The Epistemic Jerk

In physics, jerk is the rate of change in acceleration. It’s the feeling in your body when a driver slowly accelerates before slamming on the gas. While thinking about AI, I keep returning to this word: jerk. It captures the feeling of suddenness and surprise at the staggering rate of change.

Within roughly two years, many programmers have gone from writing code “by hand” to using AI to code nearly everything. And it isn’t slop anymore. Coding agents can convert mockups into websites, write bug patches for the Linux kernel, and reverse-engineer compiled binaries. Overnight, the experience of programming has changed. Many developers point to Anthropic’s release of Opus 4.5 as the moment agentic coding started to work. The model was released less than a year ago.

Similarly, a few years ago LLMs would routinely fumble basic arithmetic. Then Google’s Deep Think achieved gold-level performance at the 2025 IMO. Then OpenAI said its internal model had produced a Lean proof resolving the Navier–Stokes Millennium Prize problem. Then OpenAI dumped over 700 papers on GitHub, with results ranging from the unique games conjecture to quantum magnets. The progress is so rapid that the math community is having a collective and public reckoning.

For myself, this has been humbling. I started my PhD in machine learning in 2016, a year before the transformer. I’ve published a paper using deep learning to extract latent representations from histology slides. I’ve written about LLMs and hosted vibe-coding hackathons for friends. And I write code and solve technical problems every single day. I guess my point is this: I’ve had my eye on the puck for a while now, and I am still shocked at the rate of progress.

What happens next? AI safety researchers warn about misalignment and existential risks. Tech workers worry about job losses and a shock to knowledge work. Accelerationists argue we should move even faster to maximize technological progress. And market bears say that AI capex debt is a house of cards that will eventually crash the economy. Meanwhile, most people’s view of AI is shaped by outdated chatbots. It’s a disorienting time.

As I’ve tried to make sense of the moment, I keep circling the same inchoate thoughts. Here are some of them. We are all fitting superlinear functions to a handful of data points. In Bayesian inference, your posterior distribution can widen (more information doesn’t always make you more certain). In a world in which specialized skill becomes commoditized, people with problems to solve will thrive relative to people who want to be told what to do. It is an insidious failure mode to solve problems you don’t actually care about simply because AI has made them easier to solve. I struggle to tell the difference between cognitive surrender and a higher level of abstraction. I derive a lot of existential pleasure from deep understanding, and I don’t want to lose that. One kid is using AI to do their math homework, while another is using AI to teach themselves calculus. The transistor isn’t even a hundred years old. I have never made a value-aligned decision from a place of fear. Do not mistake the shock or uncertainty of this moment for fatalism or powerlessness.

While my thoughts are incomplete, my conviction is not. I keep telling friends about this time I watched Hikaru Nakamura play bullet chess on his live stream. He was chatting casually to the audience while moving at lightning speed when he suddenly paused and burnt most of his remaining time deciding his next move. After he won, he talked about the importance of that move. I forget the details, but he had sensed a shift in the game, a pattern that felt off. He paused because the next move was critical. That is what this moment feels like to me.

So this is a time capsule for myself, a record of this feeling. Genuinely, I don’t know what happens next, but I can feel in my shoulder blades that the game has changed.