Theo Browne argues that the models are now improving faster than we are, so the winning move is not to get better but to go bigger. He challenges the "skeuomorphic" habits and sacred cows of software engineering — env files, language identity, guilt-merged PRs — and shows how every tier of ambition has shifted down, to the point where a whole product can be a single markdown file.
Theo opens by admitting he's going through "some form of AI psychosis" and traces his path through recent models. Sonnet 3.5, he says, was a big moment — the first model that could reliably complete multi-step end-to-end tasks in the context of a codebase. He describes it as the tool-call era: not the first model to call tools, but the first that did so consistently enough for day-to-day coding work.
Opus 4.5 was the next jump — a model that could run for hours without losing track of what it was working on, testing its own work and getting it into a good state rather than needing to be walked step by step. Mythos and Fable, he argues, represent another jump to orchestration: a model that understands not just your codebase but itself, and knows how to spawn additional models, break up work, and verify it afterward — no custom software factory required, "you just need to prompt it to go a little further."
The turning point of the talk is a blunt directive: "we need to go bigger." Theo notes that most of the Jira tickets he closed at his previous job could be trivially solved by Opus 4.5 and wouldn't benefit at all from a model like Mythos. If the models keep improving — and he's now confident they will, admitting he was wrong to claim we'd hit a wall — then the humans can't keep pace by getting better.
"The models are getting better faster than we are," he says, "so we can't necessarily get better. So instead we have to go bigger." And to do that, engineers have to get over themselves and the strong opinions they've built up over years of writing code.
Theo uses the iOS 6-to-iOS 7 redesign as an analogy. Skeuomorphic apps had to be designed to convince you the phone could replace physical objects; iOS 7 dropped that pretense and became far more useful, even if people initially disliked the unfamiliar look. Developers, he argues, are stuck in their own skeuomorphic phase — clinging to the terminal, Vim, SSH, and Git because they're familiar and part of our identity.
He points at sacred cows that only look absurd once you say them out loud: why can't we commit our env files? Why do we qualify ourselves by the languages we write? Why are we so scared to delete code, and why do we "guilt-merge" a PR someone spent two weeks on rather than have an uncomfortable conversation? "One of the nice things about agents," he notes, "you don't have to feel bad when you shut down their work."
Theo shows three things he's built or is building and categorizes them as side project, startup, and "too big." But the tiers have shifted: everything is now one level lower. What used to be a startup is now a side project, and some startups at the event could arguably have been side projects. Below that sits a new tier — the markdown file.
He describes a PR-triage service that is now literally a markdown file: instructions to check four GitHub repos, assess open PRs, prioritize the work, and write a static HTML report to S3. It runs on a 9am cron and by 9:20 generates his work for the day. The gap he can't fill yet is what "too big" even means now — training your own model, building an OS, competing with npm — and he finds that uncertainty both scary and exciting.
Closing on strategy, Theo contrasts depth (many features in one area, like Vercel's edge over AWS for front-end-leaning servers) with breadth (range across many areas). Competing on AWS's breadth used to be hopeless without thousands of engineers — but now that range is viable. You can build a database platform into your product in a day or two with enough prompting.
The deeper move is to architect products so users can build the missing features themselves — the way Slack "accidentally" became the platform where half of people now run their agents, despite being, in his words, not a good product but "the right shape." His parting challenge: it's time to compete with Slack, build your own AWS, challenge Salesforce — and "if your idea doesn't feel stupid, it's because your idea is not big enough."
What I'm trying to say here is we need to go bigger.
The models are getting better faster than we are. So we can't necessarily get better. So instead we have to go bigger.
Do you know how many companies are at this event where their whole product could just be a markdown file?
If your idea doesn't feel stupid, it's because your idea is not big enough.
| Time | Topic |
|---|---|
| 00:00 | A YouTuber takes the keynote stage; "AI psychosis" |
| 01:10 | Model eras: tool calls, long tasks, orchestration |
| 03:17 | We need to go bigger — push the model further |
| 06:10 | Skeuomorphism in dev tooling: the iOS 7 analogy |
| 07:16 | Sacred cows: env files, language identity, guilt-merged PRs |
| 10:28 | Tiers have shifted: the markdown-file company |
| 13:37 | Breadth vs depth: build across the spectrum |