Nelson
Elhage

Nelson Elhage has a particular talent: taking systems that everyone else treats as black boxes and opening them up to see what's actually running. He did it with the Linux kernel at Ksplice - patching a live kernel without a reboot when everyone said that was basically impossible. He did it at Stripe with Ruby's type system, co-founding Sorbet to bring static analysis to one of the world's largest payments platforms. And at Anthropic, he did it with transformer neural networks, co-authoring research that gave the field its first real mathematical grammar for how attention mechanisms work.

None of this is accidental. Elhage is what happens when relentless curiosity meets the patience to actually read the source code. His blog, "Made of Bugs," has been running since roughly 2007, logging everything from terminal control sequences to CPU branch predictors to collaborative crossword architecture. He is, in the best possible sense, a person who thinks rebooting is a failure mode to be engineered around.

He studied Computer Science at MIT (Course 6, the hard one), taught classes on lambda calculus and game AI as an undergrad, and then went directly into Ksplice - a startup born out of research on keeping servers alive without interruption. Oracle noticed. They acquired Ksplice and brought Nelson along. Then Stripe noticed. He joined as roughly employee number 30 and stayed long enough to help define what engineering culture looks like at scale.

That focus is not accidental either. At Stripe, where he was on the path to a Staff Engineer title before departing, Elhage became known for the ability to hold company-wide context in his head - the rare engineer who understood not just his team's systems but the dependencies between teams, the history of architectural decisions, and the future problems those decisions were going to create. He called it "information routing." The rest of us call it seeing around corners.

His open-source work reflects the same instinct for making power more accessible. reptyr lets you grab a running process and move it to a new terminal - a tool born from personal frustration that's now in the default repositories of most Linux distributions. livegrep indexes entire codebases and returns regex search results as you type, and it's fast enough that Mozilla uses it as the backend for SearchFox, their code search for Firefox. ministrace is a minimal strace implementation that fits in a teaching session. These aren't hobby projects. They're infrastructure.

When Nelson joined Anthropic, he pivoted from infrastructure to something stranger and more fundamental: figuring out what, exactly, is happening inside a large language model when it thinks. The mechanistic interpretability team's job is to find the algorithms - the actual computational circuits - that transformer neural networks implement when they predict the next token.

In 2021, Elhage co-authored "A Mathematical Framework for Transformer Circuits," a paper that gave the field something it had been missing: a decomposition language for attention mechanisms. The key insight was that attention-only transformers could be analyzed as sums of interpretable functions - "paths" through the model that could be studied independently. Attention heads could be understood as having two separable computations: a QK circuit (which tokens attend to what) and an OV circuit (what effect that attention has on the output). That's the kind of clarity that lets you actually reason about what a model is doing instead of just empirically poking it.

A year later came "In-context Learning and Induction Heads" - a paper that identified a specific type of circuit, the induction head, that explains a phase transition early in transformer training. When induction heads form, the model suddenly gets much better at learning from examples in its context window. The paper traced this precisely, connecting a mechanical feature of the architecture to a visible behavioral change during training. That's what interpretability research is supposed to do, and it's extraordinarily hard.

After the interpretability work, Elhage moved to Anthropic's pretraining team - the group responsible for how Claude is trained from scratch. The jump makes sense: if you've spent years understanding what models learn to do and how they learn it, the natural next question is how to make that learning process produce better, safer, more predictable models. His trajectory from kernel hot-patching to transformer circuits to pretraining is, when you look at it, a single continuous thread: how do you make complex systems behave in ways you can actually rely on?

Joining Stripe as roughly employee number 30 means you arrive when the company is still figuring out what it's going to be. Nelson spent years there doing exactly the kind of work that doesn't fit neatly into a job description: building payment architecture, maintaining company-wide context on technical decisions, routing information between teams that didn't know they needed to talk to each other.

His biggest formal project at Stripe was Sorbet - a static type checker for Ruby built in a year with two other senior engineers. The goal was to bring the kind of safety guarantees that statically typed languages provide to one of the world's largest Ruby codebases. That meant not just writing the tool but persuading an engineering organization to adopt it, which required a year of internal deployment work after the technical work was done. Sorbet went open source and is now used across the industry.

When asked about Staff Engineering archetypes, Elhage identified three types: deep technical builders, cross-cutting initiative leaders, and team vision keepers. He moved between all three during his time at Stripe - which is probably what made him so effective, and so hard to title.

He was almost officially titled Staff Engineer. Two months short.

Nelson has been writing about software on the internet since before most tech blogs had RSS feeds. His blog, "Made of Bugs," is less a personal brand and more a lab notebook - posts on CPU branch prediction, Python interpreter performance, structured concurrency models, Z3 constraint solvers, and terminal control sequences, all written by someone who has clearly run the experiments and read the manuals.

The Buttondown newsletter, "Musings on Computer Systems," gets the more speculative material - ideas that are still taking shape, technical arguments that need an audience to push back on them. His most-shared piece, "Computers Can Be Understood," is a philosophical defense of the idea that software isn't magic - that any curious person with time and access to source code can understand how a computer actually works. It's the kind of argument that only lands when it comes from someone who has demonstrably done that, repeatedly, across five different layers of the stack.

He also maintains two Tumblr blogs, because some things require a different format: nelhagedebugsshit documents specific debugging war stories, and Accidentally Quadratic catalogs real-world cases where code quietly became O(n²) in production. The second one has become a minor institution in performance engineering circles.