Starting companies that will add 1B QALYs to humanity. Previously started CareSignal (acq. 2021), now working on physical AI @robot_health
St. Louis (freq in SF & NYC)
Joined October 2011
- Tweets367
- Following181
- Followers372
- Likes66
Pinned Tweet
It's 10x easier to point out risks than to propose solutions.
It's 10x easier to propose solutions than to build them IRL.
AI assistants are inches from being competent enough to "manage up" and hop into the driver's seat. The past and coming weeks feel like the turning point: now it's a viable choice to delegate most intelligence-requiring work. Imo working with intentional curiosity = now crucial.
The marginal cost of bits in a world of compute & energy abundance will trend to zero; value will increasingly accrue to cos moving atoms through the world in ways that help humans live better, longer lives (true utility IMO). The QALY might replace EBITDA as a proxy for EV.
Don't be afraid of that two-hour Fable stack rewrite. Those few days will be the best week of your life.
Pricing tokens on value instead of cost is getting harder than ever, even as offerings are de facto replacing huge chunks of salaries.
Imo logical conclusion is that solution providers must keep verticalizing until they reach the customer that bases purchases on value, not cost.
Deep industry expertise is more valuable than ever, especially if applied by hardcore AI-native team. Industry veterans can bring more EV to new market entrants than ever.
But the reverse is harder, given org inertia and change mgmt.
That asymmetry may do weird things to comps.
Had a dream about fixing part of a sim-to-real gap. Tried to implement it, and hit new type of sim-to-real gap.
Some companies used to (or still) capitalize their software development. Many will be tempted to capitalize token spend on AI-built software. Some will do that, AND capitalize their GPU purchases.
GPUs aren't useful for 5+ years. Software might be ~free to create next month. 😬
Right now builders are using tokens to create code that creates value for consumers/customers. Next, builders will use tokens to create products that allow consumers/customers to use tokens.
If anacyclosis is a useful model, perhaps the way out is governance by philosopher-king physical AI.
Blake Marggraff retweeted
Interesting paper coming out of THU: useful robot manipulation may come less from a single, ever-larger end-to-end policy, and more from giving a VLA a good agent harness. Let the VLA specialize in contact-rich actions like grasping, insertion, and fixture actuation, while a planner handles grounding, staging, retries, memory, and long-horizon composition.
This rhymes with what happened in coding agents: the model matters, but the surrounding system is what turns brittle capabilities into reliable work. Still simulation-only, but this feels like an interesting, practical path toward useful robots.
arxiv.org/abs/2607.08448v3
Using LLMs to do analysis of text or data without code as a middle step feels like the cardinal sin of AI in business. Like firing a machine gun at a hunk of granite to carve a statue.
Now, the main work tasks for which I'm *not* using LLMs are (i) writing natural language content and (ii) summarizing natural language content other humans have written for my direct consumption.
Corollary: my business use of LLMs is converging on deterministic work products.
First time in my career building stuff that is perhaps multiple step ahead of multiple research publications that were all published <30 days prior to today.
No joke, this will force governments to rethink immigration. Before autonomous humanoids are omnipresent, there will be real labor cost advantages to real-time tele-operated physical AI capable of real work.
Pithily: if this walks across a border, is it an immigrant?
