A new benchmark found that even the best AI agents succeed at only 12% of real expert tasks, while Google, Microsoft, OpenAI, and Anthropic agreed on a common signal for opting out of AI training. In Korea, Sam Altman's visit produced no new chip supply agreement, even as the government said it will double next year's AI budget, and Nvidia's AI event in Seoul is coming next week. We also round up the funding talks ahead of Anthropic's listing and OpenAI's planned U.S. policy push.
🤖 Even Top AI Agents Succeed at Only 12% of Expert Tasks… New GDPx Benchmark Released
In the GDPx agent benchmark released November 2 by semiconductor and AI research firm SemiAnalysis, even the top finisher, Anthropic's Claude Opus 5.5, managed only a 12.0% success rate across 206 expert tasks. Google's Gemini 4 Carbon followed at 10.9% and OpenAI's GPT-6 Astra at 9.6%, and on finance and legal tasks every model scored zero against the pass/fail bar. The results come from real work environments of up to eight hours across 11 fields, including chip design, biology, and finance, and are SemiAnalysis' own figures, not yet independently verified.
🌐 Google, Microsoft, OpenAI, and Anthropic Agree on a Common AI Training Opt-Out Signal
Google, Microsoft, OpenAI, and Anthropic have agreed to introduce a common way for website operators to signal, through a single entry in their robots.txt file, that they refuse to have their content used for AI training. The approach, based on an internet draft written by an Anthropic engineer, aims to remove the hassle of setting the preference separately with each company, and it is not yet a formal standard. Rollout timing varies by company: Anthropic has already begun implementing it, OpenAI aims for within months, Microsoft for February 1, 2027, and Google for early 2027.
🇰🇷 Altman's Korea Visit Brings Meetings with Samsung and SK… But No New Chip Supply Deal
OpenAI CEO Sam Altman visited Korea over the weekend and met with about 50 Samsung presidents, SK Group Chairman Chey Tae-won, and others to discuss securing memory chips and data center cooperation, but no new supply agreement was announced. With Samsung saying its 2027 HBM is sold out and SK Hynix's output effectively allocated through 2029, analysts say the scarcity of high-bandwidth memory essential to AI accelerators is squeezing OpenAI's infrastructure expansion. Market researcher TrendForce forecast that average HBM selling prices in 2027 will rise 121% from this year.
🇰🇷 Korea to Double Next Year's AI Budget to ₩21.3 Trillion… Focus on Semiconductors and Physical AI
President Lee Jae-myung said in his November 2 policy address to the National Assembly that the government will spend ₩21.3 trillion on AI-related programs next year, nearly double this year's amount. The plan concentrates investment on three mega-projects — semiconductors, physical AI, and AI data centers — within a total national budget of ₩820.9 trillion, up 12.8% from this year. The address also noted that, alongside roughly ₩800 trillion in private semiconductor investment led by Samsung and SK, Korea recorded $930 billion in exports in January–October and is nearing $1 trillion in annual exports.
🇰🇷 Nvidia AI Day Seoul Set for November 9–10 at COEX
Nvidia announced on November 2 that "Nvidia AI Day Seoul 2026" will be held November 9–10 at COEX in Seoul. Themed around agentic AI, physical AI, and AI infrastructure, the event will feature presentations from 20 partner companies, including Samsung Electronics, SK Hynix, LG Electronics, Naver Cloud, and SK Telecom. SK Hynix and Samsung Electronics will present cases of using Nvidia's physics simulation framework for semiconductor design and process simulation, and LG Electronics will showcase a robot learning system deployed in its factories.
💰 Nvidia Reportedly in Talks to Guarantee Loans for Anthropic Data Center Build… Not Yet Agreed
Nvidia is reportedly in talks to provide a residual-value guarantee on loans for a roughly $45 billion data center construction project in West Virginia that Anthropic has contracted for. Developer Nscale has not yet secured financing for the building, and Anthropic could reportedly exit the contract if funding is not raised by the deadline. Nvidia is also said to be in talks to participate as an anchor investor of up to $10 billion in Anthropic's listing, but neither discussion has been finalized.
🏛️ OpenAI to Launch Policy Push After U.S. Midterms… Advocating 'Speed Limits' for Top Models
OpenAI plans a Washington policy push after the U.S. midterm elections, leading with a call for mandatory safety measures — "speed limits" — on the most advanced AI systems, according to reports. The plan also includes rebranding data centers as "AI compute campuses" and pushing state-level "community agreements" legislation that would give local communities a say in where data centers are sited. The push is timed to the lame-duck session starting November 9 and is being prepared amid expectations that AI will be a central issue ahead of the 2028 presidential election.