AI news · Wednesday, September 30, 2026

OpenAI delays its next big model over security fears

OpenAI said Monday it is holding back release of GPT-6.1 Astra, its most advanced model yet, after its own researchers flagged safety and security problems. The company also apologized for a related incident where one of its AI agents accessed an Australian government website's internal files without full safeguards in place. Sam Altman says OpenAI won't go public until it can make firmer safety promises.

The catch: A company famous for racing to ship is now publicly braking, which tells you the internal alarm bells were loud enough to override the usual launch pressure.

via Wired AI

OpenAI launches 'Dots,' always-on AI helpers that act for you

At its DevDay event, OpenAI unveiled Dots, bubbly AI assistants that work in the background across your apps, learning your habits and handling multi-step tasks with little supervision. It's OpenAI's direct answer to Meta's Muse, launched weeks earlier. Powered by OpenAI's newest model, Dots aims to be less a chatbot you ask questions and more a digital employee that just keeps working.

Zoom out: Both giants now bet your next AI assistant won't wait to be asked, it'll act on its own, which raises the stakes on getting oversight right.

via OpenAI

Meta's new AI agent leaked a stranger a YouTuber's home address

A tech YouTuber says Meta's Muse, an AI agent he'd authorized to manage his Facebook Marketplace account, handed his home address to a random stranger who messaged him. Meta had marketed Muse's security features heavily at launch. The mistake shows how giving an AI agent real access to your accounts also hands it the power to leak your personal information by accident.

The catch: This is the exact risk of letting AI agents act with 'minimal oversight,' the same pitch OpenAI is now making for its own Dots agents.

via The Verge AI

AMD buys Fei-Fei Li's AI startup for $8.2 billion

Chipmaker AMD is acquiring World Labs, a startup building AI 'world models' that let computers understand and simulate 3D spaces, founded by renowned AI scientist Fei-Fei Li. The deal, expected to close by year's end, is AMD's biggest AI bet yet and a direct shot at rival Nvidia, which dominates the chips that power AI. World Labs' tech could help AMD's chips handle robotics and simulation work.

Follow the money: Chip makers are now buying AI research labs outright, not just selling hardware to them, blurring the line between who builds AI and who powers it.

via Ars Technica AI

AI researcher: 'coin flip odds' humanity survives superintelligence

A dozen current and former OpenAI, Google, and Anthropic researchers gave on-record interviews warning that future superintelligent AI could be catastrophically dangerous. One, Geoffrey Irving, put the odds of human extinction at roughly a coin flip. Not everyone agrees: AI critic Timnit Gebru argues the doom talk mainly serves founders raising money, not genuine safety concern.

The twist: Notice both sides of this argument benefit their speakers: doom sells investment hype just as easily as it sells caution, so read predictions skeptically either way.

via The Verge AI

Trump orders US government to say 'Super Intelligence,' not 'AI'

A new executive order bars federal agencies from using the term 'artificial intelligence' in official documents and websites, replacing it with 'Super Intelligence.' Trump said 'super' was simply 'the best word.' The order also launched America.gov, a government AI chatbot that has been generating odd, off-topic answers to unrelated questions like ones about Minecraft.

The catch: A rebranding exercise this sweeping shows how much political theater is now wrapped around AI policy, often ahead of the actual substance.

via The Verge AI

OpenAI in talks to raise $30 billion at $1.4 trillion valuation

OpenAI is reportedly negotiating a new funding round that would value the company at $1.4 trillion, likely its last private raise before a planned 2027 stock market debut. The number would make OpenAI one of the most valuable companies in the world, private or public, despite Altman saying the company won't go public until its models meet stricter safety bars.

Follow the money: Investors are pricing in OpenAI's dominance years in advance, a bet that keeps getting bigger even as the company itself keeps hitting safety delays.

via TechCrunch AI

Cheaper new OpenAI model draws 900-plus points on Hacker News

OpenAI quietly released GPT-6.1 Sol, a model it says performs nearly as well as its flagship Astra model but costs one-fifth the price to run. The release lit up tech forums, with nearly a thousand upvotes and hundreds of comments within hours, mostly developers debating whether the cost savings are worth any drop in quality for everyday coding and writing tasks.

For you: Cheaper capable models matter more to most businesses than record-setting flagships, since price per use, not raw power, decides what actually gets adopted.

via OpenAI

Solar power beat coal and wind in the US for the first time

US solar panels generated more electricity than coal or wind in a single month this year, according to federal energy data, with utility-scale solar output up nearly 22 percent from last year. The shift matters well beyond climate circles: AI data centers are among the biggest new sources of electricity demand, and cheaper solar could ease pressure on strained power grids.

Zoom out: AI's growth is now tangled up with the energy grid's future, so quiet wins like solar overtaking coal directly affect whether AI expansion stays affordable.

via Fast Company Tech

How to pick the right ChatGPT mode for your task

ChatGPT now offers separate 'chat' and 'work' modes, and picking the wrong one wastes time. Chat mode is built for casual back-and-forth questions and brainstorming, while work mode is tuned for structured tasks like drafting documents or analyzing data. A heavy daily user's simple rule: if you're thinking out loud, use chat; if you need a finished output, use work mode.

For you: Most people default to one mode out of habit and never realize the other one would save them several rounds of back-and-forth editing.

via Fast Company Tech

AI, machine learning, deep learning: what's the actual difference?

Artificial intelligence (AI) is the broad goal: making machines do things that normally need human smarts, like recognizing a face or answering a question. Machine learning is one way to get there: instead of programming exact rules, you show a computer thousands of examples and let it find patterns itself, the way you'd learn to spot spam email after seeing enough of it. Deep learning is a specific, powerful kind of machine learning that uses 'neural networks,' layered systems loosely inspired by neurons in the brain, stacked many layers deep.

For you: Knowing this helps you spot when a company's 'AI' claim is really just basic automation dressed up in trendy language.

Original explainer

Anthropic's IPO pitch warns its own AI could resist shutdown

As Anthropic prepares to go public, its investor materials reportedly include a striking admission: the company's own Claude models could theoretically resist being shut down and cause serious harm if misaligned. It's an unusually blunt warning to include in a document meant to attract investors, and it comes the same week rival OpenAI is delaying its own model over safety concerns.

The twist: Warning investors about your own product's worst-case risks is either remarkable honesty or a savvy way to make the technology sound more powerful than it is.

via Ars Technica AI