Building in Strange Times
It feels a bit silly to be building anything right now. Especially an “AI” company. The world feels tilted, the economy uncertain, the noise constant.
And yet here we are, teaching a little robot named Lily to move through a massive space using only what it sees, how it perceives depth, and a few fancy-smancy algorithms based on imitation and reinforcement. It sometimes feels massively disorienting. But I digress.
At the Actuate conference last month, some of the brightest people in AI were debating which would win: one massive model or an ensemble of smaller ones.
Between the presentations and the “we’re hiring” speeches, what struck me most wasn’t the technology. It was the privilege. The fact that we could all sit in San Francisco, sipping coffee, talking about which flavor of intelligence might win, while the world outside tried to stay balanced.
Weirdly, that, too, is what makes this moment precious.
Some of us are privileged to ask strange questions. Others are privileged to build things that might one day feel obvious.
This is where I continue to be, in spite of the current geopolitical tone and tenor.
If possible, keep doing the following:
Make something with maximum utility per dollar spent.
Make something that hides its technical complexities. Turns out end-users don’t care its powered by a Mixture of Experts or a single all-knowing model.
Make something that just works.
Something useful.
Something magical.
Where we are
Hardware will always matter, but it will become a commodity.
Traditional software, from navigation to manipulation, will evolve to fit the reinforcement learning era.
Learning systems, not rules, will define the next decade of computing.
That’s what I’m working on with AugustMille — not a robot, but a learning infrastructure powering the next generation of autonomous robots that work where you live.
A system where embodied agents (robots) learn by watching, not by being programmed. Where the only setup you need is a short description or a bit of show-and-tell. Where your autonomous helpers understand your preferences the way your phone already knows your face.
Lily is our proof.
Limited pre-work (no pre-mapping)
No dependency on the cloud
No human watching over.
Just camera, reasoning, and sub-second inference on a single-board computer (or, in my case, the Jetson AGX Orin).
All offline.
Why Bother
The best time to build is during uncertain times. I know that’s a privileged take.
Maybe the reward isn’t a trillion-parameter model or a billion-dollar valuation. Maybe it’s simply the joy of watching something move — gracefully and autonomously — through the same messy world we live in.
AugustMille is building the infrastructure for robots that work where people live.
Mapless. Language-driven. Learning through fun demonstration.
The operating system for physical AI.
When it’s ready, I plan to make it freely available.
I’ve avoided saying open source because that phrase comes with a level of stewardship I’m not sure I can commit to. But sharing this work, in some form, feels like the right thing to do.
A few additions
On a lighter note, here’s a short video I took today of a school of fish being magnificent — a small reminder to find community, to keep moving, to keep pressing forward. At the end of the clip is Lily moving around: no maps, just RGB-D cameras, a fine-tuned generalist policy, and plenty of creaks and cracks.
Sergey Levine — we met at the Actuate Conference. I have a few thoughts on your next-gen Physical Intelligence models you might appreciate. Lily above is running on a fine-tuned version of Pi0.5. Yes, it is running on a 64GB Nvidia AGX Orin. Yes, it was non-trivial to get going. But, it works. We should chat.
See this link for a cool article from my partners in crime: AI-Powered Spatial Experiences for AugustMille. Pretty cool. Thanks to the ManyOne team, and to Silas.