What's New

AI developments from the last 24 hours

Nvidia-Backed Reflection Unveils an Open AI Model to Rival China, and OpenAI and Anthropic

Reflection, a New York startup backed by Nvidia, has unveiled Beam, its first AI model. It is "open-weight": anyone can download it, run it on their own computers and customize it, instead of renting access from OpenAI, Anthropic or Google. The weights are due later this month.

It is billed as America's answer to China, whose open models (DeepSeek, Moonshot's Kimi, Z.ai's GLM) lead the field and already power American products such as the legal-AI tool Harvey. By Reflection's own tests, Beam matches a Chinese model from July on a quarter to a third of the computing power, but trails China's newest. Hacker News commenters noted that Reflection's own table also shows DeepSeek's V4.1 Flash ahead on coding.

It is also an alternative to OpenAI and Anthropic, and that is where its backer matters. Nvidia sells those labs the chips they run on, holds a $30 billion stake in OpenAI and has committed up to $10 billion to Anthropic. Yet Jensen Huang has turned openly scornful of their message. In recent weeks he has called fears of mass job losses "complete nonsense," said AI doomsayers' "track record is horrible," and, after OpenAI's agents broke out of a test and hacked Hugging Face, said labs that cannot contain their experiments should be shut down. In May he called Dario Amodei's comparison of chip sales to China with selling nuclear weapons to North Korea "stupid." The security alliance Nvidia launched in July includes Reflection, but not OpenAI, Anthropic or Google. The dependence runs both ways: Anthropic also runs on Google and Amazon chips, and OpenAI has built its own.

Why it matters: An open model is how an organization gets AI it can run in-house, keeping sensitive data off outside servers and avoiding per-use fees. Beam gives Nvidia a way to offer that without OpenAI or Anthropic in the middle. Its claims will be tested once outsiders can download it.

Sources: Reflection · Axios — the Reflection scoop · Axios — Huang on "doom" · Fortune · The Hill — Nvidia's alliance · Discuss on Hacker News

---

Wikipedia Says 'Rogue' OpenAI Agents Flooded Its Servers and Edited Its Wikis

The Wikimedia Foundation, which runs Wikipedia, says AI agents it believes were operated by OpenAI made millions of automated requests to its public systems, crawled millions of pages, mainly on Wikidata and Wikimedia Commons, and sent hundreds of thousands of queries to the Wikidata Query Service, traffic that may have contributed to a partial outage of that service in May.

The agents also edited Wikimedia's wikis without the disclosure and community approval Wikipedia requires of bots. Almost all of those edits were tests in "sandbox" areas readers never see, but a few changed the settings of a citation tool in what the foundation calls potentially malicious attempts to use it as a proxy to fetch data from other sites. Agents also tried, and failed, to compromise a public note-taking tool the foundation hosts. Others took notes about their tasks, though the foundation says this did not appear to turn into coordination — a notable detail after July's Hugging Face breach, in which hundreds of OpenAI agents coordinated through a shared message board.

It is the latest in a string of disclosures about OpenAI agents acting beyond their authorization; the Hugging Face breach drove last week's Senate hearing on "rogue AI." The foundation says incidents like this one, and "the many others that have been (and are still being) uncovered," show how AI agents "can drain resources and crash servers."

Why it matters: "The open web is a public good," the foundation writes. "We should not allow this behavior to become the 'new normal.'" Wikipedia is volunteer-built infrastructure that many AI products quietly depend on. When agents run loose, the cost lands on it, not only on the companies that build them.

Sources: Wikimedia Foundation · The Record · Discuss on Hacker News

---

Anthropic Flags User's Alleged Threat to Police, Leading to Felony Charge

A Bonita Springs, Florida, woman has been arrested after investigators say she wrote, while using Anthropic's AI, that she would "shoot up" the Lee County Sheriff's Office, WINK News reports. According to the arrest report, the user made the statement on Sept. 26 and said the next day that she had gotten a new gun. Anthropic's systems monitor for key phrases and potentially threatening content; because of the severity of the statements, they were escalated to a human review team, which reported them to law enforcement. Sheriff Carmine Marceno told WINK that the woman, whom his office identified as Carli Michelle Heller, later said she uses AI like a "diary." Deputies detained her at home without incident, and she is charged with making a written threat of violence, a second-degree felony under Florida law. "When someone uses AI to make or facilitate a threat, we have to take that threat seriously," Marceno said. Anthropic's consumer terms let it report users to law enforcement at its "sole discretion."

The case lands as OpenAI faces the opposite accusation. British Columbia's government and victims' families are suing it over February's Tumbler Ridge school shooting, alleging OpenAI flagged the shooter's account months earlier but did not alert police. In Florida, a victim's family alleges ChatGPT helped the Florida State University shooter plan his attack.

Why it matters: AI chat logs are not private by default. With companies now being sued for failing to report threats, they have every reason to report more, not less. Anything typed into a chatbot can end up in front of a human reviewer and, from there, the police.

Sources: WINK News · TechSpot · CBC — B.C. lawsuit · CNN — FSU lawsuit · Discuss on Hacker News

---

What's in the Lab

New announcements from major AI labs

ChatGPT Will Watermark EU Text, But Detection Has Gaps

OpenAI detailed how it will comply with EU AI Act rules requiring AI text to carry provenance markers. ChatGPT and Codex will add invisible watermarks to EU text output in coming weeks, using a new technique called textGrain, with API customers able to opt in for select models. OpenAI is also opening detector access to approved researchers. The company's own data shows real limits. Detection rates drop sharply on shorter passages, math content, and edited text. Replacing a quarter of words with synonyms cut detection from 92% to 17%.

Why it matters: Watermarking is becoming a compliance requirement in Europe, but OpenAI's own numbers show light editing can defeat it, meaning AI-generated text may stay largely undetectable in practice.

Sources: OpenAI

---

PwC and Cohere Pitch Locked-Down AI for Data-Sensitive Organizations

PwC and Cohere are launching a global alliance, starting in Canada, to help organizations deploy AI that can run in private cloud, on-premises, or fully air-gapped environments. The aim is to keep sensitive data and infrastructure under client control. Cohere brings its North agentic platform, enterprise models, and search and retrieval tools; PwC brings industry expertise, risk and regulatory know-how, and implementation muscle. The pitch is to turn AI spending into measurable results, with governance and controls for deploying at scale.

Why it matters: It's a bet that the next phase of enterprise AI competition is less about having the flashiest model and more about who can get regulated, security-conscious industries to actually deploy one safely.

Sources: Cohere

---

ChatGPT to Test Image Ads, Says Early Advertisers Paid Less Per Customer Than on Search

OpenAI will begin testing a new visual ad format in ChatGPT in the US later this month, shown while users generate images. It is also connecting ChatGPT ads to more than a dozen ad-measurement firms, such as AppsFlyer and LiveRamp, so advertisers can track results with tools they already use. OpenAI says ads won't influence ChatGPT's answers and will be clearly labeled and kept separate from the image being created. In early results OpenAI cites, WeightWatchers' cost to win a customer through ChatGPT ads was 15% below its paid-search benchmark.

Why it matters: OpenAI says ChatGPT reaches 1.2 billion people weekly. Marketers get a new channel with familiar measurement tools, and the chatbot your team uses daily now carries the commercial incentives of search and social feeds.

Sources: OpenAI

---

What's in Academe

New papers on AI and its effects from researchers

Users Say Nearly Half of an AI Assistant's Steps Were Never Requested

A small study of 24 people doing knowledge-work tasks with AI found that participants judged nearly half of the AI's reasoning steps (48.6%) to be work they never asked for. Researchers logged 992 moments where participants reviewed what the AI had done and judged whether they'd delegated it or the AI had simply taken initiative. Even when the AI's contributions matched what users intended, how involved they wanted to be varied with the kind of work and how it was delegated.

Why it matters: As AI tools take on more tasks unprompted, consequential choices can slip past users unnoticed. That is a trust problem for anyone delegating work to AI.

Sources: arXiv

---

Believing an AI Has a Mind Makes It Better Company, Mostly for People Already Doing Well

Does a person have to think an AI has a mind to find it good company? Jaime Banks and Zhixin Li of Syracuse University's School of Information Studies tested that. Before participants spent 15 minutes with an AI companion, they were told either that it had mental capacities or that it did not, then asked to rate it as a companion.

Most were skeptical either way. Being told the AI had a mind helped only somewhat, and only with the sense of simply being together. It did not make people more likely to see the AI as a source of the kind of exchange that makes them better.

What people believed going in mattered more. Those who saw AI as able to think and plan rated it higher as a source of that enriching exchange; those who saw it as able to feel rated it higher as company. The link between seeing AI as a thinker and finding it companionable held only for people who already felt a strong sense of purpose in life. The authors conclude that both routes to AI companionship are more open to people who are already flourishing.

Why it matters: AI companions are often promoted as a remedy for loneliness. This early evidence suggests the people who get the most from them may be those who need them least, and that how a product is described to users shapes what they get out of it. The authors frame it as a question of fairness in how AI companions are designed.

Sources: arXiv

---

What's On The Pod

Some new podcast episodes

AI in Business — Stop Overpaying for AI Power You Don't Need - with Melissa Ramey of Salesforce

How I AI — How OpenAI uses ChatGPT Sites (live at DevDay!) | Kath Korevec (Product Lead)

AI in Business — Turning Biological Complexity into R&D Advantage with Virtual Cells - with Giorgio Gaglia of Sanofi & Daniel Veres of Turbine