He publicly retracts his own "the models have hit a wall" call, and redraws the line for what's worth building.
Everything we knew about software has changed — Theo Browne, @t3dotgg · Theo Browne
16 min total·Actually worth watching closely: ~3 min·2 must-watch clips
- 0:17 – 2:20Listen
Admitting he called it wrong
He opens by laying it out: he said publicly that model progress had slowed, and looking back, that call was wrong. The fault wasn't the models — it was that the size of the tasks people hand them never grew.
If you only try new models with old-sized prompts, you'll never see how much better they got.
This stretch is all self-narration and reflection, mostly just him on camera — headphones are enough to catch every bit of it.▶ Jump to 0:17Speaker · Theo Browne - 2:20 – 5:55Listen
Three jumps in model capability
He walks the timeline through three watersheds: Sonnet 3.5 first made tool calls genuinely reliable, Opus 4.5 could run hours-long end-to-end tasks without losing track and test its own work at the end, and the step change with Mythos is orchestration — understanding itself, spinning up additional models to split the work, then going back to verify.
The keyword for the latest jump isn't "smarter", it's "organizes its own work" — you no longer need to build a whole tool chain to feed it.
The differences between the three generations come entirely from what he says and the examples he gives; there's nothing you need to watch, so listen while you do something else.▶ Jump to 2:20Speaker · Theo Browne - 5:55 – 7:19Skim
AI doesn't need to convince anyone anymore
He uses the iOS 6 to iOS 7 design shift as an analogy: early interfaces had to look like physical objects to convince you the phone could replace the things on your desk; once Apple had won, that layer of persuasion was no longer necessary.
AI is past proving itself too, so products should drop the persuasive packaging and get back to pure usefulness.
This part is backed by a side-by-side of the two generations of interface — one glance at the image and you know what he's pointing at, then you can just follow the argument by ear.▶ Jump to 5:55Speaker · Theo Browne - 7:19 – 9:30Listen
Tooling habits and language identity are the heaviest baggage
He names what senior engineers find hardest to put down: years of internalized tmux, SSH and Git workflows, plus the "I'm a such-and-such language person" identity. These aren't technical barriers, they're psychological ones.
Keep asking yourself: do we do it this way because it's right, or just because this is how we've always done it?
Pure argument, no demo and no code — the section of the talk best suited to listening with your eyes closed.▶ Jump to 7:19Speaker · Theo Browne - 9:30 – 10:55Listen
Stop being precious about code already written
He describes a scene everyone has lived: you know a colleague's PR takes the wrong approach, but you merge it out of guilt because they put a week or two into it. Our industry treasures written code far more than it deserves.
There's an underrated benefit to using agents — throwing away their work costs you nothing emotionally.
Carried by a highly relatable story about working on a team; his tone gives you more than the screen does.▶ Jump to 9:30Speaker · Theo Browne - 10:55 – 13:00Watch
The whole yardstick has moved down
The core claim of the talk: the side project / startup / too big yardstick has shifted down across the board because of model capability. He cites the OBS collaboration tool he built for streamers — a year ago that was startup scale, now it's a weekend project.
If your idea doesn't sound stupid, it isn't big enough.
He puts the three tiers on screen and demonstrates live how a project falls from the middle rung to the bottom one; that movement is itself the argument, so audio alone loses half the information.▶ Jump to 10:55Speaker · Theo Browne - 13:00 – 14:19Listen
Executable Markdown
He throws out a badly underrated paradigm: for some companies, the entire product can essentially be condensed into a Markdown file handed to Claude or Codex to execute.
A product's vehicle may no longer be a codebase, but one clearly written document that can be run.
Dense in argument with no demo to go with it — worth rewinding and hearing again rather than staring at the screen.▶ Jump to 13:00Speaker · Theo Browne - 14:19 – 15:43Listen
Go compete with the giants, and leave room for users
He closes by arguing that now is the time to go straight at Slack, AWS and Salesforce, and the way to do it is to build the product in the right shape — so users can put together the vertical features you never could have predicted. Slack itself isn't a good product, yet one somewhat functional bot API accidentally made it the platform a lot of agents run on.
A platform's shape decides its survival more than its polish; whatever feature a user is missing shouldn't be your problem anymore.
The ending is a dense call to action plus that counterintuitive Slack example — best heard straight through in one go.▶ Jump to 14:19Speaker · Theo Browne