{"id":875,"date":"2026-08-31T05:29:58","date_gmt":"2026-08-31T05:29:58","guid":{"rendered":"https:\/\/erasmus.ai\/?p=875"},"modified":"2026-08-31T16:11:10","modified_gmt":"2026-08-31T16:11:10","slug":"climategpt-4-3-preview-at-the-internet-archive","status":"publish","type":"post","link":"https:\/\/erasmus.ai\/?p=875","title":{"rendered":"ClimateGPT 4.3 Preview at the Internet Archive"},"content":{"rendered":"\n

Previewing ClimateGPT 4.3 on the Friday afternoon Internet Archive Tour<\/em><\/p>\n\n\n\n

The Internet Archive<\/a> hosts an awesome tour of its digital library every Friday afternoon in San Francisco. Over the last few weeks Brewster Kahle<\/a> has ended the tour with a short demonstration of our foundation’s public good AI work, on ClimateGPT<\/a>. As we are in the final tests of gen four we invited people to kick the tires on the almost final release of ClimateGPT 4.3. ClimateGPT 4 is a major improvement from model family 3, as you can see from the benchmarks of the small 9B parameter model below.<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

The group this Friday was mostly postgraduate students from Stanford, working on AI systems and engineering, digital luminaries, some former Google execs, Archive staff and volunteers. We’re at the “sitting on hands” demo level where one hands over the keyboard to the youngest people in the room, listen and not speak. What follows is the four questions four people asked. <\/p>\n\n\n\n

[First Bug: Though the screen read 3.5 everything here ran on the ClimateGPT 4.3 (preview) the number in the interface is usually the last one on the release checklist ;-).]<\/p>\n\n\n\n

A matrix puzzle<\/h2>\n\n\n\n

Ashley Malkin<\/a> is working at how AI models could gain from learning language in the same sequence as children acquire language wanted to run a standard maths test to see how a domain specific model would perform on this test.<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

The first question was a Raven-style analogical reasoning matrix: three rows of integer pairs, complete the pattern, supply the missing value. Nothing to do with climate. Domain tuning at the expense of general reasoning is a live question for any specialised model, and this is the way to poke at it. The general response gave the value and the row relation behind it.<\/p>\n\n\n\n

As climate is complex, the interface is designed to show these different viewpoints on the problem. The interface returns a general response in the main column and, beside it, the same question answered from several other positions \u2014 economic, social, scientific, risk, investment \u2014 each summarised in a few lines with an Explore control to open it out. A toggle above the input switches between a standard and a deeper response mode.<\/p>\n\n\n\n

The general panel returned the correct mathematical answer, the perspectives column did not. Economic:<\/strong> a logic puzzle with no market mechanism, incentive structure or resource allocation to analyse. Social:<\/strong> matrix completion is pattern recognition, with no climate justice or equity dimension to find. Scientific<\/strong> declined outright \u2014 I am ClimateGPT, an AI assistant specialized in climate science and environmental data, so I cannot solve abstract numerical matrix puzzles.<\/em><\/p>\n\n\n\n

Solar car encapsulation<\/h2>\n\n\n\n

The second question came from Bugatti Wu a member of the Stanford Solar Car Racing Team<\/a> who is building a solar car: what glass or coating to select for a World Solar Challenge vehicle.<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

Top Sheet Materials from Tempered Glass, to Special Coating and ETFE film and thier attributes: respective benefits, weight, durability, conformance to curved surfaces, etc. were surfaced in the context of a solar car race, without RAG references and from our ClimateGPT 4.3 very small low footprint 9B model. The exchange is listed below. <\/p>\n\n\n\n