{
  "video": {
    "id": "hVJOnuhFmTA",
    "title": "The Prompt Is Still a Punch Card - Ted Johnson, JoinIn AI",
    "duration": 1213,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": "Online"
  },
  "analysis": {
    "video_id": "hVJOnuhFmTA",
    "title": "The Prompt Is Still a Punch Card - Ted Johnson, JoinIn AI",
    "one_liner": "Ted Johnson argues that LLMs blew open what we can *express* to a machine, but the *protocol* — package a complete turn, hit submit, wait — is still the punch card's batch, and that mismatch, not the user, is why AI feels like work.",
    "summary": "Johnson separates an interface into three parts: the channel (the physical transport — keyboard, microphone, screen), expression (how much of your meaning the interface lets through), and the protocol (the shape and rules of the exchange). With LLMs, expression exploded in three years while the channel stayed the same for 15 to 180 years and the protocol didn't move at all: you still assemble a whole request, submit, and wait, exactly as with a punch-card deck — and batch itself was inherited from the weaving loom. He argues prompt engineering is just a flattering name for the rules of packaging a good batch job, shows speech-to-speech models failing because the protocol has exactly one slot, and demos JoinIn's multi-party system that tracks who is speaking, labels utterances, and takes a turn only when no one holds the floor.",
    "key_points": [
      "Three-part frame: channel = the physical medium (keyboard, mic, screen, punch card, prompt box) and its bandwidth; expression = how much of your actual meaning gets through; protocol = the rules and shape of the exchange. Carrying more signal is bandwidth, not understanding.",
      "The QWERTY layout's patent dates to about 1860 and could just as easily have been the Hansen Writing Ball; people still fight it with Dvorak, Colemak and multi-key thumb clusters. 'Nobody alive chose it. We all inherited it. And then we stopped noticing.'",
      "Expression progressed in fixed-vocabulary steps — assembly's few dozen opcodes, shell commands and flags, composable primitives in modern languages — each forcing you to pick intent off a menu the machine accepts. Natural language blew the menu open; the channel never changed.",
      "Prompting is batch: assemble the whole request, submit, wait, find one thing wrong, resubmit. Interactive features are 'batch with interactive sprinkles'; shrinking the wait from overnight to seconds fooled us into calling it interactive. Voice doesn't help — it's transcribed into the same box and submitted.",
      "'Prompt engineering' is rebranded deck assembly: think step-by-step, give examples, be an expert / don't be an expert, paste more context, paste less, only talk to it through markdown. Getting good at prompting black boxes should bother us, not reassure us.",
      "Model capacity (reasoning, speech, vision, memory, planning) curves up while the interface protocol stays flat; the human still decides what context matters, chooses timing, notices ambiguity and repairs output. When it fails, users blame themselves.",
      "Failure demo: his co-founder asked a frontier voice mode 'When is the next Timberwolves game?' — answered fine — then said 'Hey Ted, come on in' to a person in the room, and the model replied 'Sure, I'm here. What's on your mind?' A protocol with exactly one slot has no concept of who is speaking or whether words were meant for it.",
      "Field is converging: OpenAI released GPT real-time 2 in late May and its voice mode now backchannels ('Mhm', 'right'); Nvidia's PersonaPlex research model yields to an interruption mid-sentence and picks the thread back up. But making listening noises isn't the same as knowing who's in the room.",
      "JoinIn's demo sits in a multi-person requirements meeting: a utility-driven model labels each participant statement as question, proposal or answer, forms goals, and takes a turn only when no one else is speaking or holding the floor — resolving scope (expense approvals only, access requests out, $5,000 second-approver threshold later changed to $10,000) without anyone writing a prompt or hitting submit."
    ],
    "takeaways": [
      "Ask the design question that replaces prompt tuning: 'what burden are we still putting on humans only because the machine used to be too limited to carry that burden itself?'",
      "Stop treating prompt-engineering skill as mastery; it's punch-card-operator skill, and every incantation you trade is evidence the protocol, not the user, is broken.",
      "Build interfaces that participate — ask a follow-up, clarify mid-thought, notice something missing and say so — instead of waiting for a complete packaged turn.",
      "Solve turn-taking and addressing, not just latency: know who is speaking, whether the words were meant for the AI, and when no one holds the floor. Backchannels alone don't buy that.",
      "Let the AI choose channel and timing. The right answer isn't always chat, isn't always voice, isn't a wall of markdown — it can be a question, a pause, a sketch, a checklist, a quiet aside, or saying nothing at all."
    ],
    "topics": [
      "human-computer interaction",
      "voice",
      "speech-to-speech",
      "conversational ai",
      "interface design",
      "prompt engineering",
      "turn-taking",
      "multi-party meetings"
    ],
    "tools": [
      "JoinIn AI",
      "ChatGPT",
      "GPT-1",
      "GPT-2",
      "OpenAI",
      "GPT real-time 2",
      "Nvidia PersonaPlex",
      "Dvorak",
      "Colemak",
      "Hansen Writing Ball",
      "MIT (Sherry Turkle)"
    ],
    "quotes": [
      {
        "text": "Channel stayed the same. Expression exploded in the last 3 years with LLMs, but the protocol, prompting, is the protocol of a punch card.",
        "at": "06:25",
        "url": "https://www.youtube.com/watch?v=hVJOnuhFmTA&t=385s"
      },
      {
        "text": "We trade incantations. We've learned the magic words. That's the illusion. It feels like mastery, but it's the same sort of mastery a punch card operator had.",
        "at": "08:32",
        "url": "https://www.youtube.com/watch?v=hVJOnuhFmTA&t=512s"
      },
      {
        "text": "We are not bad at using AI. We are being asked to operate a brand new kind of intelligence through a protocol of a punch card. The mismatch isn't the user, it's the interface.",
        "at": "10:55",
        "url": "https://www.youtube.com/watch?v=hVJOnuhFmTA&t=655s"
      },
      {
        "text": "It needs to become what burden are we still putting on humans only because the machine used to be too limited to carry that burden itself.",
        "at": "18:19",
        "url": "https://www.youtube.com/watch?v=hVJOnuhFmTA&t=1099s"
      }
    ],
    "words": 3683
  },
  "summary_url": "/#hVJOnuhFmTA",
  "transcript": {
    "html": "/transcripts/hVJOnuhFmTA.html",
    "txt": "/transcripts/hVJOnuhFmTA.txt",
    "vtt": "/transcripts/hVJOnuhFmTA.vtt"
  }
}