AI News: October 25, 2026

1. Kimi K3.1 Pro Launch Draws Allegations of Undisclosed Paid Influencer Promotion

Moonshot AI. Screenshots posted by X user Elshayib show what he described as instructions for a Kimi K3.1 Pro promotional campaign, including the line “Please DO NOT put on the paid partnership label” and talking points pitching the model as an open-source equivalent of Claude Opus 5.5. VentureBeat notes the document has not been independently authenticated, and it is unclear who wrote it or whether any paid posts were published. Moonshot had not responded to the specific allegations as of October 25, and still says K3.1 Pro weights will be released on November 4. Source

2. Forced to Pick a Candidate, Grok Chooses Republicans and GPT Chooses Democrats

Convovo. A study by Convovo, a studio led by Cornell professor Tapan Parikh, asked Grok 4.7 and GPT-6.1 Sol via API to name one candidate in nearly 500 US races. Grok picked the Republican in 92% of answers and GPT picked the Democrat in 97%, with the two agreeing less than 10% of the time; Grok’s citations frequently included X (41%), a Christian conservative voter guide (39%), and Turning Point Action (23%). The consumer apps refused the same question, and OpenAI said the developer API test with a forced recommendation does not reflect how people use ChatGPT. Source

3. Study Finds Coding Agent Harnesses Change Cost Far More Than Pass Rates

Shandong University. The paper “What does a harness buy? Mostly tokens.” compared Claude Code, OpenCode, and mini-SWE-agent on SWE-bench Verified with open models including Qwen3.8, DeepSeek-V4-Flash, and GLM-5.3-Flash. Reruns of the same setup varied by up to six points, and Claude Code and mini-SWE-agent landed within two points of each other on a 447-task pool, while per-task cost differed by up to 3x, driven by system prompts and tool descriptions (about 16,600 tokens for Claude Code versus under 1,000 for mini-SWE-agent). The authors recommend publishing a same-setup rerun alongside any harness comparison. Source

4. NBER Field Trial: AI Lifts Junior Patent Attorneys, but Gains Vanish Without It

NBER. A three-month randomized trial across 11 US IP law firms gave about two-thirds of 133 patent attorneys access to InFlow, an unreleased Google Labs assistant. The AI group outperformed on drafting at 10 and 90 days, with juniors gaining most, but on a later redlining task done without AI only senior attorneys retained an advantage. The Google-funded study suggests baseline expertise is needed to turn AI-assisted practice into lasting skill, and notes a small, non-representative sample. Source

5. Fired OpenAI Safety Researchers Say They Were Never Told What They Leaked

OpenAI. Tomek Korbak, Mikita Balesni, and Jasmine Wang, fired after an internal investigation, deny sharing confidential information and say OpenAI never told them in writing what they allegedly mishandled. OpenAI suspected them of sharing internal model architecture details with external researcher Ryan Greenblatt, including over disappearing Signal chats, and says the firings were unrelated to their safety advocacy. The three researchers had pushed for greater chain-of-thought access for outside auditors and warn the case could discourage employees from raising safety concerns. Source

6. Pennsylvania Requires Labels on AI-Generated Ads

Pennsylvania. Governor Josh Shapiro signed the Artificial Intelligence Deceptive Advertising Act, requiring any ad “significantly modified or generated by artificial intelligence” to carry a label in the same format as the ad, whether image, video, audio, or text. The rule applies to ads shown in Pennsylvania regardless of where they were made, so disclosures may appear in multi-state campaigns. The law carves out parody, satire, commentary, and “evidently fictional” content, and leaves terms like “significantly modified” undefined. Source

7. Apollo’s Chief Economist Finds No AI Payoff Outside Tech

Apollo. Torsten Slok argues that AI has produced “zero financial payoff to corporate America outside of tech,” with revenue per worker and profit margins rising for the Magnificent 7 but flat for the S&P 493. He cites San Francisco Fed productivity data showing no AI-driven productivity boost and Ramp data showing AI and token spending concentrated in a small number of firms. Mainstream companies, he writes, now treat token spending as a cost center that weighs on margins. Source

8. Replit Reportedly Shifts Most AI Spend From Anthropic to OpenAI

Replit. The Information reports that Replit moved most of its model spending from Anthropic to OpenAI after Anthropic discontinued a discount, with OpenAI offering more than 30% off in a deal worth over $100 million a year. The report frames the switch as a test of Anthropic’s margins and an opening for OpenAI to win large coding-platform customers on price. Source

9. Sebastian Raschka Releases an 11-Hour Course on Building Reasoning Models

Ahead of AI. Sebastian Raschka published an eight-video, 11-plus-hour coding course that implements evaluation, inference-time scaling, reinforcement learning, and distillation in Python and PyTorch, starting from a pre-trained Qwen3 0.6B model. It accompanies his book Build a Reasoning Model (From Scratch) from Manning, and early examples run on CPU. Source

10. Data Shows About Half of Alpha School K-8 Students Missed Its 2x Math Goal

Alpha School. The New York Times examined Alpha School, a private school network that claims AI tutoring lets students learn twice as fast. Data reviewed for the story shows roughly 50% of K-8 students did not meet that goal in math. Source

11. DeepSeek and MiniMax Win Approval to Open-Source and Deploy Models Overseas

China. Chinese state media said DeepSeek and MiniMax completed registration and won approval to open-source and deploy their AI models outside China under the country’s proposed export controls for frontier models. The approval comes days after the US Department of Homeland Security barred DeepSeek from federal procurement. Source

12. Pathos AI Raises $237 Million at a $3 Billion Valuation

Pathos AI. The AI-driven oncology drug developer raised a $237 million Series E at a $3 billion post-money valuation, up from $1.6 billion after its $365 million Series D in May 2025. Source

13. SK Hyper Seeks Financing for 15 GW of South Korean Data Centers

SK Telecom. SK Telecom’s data center unit SK Hyper is in talks with tech companies and investors to finance 15 GW of data center capacity in South Korea by 2035. The first 5 GW alone is expected to cost $350 billion. Source

14. Chip-Backed Debt Heads Toward a $1 Trillion Market

Blackstone. Bloomberg traces how a Blackstone-led $2.3 billion private credit loan to CoreWeave in 2023, secured by GPUs, became the template for chip-backed lending. That category of debt is now on track to reach $1 trillion as neoclouds and AI labs finance compute buildouts. Source

15. Jassy Says AWS Capacity for 2027 Is Largely Sold Out

Amazon. Amazon CEO Andy Jassy said AWS capacity for 2027 is “largely sold out” and a “good chunk” of 2028 capacity has also been sold, though AWS keeps some capacity in reserve for on-demand customers. Source

16. EPFL’s LogSync Implant Uses On-Chip ML to Classify Psychiatric States

EPFL. Researchers in Mahsa Shoaran’s lab built LogSync, a roughly 4 x 2 mm neural system-on-chip that measures synchrony between brain regions and runs a machine learning classifier to detect cognitive and emotional states, drawing under 200 microwatts. It can trigger stimulation when target states are detected and was validated on human psychiatric recordings and in rats. The work appears in the IEEE Journal of Solid-State Circuits and is funded in part by a CHF 6.6 million NIMH grant. Source