3Blue1Brown Talks Machine Learning with Jane Street

Cultural Norms and Communication

• The company prioritizes "changing your mind" over defending ego; conversations are framed as collaborative truth-seeking rather than debate.
• Employees frequently adopt "Bayesian habits," explicitly updating beliefs based on new information during discussions.
• A specific lexicon exists, including phrases like "not unreasonable," which signals openness to new ideas rather than dismissal.
• Cross-group communication is highly organic; employees often approach colleagues in different departments directly without formal managerial approval to "spitball" ideas.
• While the environment is described as "chaotic" by outside standards, formal hierarchy and team leads emerge only for large, month-long projects.

Operational Structure and Compute Allocation

• Compute clusters operate on a real-time auction system where any employee can bid for resources, contrasting with the bureaucratic allocation processes of other firms.
• This auction model fosters a sense of collective responsibility; employees voluntarily monitor and pause erroneous jobs running on the cluster.
• The system is globally read-write, allowing any user to "nuke" (terminate) another user's job, a feature that enabled rapid resolution of issues without reporting up the chain.
• A past incident involved a researcher accidentally running a command without arguments in Jupyter, which altered the bid bits for every job running on the "Hive" cluster, necessitating a system fix.

Business Definition and Market Role

• Jane Street describes itself as a "generalist financial services liquidity provider," acting effectively as a "grocery store for stocks."
• The firm creates liquidity for instruments like ETFs to allow investors to access foreign markets (e.g., Chinese stocks via US-based ETFs) efficiently.
• Pricing logic is comparable to retail: setting fair prices based on supply and demand anticipation rather than arbitrary markups.
• The firm views its core function as a prediction problem, utilizing machine learning to anticipate inventory needs (supply/demand) to maintain liquidity.

Misconceptions About the Work

• Outsiders often underestimate the novelty of Jane Street's machine learning, assuming the firm primarily uses off-the-shelf models on purchased GPUs.
• In reality, the firm engages in "basic science" and works on the frontier of research, tackling problems too complex for standard models in financial domains.
• The term "Trader" is often a historical misnomer; these roles are largely research-focused and heavily systematic, involving minimal manual order entry.
• The organization is significantly more tech- and academia-oriented than traditional trading firms or banks, challenging the assumptions of industry transferees.

Evolution of Technical Work (20-Year Span)

• Twenty years ago, trading systems relied on linear models, Excel spreadsheets, and heuristic logic based on relative pricing of two securities.
• Current modeling is dominated by machine learning, specifically neural networks, Gradient Boosted Trees (GBTs), and advanced statistics.
• Technical methodologies have shifted significantly every five years, yet the underlying culture of curiosity and problem-solving has remained stable.
• The firm has invested heavily in machine learning as the single largest driver of its systematic trading capabilities.

Talent Acquisition and Retention

• The hiring strategy prioritizes generalists who are "smart, hardworking, reasonable, and adaptable" over candidates with specific technical training.
• Interviews focus on math puzzles and problem-solving scenarios designed to test learning potential rather than specific domain knowledge.
• Employees report high retention because colleagues act as genuine friends who are deeply invested in each other's professional success.
• The culture of "high agency" and low friction allows employees to re-evaluate their career fit daily, resulting in a self-selecting population of long-term employees.
• Flexible career paths are supported, including instances of employees taking leave to complete PhDs and returning to the firm.

Historical Context and Founder Trajectories

• One long-tenured employee joined in 2005 after seeing a job posting on the "Camel mailing list" for the obscure functional programming language OCaml.
• That same employee initially planned to start a PhD but chose the job after meeting the people and finding the work engaging.
• Other hires come from diverse academic backgrounds, with some initially expecting a focus on "galaxy brain" models only to find practical applications for abstract mathematical theories (e.g., Huber's M estimators).
• The interview process has evolved from single phone screens to complex on-site interactions, yet the core goal of identifying adaptable learners remains unchanged.
• Internal social and intellectual life includes ad-hoc talks on physics and mathematics, such as system mechanics and Feynman diagrams, often leading to deep theoretical discussions.