “pick up that gear.” That’s robotics in 2026: brilliant hardware, no shared language. LLMs got their platform layer in about two years. Open weights on one hub. One API to route across every model. Switch providers in a line of code. Robotics has none of that. Most robot skills are still welded to the arm they were trained on. Meanwhile, the money is flooding in: nearly $48bn into robotics this year, per PitchBook. Much of it goes to data: gyms, rigs, operators. Billions spent on data that can’t leave the building. That isn’t a market yet it's a stockpile. ft.com/content/e5fd2655-f95f… The inflection point isn’t a better humanoid. It’s portability. When a skill learned on a FANUC runs on an ABB tomorrow, and on a robot that doesn’t exist yet the day after, the economics of the whole industry flip. OEMs will read this as a threat. It’s their biggest upside. Today an arm does what it was programmed to do. Tomorrow it runs any skill from anyone. Same steel, 100x the jobs.The OEM that opens up first sells more arms than the rest combined. What the hub needs: → a common format for robot data and skills → routing: the right model for this arm, this task, this cost → provenance: where every clip came from → payment: the person paid 94 cents to record a task gets paid again every time it’s reused That last one is the edge. Hugging Face never had to pay for weights. A robotics hub will, because the data costs real hours from real people. Whoever solves micropayments for contributors wins the supply side. That’s the machine economy. It is being built right now. Here in Europe. You doing your part? My DMs are open.

ICON announced our $200M+ Series D today. I thought it was worth sharing again why we do what we do.