{
  "runtime": "Chromium / WebAssembly q8",
  "created": "2026-09-29T07:49:30.205Z",
  "runs": {
    "labels": {
      "results": [
        {
          "label": "a photo of a dog",
          "cosine": 0.2675817522627175,
          "share": 0.9932887342028608
        },
        {
          "label": "a photo of a bird",
          "cosine": 0.21208156277286253,
          "share": 0.0038612943163369067
        },
        {
          "label": "a photo of a cat",
          "cosine": 0.20904462822396316,
          "share": 0.0028499714808022563
        }
      ],
      "runtime": "Live browser inference \u00b7 WebAssembly \u00b7 q8",
      "experiment": {
        "id": "labels",
        "title": "Write a new classifier",
        "short": "New labels",
        "mode": "classify",
        "image": "newfoundland",
        "candidates": [
          "a photo of a dog",
          "a photo of a cat",
          "a photo of a bird"
        ],
        "question": "Can the same photograph answer a question the old classifier head never had a label for?",
        "try": "Replace these with Newfoundland, pug, and Persian. Then try indoors and outdoors.",
        "lesson": "The description supplies a comparison vector. Changing the candidates does not train or replace the image encoder.",
        "sourceLabel": "CLIP \u00b7 Radford et al., ICML 2021",
        "sourceURL": "https://proceedings.mlr.press/v139/radford21a.html"
      }
    },
    "attributes": {
      "results": [
        {
          "label": "a cup of coffee",
          "cosine": 0.28648920073146844,
          "share": 0.9315512681149307
        },
        {
          "label": "a cup of tea",
          "cosine": 0.2570337071206367,
          "share": 0.04897460339457629
        },
        {
          "label": "a glass of orange juice",
          "cosine": 0.24576773904544097,
          "share": 0.015874357287911072
        },
        {
          "label": "a bowl of soup",
          "cosine": 0.23092939134989823,
          "share": 0.003599771202581981
        }
      ],
      "runtime": "Live browser inference \u00b7 WebAssembly \u00b7 q8",
      "experiment": {
        "id": "attributes",
        "title": "Ask about a property",
        "short": "Attributes",
        "mode": "classify",
        "image": "coffee",
        "candidates": [
          "a cup of coffee",
          "a cup of tea",
          "a glass of orange juice",
          "a bowl of soup"
        ],
        "question": "Can a description be more specific than an object category?",
        "try": "Compare \u2018a white cup\u2019, \u2018a red cup\u2019, and \u2018a blue cup\u2019. Keep the photograph fixed.",
        "lesson": "Natural language lets us propose attributes, materials and contexts. These are hypotheses to test, not guaranteed abilities.",
        "sourceLabel": "CLIP \u00b7 Radford et al., ICML 2021",
        "sourceURL": "https://proceedings.mlr.press/v139/radford21a.html"
      }
    },
    "search": {
      "results": [
        {
          "label": "red-mug",
          "cosine": 0.2581550139215767,
          "share": 0.5638736049081738
        },
        {
          "label": "blue-mug",
          "cosine": 0.2502645857519802,
          "share": 0.2561561718203908
        },
        {
          "label": "coffee",
          "cosine": 0.245574995894994,
          "share": 0.16026493970630723
        },
        {
          "label": "Persian",
          "cosine": 0.21284188666702775,
          "share": 0.006070963938408999
        },
        {
          "label": "xray-device",
          "cosine": 0.2037125690266547,
          "share": 0.002436556159088521
        },
        {
          "label": "xray-before",
          "cosine": 0.2032279499543489,
          "share": 0.002321291530325396
        },
        {
          "label": "chest-xray",
          "cosine": 0.20291448192186823,
          "share": 0.0022496551136368453
        },
        {
          "label": "chelsea",
          "cosine": 0.1994301877523806,
          "share": 0.001587796974261997
        },
        {
          "label": "pug",
          "cosine": 0.19397261269742513,
          "share": 0.0009199738798540682
        },
        {
          "label": "portrait-hat",
          "cosine": 0.19122201802846756,
          "share": 0.0006987449598030684
        },
        {
          "label": "kiln",
          "cosine": 0.19070805088004414,
          "share": 0.0006637390640673157
        },
        {
          "label": "newfoundland",
          "cosine": 0.189043624654032,
          "share": 0.0005619688795182392
        },
        {
          "label": "portrait",
          "cosine": 0.18648373315063155,
          "share": 0.0004350484148919829
        },
        {
          "label": "kiln-chimney",
          "cosine": 0.18388442567683405,
          "share": 0.0003354680014374803
        },
        {
          "label": "retriever-sketch",
          "cosine": 0.18223655301796624,
          "share": 0.00028450172312635284
        },
        {
          "label": "fields",
          "cosine": 0.18196623723492256,
          "share": 0.00027691420581893564
        },
        {
          "label": "kiln-base",
          "cosine": 0.18163272658491375,
          "share": 0.0002678311291856165
        },
        {
          "label": "warehouse",
          "cosine": 0.1795926410131533,
          "share": 0.00021840433875110072
        },
        {
          "label": "retriever-photo",
          "cosine": 0.17837449579071205,
          "share": 0.00019335610404998312
        },
        {
          "label": "solar",
          "cosine": 0.1732177243169897,
          "share": 0.00011545218252587747
        },
        {
          "label": "rocket",
          "cosine": 0.16156403178880607,
          "share": 3.5998856977072235e-05
        },
        {
          "label": "astronaut",
          "cosine": 0.16026646030264993,
          "share": 3.161810939927729e-05
        }
      ],
      "runtime": "Live browser inference \u00b7 WebAssembly \u00b7 q8",
      "experiment": {
        "id": "search",
        "title": "Find a photo with words",
        "short": "Text \u2192 image",
        "mode": "search",
        "query": "something to drink",
        "question": "Could you find a photo without knowing its filename?",
        "try": "Search for \u2018a pet\u2019, \u2018a spacecraft taking off\u2019, then \u2018a laptop\u2019. Is the best available result always relevant?",
        "lesson": "Encode the gallery once. A new query needs only a text embedding and dot products. Search always has a nearest item even when nothing matches.",
        "sourceLabel": "Supabase \u00b7 image search, 6:54",
        "sourceURL": "https://www.youtube.com/watch?v=S7VZErcTN5Y&t=414s"
      }
    },
    "captions": {
      "results": [
        {
          "label": "a rocket launching into the night sky",
          "cosine": 0.25767250541351344,
          "share": 0.9895138651327461
        },
        {
          "label": "fireworks above a city",
          "cosine": 0.20439890063499955,
          "share": 0.00480594310710231
        },
        {
          "label": "a tall building beside a river",
          "cosine": 0.20276106043524772,
          "share": 0.004079886948696526
        },
        {
          "label": "an airplane on a runway",
          "cosine": 0.1934023088240643,
          "share": 0.0016003048114551616
        }
      ],
      "runtime": "Live browser inference \u00b7 WebAssembly \u00b7 q8",
      "experiment": {
        "id": "captions",
        "title": "Choose a description",
        "short": "Image \u2192 text",
        "mode": "classify",
        "image": "rocket",
        "candidates": [
          "a rocket launching into the night sky",
          "a tall building beside a river",
          "fireworks above a city",
          "an airplane on a runway"
        ],
        "question": "Is selecting a sentence the same as writing one?",
        "try": "Write your own competing descriptions. Include one that mentions only part of the scene.",
        "lesson": "CLIP ranks supplied text. It does not generate captions. A caption generator requires additional machinery.",
        "sourceLabel": "CLIP \u00b7 Radford et al., ICML 2021",
        "sourceURL": "https://proceedings.mlr.press/v139/radford21a.html"
      }
    },
    "neighbors": {
      "results": [
        {
          "label": "chelsea",
          "cosine": 0.8322531474649447,
          "share": 0.9999214833686297
        },
        {
          "label": "pug",
          "cosine": 0.7364182442485346,
          "share": 6.885079118982087e-05
        },
        {
          "label": "newfoundland",
          "cosine": 0.7167711164139413,
          "share": 9.65261567619247e-06
        },
        {
          "label": "retriever-photo",
          "cosine": 0.6500406967995543,
          "share": 1.2206178411170784e-08
        },
        {
          "label": "retriever-sketch",
          "cosine": 0.6247364797282084,
          "share": 9.7192226068317e-10
        },
        {
          "label": "coffee",
          "cosine": 0.5886187819144065,
          "share": 2.6245801703995373e-11
        },
        {
          "label": "red-mug",
          "cosine": 0.5858700291506097,
          "share": 1.993806591873715e-11
        },
        {
          "label": "blue-mug",
          "cosine": 0.5407614414998856,
          "share": 2.1909978793088663e-13
        },
        {
          "label": "xray-before",
          "cosine": 0.47850126813483534,
          "share": 4.332284647921412e-16
        },
        {
          "label": "xray-device",
          "cosine": 0.4738928360371655,
          "share": 2.732595323133393e-16
        },
        {
          "label": "portrait",
          "cosine": 0.43706505418182745,
          "share": 6.873296464503966e-18
        },
        {
          "label": "chest-xray",
          "cosine": 0.42910630622804646,
          "share": 3.1011376575949113e-18
        },
        {
          "label": "portrait-hat",
          "cosine": 0.4265420518860343,
          "share": 2.399699680232947e-18
        },
        {
          "label": "fields",
          "cosine": 0.3913189753731008,
          "share": 7.086603218159282e-20
        },
        {
          "label": "warehouse",
          "cosine": 0.37798708752713195,
          "share": 1.868278365028206e-20
        },
        {
          "label": "rocket",
          "cosine": 0.37160062601503596,
          "share": 9.864636468579841e-21
        },
        {
          "label": "astronaut",
          "cosine": 0.32508619535772787,
          "share": 9.418571281241472e-23
        },
        {
          "label": "kiln",
          "cosine": 0.31927513938504587,
          "share": 5.2676155937271395e-23
        },
        {
          "label": "kiln-base",
          "cosine": 0.2961642150196241,
          "share": 5.222992260339239e-24
        },
        {
          "label": "solar",
          "cosine": 0.2907977161906282,
          "share": 3.0539034276794566e-24
        },
        {
          "label": "kiln-chimney",
          "cosine": 0.28862523794829104,
          "share": 2.457569271606588e-24
        }
      ],
      "runtime": "Live browser inference \u00b7 WebAssembly \u00b7 q8",
      "experiment": {
        "id": "neighbors",
        "title": "Find visually related images",
        "short": "Image \u2192 image",
        "mode": "neighbors",
        "image": "Persian",
        "question": "Can we search without any words at all?",
        "try": "Choose a cat, dog or portrait as the query. Predict the nearest other image before running.",
        "lesson": "The image encoder is useful on its own. Similarity can suggest groups or possible duplicates; it cannot certify that two files are duplicates.",
        "sourceLabel": "Roboflow \u00b7 embedding analysis, 10:40 and 17:22",
        "sourceURL": "https://www.youtube.com/watch?v=YxJkE6FvGF4&t=640s"
      }
    },
    "hat": {
      "results": [
        {
          "label": "hat",
          "cosine": 0.09357992035256416
        },
        {
          "label": "boat",
          "cosine": 0.02688376639684986
        },
        {
          "label": "cup",
          "cosine": 0.006702530744190247
        },
        {
          "label": "cat",
          "cosine": 0.0003820618371524412
        }
      ],
      "runtime": "Live browser inference \u00b7 WebAssembly \u00b7 q8",
      "experiment": {
        "id": "hat",
        "title": "What changed?",
        "short": "The hat experiment",
        "mode": "difference",
        "image": "portrait",
        "second": "portrait-hat",
        "candidates": [
          "hat",
          "cup",
          "cat",
          "boat"
        ],
        "question": "Does \u2018with hat\u2019 minus \u2018without hat\u2019 point towards the word \u2018hat\u2019?",
        "try": "Swap the images. Then try blue mug minus red mug with the words blue, red, cup, and hat.",
        "lesson": "Inspired by CodeEmporium\u2019s hat experiment, using our own generated pair. This is an exploratory direction, not a trained change detector. Subtle image differences remain. Cosine is not a probability of the change.",
        "sourceLabel": "CodeEmporium \u00b7 hat experiment, 7:25",
        "sourceURL": "https://www.youtube.com/watch?v=FCDKn-vpn_o&t=445s"
      }
    },
    "competition": {
      "results": [
        {
          "label": "a photo of a car",
          "cosine": 0.21538369364665655,
          "share": 0.4445855611424471
        },
        {
          "label": "a photo of a bird",
          "cosine": 0.21208156277286253,
          "share": 0.31955501105116657
        },
        {
          "label": "a photo of a cat",
          "cosine": 0.20904462822396316,
          "share": 0.23585942780638636
        }
      ],
      "runtime": "Live browser inference \u00b7 WebAssembly \u00b7 q8",
      "experiment": {
        "id": "competition",
        "title": "Remove the right answer",
        "short": "Candidate trap",
        "mode": "classify",
        "image": "newfoundland",
        "candidates": [
          "a photo of a cat",
          "a photo of a bird",
          "a photo of a car"
        ],
        "question": "What happens when every offered label is wrong?",
        "try": "Run once, then add \u2018a photo of a dog\u2019. Compare each cosine with its share of the softmax.",
        "lesson": "Softmax distributes 100% across the choices you provide. It is not calibrated confidence that the best label is true.",
        "sourceLabel": "CLIP \u00b7 Radford et al., ICML 2021",
        "sourceURL": "https://proceedings.mlr.press/v139/radford21a.html"
      }
    },
    "negation": {
      "results": [
        {
          "label": "a photo of a cat",
          "cosine": 0.27420925339190855,
          "share": 0.5944378716330678
        },
        {
          "label": "a photo without a cat",
          "cosine": 0.27029147200790826,
          "share": 0.4017532390859464
        },
        {
          "label": "a photo of a dog",
          "cosine": 0.22370646753099988,
          "share": 0.0038088892809858624
        }
      ],
      "runtime": "Live browser inference \u00b7 WebAssembly \u00b7 q8",
      "experiment": {
        "id": "negation",
        "title": "Try to break the model",
        "short": "Language limits",
        "mode": "classify",
        "image": "Persian",
        "candidates": [
          "a photo of a cat",
          "a photo without a cat",
          "a photo of a dog"
        ],
        "question": "Does the word \u2018without\u2019 reliably reverse what a phrase means?",
        "try": "Try negation, counting, and two descriptions with reversed relations. Save a failure as well as a success.",
        "lesson": "A flexible vocabulary does not guarantee precise compositional reasoning. Report what this checkpoint does on these examples, not universal claims.",
        "sourceLabel": "CLIP \u00b7 Radford et al., ICML 2021",
        "sourceURL": "https://proceedings.mlr.press/v139/radford21a.html"
      }
    },
    "sustainability": {
      "results": [
        {
          "label": "an aerial image of a brick kiln",
          "cosine": 0.31949446898847267,
          "share": 0.8270048087328227
        },
        {
          "label": "an aerial image of agricultural fields",
          "cosine": 0.2969633769987064,
          "share": 0.08689507103533141
        },
        {
          "label": "an aerial image of industrial warehouses",
          "cosine": 0.2904258569791173,
          "share": 0.04519332092358141
        },
        {
          "label": "an aerial image of a solar farm",
          "cosine": 0.2894293268065749,
          "share": 0.040906799308264276
        }
      ],
      "runtime": "Live browser inference \u00b7 WebAssembly \u00b7 q8",
      "experiment": {
        "id": "sustainability",
        "title": "Search an overhead landscape",
        "short": "Sustainability",
        "mode": "classify",
        "image": "kiln",
        "candidates": [
          "an aerial image of a brick kiln",
          "an aerial image of a solar farm",
          "an aerial image of agricultural fields",
          "an aerial image of industrial warehouses"
        ],
        "question": "What would a sustainability researcher want to retrieve?",
        "lesson": "The search idea transfers to satellite tiles, but synthetic images are not evidence of field performance. Real deployment needs georeferenced imagery, independent labels, and evaluation across regions and resolutions.",
        "try": "Compare all four overhead scenes. Try \u2018brick kiln\u2019 alone, then a longer description. These scenes are synthetic.",
        "sourceLabel": "CLIP \u00b7 Radford et al., ICML 2021",
        "sourceURL": "https://proceedings.mlr.press/v139/radford21a.html"
      }
    },
    "healthcare": {
      "results": [
        {
          "label": "a chest X-ray",
          "cosine": 0.2631089829216429,
          "share": 0.9105742047464289
        },
        {
          "label": "an X-ray of a hand",
          "cosine": 0.23356698548065039,
          "share": 0.047459449449999226
        },
        {
          "label": "a photograph of a person",
          "cosine": 0.23176420207567217,
          "share": 0.03963043208607247
        },
        {
          "label": "a brain MRI scan",
          "cosine": 0.20345226025701624,
          "share": 0.0023359137174995383
        }
      ],
      "runtime": "Live browser inference \u00b7 WebAssembly \u00b7 q8",
      "experiment": {
        "id": "healthcare",
        "title": "Recognize the kind of medical image",
        "short": "Healthcare",
        "mode": "classify",
        "image": "chest-xray",
        "candidates": [
          "a chest X-ray",
          "an X-ray of a hand",
          "a brain MRI scan",
          "a photograph of a person"
        ],
        "question": "Does recognizing an imaging modality mean the model can diagnose a disease?",
        "lesson": "This public radiograph supports a modality-matching lesson, not diagnosis. General CLIP scores are not clinical evidence; medical tasks need domain-specific validation.",
        "try": "Compare modality descriptions first. Notice how different this image is from an everyday photograph.",
        "sourceLabel": "CLIP \u00b7 Radford et al., ICML 2021",
        "sourceURL": "https://proceedings.mlr.press/v139/radford21a.html"
      }
    },
    "sketch": {
      "results": [
        {
          "label": "a drawing of a dog",
          "cosine": 0.31751787158837835,
          "share": 0.9834518002749116
        },
        {
          "label": "a drawing of a horse",
          "cosine": 0.271738091741548,
          "share": 0.010105609276368228
        },
        {
          "label": "a drawing of a fox",
          "cosine": 0.26368695455708724,
          "share": 0.004517582136358367
        },
        {
          "label": "a drawing of a cat",
          "cosine": 0.2551564881641323,
          "share": 0.0019250083123618482
        }
      ],
      "runtime": "Live browser inference \u00b7 WebAssembly \u00b7 q8",
      "experiment": {
        "id": "sketch",
        "title": "Keep the concept; change the style",
        "short": "Photo \u2192 sketch",
        "mode": "classify",
        "image": "retriever-sketch",
        "candidates": [
          "a drawing of a dog",
          "a drawing of a cat",
          "a drawing of a horse",
          "a drawing of a fox"
        ],
        "question": "Can the model recognize a dog when the texture of a photograph disappears?",
        "lesson": "This is a small domain-shift experiment. A success here does not establish robustness to every drawing style.",
        "try": "Switch between the generated photo and sketch. Replace \u2018a drawing of\u2019 with \u2018a photo of\u2019 and compare.",
        "sourceLabel": "CLIP \u00b7 Radford et al., ICML 2021",
        "sourceURL": "https://proceedings.mlr.press/v139/radford21a.html"
      }
    },
    "color-change": {
      "results": [
        {
          "label": "blue",
          "cosine": 0.1180093869629936
        },
        {
          "label": "cup",
          "cosine": -0.013700523127914687
        },
        {
          "label": "hat",
          "cosine": -0.01950907051798153
        },
        {
          "label": "red",
          "cosine": -0.18614466828555015
        }
      ],
      "runtime": "Live browser inference \u00b7 WebAssembly \u00b7 q8",
      "experiment": {
        "id": "color-change",
        "title": "A second subtraction experiment",
        "short": "Color change",
        "mode": "difference",
        "image": "red-mug",
        "second": "blue-mug",
        "candidates": [
          "blue",
          "red",
          "cup",
          "hat"
        ],
        "question": "Does the embedding difference capture a color change as well as an added object?",
        "lesson": "Difference directions depend on images, normalization and wording. They are exploratory measurements, not guaranteed semantic algebra.",
        "try": "Reverse the pair. Try \u2018a blue mug\u2019 and \u2018a red mug\u2019 instead of single words.",
        "sourceLabel": "Extension of the CodeEmporium subtraction idea",
        "sourceURL": "https://www.youtube.com/watch?v=FCDKn-vpn_o&t=445s"
      }
    },
    "chimney": {
      "results": [
        {
          "label": "chimney",
          "cosine": 0.10547765895127056
        },
        {
          "label": "tree",
          "cosine": 0.051540696049011514
        },
        {
          "label": "truck",
          "cosine": 0.015133236612474517
        },
        {
          "label": "solar panel",
          "cosine": 0.003646207183224487
        }
      ],
      "runtime": "Live browser inference \u00b7 WebAssembly \u00b7 q8",
      "experiment": {
        "id": "chimney",
        "title": "Add a chimney",
        "short": "Chimney change",
        "mode": "difference",
        "image": "kiln-base",
        "second": "kiln-chimney",
        "candidates": [
          "chimney",
          "tree",
          "truck",
          "solar panel"
        ],
        "question": "Does adding a chimney produce a useful direction?",
        "lesson": "Synthetic paired scenes test a hypothesis. Shadows and fine details also change; this is not a causal proof or a validated kiln detector.",
        "sourceLabel": "Our extension of CodeEmporium\u2019s subtraction experiment",
        "sourceURL": "https://www.youtube.com/watch?v=FCDKn-vpn_o&t=445s",
        "try": "Predict the winner, run, then reverse the pair. Edit the candidate words and save the results."
      }
    },
    "medical-change": {
      "results": [
        {
          "label": "hat",
          "cosine": 0.02774713378842503
        },
        {
          "label": "stethoscope",
          "cosine": 0.003523663675473864
        },
        {
          "label": "pacemaker",
          "cosine": 0.0018934560349264445
        },
        {
          "label": "rib",
          "cosine": -0.025711801821476178
        }
      ],
      "runtime": "Live browser inference \u00b7 WebAssembly \u00b7 q8",
      "experiment": {
        "id": "medical-change",
        "title": "Add a visible medical device",
        "short": "Device change",
        "mode": "difference",
        "image": "xray-before",
        "second": "xray-device",
        "candidates": [
          "pacemaker",
          "rib",
          "stethoscope",
          "hat"
        ],
        "question": "Does the difference emphasize the added device?",
        "lesson": "Both radiographs are generated illustrations. Anatomy and device placement are approximate. This measures concept matching on synthetic pictures, not diagnostic or clinical performance.",
        "sourceLabel": "Our extension of CodeEmporium\u2019s subtraction experiment",
        "sourceURL": "https://www.youtube.com/watch?v=FCDKn-vpn_o&t=445s",
        "try": "Predict the winner, run, then reverse the pair. Edit the candidate words and save the results."
      }
    },
    "style-change": {
      "results": [
        {
          "label": "pencil sketch",
          "cosine": 0.138658188917058
        },
        {
          "label": "cat",
          "cosine": -0.0032619967625495596
        },
        {
          "label": "dog",
          "cosine": -0.006325334968683794
        },
        {
          "label": "photograph",
          "cosine": -0.014834340161318743
        }
      ],
      "runtime": "Live browser inference \u00b7 WebAssembly \u00b7 q8",
      "experiment": {
        "id": "style-change",
        "title": "Change style, keep the dog",
        "short": "Style change",
        "mode": "difference",
        "image": "retriever-photo",
        "second": "retriever-sketch",
        "candidates": [
          "pencil sketch",
          "photograph",
          "dog",
          "cat"
        ],
        "question": "Does subtracting the photo from the sketch emphasize style?",
        "lesson": "This subtraction includes background and rendering changes. Compare original image-to-text matching with the difference direction; a successful guess does not isolate a causal concept.",
        "sourceLabel": "Our extension of CodeEmporium\u2019s subtraction experiment",
        "sourceURL": "https://www.youtube.com/watch?v=FCDKn-vpn_o&t=445s",
        "try": "Predict the winner, run, then reverse the pair. Edit the candidate words and save the results."
      }
    }
  }
}