中文
AI Engineer World's Fair

Models improve by the week, and the prompt box between you and them is still running the protocol of a punch card.

The Prompt Is Still a Punch Card - Ted Johnson, JoinIn AI · Ted Johnson

20 min
AI ProductAgent

20 min total·Actually worth watching closely: ~6 min·2 must-watch clips

Orange = the 6 minutes worth watchingFor the rest, the guide is enough
Segment guide · 6 segments
  1. 0:07 2:44Listen

    Making the prompt box strange again

    Opens with the core discomfort: you type a request into a small box to a superintelligence and then you wait, and it all feels completely normal. The next 20 minutes set out to make prompting feel unfamiliar and strange again, introducing the three concepts that run through the talk: channel, expression, protocol.

    The bottleneck is not model intelligence, it's the protocol between you and it.

    Purely spoken argument, no key visuals; fine for building the framework by ear.▶ Jump to 0:07
    Speaker · Ted Johnson
  2. 2:44 7:08Skim

    An interface inherited, not chosen

    Uses the keyboard — a layout patented around 1860, under constraints that haven't existed for a century — to show that the pieces of our interface are historical legacies. Then takes stock with the three concepts: expression exploded, the channel stayed the same, and the protocol is what lags.

    Nobody alive chose the way we interact today; we inherited it and then stopped noticing.

    The keyboard history comes with a few slides (one visual worth a glance at 164s); a quick look is enough, the argument carries by ear.▶ Jump to 2:44
    Speaker · Ted Johnson
  3. 7:08 10:50Listen

    The prompt is a punch card

    Takes apart what prompting really is: assemble the whole request, submit, wait, read, do it again — the same shape as punch card batch. Shrinking the wait from overnight to seconds created the illusion of interactivity, but the protocol didn't move, and voice input is just a shorter batch.

    Lower latency is not a new protocol: prompting is structurally still batch.

    Mostly conceptual argument with no demo footage; loses nothing if you listen on the commute.▶ Jump to 7:08
    Speaker · Ted Johnson
  4. 10:50 14:10Watch

    Deflating prompt engineering, and the first signs of real-time

    Points out that incantations like think step by step, give it examples and act as an expert are black-box tricks, the same mastery a punch card operator had in assembling a deck so the job wouldn't fail. Then the turn: Nvidia's Personaplex shows real turn-taking, stopping when interrupted, yielding, and picking the thread back up.

    Getting good at black-box tricks should bother us, not reassure us — and a real-time protocol is already technically possible.

    The Personaplex turn-taking demo from 805s is an important visual moment; watch while listening to feel the difference between speaking-and-listening-at-once and batch.▶ Jump to 10:50
    Speaker · Ted Johnson
  5. 14:10 16:50Watch

    The participatory AI meeting demo

    In a live multi-person meeting the AI acts as an equal participant: it labels each statement as a question, proposal or answer and creates goals to fulfill, taking a turn only when no one else is holding the floor. It surfaces the expense-approvals-versus-general-approval-workflow ambiguity, helps the room settle it, and carries out interjected requests to change the threshold and check whether the room is free.

    The choices of timing and modality can be lifted off the human — same model, different protocol, different product.

    The densest live-demo stretch of the talk, with three key visual moments in a row at 860/941/983s; you have to watch the screen to follow the multi-person conversation and the AI's interventions.▶ Jump to 14:10
    Speaker · Ted Johnson
  6. 16:50 20:10Listen

    Conclusion: AI is an interface technology

    Ties it together: computing has been a history of humans encoding intent for machines, and every generation of interface left behind a translation tax, a precision tax, a context tax, a repair tax. The design question has to become: what burden are we still putting on humans only because the machine used to be too limited to carry it?

    If a machine can finally understand more of what we mean, we should stop reshaping ourselves to be understood.

    A spoken close and a call to action with no visual dependency; good for listening through and sitting with the argument.▶ Jump to 16:50
    Speaker · Ted Johnson