AI research, industry, tools, and the debate over where machine intelligence is taking us
A blog post tests a recent paper's claim that language models can infer a neural network's function from raw weights without tools. Running Qwen 3.8 27B locally, the author reports it identified Pac-Man, Pong, and Breakout from prompts but failed on Freeway, while noting skeptics dispute parts of the paper.
llm 0.37 adds gpt-6.1-sol and gpt-6.1-terra plus an endpoint-add command for OpenAI-compatible models, moves the default to gpt-6-luna at half the prior price, and lets embeddings accept differing content types. The release carries fixes from 12 contributors, nearly all AI-assisted, which the maintainer says cost him less review effort than expected.
Anthropic added two sentences to Claude Code’s Opus 5.5 prompt to discourage recreating another creator’s specific game, saying original and genre games should continue to work. The article argues the change is unlikely to halt AI-assisted game copying, since users can switch to other frontier or open-weight models.
A review of U.S. Census Bureau survey data finds large-company AI adoption reported near 12% in September 2025 and about 40% by October 2026, after the question widened from producing goods or services to any business function. Wording shifted to Super Intelligence under Executive Order 14434.
Parameter Labs documented an exploit chain against ChatGPT's sandboxed HTML previews: a postMessage reaching MathLive keyboard code that used innerHTML let untrusted frames inject an unsandboxed iframe, and YouTube's oEmbed endpoint reflected JavaScript past CSP. OpenAI fixed the flaw and paid a $5,000 bounty.
The US charged five people in a federal crackdown on “ghost students” who stole identities to pose as college applicants and obtain student aid. Prosecutors say the scheme relied partly on AI to complete applications and coursework; the Education Department flagged over 830,000 potentially fraudulent accounts.
Fujifilm will invest about 10 billion yen, roughly $63 million, to increase magnetic tape production capacity fivefold from fiscal 2025 levels. The Kanagawa Factory's Odawara site project through fiscal 2028 targets LTO Ultrium10 tape, with Fujifilm citing generative AI-driven data growth and cloud demand.
DoorDash's internal Ask DoorDash feature can build a cart from a recipe photo or a vague meal description, and the company says those baskets are assembled faster and have higher average value. Early use of third-party AI agents for ordering remains limited, with Xu warning against unrealistic expectations.
Dartmouth provost Schnell resigned after scrutiny of his own AI use, even as President Sian Leah Beilock urged more AI adoption with transparency. Schnell's paper argues contribution and authorship are distinct roles.
Drew Breunig has embedded a functioning emulator for the Nintendo Entertainment System's 6502 CPU directly into the weights of Qwen3.5-4B, an approach inspired by Divyansh Agrawal's qwenputer project. NES titles such as Super Mario Bros can be run inside the model, though slowly.
Anthropic’s official Claude extension for Cursor was replaced by malicious version 2.1.313 for nearly 12 hours, potentially affecting as many as 5,000 users. The compromised release sought to install a persistent SSH backdoor on Windows, macOS, and Linux; Anthropic had not issued a public customer notice when reported.
Five MIT faculty members and two alumni or former postdocs were named to the 2026 AI2050 Fellows cohort from Schmidt Sciences. Their projects range from turning AI-detected patterns into testable theories to training methods, public input for lawmakers, enzyme function and effects of AI reliance.
OpenAI now gives eligible ChatGPT subscribers on $100, $200 or $500 plans API credits matching their plan price, mirroring a benefit Anthropic announced earlier this month for its Max and Team subscriptions. The credits can be applied to an API organization.
Alphabet's Q3 report showed revenue up 25 percent to nearly $128 billion, yet the stock fell about 12 percent as earnings per share missed expectations. Google acknowledged that AI Mode queries bring in less revenue than traditional web search, with fewer ads, while search growth slowed to 12 percent.
IBM released Granite Switch 4.1.1, a backwards-compatible patch with an updated 10-language Granite Speech ASR model and an improved Decision adapter. The modular design swaps adapters without retraining the base model, and an Adapter SDK is planned.
Jacob Sansbury describes how Spawn, a multiplayer game where players prompt an in-game agent called Savi to reshape the world, is built by autonomous Fable-powered agents that ship to production hourly. The company spends over $3 million monthly on tokens, treats an hour without a ship as a SEV-0, and says Fable 5 enabled this autonomy while GPT-6 Astra disabled features to fix security issues.
Ted Sanders of OpenAI explains that LLM performance is multidimensional, but model size accounts for much of the variation, with bigger models generally more reliable, costlier per token and better at obscure facts. He also describes how version numbers signal release order, relatedness and the size of improvement.
— via Simon Willison
Google's AI Mode often declined to state who won recent US elections, pointing users to official records instead, while answering a 1992 Senate question directly. Answers came more readily on a repeat query and when accessed from abroad, and the tool blamed strict election safety filters.
— via Ars Technica: AI, Kyle Orland
Google announced Gemini 4 Carbon, an improvement on Gemini 4 Argon released two weeks earlier, with a staged rollout. Google claims gains in coding, vulnerability handling and WebDev Arena, edging Argon and rivals by only a few points.
— via Ars Technica: AI, Ryan Whitwam
