{"course":{"slug":"ai-and-you","title":"AI, Safety and What Goes Wrong","url":"https://addaly.com/learn/ai-and-you","level":"beginner","track":"start"},"lesson":{"slug":"the-demo-and-the-machinery","title":"How a demo is built","url":"https://addaly.com/learn/ai-and-you/the-demo-and-the-machinery","position":6,"of":73,"minutes":3,"answer":"A demo shows what can happen, never how often it happens.","keyIdea":"A demo shows what can happen, never how often it happens."},"body":{"format":"markdown","text":"## Rock, paper, scissors\n\nOn 6 December 2023 Google launched Gemini with a six-minute video, \"Hands-on with Gemini\". In one scene a hand makes shapes on a table and the model says, unprompted, that the person is playing rock, paper, scissors. It looks like a live conversation with a machine that watches the world.\n\nGoogle's developer blog, published the same day, shows the real prompt. The model was given three still photographs at once, with the text \"What do you think I'm doing? Hint: It's a game.\" The video's description says latency was reduced and outputs were shortened for brevity. After Parmy Olson reported the gap in Bloomberg Opinion, Google said the prompts and outputs were real and that the video showed what such tools could look like.\n\n> Figure: The scene, and the prompt behind it. What the video showed: A hand makes shapes, one at a time; Gemini names the game, unprompted; A flowing, spoken conversation. What Google's developer blog showed: Three still photographs, given at once; Typed: What do you think I'm doing? Hint: It's a game.; Latency reduced, outputs shortened. Every prompt and reply was real. A viewer was shown a harder task, live recognition without a hint, than the one the model was set.\n\nThe demo still answered a different question from the one a viewer heard.\n\n## Possibility, not frequency\n\nA demo shows that something can happen. It never shows how often it happens.\n\nThe usual edits are worth naming, because once named you see them in every launch.\n\n**Best of many.** The output shown is the best of several tries.\n\n**The prompt is hidden.** A casual request on screen may stand for a long, tuned instruction. The Gemini hint did work the viewer never saw.\n\n**Chosen inputs.** Clean photographs, good light, a familiar layout.\n\n**Waiting cut out.** A reply that takes forty seconds feels different in a meeting.\n\n**No failures.** No demo shows the refusal, the timeout or the confident wrong answer.\n\n## One request that tells you most\n\nAsk, politely, to see three failures. A team that has deployed its tool keeps a folder of them: the refusal, the timeout, the confident wrong answer, and what it cost to catch each one. The same conversation usually brings out the rest, such as the success rate on inputs nobody chose and how much human work sits inside the process today.\n\n## Running your own trial\n\nHow to build a small test set from your own work is in [what a benchmark score leaves out](/learn/ai-and-you/the-benchmark-and-the-world). A demo adds one rule to it: fix the sample before you see any results, awkward cases included, and write down in advance what success means. Have the people who will use the tool run it, not its keenest supporter. Then time the whole job, checking included. A tool that drafts in four seconds and takes eleven minutes to check saves nothing if the job took you ten by hand."},"check":{"question":"A vendor shows flawless extraction from ten invoices. Which single request tells you most about whether it will work for you?","options":[{"text":"Ask which model it runs on and whether it is tuned for your sector","correct":false,"tempting":"Sounds technical, but the answer does not constrain how it performs on your documents."},{"text":"Ask for a benchmark comparison against rival extraction tools","correct":false,"tempting":"Moves the question to a score measured on someone else's documents, the transfer problem again."},{"text":"Run invoices you pick now, before any results, once each","correct":true},{"text":"Ask for the price per document at your expected volume","correct":false,"tempting":"Necessary for the budget. It says nothing about accuracy, which is what the demo implied."}],"why":"A sample fixed in advance, one attempt each, gives the frequency a demo leaves out."},"prev":{"slug":"when-the-failure-is-invisible","title":"Failure without an error message","url":"https://addaly.com/learn/ai-and-you/when-the-failure-is-invisible"},"next":{"slug":"automation-bias","title":"Why people believe the machine","url":"https://addaly.com/learn/ai-and-you/automation-bias"},"authorship":"Founder and series editor: Dr Mustafa Mun. Written by AI models to a written standard. Source and assessment review continues. Academic review invited.","review":{"status":"Draft, open for academic review"},"cite":"Addaly (2026) 'How a demo is built', in AI, Safety and What Goes Wrong. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-and-you/the-demo-and-the-machinery (Accessed: 21 November 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"How a demo is built","courseTitle":"AI, Safety and What Goes Wrong","module":{"index":1,"slug":"how-things-fail","title":"How these systems fail"},"answer":null,"keyIdea":"A demo shows what can happen, never how often it happens.","hero":{"kind":"figure","figure":{"t":"figure","label":"Figure","figure":{"caption":"Every prompt and reply was real. A viewer was shown a harder task, live recognition without a hint, than the one the model was set.","left":{"title":"What the video showed","items":["A hand makes shapes, one at a time","Gemini names the game, unprompted","A flowing, spoken conversation"]},"right":{"title":"What Google’s developer blog showed","items":["Three still photographs, given at once","Typed: What do you think I’m doing? Hint: It’s a game.","Latency reduced, outputs shortened"]},"title":"The scene, and the prompt behind it","type":"compare"},"alt":"Figure: The scene, and the prompt behind it. What the video showed: A hand makes shapes, one at a time; Gemini names the game, unprompted; A flowing, spoken conversation. What Google’s developer blog showed: Three still photographs, given at once; Typed: What do you think I’m doing? Hint: It’s a game.; Latency reduced, outputs shortened. Every prompt and reply was real. A viewer was shown a harder task, live recognition without a hint, than the one the model was set."}},"blocks":[{"t":"h2","id":"rock-paper-scissors","runs":[{"text":"Rock, paper, scissors"}]},{"t":"p","runs":[{"text":"On 6 December 2023 Google launched Gemini with a six-minute video, “Hands-on with Gemini”. In one scene a hand makes shapes on a table and the model says, unprompted, that the person is playing rock, paper, scissors. It looks like a live conversation with a machine that watches the world."}]},{"t":"quick","prompt":"How was that scene made?","options":["Live, exactly as shown","Live, with the pauses cut","From still photos and a typed hint"],"answer":2,"reveal":"From still photos and a typed hint. Google’s own developer blog says so."},{"t":"p","runs":[{"text":"Google’s developer blog, published the same day, shows the real prompt. The model was given three still photographs at once, with the text “What do you think I’m doing? Hint: It’s a game.” The video’s description says latency was reduced and outputs were shortened for brevity. After Parmy Olson reported the gap in Bloomberg Opinion, Google said the prompts and outputs were real and that the video showed what such tools could look like."}]},{"t":"p","runs":[{"text":"The demo still answered a different question from the one a viewer heard."}]},{"t":"h2","id":"possibility-not-frequency","runs":[{"text":"Possibility, not frequency"}]},{"t":"p","runs":[{"text":"A demo shows that something can happen. It never shows how often it happens."}]},{"t":"p","runs":[{"text":"The usual edits are worth naming, because once named you see them in every launch."}]},{"t":"p","runs":[{"text":"Best of many.","b":true},{"text":" The output shown is the best of several tries."}]},{"t":"p","runs":[{"text":"The prompt is hidden.","b":true},{"text":" A casual request on screen may stand for a long, tuned instruction. The Gemini hint did work the viewer never saw."}]},{"t":"p","runs":[{"text":"Chosen inputs.","b":true},{"text":" Clean photographs, good light, a familiar layout."}]},{"t":"p","runs":[{"text":"Waiting cut out.","b":true},{"text":" A reply that takes forty seconds feels different in a meeting."}]},{"t":"p","runs":[{"text":"No failures.","b":true},{"text":" No demo shows the refusal, the timeout or the confident wrong answer."}]},{"t":"h2","id":"one-request-that-tells-you-most","runs":[{"text":"One request that tells you most"}]},{"t":"p","runs":[{"text":"Ask, politely, to see three failures. A team that has deployed its tool keeps a folder of them: the refusal, the timeout, the confident wrong answer, and what it cost to catch each one. The same conversation usually brings out the rest, such as the success rate on inputs nobody chose and how much human work sits inside the process today."}]},{"t":"h2","id":"running-your-own-trial","runs":[{"text":"Running your own trial"}]},{"t":"p","runs":[{"text":"How to build a small test set from your own work is in "},{"text":"what a benchmark score leaves out","href":"https://addaly.com/learn/ai-and-you/the-benchmark-and-the-world"},{"text":". A demo adds one rule to it: fix the sample before you see any results, awkward cases included, and write down in advance what success means. Have the people who will use the tool run it, not its keenest supporter. Then time the whole job, checking included. A tool that drafts in four seconds and takes eleven minutes to check saves nothing if the job took you ten by hand."}]}],"check":{"question":[{"text":"A vendor shows flawless extraction from ten invoices. Which single request tells you most about whether it will work for you?"}],"options":[{"text":[{"text":"Ask which model it runs on and whether it is tuned for your sector"}],"correct":false,"tempting":[{"text":"Sounds technical, but the answer does not constrain how it performs on your documents."}]},{"text":[{"text":"Ask for a benchmark comparison against rival extraction tools"}],"correct":false,"tempting":[{"text":"Moves the question to a score measured on someone else’s documents, the transfer problem again."}]},{"text":[{"text":"Run invoices you pick now, before any results, once each"}],"correct":true,"tempting":null},{"text":[{"text":"Ask for the price per document at your expected volume"}],"correct":false,"tempting":[{"text":"Necessary for the budget. It says nothing about accuracy, which is what the demo implied."}]}],"order":[3,0,2,1],"why":[{"text":"A sample fixed in advance, one attempt each, gives the frequency a demo leaves out."}]},"prev":{"slug":"when-the-failure-is-invisible","title":"Failure without an error message","minutes":3},"next":{"slug":"automation-bias","title":"Why people believe the machine","minutes":4}}}