Erwin

Truly blessed to be the grand winner of one of the biggest hackathons ever organized (@colosseum) with 3000 teams competing. Thanks to everyone for your support! Make sure to follow @crowdbrainai closely. The next moves are going to be big!

My entire timeline is doomposting robotics data companies this week. While it’s one of the fastest growing categories and it’s nowhere close to being done. Scale, Mercor and micro1 winning doesn’t mean the space has no room left. Someone, who just forces their workers to stick gopros on their heads, going out of business, doesn’t mean your approach will also fail. The category is real, with near infinite demand for a better product, a sharper angle and an execution nobody’s tried yet.

@GoingBallistic5 Scaling production (especially to millions of robots) instead of rapid iterations at this stage just sounds extremely dumb both from technical as well as business perspective.

The maxim, “Prototypes are easy, production is hard”, does not apply to humanoids No one, including Tesla, has a working prototype worth scaling in the 10’s of 1000’s, let alone millions It’s not a scaling problem, is a prototype problem Humanoid prototypes are hard Current prototypes might be scalable but are not sufficiently capable nor robust. Nor safe. Including Optimus You don’t scale a crappy humanoid. Unitree could easily scale the G1 to millions, but to what ends? Obsolescence in 6 months? Humanoid prototypes are hard Ramping up a million and 10 million annual capacity lines simultaneously is premature. We will never see that quantity of V3’s, not because the supply chain isn’t yet mature, but because the prototype isn’t Tesla is attempting to scale a moving target Tesla is impressing by flexing its scaling muscles first, rather than demonstrating any clear technological adroitness first V3 will not be revealed this year, not over piracy concerns, but because the bar for humanoid demos is beyond its current capabilities to demonstrate The million plus Optimi will be V5 or higher Humanoid prototypes worth scaling are hard

@LimX_Dynamics It's funny how chinese robot company can announce $200m round and get 8 likes and no reach at all lol

@radbackwards Send one to ASAP (rocky)

Quavo asked us for a NEO. I tried to give him one, but he still hasn’t responded…

@ericliuof97 Noveltymaxxing

“Hours of data” is a misleading metric. There are hours—and then there are hours. One hour may contain repetitive, low-quality behavior; another may cover diverse skills, objects, environments, and failure modes. Quality, diversity, and coverage matter far more than duration. Might need to report number of demonstrations × hours.

@vargastartup Goated @colosseum. As we are running onchain ops, colosseum was the best option for us. Had an yc route opportunity as well.

@PrimeErwin Which one did you get into?

If you are planning to move to sf to build a startup, getting into a top accelerator should be your #1 priority. It makes finding clients, partners, VCs dramatically easier and it saves you years of grinding for the same opportunities.

@adelbucetta True. If you are sure you can pull it off yourself and got that kind of motion, could be negative. But for absolute majority of founders, accelerators are huge ev+

@PrimeErwin the honest answer is that accelerators have their advantages, but they also come with strings attached you're often sacrificing equity and autonomy in favor of a short-term boost

@NimaZeighami Thinking about remote exoskeleton teleop as well. Want to experiment and try as many things as we can to find out what makes the most sense. What would you recommend

@PrimeErwin Good luck, good teleop is hard but it’s still early enough that there’s still MANY areas for improvements.

We just bought 50 Meta Quest and Pico VR headsets for our teleop operators around the world. Building a qualified global workforce, gamifying their experience and nailing the distribution ahead of massive demand that’s coming is the real moat.

@NimaZeighami Yes and currently also testing our own custom solution.

@PrimeErwin What’s the base teleop stack you’re using? XRobotToolkit?

@JTEIII Home and factory tasks. We have that already set up and running

@PrimeErwin What types of tasks are you looking to start your teleop operators training on first?

@IsaacSin12 Isaac sim vs Isaac sin

same same but different

@oyhsu Decent article. The shift toward application focused teams and neo integrators feels like the next phase that we’ve been missing for a while.

A review of patterns we've observed in physical AI over the last 2-3 years, as the field has continued to make technical progress and aggregate talent and capital across new and existing companies: In 2023-2024, we started noticing more teams building for a given layer of the robotics stack. Whereas historically teams built all their own infrastructure, the promise of scaling laws for robotics meant generality was on the table, and what generality enables is for teams to build on top of a common platform — a huge advantage given the long tail nature of the physical world. We also posited that given the variance of the physical world, the proliferation of such a platform might resemble more of an open ecosystem in the vein of Android. This dynamic could kick off an apps>infrastructure cycle that robotics has historically lacked, but every other major technology cycle has seen (web, mobile, crypto, etc.). This is partially because there has been no breakout robotics app, but maybe, like with language models, this would be a case where infrastructure (and generality) kicks off the cycle. I wrote about that idea here 2 years ago: a16z.com/toward-a-gener… Much of 2024-2025 was focused on the infrastructure side of this ecosystem, particularly around different methods of scaling robot data. Concurrently, there was also more activity in building the hardware platform and components of robot infrastructure, as well as the emergence of tools built for a 'robotics developer' persona. 2025-2026 saw more entrepreneurs and engineers shift their focus towards the application layer. Specifically, the idea of a 'neo-systems integrator', or a company focused on deploying the new generation of learning-based robots, started coming into focus as more early stage teams began to pursue this. Concurrently, large robotics labs began pursuing both first and third party deployments, and the app<>infrastructure relationship started taking shape. I wrote about the deployment layer and some of the problems to be solved in deploying learning-based systems here: a16z.news/p/the-physical… In parallel, throughout 2024-2026 we saw the proliferation of research around different elements of solving generalist robotics. Beyond overarching methods built around VLAs, world state prediction, and sim, there were key building blocks in areas like human motion transfer, online RL, domain randomization, and bridging the high and low levels. Moreover, we continue to see more evidence for scaling laws for robot actions. This research progress across the field forms the building blocks for the pursuit of generalist robotics and a robust robotics developer ecosystem. I've also started focusing more on the frontiers of physical AI beyond robotics, specifically on incorporating physical reasoning and physical modalities into AI in general. Much of this work is still somewhat nascent, and I wrote about some of these areas here: a16z.com/frontier-syste… Some of the things I think are interesting to observe now: - Where we are in the Perezian financial/technological cycle with robotics. - Without knowing the thresholds we have to scale to see inflections in robot capabilities, robustness, and generalization, what is a minimum deployable robot? - Signals around adoption of common hardware platforms for robotics developers. - The combination of scaling less-than-reliable deployments + continual learning for robots. In general, I continue to believe that many of the most interesting things in robotics in the coming years will be emergent behaviors and properties. This is exciting and also involves working with a lot of unknown unknowns.

@BostonDynamics We need more of this kind of content

World Cup fever continues! Viking Row at Boston Dynamics HQ!

@adambcohen93 We need to bring toxicity back. Good for both - founders and VCs.

@JunyaoShi Deploy memo in our businesses!

I wonder🤔

While toxicity can sometimes cross the line, it brings radical openness, direct feedback and a shared obsession with what actually matters. That’s how the strongest and long lasting relationships of any kind (including founder-vc one) are built.

@allgarbled 100% youtube.com/shorts/uIEUNBN…

Extremely overhyped sandwich and very mid

Founder friendly VCs in the 90's