Jane Street engineering blog

Benjamin Devlin and Anish Singhani reveal how their ASIC puzzle implements an 11x11 Star Battle checker and explain how solvers reverse-engineered it from GDS layout. The post covers netlist extraction, simulation, SAT solving, LFSR-obfuscated outputs, debugging flawed models, and verification techniques.

Adam Yi traces six interacting bugs exposed by recurring network partitions in Jane Street’s Kafka infrastructure, spanning glibc DNS resolution, OCaml networking, Async timeouts, retry cancellation, socket leaks, and file-descriptor limits. The post demonstrates how failure injection, resource accounting, and quantitative predictions connected the incidents and guided fixes.

Jane Street details how it built an OCaml-native FUSE library and read-only server to expose its Depot object store through a POSIX filesystem interface. The post covers protocol testing, object-store-to-filesystem semantic mapping, and performance optimizations that increased throughput from 30 MiB/s to 1.3 GiB/s.

Jane Street describes a resilient Linux kernel-log reporter that continues operating when local storage, networking, or normal crash-dump mechanisms fail. The design uses /dev/kmsg, memory locking, cached DNS, UDP replication, EFI-backed pstore, and post-boot recovery to preserve and diagnose otherwise lost crashes.

James Somers details three Jane Street intern projects: matching voice trades to anonymous public prints, implementing exchange-specific primary-peg order types, and reconciling global ETF holiday calendars. The projects illustrate how engineers translate complex domain rules into tested OCaml, FIX, and internal workflow tools.

Jane Street explores autoregressive diffusion for generating realistic market-data events, including order-book updates, trades, prices, and event timing. The post compares DDPM with flow matching and explains how categorical modeling and atom smoothing address the discrete discontinuities that make financial data difficult to synthesize.

Jane Street surveys interns’ 2026 projects across ML research, software engineering, Linux, IT, and trading operations. The roundup highlights practical techniques and measured results, including activation checkpointing, Aria indexing that reduced CPU use by 30%, resilient kernel logging, and a 1.3 GiB/s OCaml FUSE server.

Jane Street investigates how LLMs rely on memorized pretraining data when predicting cricket match outcomes. The post evaluates divergence decoding, distillation, prompt rewriting, natural-language autoencoders, and model-difference probing, finding that divergence decoding and podcast-style rewrites can reduce memorization without harming out-of-sample performance.

Jane Street describes an activation-checkpointing planner for PyTorch training that trades recomputation for lower peak memory. The approach models full-graph memory lifetimes and recomputation costs, then greedily removes saves under an absolute memory budget, outperforming existing policies across tested models.

Joanna Lu highlights several 2026 Strategy and Product intern projects at Jane Street, including financing-cost allocation, risk-limit lifecycle management, automated Linux fleet upgrades, AI-agent workflow improvements, and alerting for late reference data. The examples show how interns combined system analysis, data, and cross-team collaboration to improve operational tooling.

Alex Renda and Nitya Mani study how sequence weighting affects language-model training across model families ranging from millions to hundreds of billions of parameters. Their experiments reveal a non-monotonic pattern: weight sensitivity rises at medium scales before falling at larger scales, challenging straightforward scaling-law extrapolation for data mixing.

Jane Street announces an open-source ASIC competition to design a reprogrammable protocol emulator supporting UART, SPI, and I2C. The post outlines the Tiny Tapeout CMOS5L constraints, area and timing considerations, FPGA prototyping, verification approaches, submission deadline, and fabrication prizes.

The post introduces a Jane Street ASIC reverse-engineering puzzle, providing only the chip layout, sample inputs and outputs, and a repository of files for participants to analyze. It explains the basics of how chips are synthesized, placed, routed, and fabricated, then outlines the steps required to recover a netlist, determine the circuit’s purpose, and derive the final string answer from the chip’s behavior.

Jane Street is known for being an OCaml shop, but for years now Python has been our second major programming language, acting as the primary...

I’ve been telling people for the last 25 years that Jane Street as an organization was just not interested in formal methods.

We’ve always found strace useful but somewhat hard to work with, especially because its output can be inscrutable and subprocesses or threads are difficult to follow. This post appears to discuss tooling and the renewed interest in terminal user interfaces.

Attention is a computational primitive at the core of modern language models, allowing internal representations to reference and influence each other. The post explores positional encodings through the lens of group theory and how that mathematical structure helps understand attention.

Many CTF-style ML puzzles give you a black-box neural net and ask you to determine what it does. This post walks through approaches to reverse-engineering such networks and illustrates techniques for understanding their behavior.

Jane Street ran a challenge alongside Advent of Code focused on FPGA-related problems. This post presents the results and reflections from that Advent of FPGA Challenge.

The author describes a shift from using Figma to relying on the Claude LLM for parts of the design process after earlier skepticism about LLMs. The post recounts initial disappointment with models like Copilot and explains how recent interactions changed their workflow and tooling choices.