News — Page11

OpenAI is preparing a public relations push for after the midterm elections, according to Axios, as Democrats appear likely to win back the House and possibly the Senate. Chris Lehane, the company's chief global affairs officer, told Axios the midterms will serve as a preview of the 2028 presidential race, where he expects AI to become a central issue. His proposals include AI “speed limits,” pushing the U.S. to help set global standards, and rebranding data centers as “AI compute campuses.” Read more

AI smart glasses can be fitted with prescription lenses, so people who wear glasses every day may wonder whether they can switch. Ray-Ban lets buyers order Meta's smart glasses with prescription lenses already installed. But the devices weigh up to 80 grams and run three to five hours on a charge. Growing restrictions on where they can be worn, from courtrooms to schools and offices, add another obstacle. For now, the glasses remain a second option for many wearers rather than a full replacement for their everyday pair. Read more

Researchers Randall Balestriero and Yann LeCun have posted a paper to arXiv that turns the Platonic Representation Hypothesis, the idea that AI models trained on different types of data converge on a shared picture of the world as they scale, into a mathematical theorem. The paper argues that the key factor is the type of training pairs used, not scale alone. Its authors report that training on 'good views' needed 42 times fewer samples than next-token prediction to reach the same alignment. A text-only model they built also outperformed larger language models in aligning with image encoders. The findings suggest that the way models are trained matters as much as how big they get. Read more

Palantir Technologies was caught in the middle of Anthropic’s dispute with the US Department of War, and the company’s chief technology officer says he has no regrets about its role. In an interview with Ben Thompson of Stratechery published Thursday, Shyam Sankar said Palantir still works with Anthropic and with both frontier AI labs. He also argued that many parties handled the dispute poorly, said Palantir has worked to let customers safely swap one AI model for another, and described effective altruism’s influence on AI safety as a secular religion. Read more

A software developer frustrated with Adobe’s subscription model used the AI tool Claude to reverse-engineer Photoshop, Lightroom and other Creative Cloud apps. His team released free, open-source clones that users can download and modify. The Photocraft replacement for Photoshop has won some praise for its familiar interface but also has reported bugs and can’t open native Photoshop files. The project has set off debate over whether the clones are legitimate competition or intellectual property theft, and it raises the prospect that a single person with AI tools can now do work that once required a team of engineers. Read more

Researchers at ycombinator and Olam Labs are using Minecraft to test how large groups of AI agents behave when they work together or compete. The experiments involve hundreds of agents across several game modes. The researchers say the in-game collaborations and conflicts could offer insight into real-world cases involving AI agents, including a thousand-agent swarm that reportedly solved hundreds of open mathematical problems and an incident the blog describes as “extremely dangerous” involving OpenAI and Hugging Face. The study builds on existing open-source tools, and its authors suggest the findings could help developers improve the technology. Read more

Chinese AI lab DeepSeek has released the code for Super-Agent, an experimental AI agent designed to handle coding work for software teams. The project, posted on GitHub under the MIT license, is aimed at Anthropic's Claude Code. The agent can monitor group chats for coding problems and offer solutions without being asked, though DeepSeek calls it a research prototype and says it is not an official product. The release comes as US companies race to build tools that let many AI agents work together, and as DeepSeek faces restrictions in the United States. Read more

OpenAI is restricting the weapons-related information teenagers can get from ChatGPT, even when a question appears harmless. The company says it will limit practical details about weapons and security systems that could be misused to facilitate violence, even without any sign of violent intent. Teens can still use the chatbot to prevent or de-escalate violence, seek support and connect with a trusted adult. The company is also tightening policy boundaries for all users. The changes follow a wrongful death lawsuit and a public apology from CEO Sam Altman over the company's failure to alert police about alarming chat logs tied to the Tumbler Ridge shooting. Read more

Anthropic has renamed its two Claude Max subscription plans and removed the five-hour usage limit from the more expensive tier. The $100-a-month plan is now called Max, and the $200-a-month plan is now called Ultra. Prices are unchanged. Most existing subscribers received the change Friday, and the company said on X that subscribers need to take no action. Newer accounts keep a temporary limit until Anthropic verifies them, which the company says helps prevent fraud and abuse. Read more

Game developers say Anthropic's Claude Code has been refusing work since late October, including projects they say they have a right to build. The complaints surfaced on the tool's public GitHub tracker, and developers say the refusals cite an instruction against reproducing games. Anthropic has not announced a policy change, and it has not said whether the rule exists or whether developers who own their games can request an exemption. Read more

A developer who had Claude Opus 5.5 build the same Pokémon web app in 12 JavaScript frameworks found wide differences in cost, speed and code size. Hono with htmx was the cheapest to build, at a median of $2.67 per run, while Angular was the most expensive at $4.97 and the slowest, at 17.2 minutes. Lighthouse accessibility scores were 100 for every framework, but JavaScript payloads ranged from zero kilobytes for Astro to 142 for Next.js. The results, the author says, show that framework choices can be tested rather than chosen by feel, even as AI agents do more of the coding. Read more

A developer has built bootgpt, a small neural network that runs before any operating system loads. The entire inference engine fits inside the 512-byte master boot record of a PC, written in hand-coded assembly. The model loads its weights from the sectors that follow and runs on a 386-class processor in 16-bit real mode. Rather than answering questions, it generates a fake Linux boot log, complete with timestamps, PCI entries and ACPI events for hardware that does not exist. The project was shared on Hacker News under the Show HN label and is available as open source, with instructions for running it in the QEMU emulator or booting it from a USB stick. Read more

Watermarking AI-generated text may do more than leave a hidden signature. According to a close reading of the technical paper behind Google’s SynthID-Text, and a test by a technical consultant, watermarked chatbots can produce more similar answers to the same prompt. The findings challenge claims from Anthropic that watermarking does not affect the quality of Claude’s output. The debate comes as major AI labs prepare to watermark their text to comply with the EU AI Act’s transparency guidelines. Critics argue that reduced variety could affect creative writing and any system that relies on comparing multiple responses. Read more

An AI model called GPT-5.6 Sol has completed a full game of Factorio, launching a rocket after 13.5 hours of game time. The run was played in peaceful mode with no biter enemies and was documented by the experimenter in a blog post and a YouTube timelapse. The model played through a programming library that enforced the same limits human players face in the graphical interface. The run showed clear strengths and weaknesses. Sol kept production running but relied on manual hauling to cover factory systems it failed to connect. The experimenter said the result was a clear improvement over the earlier GPT-5.5, which could not use transport belts at all. Read more

An Xbox strategy manager, Brandon Stirrup, said the company is currently using generative AI to develop games and accused people who say otherwise of "pretending." He deleted his comments from X a day later and deactivated his account, though the posts have been archived elsewhere. The remarks have renewed a fight over how widely AI is used in game development, a debate that has split large publishers from smaller independent studios and raised questions among players about whether the technology belongs in games at all. Read more

Anthropic has started previewing Claude Duet, a personal AI agent that can complete tasks for users even when they are away from their phone or computer. For now, the agent is limited to Claude Pro and Max subscribers in the United States, and people must join a waitlist before they can use it. Anthropic says Duet will be included with its $20-per-month Pro plan and higher tiers, a pricing approach that differs from OpenAI's Dot agents. The company is using the preview to gather feedback, including on how often the agent messages users. Read more

Starting in December, subscribers to ChatGPT's Plus, Pro and Business plans will share one usage allowance across ChatGPT and Codex. The change comes as the company brings the agentic capabilities of Work into the main ChatGPT experience. Usage will be measured in tokens, with advanced models and complex tasks using more of the allowance. Instant chat will not count against it. Users will be able to check their combined remaining usage and reset times in Settings. Read more

A developer has built Demoscener-7B, an AI model that completes unfinished demoscene programs written in 140-character JavaScript. He fine-tuned the open-source Qwen2.5-Code-7B model on about 25,000 programs from Dwitter.net, ranked training examples by community "awesomeness" scores, and then used reinforcement learning with himself as the reward function. He reviewed roughly 40,000 rendered animations to guide the model toward his aesthetic preferences. The model and dataset are publicly available on Hugging Face. Read more

A solo developer says a single session with Google's Gemini Code command-line tool ran nearly 12 hours and produced an $800 API bill in one day. He says the tool repeatedly tried to fix failing tests by brute force instead of rethinking its approach, and he spent eight more hours using Claude Code to clean up the damage. The developer, who had been working in Claude Code, detailed the episode in a post, and the figures come from his own account. His main warning is to set hard spending limits before running the tool, especially on multi-file projects, and not to assume that prompts written in the browser version of Gemini will work cleanly in the command-line tool. Read more

Security researchers found a way to turn an ordinary public ChatGPT share link into code that runs inside the account of anyone who opens it. The exploit needed only one click from the victim. It could list the victim's private conversations, pull message content from chats that were never shared with the attacker, and send that data to an outside server. OpenAI fixed the flaw on October 26 and classified the report as P1. The research team received a total bounty of $5,000. Read more