{
  "video": {
    "id": "8KkibGU_DDY",
    "title": "Mousepower: agents that can’t be measured, can’t be managed. — Maximillian Piras, Yutori",
    "duration": 1238,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer"
  },
  "analysis": {
    "video_id": "8KkibGU_DDY",
    "title": "Mousepower: agents that can’t be measured, can’t be managed. — Maximillian Piras, Yutori",
    "one_liner": "Yutori's founding designer argues agents have a measurement problem — like James Watt inventing \"horsepower\" to sell steam engines to people who thought in horses, if you ship an agent you also have to ship the rubric that lets customers verify its work and justify the token spend.",
    "summary": "Maximilian Piras opens with the now-familiar workflow of backgrounding and fan-out-parallelising agents — fun \"until you get the bill\" — and uses it to set up his thesis that agents can't be sold or adopted without a credible measure of value. He retells James Watt's invention of horsepower: an unscientific, arguably inaccurate metric whose real job was to let horse-minded buyers calculate an ROI and get over the line to try a steam engine. He argues tokens are an internal output, not an outcome, and that the industry's real bottleneck has shifted to verification — Anthropic has claimed to solve coding but admits it hasn't solved code review. He closes with \"mouse power\" as an idea rather than a metric, plus an entropy-based 2x2 for choosing which tasks deserve an agent at all.",
    "key_points": [
      "Thesis: agents have a measurement problem. The room of early adopters is biased (cue the \"mandatory Upton Sinclair quote\") and doesn't represent the people who still copy-paste into ChatGPT — and in some way or another everyone here is selling tokens, directly or indirectly.",
      "James Watt studied horse gins — a horse hooked to a rotary arm walking in a circle to power a mill — and derived horsepower. It wasn't scientific or even accurate; its job was to communicate an increase in value and get people to try the steam engine at all. Watt's barrier was cognitive dissonance in people who think in horses.",
      "Piras works as founding designer at Yutori for the past year and a half on computer-use models — agents that use a computer like a human when the data isn't reachable via an API or MCP; less efficient than APIs/MCPs, so it's a last resort. He demos the agent browsing Yutori's own site to check its own benchmark. Customers are excited but keep saying they're \"just scratching the surface.\"",
      "We're in an \"overspending and underusing\" doom loop (term borrowed from Ramp, who have a blog post on it): token-max into austerity, drop out, then FOMO back in. Coinbase's CEO posted a chart on X showing AI spend diverging from token usage after they changed default models and reserved frontier models for the hardest tasks — a good start, but still too focused on tokens.",
      "Tokens are a fine internal-system measurement but are just an output; they must trace cleanly to outcomes (bugs squashed, support requests closed) and then to progress on objectives.",
      "We're \"dying by a thousand pull requests\": Anthropic team members have claimed coding is solved but admit code review is not, so the bottleneck moved to human review and verification. Piras cites Noah Hein's post arguing the assumptions underneath code review are what need revisiting; code review feels solvable because the culture converged on shared assumptions, which is what makes a clear rubric — and those assumptions now need adapting for the agentic age.",
      "\"Mouse power\" — a horsepower equivalent for agents. He had Claude vibe-code a device to measure cursor movement speed to compute a human-vs-agent delta, but calls it a joke and a fool's errand: information space is too high-dimensional, so mouse power is an idea, never a metric.",
      "A Shannon-entropy-flavoured 2x2 for task selection — x-axis: uncertainty in the steps to perform the task (booking a flight has a departing destination, arriving destination, a seat chosen; painting a masterpiece has no knowable steps); y-axis: uncertainty in the acceptance criteria. Low step uncertainty → just write a script. High step uncertainty → likely out of distribution in pre-training and sparse rewards for RL. High acceptance-criteria uncertainty → verification is indistinguishable from execution, a waste of tokens. The sweet spot is the middle: NP-shaped tasks, easier to verify than to execute — and a repeatable verification pattern means you can throw agents at the verification too."
    ],
    "takeaways": [
      "Ship a verification rubric alongside every agent. It is not enough to build the agent; you have to give the customer a method for checking the output is good, because without it they can't calculate ROI or justify the spend.",
      "Stop reporting tokens as the measure of value. Trace token spend to clean outcomes — bugs squashed, support requests closed — and then to objectives; a token leaderboard is a sign the incentives aren't aligned.",
      "Change your model defaults: reserve frontier models for the hardest tasks, per the Coinbase chart where AI spend diverged from token usage.",
      "Screen candidate agent tasks on two axes before building: if the steps are predictable, write a script instead; if verifying the result means a human redoing the work, don't build the agent; target the NP-shaped middle that's easier to verify than to execute.",
      "Where verification has a repeatable pattern, build the agent that verifies the work of the agent — that's how you get from execution at the speed of computer to measurement at the speed of computer."
    ],
    "topics": [
      "agents",
      "evals",
      "computer-use",
      "roi",
      "token-cost",
      "code-review",
      "product-design",
      "verification"
    ],
    "tools": [
      "Yutori",
      "Claude",
      "Anthropic",
      "ChatGPT",
      "Ramp",
      "Coinbase",
      "X"
    ],
    "quotes": [
      {
        "text": "And it's a lot of fun, of course, until you get the bill. And then you start to wonder, was it all worth it?",
        "at": "02:07",
        "url": "https://www.youtube.com/watch?v=8KkibGU_DDY&t=127s"
      },
      {
        "text": "let's be honest regardless of how efficient this was horses just have great vibes. So like it's kind of hard to beat the vibes of horses",
        "at": "08:31",
        "url": "https://www.youtube.com/watch?v=8KkibGU_DDY&t=511s"
      },
      {
        "text": "So when you have high uncertainty in the acceptance criteria, you pretty much are in a spot where verification is indistinguishable from execution. So, why would you build an agent for something that to verify was useful, a person pretty much has to do the work again.",
        "at": "19:14",
        "url": "https://www.youtube.com/watch?v=8KkibGU_DDY&t=1154s"
      },
      {
        "text": "And so of course you don't just build the agent, you perhaps build the agent that verifies the work of the agent.",
        "at": "20:11",
        "url": "https://www.youtube.com/watch?v=8KkibGU_DDY&t=1211s"
      }
    ],
    "words": 4816
  },
  "summary_url": "/#8KkibGU_DDY",
  "transcript": {
    "html": "/transcripts/8KkibGU_DDY.html",
    "txt": "/transcripts/8KkibGU_DDY.txt",
    "vtt": "/transcripts/8KkibGU_DDY.vtt"
  }
}