charts: 6970
Data license: CC-BY
This data as json
id | slug | type | config | createdAt | updatedAt | lastEditedAt | publishedAt | lastEditedByUserId | publishedByUserId | isIndexable | title | subtitle | note | title_plus_variant | configWithDefaults |
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
6970 | ai-performance-knowledge-tests-vs-training-computation | ScatterPlot | { "id": 6970, "map": { "columnSlug": "737492" }, "note": "The values for training computation are estimates and come with some uncertainty, especially for models for which only minimal information has been disclosed, such as GPT-4.", "slug": "ai-performance-knowledge-tests-vs-training-computation", "type": "ScatterPlot", "title": "Artificial intelligence: Performance on knowledge tests vs. training computation", "xAxis": { "max": 25000000000, "scaleType": "log", "canChangeScaleType": true }, "yAxis": { "max": 100, "min": 0 }, "$schema": "https://files.ourworldindata.org/schemas/grapher-schema.004.json", "version": 57, "subtitle": "Performance on knowledge tests is measured with the [MMLU benchmark](#dod:ai-MMLU), here with 5-shot learning, which gauges a model\u2019s accuracy after receiving only five examples for each task. Training computation is measured in total petaFLOP, which is 10\u00b9\u2075 [floating-point operations](#dod:flop).", "originUrl": "https://ourworldindata.org/artificial-intelligence", "colorScale": { "legendDescription": "Developer of AI system" }, "dimensions": [ { "property": "y", "variableId": 820642 }, { "property": "x", "variableId": 820644 }, { "property": "color", "variableId": 820643 } ], "entityType": "system", "isPublished": true, "hideTimeline": true, "comparisonLines": [ { "label": "Expert human performance", "yEquals": "90" }, {} ], "entityTypePlural": "systems", "hideAnnotationFieldsInTitle": { "time": true } } |
2023-07-14 21:53:53 | 2024-04-08 12:13:54 | 2024-03-10 17:04:36 | 2023-08-30 14:49:44 | 72 | 46 | 1 | Artificial intelligence: Performance on knowledge tests vs. training computation | Performance on knowledge tests is measured with the [MMLU benchmark](#dod:ai-MMLU), here with 5-shot learning, which gauges a model’s accuracy after receiving only five examples for each task. Training computation is measured in total petaFLOP, which is 10¹⁵ [floating-point operations](#dod:flop). | The values for training computation are estimates and come with some uncertainty, especially for models for which only minimal information has been disclosed, such as GPT-4. | Artificial intelligence: Performance on knowledge tests vs. training computation () | { "$schema": "https://files.ourworldindata.org/schemas/grapher-schema.004.json", "map": { "projection": "World", "hideTimeline": false, "colorScale": { "baseColorScheme": "default", "equalSizeBins": true, "binningStrategy": "ckmeans", "customNumericColorsActive": false, "colorSchemeInvert": false, "binningStrategyBinCount": 5 }, "timeTolerance": 0, "toleranceStrategy": "closest", "tooltipUseCustomLabels": false, "time": "latest" }, "maxTime": "latest", "baseColorScheme": "default", "yAxis": { "removePointsOutsideDomain": false, "scaleType": "linear", "canChangeScaleType": false, "facetDomain": "shared" }, "tab": "chart", "matchingEntitiesOnly": false, "hasChartTab": true, "hideLegend": false, "hideLogo": false, "hideTimeline": true, "colorScale": { "baseColorScheme": "default", "equalSizeBins": true, "binningStrategy": "ckmeans", "customNumericColorsActive": false, "colorSchemeInvert": false, "binningStrategyBinCount": 5 }, "scatterPointLabelStrategy": "year", "selectedFacetStrategy": "none", "isPublished": true, "invertColorScheme": false, "version": 57, "logo": "owid", "entityType": "system", "facettingLabelByYVariables": "metric", "addCountryMode": "add-country", "compareEndPointsOnly": false, "type": "ScatterPlot", "hasMapTab": false, "stackMode": "absolute", "minTime": "earliest", "hideAnnotationFieldsInTitle": { "entity": false, "time": false, "changeInPrefix": false }, "xAxis": { "removePointsOutsideDomain": false, "scaleType": "linear", "canChangeScaleType": false, "facetDomain": "shared" }, "hideConnectedScatterLines": false, "showNoDataArea": true, "zoomToSelection": false, "showYearLabels": false, "hideLinesOutsideTolerance": false, "hideTotalValueLabel": false, "hideScatterLabels": false, "sortBy": "total", "sortOrder": "desc", "hideFacetControl": true, "entityTypePlural": "systems", "missingDataStrategy": "auto", "id": 6970, "note": "The values for training computation are estimates and come with some uncertainty, especially for models for which only minimal information has been disclosed, such as GPT-4.", "slug": "ai-performance-knowledge-tests-vs-training-computation", "title": "Artificial intelligence: Performance on knowledge tests vs. training computation", "subtitle": "Performance on knowledge tests is measured with the [MMLU benchmark](#dod:ai-MMLU), here with 5-shot learning, which gauges a model\u2019s accuracy after receiving only five examples for each task. Training computation is measured in total petaFLOP, which is 10\u00b9\u2075 [floating-point operations](#dod:flop).", "originUrl": "https://ourworldindata.org/artificial-intelligence", "dimensions": [ { "property": "y", "variableId": 820642 }, { "property": "x", "variableId": 820644 }, { "property": "color", "variableId": 820643 } ], "comparisonLines": [ { "label": "Expert human performance", "yEquals": "90" }, {} ] } |