← back to choices

The problems I care about share a common bottleneck: the speed at which we can traverse design space.

Every project I've worked on (batteries, robotics, medical devices) has been constrained not by a lack of ideas, but by the cost of testing them. Physical iteration is slow and expensive. In-silico iteration is not nearly as fast and cheap as it should be. For the wicked problems I care about, time is in short supply: clean energy abundance, ocean stewardship, vaccines available to all.

In physical design, CAD formats don't talk to each other, simulations require manual setup, and manufacturing constraints live in tribal knowledge rather than computable rules. In science, experimental data is siloed, hypotheses aren't machine-readable, and the path from observation to insight is rarely automated.

The result is that most of the design space, and most of the hypothesis space, remains unexplored: the battery chemistry that makes storage cheap, the antibiotic for the next resistant strain, the diagnosis that comes early enough.

I'm interested in latent spaces for 3D geometry and scientific knowledge. Representations that let us search, interpolate, and optimize across designs and experiments. Wicked problems, faster.

← back to choices

When Deep Blue beat Kasparov, some declared chess finished — what was the point of mastering a game a brute-force algorithm could already play better? The answer required letting go of needing to be the best. What followed was centaur chess ↗ — human + machine — and after more oscillation through self-play and reinforcement learning, the game is richer than it's ever been. Players now see patterns that were computationally unreachable a generation ago. The space to master has never been more open.

Post-AI education doesn't yet have its centaur moment.

It's all fun and games when a kid turns in an essay they didn't write. No one's laughing when the resident doesn't know where your spleen is.

Nobody runs a marathon for the mileage. Why are we building learning systems as if the artifact at the end (the essay, the PSet, the grade) was the point?

That's the default mode of almost every LLM interaction around learning. Every stage of training, pretraining, RLHF, instruct-tuning, grows these models into obsequious, extremely capable genies. Having a genie is wonderfully irresistible… too bad those who wish to fly tend to wake up strapped to a goose.

I've been building self-tutoring tools for myself since graduation, because the time I had for mastery kept shrinking while the gap between what I'd studied and what I could reproduce kept widening. The current one, idactic ↗, is a personal experiment in teaching myself to do math derivations. What I didn't expect, after dozens of attempts, is that the hard part is never the AI. It's teaching the AI when not to help. Every feature added is another place the learner can outsource the wrong skill. Surprise, surprise; less is more.

In addition, it's never been easier to connect a learner to a tutor, a peer, or a teacher. Education research clearly shows massively improved outcomes simply by having the proper community and feedback loops in place. These tools enable rapid parsing, storage, and transmission of learning through video and audio embeddings, which can give teachers a window into the mastery (and also misconceptions) of their students.

I am concerned that we are pulling up the ladder on the next generation of learners by unleashing unrestrained helpful, honest, and harmless AI into education. While that philosophy avoids many first- and second-order harms of AI deployment, education and human capability are unmitigated casualties.

What I want to build is a learning harness: a system that absorbs every task surrounding the cognitive work — scheduling, decay tracking, reformatting scribbled notes, surfacing what to interleave with what — and strategically refuses to do the deriving, explaining, or recognizing itself. I also believe very simple, well-known patterns can be merged with this philosophy to create learning communities that unburden our education system.

The harness that knows when not to help is the hardest to build, and the most worth building.