jax.random.key(42)

August 31st is my last day at Google. Before I get to what comes next, the exit interview nobody asked for.

Top 10 most impactful things I did during the last eight years

10. Scaled content moderation forecasting from a handful of time series to a couple of thousands, running in production with 0 SRE.

9. Mentored people inside and outside of Google, and co-supervised a PhD student in cosmology. Very rewarding experience.

8. Did not use Google’s internal Year in Search data to clear $1.2M on Polymarket.

7. Did not personally take down half the internet with a null pointer.

6. Combined nowcasting with forecasting and built a dynamic model selection system that accounts for data drift, because our stakeholders love to retrospectively change their historical metrics.

5. Was the data science go to person for Trust and Safety business planning these last few years. Honestly, sick of arguing about million dollar questions.

4. Got my reports promoted from L4 to L5, all on the first try. Three for three.

3. Co-wrote a book on Bayesian modeling and computation. People bought it. Some even read it.

2. Gave a healthy stack of Bayesian modeling and time series talks, internally and at conferences.

1. Did not force a single frequentist to become Bayesian, including the ones who reported to me.*

* In fairness, I built zero Bayesian models and looked at zero posterior samples at work in the last two years.

I hope you read that in Letterman Top 10 voice.

Wait, why is Jeff Dean also following what I am doing?

In The Hitchhiker’s Guide to the Galaxy, a computer called Deep Thought is built to work out the Answer to Life, the Universe, and Everything. It returns 42. The problem, which takes the rest of the book, is that nobody ever knew what the question was.

That gag has aged strangely well, though the shape is inverted. Deep Thought spent seven and a half million years to produce two characters. Ask a LLM almost anything and two thousand words come back before you finish typing, fluent, fast and cheap. What has not gotten cheaper in either case is knowing which question was worth asking, and whether the answer you got is warranted.

That is the part I want to work on. In September I am joining my best friend Josh and my new best friend Dan at rekursiv.ai as Founding Scientist and Chief Epistemologist. The mission is to autonomously create knowledge, using AI, LLMs, agents, whichever word you prefer (interchangeable today, probably not for much longer).

My pet theory about LLMs is that they are excellent (meta) compression algorithms. The compression half is obvious, they are very good at summarizing. The decompression half goes beyond the traditional sense, because it is active: you give it a research question and it does not only unpack what it already holds, it goes and gathers, pulling in fresh information from web search and papers, and running its own experiments. Then it hands you end to end reasoning. What makes the decompression almost magical is when your question connects two concepts that look unrelated. Sometimes you add the right hint or heuristic as a follow up, and it draws a path between two points sitting far apart in whatever manifold it learned. That is when the thing unfolds into a coherent picture with real sparks in it. A recent example is Terence Tao digesting the Jacobian Conjecture counterexample, with a chatbot to talk it through and confirm the calculations.

So what is missing? Having the model ask that starting question. Having it produce the smart follow up. Having it go past (de)compression into something that actually discovers.

Josh and Dan have been calling this the next frontier since late last year, and I see it more clearly every day I work with agents. Then Jeff Dean and Sanjay Ghemawat left Google to do it too, which I choose to read as confirmation rather than competition.

Five minutes early

The part people forget about that joke: in the book, the Earth is the computer. It was built to work out the question, it ran for ten million years, and the Vogons demolished it five minutes before it finished.

I have a much worse computer and considerably less than ten million years. Starting in September.