On Astro Mechanica, Compute, Molecular Nanotechnology, Biotech, Enterprise SaaS, AGI Media, and Robotics.
Disclaimer: The views and opinions expressed are solely my own and do not necessarily reflect the views of any current or former employer, colleague, or affiliated organization.
On Astro Mechanica:
They’re really quite good and I believe they’re undervalued as a company given their current progress, but enormously undervalued if they end up achieving what I believe they’re going to achieve.
The markets they’re targeting are big enough for them to be a large company, there are many unfair advantages they can build up, it is clearly Ian’s life mission, and I just think supersonic jets are awesome.
On Compute:
I’ve spent the last month or so seriously looking into various compute strategies that startups are doing, trying to see if anything is interesting as it could be quite impactful to advance AI and provide enormous useful value to the world.
I have a low opinion of startups that are trying to directly compete with Nvidia on transformer AI workloads with some small differentiation. I don’t think they’re going to make it. You need a true paradigm shift.
Thermodynamic computing isn’t a scam, despite many prominent voices not doing much to help that perception. It’s more of a bet than a certainty though, so we’ll shortly see where the coin ends up. The justifications around components being small enough at recent CMOS nodes to now exist in the stochastic regime combined with many generative AI workloads being mappable to Langevin dynamics is quite intriguing. The question though is whether the approach that Extropic is taking is the correct one or if there are better ways to do thermodynamic computing.
Reversible computing is similarly not a scam but it has less going for it in terms of how big of a gap it can have over NVIDIA. It also seems to suffer from the perpetual “it’s inevitable but still oh so early” feeling. I could probably ignore it for another five years and still be fine.
Timing matters just as much as courage or brilliance to do useful things for the world but doesn’t get discussed enough.
On Molecular Nanotechnology:
I think that, currently, machine phase systems is the most serious company going after molecular nanotechnology
I understand why there’s skepticism around the industry of molecular nanotechnology though: the markets scanning probe microscopy (SPMs) target are initially quite small and anything with the words quantum or nanotechnology in it sets off grifter alarm bells.
I think that this is an important area of technology that should be funded and run through to its logical conclusion though. If I was a multi-billionaire I’d fund this area and see it through.
However I don’t think it’s currently a good VC investment.
On Biotech:
The markets are fundamentally bad for many reasons, partly because of success being seen as selling to pharma and partly for many other reasons (population growth, IP cliffs, long feedback loops, significantly increased competition from China, etc.)
This is why even some of the best biotech VC firms in silicon valley with large AUMs tend to have mediocre returns and there’s essentially no true breakout biotech companies. (Moderna is worth less than Instacart.)
Despite that, there are a lot of fascinating things in biotech that would be useful to society that aren’t being done but should be: tooth regrowth, needing less sleep, next-generation beautification, advances in who can have kids, and so on. (In addition to base-editing beginning to take off!)
I’m not sure how to square the circle here though, in that there are many important things in biotechnology that could be done but aren’t done super well and even if done super well would not be great investments.
On Enterprise SaaS:
If you don’t integrate with Windows, SAP, Salesforce, and other similarly high distribution older enterprise companies, then you’re probably not at the big leagues yet.
It’s such a dumb little heuristic but tends to work more often than not. It’s also a clever trick for seeming (and being) more serious: how can you be a large enterprise company if you can’t even work with large enterprises?
On AGI Media:
High quality AGI media is quite scarce, even in San Francisco, the heart of the cultural boom around AGI. Some of this is due to people having real jobs with real NDAs that prevent them from speaking out, but some of it also comes from the people with traditional cultural authority not taking “this stuff” seriously.
The main person I see with cultural authority taking “this stuff” seriously is Kevin Roose, and I think his book on AGI coming out in 2026 will be incredibly important.
I’m very proud of helping him find voices, stories, and perspectives for this book.
On Robotics:
Hardware isn’t the issue, generalization via software is.
Inference magazine wrote about how there are hardware barriers for robots taking over 100% of human labor, but I view that as not terribly relevant for whether robotics as an industry gets enough momentum to steamroll all remaining barriers.
The correct thing to do back in the roughly 2018-2020 era of AI was to switch from robotics or agentic reinforcement learning theory to instead scale up large language models to serve as a base to then do reinforcement learning (“reasoning models”) or robotics on top of. (This is what OpenAI did and why they became so successful: they’re great at doing the thing that works over and over and over again. This is also the secret to being a great investor I believe. Just do the things that work and then keep doing them. The point isn’t to be contrarian, it’s to be right, being contrarian is just the cherry on top.)
The correct thing to do now is to switch from niche hardware concerns like what ARIA is funding and instead get the software in humanoid robotics to generalize enough to make the industry go from 0% of labor automated to even a couple percentage points, causing a massive capital influx which puts you in a “boom loop” that steamrolls all remaining problems.
Funnily enough, humanoids/robotics companies may not capture much of the value here.
Many companies are betting that the value of humanoid robotics will come from being able to have a robot that performs a diverse set of tasks. Eventually though, your robots all level out at roughly the same level of capability. Maybe you have a slight lead to start, but how good can a humanoid robot be at cleaning toilets, let’s say? At that point, the robots and their intelligence become commodities, differing from the LLM labs which can have a comparatively near infinite skill ceiling on white-collar work.
The value might end up accruing either to small companies which maintain a luxury home brand (such as 1X) or to potentially aggregators that are able to take this commodity and solve the various coordination problems inherent to society.
In other words, Uber-For-Robots will likely exist. Now, will Uber/DoorDash/Instacart/etc. be the Uber-for-robots? That’s hard to say. As we’ve seen in the data labeling industry, the Scale AI of XYZ is Scale AI. The something of somewhere is the nothing of nowhere. If they remain live players, then that’s quite likely.
Right now, though, it’s too early to tell. We should come back in a few years when the robots begin to work.