{
  "video": {
    "id": "xUnRQ9vLXxo",
    "title": "Everything we knew about software has changed — Theo Browne, @t3dotgg ​",
    "duration": 962,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": "In person"
  },
  "analysis": {
    "video_id": "xUnRQ9vLXxo",
    "title": "Everything we knew about software has changed — Theo Browne, @t3dotgg ​",
    "one_liner": "Theo Browne argues that model capability has jumped a tier — Sonnet 3.5 made tool calls reliable, Opus 4.5 made hour-long tasks reliable, and Mythos orchestrates other models — so every ambition level shifts down one notch: yesterday's startup is today's side project, and a lot of products are now just a markdown file on a cron.",
    "summary": "Theo frames recent models as eras — Sonnet 3.5 as the tool-call era, Opus 4.5 as long-running-task era, Mythos as the orchestration era where the model understands itself well enough to spawn sub-models and verify their work with nothing more than a prompt. Because the models are improving faster than developers are, he argues the only response is to go bigger, which means shedding the identity attachments (terminals, vim, git habits, languages, sunk-cost code) that he calls software's skeuomorphic phase, echoing Apple's shift from iOS 6 to iOS 7. He shows his own project tiers — Reddit meme scraper, a YC startup, a full-stack cloud — all sliding down one tier, with a new bottom tier below side project: a markdown file you pipe to Codex or Claude. He closes on breadth vs depth: startups no longer have to pick a narrow vertical against AWS-scale breadth, because range is now buildable and users can fill the gaps themselves if you architect for it.",
    "key_points": [
      "Three model eras: Sonnet 3.5 was the first to do tool calls consistently enough for day-to-day codebase work; Opus 4.5 ran multi-hour tasks without losing track and tested its own work; Mythos is the first model that 'understands itself' and will spawn additional models, split up work and verify it — no custom software factory needed, just prompt it further.",
      "He retracts his earlier claim that models were hitting a wall: 'the models are getting better faster than we are,' so the response is to go bigger rather than expect to get better.",
      "Most Jira tickets from his previous job could be trivially solved by Opus 4.5; that work wouldn't even benefit from Mythos — the bottleneck is the size of the task you hand the model.",
      "Developers are in a skeuomorphic phase, like iOS 6's leather-and-compass apps before iOS 7: terminals are treated as the ultimate interface when natural language has no place in a terminal, and habits like not committing .env files are Git's accident, not a truth.",
      "Identity attachments to blame: qualifying yourself by the language you write, fear of deleting code, sunk-cost 'guilt merged' PRs — problems that vanish with agents, whose work you don't feel bad shutting down.",
      "The tiers shifted one level down: his Reddit meme scraper (side project, 2021), Ping / 'Zoom for streamers' from YC (startup, 2021), and a full-stack cloud — 'Vercel but it goes further each direction', with auth and databases built in (the too-big one he's building now).",
      "A new bottom tier below side project — the 'G brain tier': a markdown file. His PR triage service is now a markdown file that reads open PRs across four GitHub repos, prioritizes them, writes a static HTML file to S3 and returns the URL; it runs on a 9:00am cron and finishes around 9:15–9:20.",
      "Breadth vs depth: Vercel beat AWS by being deeper in front-end-leaning full-stack servers ('even the agents prefer it'), but now breadth is viable — you can build a database platform into your product in a day or two of prompting; not as reliable as RDS, but enough to let users start, and they can build the missing vertical features themselves if the product is the right shape (Slack accidentally became the platform people run agents in)."
    ],
    "takeaways": [
      "Push the model past what feels reasonable: ask a single prompt to break work into sub-agents and verify the results, instead of stepping it through tasks or building custom orchestration tooling.",
      "Try replacing a small service with a markdown file piped to Codex or Claude on a cron — Theo's PR triage, prioritization and static-site publish is literally one prompt file.",
      "Re-tier your project ideas one level down: what you'd have called a startup is now a side project, so pick the thing you previously wrote off as too big.",
      "Audit the practices you hold for identity rather than results — terminal-only workflows, language tribalism, not committing env files, refusing to delete or reject code — and ask whether each is right or just habitual.",
      "Architect for breadth plus extensibility: cover enough surface that users can start, and expose APIs shaped so they can build the vertical features you skipped, rather than narrowing to one deep vertical."
    ],
    "topics": [
      "agents",
      "orchestration",
      "ai-coding",
      "developer-culture",
      "startups",
      "model-capabilities",
      "product-strategy"
    ],
    "tools": [
      "Claude Sonnet 3.5",
      "Claude Opus 4.5",
      "Mythos",
      "Fable",
      "Claude",
      "Codex",
      "Vercel",
      "AWS",
      "RDS",
      "S3",
      "GitHub",
      "Git",
      "Slack",
      "Salesforce",
      "Jira",
      "OBS",
      "Y Combinator",
      "Ping",
      "Reddit",
      "Twitter",
      "tmux",
      "GNU screen",
      "Vim",
      "npm",
      "Node",
      "Apple / iOS"
    ],
    "quotes": [
      {
        "text": "The models are getting better faster than we are. So, we can't necessarily get better. So, instead we have to go bigger.",
        "at": "03:30",
        "url": "https://www.youtube.com/watch?v=xUnRQ9vLXxo&t=210s"
      },
      {
        "text": "Natural language has no place in a terminal, but we pretend it does because the terminal's familiar.",
        "at": "06:30",
        "url": "https://www.youtube.com/watch?v=xUnRQ9vLXxo&t=390s"
      },
      {
        "text": "Do you know how many companies are at this event where their whole product could just be a markdown file?",
        "at": "11:21",
        "url": "https://www.youtube.com/watch?v=xUnRQ9vLXxo&t=681s"
      },
      {
        "text": "If your idea doesn't feel stupid, it's cuz your idea's not big enough.",
        "at": "15:35",
        "url": "https://www.youtube.com/watch?v=xUnRQ9vLXxo&t=935s"
      }
    ],
    "words": 3723
  },
  "summary_url": "/#xUnRQ9vLXxo",
  "transcript": {
    "html": "/transcripts/xUnRQ9vLXxo.html",
    "txt": "/transcripts/xUnRQ9vLXxo.txt",
    "vtt": "/transcripts/xUnRQ9vLXxo.vtt"
  }
}