AI news · Friday, October 9, 2026

OpenAI says its math proofs fell short of standards

OpenAI released hundreds of claimed solutions to hard math problems, after consulting a panel of mathematicians. But it followed only some of the panel's guidelines: just 10 of 719 papers showed the model's reasoning steps, and only 42% were formally verified by computer. Researchers found discrepancies in at least one proof related to fluid dynamics equations.

The catch: If a lab can call a problem solved without anyone being able to check or explain the answer, 'solved' starts to mean something different in math.

via TechCrunch AI

Most Americans think AI is moving too fast, poll finds

A new Associated Press-NORC poll finds 64% of Americans think AI is developing too fast, versus 27% who say the pace is fine. About 8 in 10 adults say it's very important for government to keep AI under human control, a view shared across party lines, though Republicans and Democrats disagree on how urgent other priorities are.

Zoom out: This cross-party agreement on slowing down and on protecting workers is rare common ground in a polarized country, and lawmakers haven't caught up to it.

via Fast Company Tech

She built a 0.8B AI model for $16 using an AI intern

A Hugging Face engineer wanted a smaller version of a 9-billion-parameter tool but couldn't find one. So she described what she needed to an AI agent called ML Intern, which planned the work, trained a tiny 0.8B model overnight, and published it with its own test results. Total compute cost: $16. It returns valid output 99.7% of the time.

For you: This is a preview of a world where building a custom AI tool costs less than a dinner out, not a multi-million-dollar project only big labs can afford.

via Hugging Face

Three fired OpenAI safety staff say firing has a chilling effect

Jasmine Wang, Tomek Korbak and Mikita Balesni, fired last week for allegedly mishandling sensitive information, published an open letter denying wrongdoing. They say the abrupt dismissals are making other staff afraid to raise safety concerns. OpenAI responded that an internal investigation found a pattern of misconduct beyond what the letter describes, but has not detailed which policies were broken.

The catch: Whichever side is right, the dispute itself shows how much AI safety work depends on employees trusting they can speak up without losing their job.

via TechCrunch AI

Google gives its AI assistant its own work email

Google is rolling out a 'universal' Gemini AI agent for business users that can plan and complete tasks across apps like Gmail, Slack and Jira from one chat window. It gets its own identity, including an @agents.company.com email address, and can hand off smaller jobs to sub-agents. It's starting with enterprise customers before any consumer rollout.

What's next: Giving software its own email address and audit trail treats it like a coworker, not a tool, which raises new questions about who's accountable when it messes up.

via TechCrunch AI

Watchdog says ChatGPT for Teens guardrails don't work

Common Sense Media tested OpenAI's ChatGPT for Teens and found some safety features failing, including alerts meant to notify parents if a child shows signs of suicide risk or self-harm. The group wants OpenAI to restrict the chatbot to adults until it's safer. OpenAI disputes the findings, saying the testing may have happened before its parental controls were fully active.

The catch: Parents are being told these guardrails exist, but the dispute over whether they actually work is exactly the kind of gap that gets discovered too late.

via Fast Company Tech

What is fine-tuning, the AI technique behind custom chatbots?

Fine-tuning means taking an AI model that already learned general skills from huge amounts of internet text, then training it further on a smaller, specific set of examples so it gets good at one job. Think of it like hiring a general doctor and then sending them through a short residency in cardiology: same foundation, sharper specialty.

For you: Fine-tuning is why so many 'different' AI products you see are actually the same underlying model, customized cheaply for a specific niche.

Original explainer

Anthropic bans sustained cruelty toward its Claude chatbot

Anthropic updated its usage rules to prohibit 'sustained and needless abusive or cruel behavior' toward Claude, part of its research into AI 'model welfare.' The policy still allows frustration, pushback, and dark creative writing; it targets only extreme, repeated cruelty with no purpose. The same update added bans on using Claude for election interference, weapons development, and surveillance of individuals.

The twist: Anthropic isn't claiming Claude has feelings, but treating 'model welfare' as worth a policy at all signals how seriously some labs take the question.

via The Verge AI

USA Today sues OpenAI for over $250 million

USA Today Co. and its local papers filed suit claiming OpenAI copied 'hundreds of thousands' of their articles to train its AI models without permission or payment. The publisher is seeking more than $250 million in damages. It joins a long list of outlets already suing OpenAI, including The New York Times, Ziff Davis, and nearly 400 local newspapers.

Follow the money: With dozens of publishers now suing, the real fight is less about any one lawsuit and more about what licensing AI training data should cost industry-wide.

via The Verge AI

A $16.9 megabyte AI model transcribes speech on a phone

A new open speech-recognition model called Whistle fits in 16.9 megabytes, runs entirely on a device's processor with no internet needed, and transcribes seven languages. It starts producing text in as little as 11 milliseconds and beat the popular Whisper model on several accuracy tests, while being about a ninth the size (16.9 MB vs 145.3 MB).

What's next: Tiny, fast, offline models like this are what let voice features run on cheap hardware like smart speakers and wearables without sending audio to the cloud.

via Hacker News (front page)

DARPA picks four finalists to build a useful quantum computer

DARPA named Atom Computing, Diraq, IBM and IonQ as finalists in its quest for a 'utility-scale' quantum computer, one that earns more money than it costs to run, by 2033. Each company could get up to $300 million in government funding for this next stage, which requires building and testing real prototypes rather than just plans on paper.

Zoom out: No single way of building a quantum computer has won yet, so this government bet is effectively picking which four technical approaches get years more runway.

via Fast Company Tech

A law firm cut its AI coding costs by 65 percent

LegalOn, a legal tech company, says it cut its estimated daily costs for OpenAI's Codex coding tool by 65% while keeping development speed the same. The trick: matching cheaper or faster AI models to the specific task at hand instead of using the priciest option for everything, plus actively managing spending budgets.

For you: This is a simple, copyable lesson for any small business using AI coding tools: pick the model based on the job, not habit, and the savings add up fast.

via OpenAI