Personal relationships are what will make or break robotics over the next decade.

Right now, due to the unusual shape of technical progress in robotics, deployments are very “jagged.”

Portions of this “jaggedness” will remain, but the two key components, which I think will determine much of how robotics will play out, BOM and speed of deployments, will heavily rely on personal relationships.

Your BOM is a function of your supply chain, which is actually a function of trust, which is a function of, you guessed it, personal relationships.

Your speed of deployment is a function of your Go-To-Market function, which is actually a function of trust, which is a function of, well, I don’t need to say it again, you get the point.

So, given that personal relationships are going to be such a massive part of how big the robotics industry gets, it’s very strange to me that so few folks are trying to build these networks at scale.

(Note: when I say networks, I don’t mean in the highly transactional sense, but more so the “trust-building” variety. How are we, the robotics community, supposed to transform the world if we don’t even know each other?)

Last week, when I started the robotics papers groupchat, which unintentionally started “self-organizing” meetups in San Francisco, it shocked me how much of a hit it was. People really want community, but it’s hard to just get started.

I’m making a “real-world robotics gc” for anyone interested in buying, deploying, and building robots.

If you’re in construction, retail, logistics, manufacturing, eldercare, energy, or data centers, please come on in! The more the merrier.

Yes, I am aware that many Chinese OEMs are primarily on WeChat, so there will naturally be gaps in this group chat. However, I think a partial solution is better than no solution, and we can creatively iterate until the best set of solutions is found.

I’ve lived my entire adult life at the center of the AI boom, first after moving to Palo Alto from Northeast Ohio to attend Stanford in the fall of 2022, and now through living in San Francisco.

Here is (some of) how I think about AI.

1. Leopold Aschenbrenner is right: the markets still have not fully priced AI in. I know that sounds incredulous, and, yes, the timing will matter on many trades, so an unsophisticated leveraged long position will not do well. However, you should be default “skeptical of your skepticism.” Some of the most fervent AI bulls may overshoot their predictions, but most of the world will dramatically undershoot theirs.

2. Yes, I do use AI constantly, but not for deep thinking and writing.

3. Why? Elegance is difficult. The most brilliant insights are often the simplest. The ability to say the one thing that matters and then “move on with it” is rare, hard to articulate, and thus difficult for researchers to incorporate into their post-training stacks. It’s part of why Elad has made so much money and also why I think (in the short term) AI could remain reasonably jagged.

4. So, what do I “use AI for?” These days, mainly building custom websites and getting up to speed on new fields. While Claude Code probably shouldn’t be trusted with writing global-scale distributed systems yet, I’ve used it to help make beautiful websites for essays I’ve written. Furthermore, getting an 80/20 on the “new vernacular the kids are saying these days,” what recent political event happened, or learning your way around a codebase, has never been easier.

5. However, for the best insights, you’ll want to talk with and work with those on the frontier. The most interesting things about a field, company, or person tend to be “materially non-public” for a reason.

6. Yes, most people are bad at prompting/context engineering. However, you don’t need to buy a convoluted course to get good at it. Instead, imagine you are the AI, trying to complete a task someone gave you. What information would you need to do your job best? Then give yourself (the AI) that information.

7. Although I am no longer a student, I do think professors are right to be wary of AI. In particular, classes that teach technical fundamentals should be very wary of AI, as not only do you need the foundations of your field to advance the frontier, but many great breakthroughs have been shockingly simple. Again, elegance is difficult, and one can learn it only through complete and total mastery.

8. However, the way forward there is to raise the ceiling for what an education can be. Essays should be much more deeply researched. Final projects could be much more ambitious. What would you do if you had 10x the agency you do now?

9. Finally, I think most (even in AI) are underestimating how big robotics could be as a market. It’s why I’ve very recently started writing on Substack in the first place: there is such an enormous arbitrage right now, similar to what I have been told by the LLM researchers who saw the GPT-3 results in 2020 and briefly glimpsed the world of today. I made a GitHub repository on the current frontier of robotics, which I plan to massively expand next week, but if you feel as though you “missed AI” (which, by the way, isn’t true: there’s still a lot of room left for AI to run), now’s the time to start thinking about robotics.

This is a very provocative claim, as robotic manipulation is neither human-speed nor generally 90% reliable in the real world, and any evaluations we could use to rigorously measure it remain primitive.

However, 1X’s recent paper, along with Skild’s Series C announcement, shows that there are dozens of avenues beyond just the doomed “data sweatshop model1” that could be pursued in parallel to rapidly improve the capabilities and efficiencies of robotic systems.

First, robotics could leverage large-scale simulations, internet videos, egocentric data, teleoperation/UMI data, real-world deployments, caption upsampling, preference learning, and test-time compute all as separate axes to “scale capabilities.”

Some of these axes, such as the 900 hours of egocentric data and 70 hours of NEO teleoperation 1X used, along with test-time compute, could feasibly be scaled up more than 100x this year.

Meanwhile, others, even when combined in small quantities, enable previously impossible behaviors, such as scrubbing dishes autonomously.

Additionally, we haven’t thrown in speculative decoding, quantization, distillation, framing DepthAnything as a depth prediction pretraining auxiliary objective,2 or the million-and-one optimization techniques that top game developers and LLM hackers use every day.

But What Does This All Imply?

Eric Jang once wrote that robotics will create an API to the physical world and make physical reality programmable the same way software made information programmable. I think that as robotic manipulation becomes reliable and human-speed, then this vision of the future will come to pass. However, if we are going to now have an “API to the physical world,” then…

We will also need new abstractions for the physical world that don’t exist right now.

What is the Map-Reduce for robotics? How do we coordinate 10,000 robots building a building together? Do we need to learn lessons from Cursor's recent long-horizon multi-agent research? Or perhaps we should revisit Edward Dijkstra’s 1972 Turing Lecture. I don’t know, but…

We’ll also need to start caring about the “Physical AI Deployment Gap” sooner than most think if reliable manipulation starts working.

Oliver Hsu already masterfully covered this in his linked essay above, but I also suspect that CommaCon and the broader AV community will offer further interesting insights here on how to close this gap. Companies such as Auki, which are building spatial “maps” of the world, will also have a head start on this problem, but there are many other angles to consider as well.

Finally…

As the underlying cost and capabilities of individual robots get commoditized, companies will need to shift up the value chain to grow. Coordinating a single robot is easy, but coordinating the construction and operation of ever-larger factories, datacenters, and, eventually, even orbital megastructures, will be what the best robotics companies evolve into.

Pushing the limits of programming the digital world has led us to construct the cathedrals of code we call the modern internet.

Pushing the limits of programming the physical world may lead us to something even greater.

⁂