“The machines have proved something about the equations, but turning verified facts into understanding, and understanding into prediction – that’s still a human job.”
– Justin Beroz, ReynKo
ReynKo is a lab focused on applying their solution to Navier-Stokes equations, and a Third Sphere portfolio company. The promise is first-principles physics and fully testable predictable solutions. That testable, predictable part is what’s interesting. It’s great to generate more solutions to math problems, but it’s prohibitively expensive to build models for each one and test them in a wind tunnel. ReynKo’s is the only solution that avoids those costs and with better outcomes.
Prove that it’s useful
OpenAI put roughly ten thousand agents on proving that the Navier-Stokes equations can break down. If it survives verification, it settles a question the field has chased for a century. But it will not make a single wing, turbine, hull or heat sink easier to design. Solving CFD problems in software, is what Justin Beroz, ReynKo’s founder, calls the trillion-dollar questions.
Until this month, in mathematics, proving and understanding were nearly the same thing. A proof was the written record of human understanding. And understanding was the precursor to making useful things with that knowledge. Now there are results that are machine-checkable and not yet understood. Interestingly the universe of people who can check claims for this math AND the physics is very, very small. And even with verification, it doesn’t mean there is anything useful that can be done with the result. NVIDIA’s own CFD team, writing about AI surrogates: a model can achieve a low mean squared error while still violating important physical constraints.
Violating physical constraints is a bit like an unplanned rapid disassembly. It’s not really a thing you want. Our investment case for ReynKo was to let designers and engineers see a clear picture of reality in software so they can make and test digital designs and know they can expect the same result when they bring the designs into the physical world. Here’s Justin on TBPN explaining the difference.
Alignment is a systems resilience problem
Anthropic researcher Jacob Coxon resigned publicly over concerns that AI could kill us all by the end of the decade. His colleague Evan Hubinger put it above 10% within the decade and said plainly that there is no plan yet for aligning superintelligence. It’s an interesting conversation for some, and hypothetical deaths are easier to talk about than the measurable and avoidable deaths that are already here. Researchers attribute more than 178,000 deaths to the 2023 global heat wave, over half of them to human-induced climate change. WHO projects roughly 250,000 additional deaths a year between 2030 and 2050. And that’s not including floods due to warming. One of the fastest-growing new sources of emissions is hyperscalers spinning up fossil-fuel-backed datacenters to run AI. In the US alone, data centers are on track to use as much gas as Germany and Japan combined. So yes, in a way, AI is already killing us.
Beyond killer super intelligence and planet warming resource use, we can also talk about alignment as a resilience problem. Even with perfect alignment on people and planet, the tests of how fragile our human systems are does not go away. No rogue models, no bioweapons needed. Just everyone holding a personal agent with their credentials, and ordinary people having ordinary bad days: Mehrad Yaghmai’s examples run from the road rage that used to decay on the drive home to the neighbor dispute that becomes a night of code complaints and FOIA requests, every one of them legitimate, every one something a motivated person with a free week could have done, and now everyone has agents with unlimited free weeks. All of it is aligned use, which is the problem: what made petty grievance harmless was never intent, it was that acting on it was expensive, and as he puts it, “we are about to find out how much of civil society was held together by the fact that revenge was inconvenient.”
One remedy for both big Alignment (AI mass annihilation risk) and little alignment (society breakdown from death by a thousand grievances) is liability. Thomas Claburn
thinks we should imprison AI execs for releasing badly behaving models and agents. Oliver Schmidt of VW got seven years for Dieselgate. Mehrad’s remedy is institutional: rate limits on filings, provenance on complaints, an agent-of-record standard so a recipient of forty complaints in one night can … know who to punch? Societal resilience might require making humans even more liable for their actions, and their AI agent’s actions. And it’s likely their AI agent that snitches on them.
Plan A still isn’t Series A
On August 27, Hugging Face and Pollen Robotics opened pre-orders for Microduck, a 25cm bipedal robot duck at $399. It passed $1M in about six hours and $2.6M within a day. By the following week it had cleared 10,000 units and roughly $5M, at a peak rate of about one every four seconds, with delivery pushed out four to six months. Thomas Wolf’s own write-up is the founder’s version of how they did it, and CNBC, Quartz and Yahoo Finance have the outside numbers, including the Rockchip part that makes the price work. This is hardware, and the customers funded the production run.
It is also a choice more companies are making deliberately. Fi has raised roughly $40M but is now turning away investors clamoring to write new checks. This is an old idea, making a comeback. CIA money kept NeXT alive in the late ’80s, and NeXT’s code became the foundation of the iPhone. Non-dilutive government capital underwrote the most valuable consumer product ever built. It is not the part of the story anyone tells but is why we built Capital Stack Compass.
The non-equity layer comes with its own bill, though. Jigar Shah, the founder of Miles wound it down this year after a decade. Miles was one of our early portfolio companies, and we rooted for it the whole way: roughly $30M raised, 7 million registered members and about $5M in revenue, at break-even. A viable business that was no longer a viable venture investment. He also had debt to service, and among the lessons he lists is what founders should understand about debt before taking it, which is a fair warning to anyone taking our advice to go find the non-equity layer. More capital sources means more stakeholders to align, and at the end that is what narrowed his options rather than the business.
Portfolio Updates
Google is putting 25 electric semis on a new all-electric Houston–Dallas route with Nevoya and the Center for Green Market Activation, announced Sep 8. The full project is 63 trucks, the largest known Class 8 zero-emission deployment in Texas, roughly 11 million miles a year, roughly 92,000 tCO2e avoided against diesel over the contract terms.
Wildfire-fighting technology evolves with driverless drones and AI tools, in Bloomberg’s coverage of Sonic Fire Tech.
Metalmark launched Pasoh, a room-by-room indoor air quality sensor, and the first units are going into Mayo Clinic emergency departments this month under the ARPA-H BREATHE program Metalmark joined last year.
Wonderfil keeps adding campuses: Lewis & Clark gets a 24-hour refill station this fall, joining CSUMB and UC Santa Cruz. Bring your own bottle, pay about 40% less than the shelf price.
Opportunities
Mill wants a Marketing Lead for Growth & Acquisition in San Bruno, and an Operations Program Manager for manufacturing. Revivn needs a Logistics Coordinator in Fresno. Mark43 is looking for a Strategic Account Executive covering Colorado, New Mexico and Utah. Circuit is hiring Driver Ambassadors for its electric shuttles in Pompano Beach, Woburn and West Dallas. ChargeLab is hiring a Senior Product Manager, Charging Experience in Toronto, and Sention wants an Applications Engineer for battery diagnostics in London. See all 229 open roles at jobs.thirdsphere.com.
Third Sphere
Come Touch the Future: Hard Tech Expo. Thursday, October 8, San Francisco. We would love to have you at Third Sphere’s SF Tech Week event. Doors open at 1pm and hardware demos from our portfolio founders run all afternoon. You’ll hear from speakers including Nancy Pfund and sci-fi author Ramez Naam, we’ll give a talk with portfolio highlights and our take on the AI correction, and we’ll launch the Speedstrap book. RSVP here. We’re opening a spot or two for guest demos, so if you’re building something you want in front of the room, reply and tell us what you’d show. Know people who should be there? Just hit forward. We’re also holding our AGM earlier that week for current and prospective LPs; reply if there’s someone we should invite.
Shaun and Miela will be in New York for Climate Week. Shaun is joining Capital at the Crossroads on September 22, a New York Climate Exchange and OxCAN session at Moody’s on how climate capital actually gets allocated. Shaun will also be at the NYCW kickoff at PENN 2 on September 21 and Small Funds, Big Alpha on the 23rd. Miela has a very comprehensive shortlist of events she’ll be popping around through. If you’re in town, reply and we’ll share more or find time for a 1:1.
The Speedstrap book is up for Pre-order on Amazon. It is the single most useful thing you can do for the book and takes about nine seconds. You can still sign up to help on the working draft too.
Looking forward to aligning with you 😉
Best,
The Third Sphere Team: Shaun, Stonly, Yana, Miela & Roscel
P.S. WK654 = 654 weeks since we sent our first Third Sphere newsletter.
P.P.S. Know founders we should meet? Forward this to them.