Every 10 Milliseconds Cost Copilot 1% of Completions

Alex Graveley tells Sarah Guo and Elad Gil that he got into Linux at fourteen, around the moment the web still took work to reach. The GPL struck him early: helping in the open, so other people could learn the way he had, seemed obviously good. He spent years compiling kernels while Windows still ran the data centers, then did four or five years at VMware before leaving to start an education company. Nine months later he killed it. The lesson he kept was simple enough to sound like a tautology: you have to make something that is both achievable and that people will pay for, or at least spend time on.

He was living in a San Francisco warehouse with Burning Man people who could not organize large camp builds, so he forked Etherpad and started Hackpad. It was a real-time editor in the Google Docs mold, with one extra that mattered in pseudo-anonymous groups: it highlighted who wrote what, so you could hunt down the person who owned a decision. He used Y Combinator itself as a distribution hack, taking notes on the product and getting fellow batch companies onto it. Stripe used Hackpad for years as what Graveley believes was its first knowledge base. Dropbox acquired the company, and the work became Dropbox Paper.

After Paper he knew he wanted a robot that did things for people. That search landed him at Magic, a text-based personal assistant that looked like AI and was, in practice, ops. Teams cycled around the clock, lost context, and drowned in requests. Graveley calls it a crash course in human behavior: what people do under stress, what you can train, and what you cannot. Text has an annoying property the web does not. You can be at the 99% mark of a task and then change the goal entirely. Flights get delayed. Passwords get lost. Elad Gil asks whether this is the last generation that will handle those failures by hand.

The hosts tick through a whole era of ops-heavy almost-agents: Aardvark's social search, Clara Labs for scheduling, Operator, the Q&A app Jelly, Facebook M. People wanted rich bots; the technology was not there, so they stacked humans. At Magic, Graveley still got a first shipping AI product out of it. He trained a small sequence-to-sequence model on assistant-user chat history and showed gray text in the operators' input bar. Measured the old-fashioned way, it saved about half an hour of typing across 100 people on an eight-hour shift. After that he helped launch hCaptcha, now one of the top captcha services in the world, then built MobileCoin with Signal's Moxie Marlinspike, aiming at Venmo-quality transfers that settled in a few seconds.

Gil asks whether crypto matters for agents as programmable money, or as cryptographic identity that can tell a person from a bot. Graveley's honest answer is darker than either framing. He expects a multi-year stretch of extreme discomfort in which AIs pretend to be things, confuse people, extract money from grandparents, and drain life savings. OpenAI is trying, he says, but the public focus has landed on that one lab doing everything. The work that would actually help — systems that prevent the scams, legislation that drops a hammer on the people running them — is not what is being built.

He is not especially scared that AI will kill us, though he is grateful people are thinking about it. He is more worried about bad people using new technology to hurt us. Gil cites Illia Polosukhin, last author on the Transformer paper and later founder of the NEAR blockchain, and Polosukhin's idea of stress-testing society against the coming wave. Graveley likes the frame. It is not as bad as it could be: most things you would want to spam already have blockers, and accounts are somewhat hard to create. Sock-puppet filtering will matter more. He likes what Cloudflare is doing with device fingerprinting now that visual captchas are no longer good enough.

Cost is a weak brake, not a wall. Calling everyone costs money; texting should be legally constrained and also costs money, maybe not enough once agents can earn. Grandmothers, he notes, do not have crypto — they have bank accounts, which can be traced. His middle-ground image is blunt: North Korea has been trying this for a long time, and now there is a version with more resources, or a more distributed one. Some mitigations exist. We need more, and we need to think about it a lot more. He suspects really bad things will have to happen before the alignment of incentives is correct.

The path into GitHub was not a career plot. While Graveley was on MobileCoin his father's kidneys failed; he tried to donate, the scans found a lump in his chest, and in 2018 most of his right lung was removed. Healing from internal injuries, he says, takes longer than you think. The happier ending is that he had been cancer-free for more than four years at the time of the taping, and his father got a transplant. After a long recovery he begged a friend for a job, landed at GitHub, and first converted the company to building GitHub on its own Codespaces. Then the OpenAI opportunity showed up, and because he had been tracking the field, he jumped on it.

There was no romantic origin story. OpenAI and Microsoft were negotiating a supercomputer cluster; software provisions got thrown in — Office and Bing, probably — and GitHub asked whether there was something it could do. OpenAI tossed over a small fine-tune trained on some code and said, see if this is interesting. It was not a Davinci-sized GPT-3. Graveley later learned it was basically a training artifact: OpenAI had wanted to know what introducing code into base models would do, and it seemed to help chain-of-thought reasoning, because code is linear and earlier lines constrain later ones. In the product sense it was very bad, Python-only, and trained on a small crawl of GitHub.

A colleague got the model first and showed that it did not work most of the time — and then, occasionally, generated something useful. That was enough. Graveley and Albert Ziegler became the first two engineers on the project. They crowdsourced in-house Python problems they knew were not in the training set, then built an in-the-wild harness: download a repository, run tests, find the functions those tests called, zero the function bodies, ask the model to regenerate them, and rerun. A new pytest feature even told them which functions a test had touched. Up front, something like less than 10% passed. Most teams would have stopped. They judged there was a lot more juice to squeeze.

They poured all of GitHub's code into the model, improved prompting, and trained on every version of a file plus diffs so the model could see small changes, not just the latest snapshot. The in-the-wild pass rate climbed from under 10% to over 60%. That is the number Graveley still sounds slightly stunned by: one in two tests, the model can just generate the code. Along the way it was 10% to 20% to 35% to 45%, a grind rather than a single leap. When they started, they did not know what they had. They were just trying to find out.

OpenAI's side of the house was thinking in terms of a Stack Overflow competitor: they had all this code, would it not be nice to leverage it as Q&A. Graveley's honest answer is that he did not know how to beat Stack Overflow with this thing. Early UIs made the failure rate painfully visible. You would wait some number of seconds, watch a stream of bad generations, and cheer for the one success. They tried to cover up the one-in-ten reality with interface. Autocomplete popups mixed generated code with ordinary identifier suggestions and felt like the wrong metaphor. A button over empty functions, a hotkey that dumped a list of candidates in a sidebar, little popup pickers — they tried every VS Code UI they could think of. Lists were the worst: maybe one generation per person per day in a tiny internal sample of language nerds and OpenAI folks.

Graveley had been enamored for years with Gmail's gray-text autocomplete. He calls it the first large language model deployment in the wild: fast, a little bad, and backed by a paper that actually explained the workarounds. That was always in the back of his head. VS Code did not support inline ghost text, so he hacked it in far enough to shoot a demo video, then had to politick a Skunk Works project almost nobody knew existed into a short-timeline feature from the VS Code team. People on that team, reasonably, did not know who they were. From beginning to public launch, it was still less than a year.

The interaction that finally fit how programmers think was not a menu. You get gray text, you hit Tab, it inserts; ignore it and it disappears. Single-line autocomplete could not show what the model could do, so they fought for multi-line ghost text that completed the rest of the current block — an if statement, a loop — because programmers judge a block quickly. If the completion does the right thing, they know. Once that mechanism worked, they never found the bottom of the performance curve. They made it as fast as they could, and completions kept improving.

A long free public-access window before general availability let them slice usage by experience, geography, and everything else. The finding that survived every cut was almost rude in its simplicity: speed is the only thing that matters. Every 10 milliseconds was 1% fewer completions, and 10 milliseconds is fast. They learned it because Indian completions were mysteriously low in the first months. There was one data center, in Texas. A keystroke from India went through Europe, across the Atlantic, down to Texas, and back. By the time the gray text returned, the user had typed past the request, and the completion was worthless.

People close to Texas, meanwhile, were posting in Slack that it was freaking great. The posts Graveley loved most were from people who did not program: they had just written a hundred-line script that did the thing they needed. Guo says she now speaks languages whose syntax she does not actually know. The models, Graveley notes, are extremely good at finding patterns. Azure had a six-month plan to run OpenAI; the Copilot team needed GPUs in Europe within a month so Asia would not keep missing the window. Microsoft stepped up, they got it running, and then they launched.

Gil is not surprised. At Google, on mobile products and on search, a 100-millisecond shift moved market share. Graveley was not surprised by uptake either. Retention sat at 50% and, months later, was still above 50% by weekly cohort, which he calls insane. What they did not know was whether people would pay. Early inference estimates were about $30 per user per month with no margin, enough that the team seriously discussed an enterprise-only product. Graveley lobbied hard to go cheap and capture the market. Once Microsoft's Azure infrastructure let them fork more accurate projections, the first big number came back: $10 a month. They had not even started optimizing price. After they did, it was less than that. For a tool that writes about 40% of a developer's code — and 80% for some whales — that is not a lot.

Asked to look three to five years out — 95% of code written by copilots, humans directing — Graveley says he cannot quite picture the world because the psychological jumps are so large. Past: your hands are on the code and you know it is right. Now: you mostly know, but you have to test it and see it run. Next: you write it and trust that it will work. If that is the goal people want, he thinks every barrier is achievable. Review only the dangerous or confusing parts. Train on functions before and after edits so a model can propose a more polished version. Start from a basic main loop, add features piece by piece with tests, and keep generating the logical next one. If that is what we want to do, that is what is going to happen.

Minion, still in stealth at the time, is the same instinct moved out of the editor: bots that do stuff for you. Not just answering questions or writing copy, but organizing a schedule, booking flights, finding a trip, doing taxes, flagging which contacts you have not talked to in a long time. Copilot gives a little help in code; he wants Minion to give a little help in daily life, with access to information and actions in a controlled, aligned way. He quit Magic because he could not hook AI up to real-world data. If improving quality required a math PhD, he did not know what to do. The models got better. That specific problem started to look solvable.

The real world is still hostile to a frozen training set. Flights delay, prices move constantly, the concert is missing the seat you wanted. AIs compress everything in the data; they are not yet a real-time mechanism. Graveley's bet is that you break tasks the way humans do: write a list, notice missing information, maybe write code, maybe ask questions, maybe query data. People do not usually write code to do their taxes, but walking through pay stubs is still executing a for loop. Code was special for Copilot because you can run it and learn from success or failure; agent work on the web and in apps has some of the same property. The sci-fi end state is saying "computer, file my taxes" and having it do the right thing. He thinks that is in the next few years. He also likes datasets that do not exist yet — people clicking on the web is a gigantic, currently unowned one.

Minion is hiring, and Graveley has noticed a strange bimodal. Very senior people get it. Very young people get it. The middle has not caught up. It is not an age thing so much as an experiential midwit curve. Tell a naive person you can make magic and they say that is awesome; tell a deeply experienced person and they say maybe; almost everybody else says that is not possible. Gil has been watching founders of hundred-person companies dive back into writing code because their teams do not have intuition for what the new models can do, and therefore do not know what to build. The founder mindset — learn the capability, then apply it — is suddenly scarce inside the org chart.

Comfort with "I do not know what will work, let me try some stuff" looks a lot like web and product work: pick metrics, test, look again. What is still rare is the tenacity to face something that does not work at all. The natural reaction to a system at 10% is why even bother. People try ten things, watch them fail, and declare the problem impossible. The crazy fact, Graveley says, is that these things work at all — and not only that, they scale and they improve at scale. Most technologies break every time you double. That is usually what you hire people to fix. Here the curve goes the other way.

He is in favor of pushing the limits precisely because even the best people cannot explain what happens if the models get ten times bigger. We built the Large Hadron Collider for the same reason: maybe we will find some stuff. Those endeavors are valuable. He is grateful to be alive during this time. Guo takes it as a closing note, and the episode ends there — curiosity as the right posture when a technology still has unknown emergent properties at the next order of magnitude.