Tap Notes: Exit 0 Is Not the Same as Done

Three items from three directions. What connects them: the gap between what a signal reports and what actually happened. An agent that loops but exits cleanly. An audience that discards most of a talk. Intelligence becoming cheap enough that it stops being the differentiator. The measurement problem shows up everywhere. The tooling to catch it usually doesn’t.

607 Todos, Zero Real Work

A developer’s tool (MicroJack) entered a generation loop and produced 607 todos — 246 unique, the rest duplicates. The deduplication ratio caught the bug. Nobody checked whether the work was meaningful. Detection was structural, not semantic.

This is the failure mode worth naming: exit-0 with plausible prose. The agent completes its run, produces output that looks like progress, and triggers the next step in the chain. No crash. No error. Just hollow output moving downstream until something catches it — or doesn’t.

“Case 2 is worse than a crash, because a crash is honest.”

Post to X

The fix isn’t exotic: diff current run output against recent work history, flag when the delta falls below a threshold — the proactive version of what the deduplication ratio did, applied before the chain runs. The harder part is admitting most agent pipelines don’t do this. Structural failure gets caught. Semantic failure — technically non-empty, functionally inert — mostly doesn’t.

What’s Important in a Presentation? Peaks and Endings

Chris Lema on the peak-end rule: people evaluate experiences based on two moments — the emotional peak and the ending. Everything in between gets compressed. Duration neglect: the middle doesn’t count.

Kahneman tested this with colonoscopy patients. Patient A had a shorter procedure with less total pain — but it ended abruptly, rated the experience an 8. Patient B had a longer procedure with a gentler close — rated it lower despite more total suffering. The ending rewrote the evaluation.

The implication is uncomfortable: optimizing for comprehensiveness is mostly waste. You’re filling time that gets discarded. A product that ends mid-flow, a blog post that closes with a summary bullet list, a game level that ends on frustration instead of flourish — the peak-end rule doesn’t care what you intended. It only records what the brain felt at the two moments that matter.

Lema’s five techniques for closing a talk are really five ways to make the audience the agent of the ending rather than a passive recipient. The most durable: make them the hero. Flip the emotional attribution so the audience associates their own feeling with your work. That’s close to how good game endings function — the player feels like they did something, not that they watched something.

Can AI Be Conscious? — Jensen Huang

Jensen Huang describes himself as a dishwasher in a room of superhumans. The framing is deliberately provocative, but the structural claim underneath it is the one worth carrying: intelligence is now a layer, and that layer is becoming commodity infrastructure. What sits above it — taste, judgment, relationships, tolerance for ambiguity — can’t be compiled.

This inverts the standard AGI-panic narrative. The real story isn’t “AI will surpass human intelligence.” It’s “intelligence is becoming cheap enough that something else becomes the differentiator.” When reasoning stops being the scarce resource, what you do with your judgment, your relationships, and your bets starts to matter more, not less.

His clearest point: two humans given identical context produce statistically different outcomes — not because of compute variance, but because humans carry irreducible subjective variance. He doesn’t romanticize the gap. He treats it as a fact of architecture. Intelligence goes lower in the stack. The human layer operates at higher altitude.

🪨