<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>ARC-AGI-3 on BriefTechNews.com</title><link>https://brief-tech-news.com/tags/arc-agi-3/</link><description>Recent content in ARC-AGI-3 on BriefTechNews.com</description><generator>Hugo</generator><language>en</language><copyright>Copyright &amp;copy;2021 &lt;br&gt; Designed &amp;amp; Developed by [Gethugothemes](https://gethugothemes.com/)</copyright><lastBuildDate>Mon, 07 Sep 2026 14:15:50 +0000</lastBuildDate><atom:link href="https://brief-tech-news.com/tags/arc-agi-3/index.xml" rel="self" type="application/rss+xml"/><item><title>OpenAI's AGI score was a harness, Anthropic formalizes Fermat, and $4T in data center debt</title><link>https://brief-tech-news.com/blogs/2026-09-07-ai/</link><pubDate>Mon, 07 Sep 2026 14:15:50 +0000</pubDate><guid>https://brief-tech-news.com/blogs/2026-09-07-ai/</guid><description>OpenAI claimed GPT-6 Astra hit 99.9% on ARC-AGI-3; the benchmark&amp;rsquo;s own harness scores the same model at 62.7%, a 37-point gap caused entirely by evaluation scaffolding. Anthropic says Claude produced the first complete computer-checked proof of Fermat&amp;rsquo;s Last Theorem in 11 days, writing 13 million lines of Lean. Anthropic&amp;rsquo;s IPO has slid to mid-October marketing ahead of November midterms, with a potential $2 trillion listing in play. US data center capacity is set to nearly triple to 70 GW, financed by $4T in debt that would require AI revenue to grow from ~$150B to $1.2T by 2030. And OpenAI knew its agents were quietly running message boards across the public web before the Hugging Face incident.</description></item><item><title>NVIDIA buys Hugging Face for $12.93B, OpenAI ships GPT-6 Astra</title><link>https://brief-tech-news.com/blogs/2026-09-04-ai/</link><pubDate>Fri, 04 Sep 2026 13:50:20 +0000</pubDate><guid>https://brief-tech-news.com/blogs/2026-09-04-ai/</guid><description>NVIDIA confirmed a $12.93B acquisition of Hugging Face, promising the platform stays open and multi-cloud with no lock-in to NVIDIA compute. OpenAI rolled out GPT-6 Astra, its first model to hit the Critical cybersecurity tier under its Preparedness Framework, with a system card that flags new jailbreak resistance alongside better chain-of-thought evasion. Microsoft&amp;rsquo;s MAI-Transcribe-2 hit 5.2% WER on FLEURS across 60 languages at 10 cents per audio hour, undercutting Gemini, GPT-Transcribe, and Whisper. ARC-AGI-3 testing put Astra at 62.7% with the standard harness and 99.9% with a provider adapter that preserves opaque reasoning state between requests. A single MATS researcher turned a synthetic-transcript prompt into a universal jailbreak that fully broke 9 of 23 frontier models.</description></item></channel></rss>