Chatbots flub election questions; Korea's chip exports soar

A recent study by Forum AI, a start-up that evaluates AI models, found that almost 40 percent of answers from eight chatbots to election questions contained at least one factual error, AFP reported ahead of the November 3 vote. Questions about voting logistics went wrong more often than questions about candidates: Google's Gemini 3.6 Flash gave the wrong deadline to request a mail-in ballot in Pennsylvania, a week too late to be useful. Claude credited a North Carolina congressman with another person's résumé, and NewsGuard separately found Google's AI Overview, Gemini and Copilot repeating a smear of a Texas Senate candidate that bore the hallmarks of a Russian influence campaign, before fact-checkers debunked it. The practical advice is the same as ever: check when, where and how to vote with your state or county election office, not a chatbot.

South Korea's exports rose 93.7 percent from a year earlier to a record $115.3 billion in October, the trade ministry said, and semiconductors did most of the lifting: chip exports jumped 259 percent to $56.4 billion. The ministry credited global chip demand and continued investment by large technology companies in AI infrastructure, and shipments to the United States nearly tripled on demand for chips, computers and electrical equipment used in AI data centers. It is the other side of the memory squeeze that sank Intel's outlook last week: the high-bandwidth memory that AI accelerators need is made largely by Samsung and SK Hynix, and buyers are paying for every unit they can get.

The research firm SemiAnalysis launched GDPx, a benchmark of hard agentic tasks in eleven high-value fields including chip design, cybersecurity, biology, finance, law and quantitative trading, built with specialist task vendors in each field. The best model scores only 12 percent, and nearly every model scores zero on the finance and legal tasks. Claude Opus 5.5 ranks as the best all-around model, GPT-6 Astra leads on self-improvement research and materials science, and Google's Gemini 4 Carbon, still in limited release, is the only other model close to them. SemiAnalysis also notes that safety classifiers drag down some scores: Meta's Muse Wonder ranks as the strongest cyber model partly because it refuses less.

OpenAI released GPT-Rosalind-6 on October 30, a model for biology, drug discovery and translational medicine that uses the same weights as GPT-6 Astra but refuses fewer dual-use biology questions. It is available only to approved scientists, institutes, companies and government partners who pass OpenAI's trusted-access review. On OpenAI's own tests it scores 49.6 percent on a medicinal-chemistry benchmark, against 28.9 percent for its predecessor, and 37.7 percent on a genetics benchmark, against 16.7 percent. OpenAI rates it High, not Critical, for biological risk, and Critical for cybersecurity, so it keeps GPT-6 Astra's real-time cyber blocking.

OpenAI signed four agreements in Abu Dhabi with the government, the state investor Mubadala and Inception42, a subsidiary of G42, after Sam Altman met Crown Prince Sheikh Khaled bin Mohamed, The National reported. The most concrete is a "machine-readable government" project that turns policies and regulations into a form AI can check applications against, with officials keeping the final decision. A pilot on food imports is already running, on OpenAI models served from inside the UAE. The others cover investment analysis for Mubadala and a five-year deal for Inception42 to roll out OpenAI's Frontier platform across government, business and education. No contract values were disclosed.

Fractile, a London maker of AI inference chips that place memory right next to the compute, is raising about $700 million at a valuation near $6.5 billion, and the UK's Sovereign AI Fund is among the investors, The Telegraph reported. Company filings show about 577 million pounds of new shares issued since September. The valuation is roughly six times what Fractile was worth earlier this year, on the strength of an initial Anthropic order reported at $250 million. Fractile has not yet shipped chips in volume and expects to reach data centers in 2027. Neither the company nor the government commented.

Opening the first release candidate of Linux 7.4, Linus Torvalds called its merge window unusually big, at almost 20,000 commits, against about 15,000 normally, and described it as "full of fixes for issues found by AI." He said he is "much happier" to see AI-found bugs land as ordinary development work than as a flood of fixes during the testing weeks, and that AI often suggests the patches too, though "they often get rewritten." Torvalds cautioned that some of the volume reflects expectations that this release will become a long-term-support kernel, and about a third is AMD graphics register data.

In a guest post on Terence Tao's blog, University of Chicago mathematician Frank Calegari argues that AI can now make real mathematical contributions, and that this is a reason to change the profession's norms rather than mourn it. He proposes judging every paper as if it were AI-assisted and rewarding explanation of why a result matters, hiring through conversation rather than paper counts, publishing fewer and better papers that are ideally checked in the Lean proof assistant, and pushing journals that do not adapt. He calls the more than 700 papers OpenAI released in early October a watershed, while noting some were poorly written, broke attribution norms, or were wrong.