Blog · November 19, 2026 Research

Code like a millionaire: what Anthropic's early hires built before they joined

CW
Cam Whiteside · 20 min read

Anthropic's IPO could make about 4,250 employees millionaires, by one estimate. Here are the problems its early hires were solving in public before they joined, and what they built.

Code like a millionaire: what Anthropic's early hires built before they joined

Anthropic priced its IPO at $83 a share on November 18, and the shares start trading on Nasdaq today (Anthropic, CNBC). Hill, a platform that tracks private-market trades, estimates the listing could make about 4,250 of Anthropic's roughly 5,300 employees millionaires (Money.it). That is an estimate about the company as a whole, not about any one person.

So we asked a different question about the people who got there early, in 2022, 2023 and 2024: what problems were they trying to solve before they joined? Their public GitHub work answers it pretty well. Some of them were mapping the wiring of brains or simulating stars. Some kept a tool alive that millions of programs run every week. A lot of them built games, gadgets and things nobody asked for.

Vamo is built on the idea that what someone builds in public is the best evidence of what they will build next, so this was a fair test. Here is the short version.

were building in public in the year before they joined
67%
of those worked on someone else's project
3 in 4
wrote no machine learning code in the two years before
55%
had a side project nobody assigned them
55%

Six things worth building, if you want to try this

If you're wondering what to build, here are six patterns from the data, each with the number behind it and one real project. They describe one group's public record, not a recipe for getting hired: a third of the group had no public activity at all in the year before they joined, and nobody in it has said a repository got them the offer. The stories behind each one follow.

  1. Build a tool other builders use

    Tools for programmers were the most common side project, at 16% of the active.

    A converter for API descriptions, now 6.5 million downloads a week: bcherny/json-schema-to-typescript  by Boris Cherny

  2. Fix something in someone else's project, then come back

    3 in 4 worked on projects they didn't own, and returners outnumbered drive-bys three to one.

    Some of the first changes to the main library for looking inside language models: TransformerLensOrg/TransformerLens  by Ansh Radhakrishnan

  3. Explain a hard idea with a visual

    11% worked on teaching, docs or visual explainers.

    Interactive essays about how models behave, then Anthropic's feature browser: PAIR-code/ai-explorables  by Adam Pearce

  4. Bring the science you already know

    22% had science code, and 63% of the interpretability paper authors in the group did.

    Brain-mapping methods, pointed at a language model: nicholasturner1/gpt-omics  by Nicholas Turner

  5. Scratch your own itch

    10% built tools for their own day: reading, recipes, workouts, notes.

    Typing shortcuts mined from your own emails and Slack messages: eschluntz/compress  by Erik Schluntz

  6. Finish the weird side quest

    55% had a personal project, and games and puzzles came second only to tools.

    A 1970s game emulator, finished five years after it was started: kevinji/chip-8-emulator  by Kevin Ji

Shares are of people active on GitHub in the year before joining, over the two years before their start year.

Three in four worked on other people's projects

Start with the basics. Two in three of the people we could match were building in public in the year before they joined. Everything below is about them.

67%:made public contributions on GitHub in the year before they joined

The rest of the numbers here are about those people.

Most of that work happened in projects they didn't own. Three in four committed code or proposed a change to someone else's project in the two years before they joined, and most got a change accepted. Nearly half worked on a project popular enough to have 1,000 or more stars on GitHub.

3 in 4:worked on a project they didn't own in the two years before joining

  • Committed or opened a pull request 75%
  • Had a change accepted 62%
  • Worked on a project with 1,000+ stars 45%
  • Worked on a project with 10,000+ stars 30%

Two stories show the range. In 2022 a small library called TransformerLens appeared that let researchers open up a language model and read what was going on inside. Ansh Radhakrishnan sent some of its very first changes, the ones that load a trained model and record what it does, and two other future Anthropic hires followed within a year.1 At the other end of the range, Justin Spahr-Summers and, years later, another future hire helped keep Tigerbrew, a version of Homebrew, running on Macs from 2005.2

  • TransformerLensOrg/TransformerLens

    The problem

    Researchers who wanted to look inside a language model had to write their own plumbing just to read what it was doing.

    What it does

    TransformerLens is the shared library that loads a GPT-style model and exposes everything happening inside it, so a researcher can start experimenting in a few lines.

    Who uses it

    4,000+GitHub stars

    It became a common starting point for this kind of research.

    Ansh Radhakrishnanone of three early contributors who joined Anthropic; there by 2023

  • mistydemeo/tigerbrew

    The problem

    Apple stopped supporting its 2005 PowerPC Macs, but people still use them and still want new software.

    What it does

    Tigerbrew is a version of Homebrew, the popular Mac app installer, that still works on those machines.

    Who uses it

    670+GitHub stars

    A small, loyal crowd of vintage-Mac users.

    Justin Spahr-Summersone of two contributors who later worked at Anthropic

They stuck around: returners outnumbered drive-bys three to one

A one-line typo fix counts as a contribution. That isn't what most of this group did. 29% were maintainers or core contributors on a project with real users, and another 28% kept coming back to the same outside project. Only 19% passed through once and never returned.

57%:kept coming back to the projects they worked on

Maintainers and repeat contributors outnumbered drive-by contributors three to one.

Maintaining is the unglamorous part of open source: answering issues, reviewing other people's changes, shipping releases for years. Two of the longest-running examples in the group show what it looks like at scale. Ned Batchelder has maintained coverage.py, the tool most Python projects use to see which lines their tests miss, since 2004. It now gets about 82 million downloads a week, and he kept shipping it right up to his move to Anthropic in 2024; he worked there until 2025.3 Justin Spahr-Summers was the top committer to ReactiveCocoa and Mantle, two of the most popular libraries for Apple developers in the 2010s, then made the very first commit to Anthropic's TypeScript SDK in early 2023. He worked at Anthropic and has since left to make games.4

  • coveragepy/coveragepy

    The problem

    You write tests to catch bugs. But which lines of your program do the tests never run?

    What it does

    coverage.py measures exactly that for Python: run your tests and it shows every line they missed. One person has maintained it since 2004.

    Who uses it

    82 milliondownloads a week

    GitHub counts more than 515,000 repositories that depend on it.

    Ned Batcheldermaintainer; worked at Anthropic, 2024 to 2025

  • ReactiveCocoa/ReactiveCocoa

    The problem

    An iPhone app reacts to taps, network replies and timers all at once, and wiring those together by hand gets tangled fast.

    What it does

    ReactiveCocoa treats each of those as a stream you can combine and transform. Its sister project Mantle turns web data into app objects.

    Who uses it

    31,000GitHub stars, the two together

    Two of the most popular open-source libraries for Apple developers in the 2010s.

    Justin Spahr-Summerstop committer; worked at Anthropic from 2023

Most built small, useful tools, and a few got huge

Most of what these people built on their own was small. Three in four of their side projects have fewer than 10 stars, and only one in ten reached 100. The big, famous projects they touched usually belonged to someone else.

76%:of their own side projects have fewer than 10 stars

Most of what they built was small. The big projects they touched usually belonged to someone else.

  • Worked on someone else's project with 1,000+ stars 45%
  • Own side projects with no stars at all 48%
  • Own side projects with 100+ stars 10%

A few of their own did get huge, and each one started with a specific annoyance.

The biggest is a converter. Web services describe the shape of their data in a format called JSON Schema, and programmers using TypeScript had to retype every one of those descriptions by hand. In 2016 Boris Cherny wrote json-schema-to-typescript to do it for them, and he kept maintaining it for eight years. It now gets about 6.5 million downloads a week, and more than 32,000 projects on GitHub use it. In September 2024 he joined Anthropic and went on to build Claude Code, which is also written in TypeScript.5

Another is the math under modern AI. Training a model means nudging millions of numbers in whatever direction makes it a little less wrong, and finding that direction is calculus nobody wants to do by hand. Autograd, from a Harvard lab, does it automatically for ordinary Python code, and Google's JAX, now one of the main tools for training large models, grew out of it. Its co-author David Duvenaud worked at Anthropic from 2023 to 2024, leading a team that tested models for safety.6

And some stayed small and charming. Erik Schluntz, a robotics startup co-founder, fed a pile of his own emails and Slack messages into compress, which finds the phrases you type most and suggests shortcuts for them. "Just reading through the results was amusing to see how repetitive some of my writing is," he wrote in the README. He joined Anthropic by 2024.7

  • bcherny/json-schema-to-typescript

    The problem

    Web APIs describe their data in JSON Schema. TypeScript programmers had to copy each description into types by hand, and the two drifted apart.

    What it does

    json-schema-to-typescript reads the schema and writes the types for you. Its author kept maintaining it for eight years.

    Who uses it

    6.5 millionnpm downloads a week

    More than 32,000 repositories on GitHub use it.

    Boris Chernyjoined Anthropic in 2024 and built Claude Code

  • HIPS/autograd

    The problem

    Training a model means nudging millions of numbers in the direction that lowers its error. Working out that direction by hand, for every new model, is slow and easy to get wrong.

    What it does

    Autograd works it out automatically for ordinary Python code. Google's JAX, now one of the main tools for training large models, grew out of it.

    Who uses it

    1 milliondownloads a week

    Autograd still gets about a million downloads a week, and JAX has more than 36,000 stars.

    David Duvenaudco-author; worked at Anthropic, 2023 to 2024

  • eschluntz/compress

    The problem

    You type the same phrases all day, in email and in Slack.

    What it does

    compress reads a pile of your own writing, finds the phrases you type most and suggests keyboard shortcuts for them, like "the" to "t".

    Who uses it

    1,400+GitHub stars

    A small tool that more than a thousand people starred.

    Erik Schluntzjoined Anthropic by 2024

More than half wrote no machine learning code

You might expect the people building Claude to have spent the years before writing AI code. Most didn't. 55% touched no machine learning code at all in the two years before they joined, and three in four of the repositories they owned had nothing to do with it. Anthropic says something similar on its careers page: "About half our technical staff had no prior ML experience" (anthropic.com/careers).

55%:wrote no machine learning code in the two years before joining

Anthropic's careers page says about half its technical staff had no prior ML experience.

So what were they working on? Mostly the problems every software team has: websites and apps, servers and data pipelines, tools for other programmers, science. Machine learning libraries were as common as web frameworks, and the research Anthropic is best known for, looking inside models, was among the rarest.

The kinds of problems they worked on, in other people's projects

Share of people active before joining who contributed code to each kind of project in the two years before. One person can count in several.

Categories are ours, from an organization and repository map, keywords and languages, with hand labels for the long tail.

Show the numbers
Kind of outside projectShare of the activeLeaving out their employer's repos
Web, browsers and frontend26%23%
ML frameworks and model libraries26%23%
Infrastructure, cloud and data22%19%
Scientific computing and research code17%17%
Developer tools, editors and languages17%16%
Teaching, docs and explainers11%11%
Python language and tooling10%9%
LLM apps, SDKs and agents9%9%
Interpretability, evals and alignment8%7%
Games, hardware and hobby projects7%7%
Security, cryptography and blockchain6%5%
Rust ecosystem5%5%

One path in started with charts. Adam Pearce made graphics at Bloomberg and dozens of interactives at The New York Times, from a tax calculator to a map of where refugees come from. Then at Google he made AI Explorables, interactive essays that let anyone poke at a model and watch what it does. His reason, from his own talk notes: "Well-crafted visualizations are the highest bandwidth way of downloading information into our brains." By May 2024 he was at Anthropic, building the interactive feature browser for one of its best-known research papers. hot-server, a tiny live-reload server he wrote for himself in 2016, is still how Anthropic's circuit-tracing website starts up.8

Most came from product companies, not AI labs

Only 9% came straight from another AI lab. Half came from late-stage tech companies like Stripe and Asana, or from startups. Stripe sent more people straight to Anthropic than any other single company, with Google close behind.9

Most paths into Anthropic started at product companies

Kind of job right before Anthropic, share of people with work history. Band thickness is the share of people.

Buckets are hand-mapped. Late-stage tech means large private or recently public product companies; big tech means the largest public ones.

Show the numbers
Kind of job right before AnthropicShareCompanies in it
Late-stage tech29%Stripe, Asana, Shopify, Robinhood, Airbnb, Linear, Waymo, Retool, Airtable, Sentry, Loom, Scale AI, Skydio
Startups21%Cobalt AI, AssemblyAI, Playground Education, The Browser Company, Filmhub, Monograph, Standard AI, Artifact.
Big tech18%Google (outside DeepMind), Meta, Apple, Microsoft, TikTok, Netflix, Amazon, Twitter
Academia11%Stanford, MIT, Toronto, Cambridge, Oxford, Cornell, RIKEN, the Kavli Institute for Theoretical Physics
Another AI lab9%Google DeepMind, OpenAI, EleutherAI, Adept, Contextual AI, Covariant
Independent, other7%Freelance and independent work, trading firms, CPython, government and nonprofit groups
VC, fellowship, EIR3%Sutter Hill Ventures, O'Shaughnessy Ventures, Anthos Capital, AI House, Y Combinator, Thrive Capital
AI safety orgs2%Redwood Research, METR, MIRI, the UK government's AI work

The mix shifted by year. The 2022 hires came most often from late-stage tech companies and universities. By 2023 and 2024, startups and big tech had caught up.

Startups and big tech grew as a path in after 2022

Kind of job right before Anthropic, by start year, share of people with work history who started that year.

2022 hires are the smallest of the three groups, so read their bars loosely.

Show the numbers
Job right before Anthropic2022 hires2023 hires2024 hires
Late-stage tech40%28%29%
Startups7%21%22%
Big tech7%17%19%
Academia20%11%10%
Another AI lab7%6%10%

2023 brought the scientists, 2024 the product builders

Split by start year, the group looks like two different hiring waves. The 2023 hires were the busiest in public and the most scientific: 37% had science code, twice the 2024 share. The 2024 hires look more like product engineers, and 57% of them touched no machine learning code before joining. Working on other people's projects held steady across all three years.

2023 hires were the most scientific; 2024 hires the least likely to touch ML

By Anthropic start year. The first row is a share of everyone who started that year; the rest are shares of people active before joining, in the two years before.

2022 hires are the smallest of the three groups, so read their bars loosely.

Show the numbers
In the two years before joining2022 hires2023 hires2024 hires
Active on GitHub in the year before75%72%64%
Contributed code to someone else's project83%76%74%
Touched no ML code42%55%57%
Owned a personal project58%61%52%
Science code8%37%18%
Interpretability or alignment code0%16%11%

The scientists brought stars, brains and viruses

One in five had science code before joining, and the fields were far from AI: stars, brain tissue, the spread of viruses, medical physics. For the people who later wrote Anthropic's papers on reading a model's mind, science was the most common thread of all.

22%:had science code: stars, brains, viruses and the like

Among the authors of Anthropic's two Monosemanticity papers in our group, it was 63%.

  • Everyone active before joining 22%
  • 2023 hires 37%
  • Monosemanticity authors 63%

Those papers need a little setup. A language model's neurons don't line up neatly with ideas: in one small model Anthropic studied, "a single neuron responds to a mixture of academic citations, English dialogue, HTTP requests, and Korean text." In 2023 and 2024 Anthropic published two papers, Towards Monosemanticity and Scaling Monosemanticity, that pulled those mixtures apart into features a person can name, like DNA, Arabic script or the Golden Gate Bridge. Here is what some of their authors were doing first.

A physicist who studied stars. Stars live for billions of years, so astronomers simulate them instead of watching. Adam Jermyn worked on MESA, the open-source code astronomers around the world use to do that, on the part that models how heat churns inside a star. In October 2022 he built the smallest model he could that still showed the effect he was after: a two-layer toy network "primarily meant for interpretability experiments." A year later he was a core contributor on the first paper.10

A neuroscientist who mapped brain wiring. To understand a brain, scientists slice tissue thinner than a hair, image it and trace every connection between neurons, far too many to do by hand. Nicholas Turner built Synaptor at Princeton to find those connections automatically. In April 2022 he pointed the same idea at a language model, in gpt-omics: "Analyzing the virtual weight connectomes in GPT-like transformers." He was an author on both papers.11

A brain-mapping researcher's year before joining Anthropic

Nicholas Turner's public GitHub contributions in 2022, day by day. Most went into the project that treats a language model like brain tissue.

A mathematician who chased viruses. Joshua Batson describes himself as a "Mathematician working on machine learning, computational biology, and infectious disease," and in 2022 kept a small home project in genomic epidemiology, which tracks how a disease spreads by reading its genomes.12

And Wes Gurnee, still in a PhD at MIT, asked whether a language model just memorizes word patterns or builds something like a map. His world-models code found that open models keep track of where places are and when events happened, down to individual "space neurons" and "time neurons." He now works on Anthropic's interpretability team: "I study how Claude's thoughts are organized so that we can read and steer them."13

  • MESAHub/mesa

    The problem

    A star lives for billions of years, so nobody can watch one. Astronomers simulate them instead.

    What it does

    MESA is the free, open-source code astronomers around the world use to model how stars are born, burn and die. Adam Jermyn worked on how heat churns inside them.

    Who uses it

    Built and used by a global community of astronomers.

    Adam Jermyncontributor; joined Anthropic by 2023

  • nicholasturner1/Synaptor

    The problem

    To understand a brain, scientists slice tissue thinner than a hair, image it and trace every connection. There are far too many to trace by hand.

    What it does

    Synaptor finds the connections between neurons automatically, at scale. In 2022 its author pointed the same idea at a language model, in a project called gpt-omics.

    Who uses it

    Used in brain-mapping work at Princeton's Seung Lab.

    Nicholas Turnerjoined Anthropic by 2023

  • wesg52/world-models

    The problem

    Does a language model just memorize word patterns, or does it build something like a map of the world?

    What it does

    world-models probes open models for where places are and when events happened, and finds them, down to single "space neurons" and "time neurons".

    Who uses it

    260+GitHub stars

    The code behind a paper at ICLR 2024, a top AI conference.

    Wes Gurneejoined Anthropic by 2024

The Monosemanticity authors were nearly three times as likely to have science code

Authors of Towards Monosemanticity or Scaling Monosemanticity who are in our group, against everyone in it. The first row is a share of all people; the rest are shares of people active before joining, in the two years before.

Show the numbers
In the two years before joiningMonosemanticity authorsEveryone
Active on GitHub in the year before89%67%
Contributed code to someone else's project88%75%
Science code63%22%
Owned a personal project63%55%
Touched any ML code38%45%
Interpretability or alignment code25%11%

Our reading: the interpretability team hired scientists who had already built small models to see inside something, and people who could make the result visible. That is our inference, not something Anthropic has said.14

Nobody assigned the best part: games, gadgets and Advent of Code

Boris Cherny has said he looks for "side quests" when he hires: "cool weekend projects," like someone "really into making kombucha" (Business Insider via AOL). That is how he hires now, not how this group was hired. But the group had plenty of side quests.

55%:had a side project nobody assigned them

49% had one with nothing to do with machine learning or science.

We didn't find any kombucha on GitHub. We did find a bike that plays Mario Kart, an ad blocker for anything with an HDMI plug, Ziyad Edher's homemade LED wall, a tool that dates an audio recording from the hum of the power grid behind it, and a lot of Advent of Code.

  • Games A 1970s game machine, in the browser A CHIP-8 emulator, started to learn Rust in 2018 and finished in 2023. Pong flickers, as it should. kevinji/chip-8-emulator Kevin Ji2018 to 2023
  • Hardware A bike that plays Mario Kart Arduino code that turns a bike into a Nintendo 64 controller.
  • Music Pull a song apart in Ableton Split any track into vocals, drums and bass inside a music app, with an open-source model. 600+ stars
  • Hardware A 64 by 64 LED wall Type a message on a web page and it scrolls across a homemade LED matrix, written in Rust. ziyadedher/led Ziyad Edher2023 to today
  • Games Routes through EVE Online A library for finding your way across the space game's star map. dsp/neweden David Soria Parra2019 to 2023
  • Life tools Hide the retweets A Chrome extension that clears retweets and threads out of a Twitter feed. alexalbertt/twitter-hider-extension Alex Albert2022
  • Games Advent of Code, every December About a third of the game repos in the group are answers to the yearly puzzle calendar.
  • Hardware An ad blocker for HDMI An ad blocker for anything with an HDMI plug.
  • Life tools A Kindle reading tracker For someone who also wrote a tool to turn Google Docs into Markdown. 1wheel/kindle-tracker Adam Pearce2017 to 2021
  • Science hobby Date a recording by its hum The power grid's hum drifts over time, so the background buzz on a recording says when it was made.
  • Games Riichi mahjong for the PICO-8 A full Japanese mahjong game for a tiny fantasy game console.
  • Hardware Program a hobby servo An Arduino library for setting up the small motors in model planes and robots. timmaxw/HitecDServo Tim Maxwell2022
  • Life tools Tides for one stretch of coast Plus a live surf poster with dials for the wave conditions.
  • Games A doctor for Game Boy emulators Compares your emulator against a known-good one, step by step, to find the bug. 290+ stars
  • Science hobby "Convection, in a moose" A stellar convection simulation, named as described.
  • Hardware A LEGO CNC machine A computer-controlled cutter, built from LEGO.

Games and puzzles were the second most common kind of side project, right behind tools for other programmers, and about a third of the game repos are answers to Advent of Code, the puzzle calendar programmers race through every December. Kevin Ji started a 1970s game emulator in 2018 because he wanted to learn Rust and was "specifically excited about the ability to compile to WebAssembly," which lets it run in a browser. He came back and finished it in July 2023. A year later he was at Anthropic.15

An emulator builder's year before joining Anthropic

Kevin Ji's public GitHub contributions in 2023, day by day: the emulator, a compression library written from scratch, and fixes to a large open-source blockchain.

The life tools are the most personal: Adam Pearce's Kindle reading tracker, Ziyad Edher's to-do app for the terminal, Alex Albert's Chrome extension that hides retweets, plus recipe managers, a random workout generator and a tide tracker for one stretch of coast. Hobby code was also where people tried new languages. Rust showed up in their side projects five times as often as in their work.16

Tools, games and dotfiles came first

Kind of personal project in the two years before the start year, share of people active before joining. One person can have several.

Categories are ours, one per repository, labelled by hand. Personal sites, CVs and scratch repos are left out.

Show the numbers
Personal project categoryShare of the active
Developer tools and libraries16%
Games, puzzles and emulators15%
Dotfiles and editor configs15%
Teaching and study notes11%
Life tools: reading, recipes, workouts, notes10%
Web apps and sites9%
LLM apps and agents7%
ML side experiments7%
Art, music and photo6%
Systems and performance6%
Data analysis5%
Science hobbies5%
Security and privacy4%
Hardware and home4%
ML libraries3%
Interpretability experiments3%
Finance, trading, crypto2%

Some side quests turned into the job

The most interesting thread in the data runs from the side project to the job. For a handful of people, the thing they built for fun in 2023 was close to what they did at Anthropic a year later. People who had built apps on language models in their spare time went on to write Anthropic's public code at about four times the rate of everyone else.

4x:People who built LLM apps on the side were about four times as likely to write Anthropic's public code

  • Built LLM apps or agents before joining 27%
  • Everyone else 7%

Zack Lee spent early 2023 publishing starter kits for apps built on language models, like create-t3-turbo-ai. Two years earlier he had written about how he landed a job: "If you really want to work for someone, don't ask for permission or submit a resume first. Send them a demonstration of what you can do for them." At Anthropic, by 2024, he worked on Anthropic's own starter apps.17

Barry Zhang spent 2023 building agent side projects, plus a joke repo about stacking agent frameworks: "For the memes until someone actually creates this." In December 2024 he and Erik Schluntz wrote Anthropic's essay Building effective agents, which argues "the most successful implementations weren't using complex frameworks." David Hershey built ScrapeGPT, a web scraper that "Self-heals to debug itself," in 2023, joined Anthropic, and in 2025 set Claude loose on Pokémon Red, live on Twitch.18

  • zckly/create-t3-turbo-ai

    The problem

    In early 2023, getting from nothing to a working app on a language model took a weekend of plumbing.

    What it does

    create-t3-turbo-ai is a starter kit that does the plumbing for you. Its author shipped two more that spring: one for long-running AI workflows, one that generates user interfaces.

    Who uses it

    700+stars across the three

    Then, at Anthropic, he worked on Anthropic's own starter apps.

    Zack Leejoined Anthropic by 2024

  • ItsBarryZ/langchain-chain-chain

    The problem

    Everyone was stacking agent frameworks on top of agent frameworks.

    What it does

    langchain-chain-chain is a joke repo about it: "For the memes until someone actually creates this." Its author spent 2023 building agent side projects.

    Who uses it

    In 2024 he co-wrote Anthropic's essay Building effective agents, which says the best agents skip complex frameworks.

    Barry Zhangjoined Anthropic by 2024

  • davidhershey/ScrapeGPT

    The problem

    Web scrapers break every time a site changes its layout.

    What it does

    ScrapeGPT has a language model write the scraper, and when it breaks, "Self-heals to debug itself."

    Who uses it

    A year later its author was at Anthropic, and in 2025 he set Claude loose on Pokémon Red, live on Twitch.

    David Hersheyjoined Anthropic by 2024

The Model Context Protocol, which many AI apps now use to plug into other tools, has a story like this too. David Soria Parra, one of its creators, says it started with an annoyance: copying text "back and forth" between the Claude desktop app and his code editor. "This is not part of a big strategy," he said; "it was literally just me going to a room with Justin." Its first public code was a pull request on someone else's open-source code editor, Zed, in August 2024.19

Their public record peaked the year before they joined

Public activity rose in the year before people joined, then fell off. The median went from about 16 public contributions two years before, to 28 the year before, to 5 the year after they started. The work most likely moved rather than stopped: once people join, most of their new code lives in private repositories or separate work accounts.

Public activity peaked in the year before people joined

Median public GitHub contributions per person, by calendar year relative to the Anthropic start year, for people active in the year before.

Years are calendar years, because work history gives start years only.

Show the numbers
Public contributions per personTwo years beforeYear beforeStart yearYear after
Median15.528205
Middle half, 25th to 75th percentile2 to 595 to 1033 to 600 to 29

What this means if you hire engineers

Read the work before the titles. In this group, the person who built Claude Code had spent eight years maintaining a TypeScript tool, the interpretability team was full of scientists, and more than half the group wrote no machine learning code before joining.

Read how people work on code they don't own. Most of this group did, and the accepted changes and repeat visits show how someone works with maintainers and reviewers.

Read the year before a move, and read the side projects. Once someone is inside a lab, most of their work leaves the public record, and the side projects show what they build when nobody assigns it. If you want to look for yourself, browse engineers on Vamo's public lists and what each one builds, or see how Vamo ranks people by what they've shipped. For more on reading a person's public work, see how to evaluate an engineer's GitHub work.

How we did this

We took the 225 people on Vamo whose Anthropic role started in 2022, 2023 or 2024, including people who have since left, and left out bots, organization accounts and people whose only Anthropic titles were outside engineering and research. About 21% of the group has since left Anthropic, judging by the current employer on their profiles. We counted their public GitHub contributions for each calendar year around their start year, with private repositories removed: commits, pull requests and whether they were merged, issues and repositories created.

A person is "active" if they made at least one public contribution in the calendar year before their start year; 67% were. Unless we say otherwise, every share in this post is a share of the active people, over the two calendar years before their start year. "Maintainer or core contributor" means 10 or more commits or merged pull requests to a project with 100 or more stars, their own or someone else's. "Repeat contributor" means more than three to at least one outside project, below that bar. "Drive-by" means contributing outside but never more than three times to any one project. "Side project" means a repository they owned that was not their job, their research or their teaching material; personal sites, CVs and scratch repos are left out. Project categories are ours, from an organization and repository map, keywords and languages, with hand labels for the long tail.

The group is the people we could match to a GitHub account and an Anthropic role, not every hire from those years. Work history gives start years only, so every "year before" is a calendar year. Commits from separate work accounts are not counted, and star counts are as of November 19, 2026. Download counts are for the week ending November 18, 2026, from npm and PyPI; dependent counts are GitHub's. People with no public activity are not weaker engineers; most professional code lives in private repositories.

We name a person only when a public source shows when they got to Anthropic: their own site or CV, a paper byline or acknowledgment, an interview, or a first commit to an Anthropic repository from an anthropic.com address. The footnotes give each source. Dates marked "by" come from the earliest public evidence we found, so the real start may be earlier. We checked whether each named person still works there against their GitHub profile, their own site and Vamo's current-employer record: people who have left are described as having worked there, and where the sources disagree we only say when they joined. Everyone else is described without a name.

FAQ

What did Anthropic's early engineers build on GitHub before they joined? Two in three of the 2022 to 2024 hires we could match made public contributions in the year before. Three in four of those worked on someone else's project, 55% wrote no machine learning code, and 55% had a side project. Their work ranged from a TypeScript tool with 6.5 million weekly downloads to star simulations, brain-mapping software, game emulators and LED walls.

Did Anthropic's engineers contribute to open source? Most of the active ones did. 75% committed to or proposed a change on someone else's project in the two years before joining, and 62% had a change accepted. Maintainers and repeat contributors outnumbered drive-by contributors three to one.

Did Anthropic's engineers have side projects? 55% of the active people had a personal project in the two years before they joined, and 49% had one that was not machine learning or science. Tools for programmers, games and puzzles, and shell and editor setups were the most common.

Did Anthropic's early hires have machine learning experience? More than half didn't, in public. 55% of the active people touched no machine learning code in the two years before joining. Anthropic's careers page says about half its technical staff had no prior ML experience.

Where did Anthropic's early hires work before? For the job right before Anthropic, 29% came from late-stage tech companies such as Stripe, Asana, Shopify and Robinhood, 21% from startups and 18% from big tech. 9% came straight from another AI lab. Stripe sent the most people of any single company.

How do you get a job at Anthropic? Anthropic's careers page asks applicants to put open-source work at the top of their resume. In this group, the public work that lined up with later jobs was usually a side project or an outside contribution in the same area. No one in the group has said a repository led to their offer.

Sources and credits

json-schema-to-typescript and Claude Code. Built by Boris Cherny (GitHub). Repository, live demo, npm package, GitHub dependents; Ink and Ink #682; Claude Code. Joined Anthropic in September 2024 (The Pragmatic Engineer, and again); side quests quote from Business Insider via AOL. Status: at Anthropic, per his GitHub profile.

coverage.py. Maintained by Ned Batchelder (GitHub). Docs, repository, history, PyPI downloads, GitHub dependents, watchgha. Joined Anthropic in 2024 (his blog, July 2024). Status: has left; he joined Netflix in April 2025 (his blog).

ReactiveCocoa and Mantle. Top committer Justin Spahr-Summers (GitHub). ReactiveCocoa, Mantle, Tigerbrew. At Anthropic by January 2023 (first commit to the TypeScript SDK); co-creator of MCP and top contributor to its TypeScript SDK. Status: has left to make games (jspahrsummers.com).

Autograd and JAX. Co-author David Duvenaud (GitHub). Autograd, PyPI downloads, JAX. Worked at Anthropic from June 2023 to October 2024 as Alignment Evaluations team lead (his CV). Status: has left; now at the University of Toronto.

compress and SWE-bench. Built by Erik Schluntz (GitHub, site). compress, SWE-bench submission, Raising the bar on SWE-bench Verified, Building effective agents. At Anthropic by October 2024. Status: unclear; his GitHub profile lists Cobalt Robotics and Vamo's record lists Anthropic.

Interactive graphics, AI Explorables and hot-server. Built by Adam Pearce (GitHub, roadtolarissa.com). AI Explorables and source, the grokking explorable, talk notes, hot-server, attribution-graphs-frontend, kindle-tracker. At Anthropic by May 2024, credited in Scaling Monosemanticity. Status: at Anthropic, per his GitHub profile.

TransformerLens. Repository, early pull requests #6, #7, #8 and #13 by Ansh Radhakrishnan (GitHub), at Anthropic by July 2023 (commit from an anthropic.com address). Status: at Anthropic, per his GitHub profile. The other two contributors are not named.

MESA and the toy model. Adam Jermyn (GitHub, site). MESA and source, Skye, toy_model_interpretability, Engineering Monosemanticity in Toy Models. At Anthropic by October 2023, as a core contributor on Towards Monosemanticity. Status: at Anthropic, per Vamo's current-employer record.

Synaptor and gpt-omics. Nicholas Turner (GitHub, site). Synaptor, gpt-omics. An author on both Monosemanticity papers, so at Anthropic by October 2023. Status: at Anthropic, per his site.

Genomic epidemiology at home. Joshua Batson (GitHub). gen-epi-at-home, cox. An author on both Monosemanticity papers. Status: at Anthropic, per Vamo's current-employer record.

world-models. Wes Gurnee (GitHub, wesg.me). world-models, Language Models Represent Space and Time. At Anthropic by October 2024, thanked as a colleague in Sparse Crosscoders. Status: at Anthropic, per his site.

The CHIP-8 emulator. Kevin Ji (GitHub, kevinji.com). chip-8-emulator and his write-up, deflate-rs, Solana. At Anthropic by October 2024 (first commit). Status: at Anthropic, per his GitHub profile and site.

The LED wall and other side quests. Ziyad Edher (GitHub): led, led.ziyadedher.com, todo, evm-bench; at Anthropic by April 2024 (his site, archived); status: at Anthropic, per his GitHub profile. David Soria Parra (GitHub): neweden; joined Anthropic in April 2024 (experimentalworks.net); status: at Anthropic, per his GitHub profile. Alex Albert (GitHub): twitter-hider-extension; at Anthropic by September 2023 (commit); status: at Anthropic, per his GitHub profile. Tim Maxwell (GitHub): HitecDServo, os2cx; an author on Towards Monosemanticity, so at Anthropic by October 2023; status: at Anthropic, per Vamo's current-employer record. The other side quests are described without names.

Starter kits. Zack Lee (GitHub). create-t3-turbo-ai, dynamic-interface-web, wove, his 2021 post, claude-quickstarts. At Anthropic by October 2024 (his site, archived). Status: at Anthropic, per his GitHub profile.

Agent side projects. Barry Zhang (GitHub). agent-sim, langchain-chain-chain, Building effective agents. At Anthropic by November 2024 (first commit). Status: at Anthropic, per his GitHub profile.

ScrapeGPT and Claude Plays Pokémon. David Hershey (GitHub). ScrapeGPT, feudal_networks, ClaudePlaysPokemonStarter, TechCrunch. At Anthropic by March 2024 (first commit). Status: at Anthropic, per his GitHub profile.

The Model Context Protocol. David Soria Parra and Justin Spahr-Summers (above). Zed #16103, specification, Introducing the Model Context Protocol, The Creators of Model Context Protocol, Latent Space.

Papers. Towards Monosemanticity, Scaling Monosemanticity, Sparse Crosscoders, Toy Models of Superposition and Circuit Tracing, on transformer-circuits.pub.

The IPO and the company. Anthropic prices its initial public offering, CNBC, Reuters via Investing.com, Money.it, citing Hill's estimate, anthropic.com/careers, checked November 19, 2026.

Footnotes

  1. Pull requests #6, #7, #8 and #13, September 2022. Ansh Radhakrishnan was at Anthropic by July 2023, when he committed paper code from an anthropic.com address. The other two contributors had changes merged in October 2022 and April 2023. TransformerLens has 4,085 stars as of November 19, 2026. ↩
  2. Justin Spahr-Summers made 14 commits to Tigerbrew from December 2014 to June 2015. Another future hire made 10 commits in May and June 2021, and in November 2021 had a NumPy fix merged that restored support for i386 and PowerPC Macs. ↩
  3. Ned Batchelder has maintained coverage.py since 2004 (its FAQ) and made 342 commits to it in 2022 and 2023. Downloads: 81,975,749 in the week ending November 18, 2026 (pypistats). Dependents: 515,498 repositories (GitHub). He announced the Anthropic job in July 2024 and joined Netflix in April 2025. ↩
  4. Justin Spahr-Summers is the top committer to ReactiveCocoa (2,495 commits, 19.8k stars) and Mantle (11.2k stars). He made the initial commit of Anthropic's TypeScript SDK on January 30, 2023. His site says he has since left to make games. ↩
  5. Started March 2016; 3,352 stars. Boris Cherny made 20 commits in 2023 and 38 in 2024, through v15.0.3. Downloads: 6,505,787 in the week of November 12 to 18, 2026 (npm). Dependents: 32,312 repositories and 2,582 packages (GitHub). Joined Anthropic in September 2024 (The Pragmatic Engineer). He sent a fix to Ink, the terminal UI library Claude Code uses, on November 27, 2024. ↩
  6. Autograd has 7,557 stars and had 1,012,786 downloads in the week ending November 18, 2026 (pypistats); JAX has 36,654 stars. David Duvenaud committed to JAX from 2018 to 2020. His CV lists him as Anthropic's Alignment Evaluations team lead from June 2023 to October 2024. ↩
  7. compress, October 2021, 1,415 stars. Erik Schluntz opened Anthropic's SWE-bench submission on October 22, 2024. His GitHub profile lists Cobalt Robotics and Vamo's record lists Anthropic, so we don't say whether he is still there. ↩
  8. Adam Pearce's site lists about 30 Bloomberg graphics and 69 New York Times interactives. He made 143 commits to AI Explorables, 31 in 2022 and 45 in 2023. Quote from his talk notes. Scaling Monosemanticity (May 2024) credits him with the feature browser. hot-server (2016, 103 stars) is the run step in attribution-graphs-frontend. ↩
  9. Stripe was the job right before Anthropic for 8% of people with work history, Google for 7%. ↩
  10. Adam Jermyn's commits to MESA ran through 2022, including its convection code; he also wrote Skye. toy_model_interpretability dates from October 2022, next to Engineering Monosemanticity in Toy Models (November 2022). He was a core contributor on Towards Monosemanticity, October 4, 2023. ↩
  11. Synaptor dates from 2016; gpt-omics from April 2022. Nicholas Turner is an author on both Monosemanticity papers; his site describes the Seung Lab work. ↩
  12. gen-epi-at-home, May 2022. Joshua Batson is an author on both Monosemanticity papers. ↩
  13. world-models, October 2023, 263 stars; the paper is Language Models Represent Space and Time (ICLR 2024). Wes Gurnee was thanked as a colleague in Sparse Crosscoders, October 2024. Quote from wesg.me. ↩
  14. Among Monosemanticity authors in our group, 63% of those active before joining had science code, against 22% of everyone, and 89% were active in the year before, against 67%. ↩
  15. Kevin Ji's write-up gives the dates and the quote. He had 9 pull requests merged into Solana in 2023 and made his first commit to an Anthropic repository in October 2024. ↩
  16. 30% of the game repositories are Advent of Code. 5% of personal repositories were in Rust, against 1% of work ones. ↩
  17. Zack Lee's three kits have 352, 275 and 112 stars. Quote from his 2021 post. His site said Anthropic by October 2024, and he made the initial commit to the MCP docs. ↩
  18. Barry Zhang made 57 public commits in 2023, against 2 the year before, and his first commit to an Anthropic repository in November 2024. David Hershey's first commit was in March 2024; TechCrunch covered the Pokémon stream. ↩
  19. Zed pull request #16103 by David Soria Parra, opened August 12, 2024 and merged three days later. Quotes from Latent Space. He joined Anthropic in April 2024. ↩
Problem with this page? Report it
What is wrong?