Trevor Chow
I am interested in online learning: taking imperfect information and make better decisions in real-time.
I work on one instance of this, browser agents, at Moonglow (YC S24), and write about two others, AI and financial markets, at Bunnyhopping (supported by EV).
My background is in mathematics at Stanford (esp. numerical methods), where I taught models to learn at inference time with Hazy Research Lab.
I’ve also worked on trading index options at Optiver, and building market making algorithms for the SPX complex that react to orderbook microstructure dynamics.
When I was younger, I was a King’s Scholar at Eton College and was active in the British competitive mathematics and debating circuits.
Beyond work, I split my time between ski slopes, skydiving dropzones and the Eras Tour. Say hi!
Essays
From Apples to Strawberries
Three Kuhnian Revolutions in ML Training
The Vanguard of the Revolution
This Time isn't Different
Storytelling in Kennywood Park
Line Goes Up

From Apples to Strawberries
ChatGPT was a big deal, but o1 is a bigger deal. It marks the start of the "search" paradigm for LLMs.
Three Kuhnian Revolutions in ML Training
The science of pretraining LLMs has faced three Kuhnian revolutions in the last 4 years. The dominant theme is the growing importance of data.
The Vanguard of the Revolution
Index funds are socialism, in the best and worst sense.
This Time isn't Different
The anatomy of the 2022 crypto crash mirrors the 2008 financial crisis, involving an unmet demand for safe assets and an accompanying shadow banking system.
Storytelling in Kennywood Park
Monetary policy works in practice, not in theory. We don't understand its exact mechanisms, even as we can see its effects.
Line Goes Up
Economic growth is the phenomenon that underwrites all other value, but it is fragile and not an inevitability.Research
Smoothie: Label Free Language Model Routing
What Causes Polysemanticity?
Explainable Detection of Online Sexism
Randomised Serial Dictatorship with Transfers
Transformative AI, Existential Risk, and Real Interest Rates

Smoothie: Label Free Language Model Routing
LLM routing scores can be learned without labelled data using a weak-supervision-inspired technique. Accepted to NeurIPS ‘24. Blog post here. Code here.
What Causes Polysemanticity?
Polsemanticity can arise incidentally, due to regularisation pressures and random noise during training. Accepted to BGPT & Re-Align @ ICLR ‘24. Blog post here. Code here.
Explainable Detection of Online Sexism
Fine-tuned BERT models can provide explainable classification of sexist text. Accepted to SemEval @ ACL ‘23.
Randomised Serial Dictatorship with Transfers
Adapting RSD to allow trading to occur Pareto-improves the allocation of indivisible goods, at the expense of strategy-proofness.