Random market thought after earnings: Meta and Tesla both sold off hard despite very strong revenue growth. Tesla revenue rose 26% YoY while robotaxi deployment slowed down and Meta's revenue rose 28% but free cash flow collapsed as AI infrastructure spending rose. The common theme seems to be discomfort with enormous AI spending when the payoff still looks uncertain or distant.

What makes me more bullish is my own experience with the technology.

On coding, I still see a surprising amount of skepticism, including from professional programmers who seem to think AI is mostly glorified autocomplete. My experience has been completely different. I went from writing everything manually, to autocomplete, to having AI write functions and classes, to whole features that I reviewed. Now I routinely give coding agents substantial engineering work and often barely look at the implementation. In my experience, they can comfortably do the technical implementation work companies pay a $350k/year senior developer to do. What’s remarkable is how quickly that progression happened.

I’ve seen something similar with Tesla autonomy. I’ve used Tesla’s driver-assistance/FSD technology for roughly eight years, and over that time I’ve needed to intervene less and less. It’s now at the point where in many situations I trust it more than I trust my own driving — and I love driving and consider myself very competent. Tesla also just reported rapidly increasing FSD adoption, with roughly 1.5 million paid users and an attach rate of around 55% on new North American deliveries.

I think Tesla’s technical approach is especially important. At MIT, my lab worked on robotics problems involving things like localization, mapping and perception, so I’m familiar with the traditional approach: detailed maps, expensive high-precision sensors like lidar, and a heavily engineered autonomy stack.

Tesla made a very contrarian bet early on: use inexpensive cameras and solve the hard parts with neural networks. That makes the AI problem harder, but if it works, the deployment model is dramatically more scalable — you can put the hardware on millions of ordinary cars rather than relying on expensive sensor suites and highly mapped operating areas.

That’s why I’m less concerned by the slow robotaxi rollout than the market seems to be. I see it as a hard engineering problem that still needs iteration, not evidence that autonomy doesn’t work. Tesla has historically been unusually willing to brute-force problems like this — the Model 3 production tent is still my favorite example.

And the payoff is much bigger than selling a nicer car. Imagine a large portion of taxis carrying passengers, and eventually trucks carrying cargo, no longer requiring a human driver. You’re automating an enormous category of labor and changing the economics of transportation. Cybercab is explicitly designed around that future, while Optimus is the much more distant extension of the same idea into physical labor more broadly.

Meta feels familiar for a different reason. Around Facebook’s IPO, investors worried that usage was moving to mobile and Facebook wouldn’t figure out how to monetize it. That concern looks almost quaint now.

And I don’t think programming and autonomous driving are isolated examples. Frontier models are beginning to contribute to genuine mathematical research: an OpenAI model recently disproved a longstanding conjecture related to Erdős’s unit-distance problem. In cybersecurity, the latest models can find serious vulnerabilities and, in controlled environments, turn some of them into working exploits.

There’s also a more mundane example sitting right in front of you: I used ChatGPT extensively to help write this essay. For this kind of writing, I’ve found it more effective than I am at organizing and expressing my own ideas.

That's basically my point. AI isn't some future technology I'm waiting to become economically useful. I'm already delegating large portions of both my programming and my writing to it, while researchers are beginning to use the same class of systems on problems in mathematics and security that require genuine technical reasoning.

So when I see investors looking at hundreds of billions of dollars of AI investment and asking where the return is going to come from, my reaction is almost the opposite: I think we're still underestimating how many forms of expensive human cognitive work this technology is going to touch.

Obviously enormous AI capex can still be wasted, and execution matters. But my impression from actually using these technologies is that the market is asking “where is the return?” while underestimating both how economically valuable AI already is and how quickly it’s still improving.