3Blue1Brown and Jane Street on Trading, Machine Learning and Culture

Grant Sanderson of 3Blue1Brown asks Jane Street participants what their culture is like and what the firm actually does. One participant describes conversations aimed at finding the truth rather than proving how smart someone is: people expect their views to change when they encounter new information.

In meetings, participants sometimes share their initial views before hearing everyone else and then compare their views afterward. The discussion tends to bring them closer together. The group jokes about a Bayesian habit of conversation and familiar Jane Street phrases such as ‘that’s not unreasonable.’

Another participant describes a sense of being one company: people walk over to other groups to brainstorm rather than always going through managers. Sanderson asks how that scales without chaos. The answer is that managers and project leads provide structure when a longer, multi-person project needs it, while informal conversations remain comparatively fluid.

The group offers its real-time compute auction as an example. Researchers value their jobs relative to other work and bid for a shared cluster. One participant says the system encourages collective responsibility: a colleague might notice a failing job and pause it, or discuss why another job’s utilization is poor.

The openness can have surprising consequences. A researcher once changed the bids on every job in the cluster by running a notebook command without arguments. The group says the problem was fixed and presents it as an example of building a system around current needs and adjusting it as experience exposes problems.

Sanderson asks for an explanation of Jane Street that lies between vague descriptions of quantitative finance and proprietary details. One participant calls market making a core business and compares it to a grocery store for stocks: the firm provides access and must offer prices customers consider fair.

An ETF can give an investor access to companies in a foreign market; pricing ETFs in relation to their underlying stocks is one example of the business. Another participant extends the grocery-store analogy: anticipating demand for pumpkins at Halloween is a prediction problem, just as a liquidity provider needs to anticipate demand and manage inventory.

The participants say people can misunderstand both the scale and novelty of Jane Street’s machine learning. Financial problems are noisy, and simply taking a model off the shelf and running it on GPUs does not generally work. They describe research into fundamental questions rather than only applying existing models.

One participant says the firm has seen significant success with machine learning in recent years and feels it is at the beginning of an S-curve. In that speaker’s view, ML is already the single largest driver of the firm’s systematic trading. Another says hires from banks or trading firms can be surprised by how much the environment resembles technology or academia.

Job titles can also mislead. A participant says the role called trader is often substantially a research role, and the business is more systematic than outsiders may assume. The group itself jokes that colleagues do not always know one another’s formal titles.

The participants describe a spectrum of roles: some developers work close to trading, while others build tools for developers, including code-review and release infrastructure or systems for managing petabytes of market data. Traders and researchers likewise overlap, from attention to today’s securities and trading systems to longer-cycle model training and research tooling.

Craig, who joined in 2005, recalls that logging data and training a linear model on a compute cluster once felt advanced. Some trading systems literally ran in Excel, and a person could sometimes see that two securities were out of line without elaborate statistics.

Today, he says, much of the modeling work involves machine learning, including neural networks and gradient-boosted trees. Techniques have changed substantially every five years or so, even as the culture has felt familiar and many early employees have stayed.

Asked why people stay, a participant emphasizes colleagues who are also genuine friends and who care about one another’s success. That support makes difficult work and entry into new areas feel less isolating. Another says staying has been a continual reassessment of learning, enjoyment, conversations and agency; the pace of learning has even increased over time.

Nitya says she expected to become a mathematician, but family and friends encouraged her to try a job outside teaching mathematics. She enjoyed a Jane Street internship, returned the next summer and then joined full-time. After a few years she took leave to earn a PhD before returning; she says that particular path was unusual but the firm accommodated individual circumstances.

Nitya recalls interview questions about probability and general problem solving. The onsite included games rather than the advanced mathematics she expected. She found it interesting in hindsight but stressful at the time, especially when recruiters rushed her out to catch a flight during her final interview; despite her fears, the outcome was positive.

Asked how hiring has changed, a participant says the main strategy remains finding smart, hardworking, reasonable and adaptable people who like learning. Machine learning skills are a plus, but interviewers seek general abilities and use unfamiliar puzzles as material for discussion rather than relying on knowledge of a particular course.

The group notes that teaching an employee a new field makes more sense when people stay for a long time. But evaluating someone’s future effectiveness from a few hour-long interviews is difficult. The approach of finding adaptable generalists has helped the firm respond as its technical needs change, without eliminating the difficulty of prediction.

Alok says he applied for a research internship in summer 2018 while studying mathematics and computer science. He enjoyed both but assumed his future job would mostly use computer science, with mathematics as a possible bonus. He saw machine learning and finance as two areas that might use both.

During his internship, Alok found that the work was more like data science elsewhere than he had expected. Abstract ideas from college classes had concrete uses in analyzing models, which surprised and pleased him.

As an example, Alok describes studying M-estimators and the broader family of robust estimators in a statistics class. What seemed like abstract asymptotic theory proved useful when dealing with malformed data. During the internship he tried applying that theory to a concrete modeling problem.

Craig explains that his own path was unusual. Planning to start a PhD, he spotted a job posting on the OCaml mailing list and sent in a practice résumé. After one phone interview he visited New York, liked the people he met and accepted the job in 2005.

Craig had used OCaml for personal projects such as drawing fractals and building an Othello bot. He describes these as fun experiments rather than serious professional projects. Sanderson then asks the group what they do for fun.

Alok describes occasional informal math and physics talks in a side room. A recent discussion moved from statistical mechanics and partition functions to their close relationship with Feynman diagrams, and eventually to the playful idea of an exponential of graphs. Sanderson jokes that it sounds like a 3Blue1Brown video and thanks the group for clarifying the firm’s work and culture.