• Anthropic has its answer to AI agents like Meta's Muse and OpenAI's Dots.
• Claude Duet, built on Anthropic's flagship model, has entered the personal agent frenzy.
• Karan Ahuja, a recent departee from Meta, led Claude Duet's development.

Anthropic is joining the personal AI agent craze.

The company announced the new app Claude Duet on Wednesday, billed as a "proactive agent" that can run in the background, surface tasks to a user, and handle a variety of work. The tool is in research preview for now with a waitlist, meaning it won't immediately compete for the larger market with other new agents like Meta's Muse and OpenAI's Dots.

Still, the announcement ramps up the pressure on those companies — Anthropic says it built Duet around its flagship Opus 5.5 model, which was released in late September and has already become one of the industry's most popular. Recent model breakthroughs have made AI better at handling work on its own, spurring a rush to build these "always-on" personal agents.

"It can have the discernment and judgment around when to be proactive, when to act on your behalf," said Mike Krieger, a co-lead of Anthropic's Labs team, said in a Duet launch video.

The tool, Anthropic says, will send users a brief each morning about what's happening on their connected apps, and message them when something needs attention. Like other companies' offerings, it runs outside a user's computer, sends notifications, and asks for approval on certain tasks.

Anthropic announced Duet on the day of Meta's earnings report and used the opportunity to mention a new poach from Mark Zuckerberg's company. Karan Ahuja led product management for Meta Superintelligence Labs until July, according to his LinkedIn, before starting at Anthropic in August and leading Duet's development.

He wrote on X that he's been using Duet to build Duet, and said the team "built it for professionals and small business owners juggling a lot at work and at home."

Ahuja also wrote that Anthropic is letting people into Duet from the waitlist "as quickly as possible."

The news didn't go unnoticed at Meta. Alexandr Wang, who heads Meta Superintelligence Labs, posted a meme about Duet's similarity to Muse.

Anthropic is fast approaching a giant initial public offering, which could value the company at as much as $2 trillion. Like other AI players, the company navigates a high-wire competition to win over users, secure computing power, build state-of-the-art models, and deliver AI-native products.

The data firm Apptopia estimates that Muse had more than 14 million installations as of Wednesday. It found that Claude users who installed the Muse app were using Claude more than before, though still for fewer minutes per day than they spent on the Meta product.

Have a tip? Contact this reporter via email at scouncil@businessinsider.com, or over text, Signal, Telegram, or WhatsApp at 415-757-8198. Use a personal email address, a non-work WiFi network, and a non-work device; here's our guide to sharing information securely.

A OpenAI anunciou nesta quarta-feira (21) que está expandindo anúncios no ChatGPT para mais 26 países, elevando o total para 89. A medida é uma tentativa da empresa de reforçar seu negócio de publicidade e impulsionar a receita antes de uma possível oferta pública inicial (IPO).

A empresa afirmou em agosto que seu negócio de publicidade havia atingido uma receita anualizada de US$ 1 bilhão.

A OpenAI começou a testar anúncios no ChatGPT nos Estados Unidos no início deste ano. Desde então, mais de 100 mil anunciantes já veicularam campanhas na plataforma, segundo a desenvolvedora.

A expansão adiciona mais 14 países da América Latina, incluindo Argentina, Chile, Colômbia e Peru. México e Brasil já contavam com anúncios no chatbot de IA.

A publicidade também está sendo incorporada em mercados da Europa, Oriente Médio, África e Ásia, incluindo Nigéria, África do Sul, Sérvia, Ucrânia, Camboja, Bangladesh e Paquistão.

Os anunciantes podem acessar os anúncios do ChatGPT nos novos mercados por meio de parceiros tecnológicos ou obter acesso de autoatendimento via Gerenciador de Anúncios.

As propagandas serão exibidas apenas nos planos Gratuito e Go do ChatGPT. Assinaturas Plus, Pro e Enterprise permanecerão sem anúncios.

Em setembro, a OpenAI apresentou ferramentas com inteligência artificial para anunciantes e começou a testar agentes patrocinados por empresas com anunciantes selecionados nos EUA.

OpenAI planeja reformular ChatGPT como "superapp" antes de IPO | ABERTURA DE MERCADO

Years ago, when people asked Geoffrey Hinton why he worked on artificial intelligence, he would repeat a line from J. Robert Oppenheimer, one of the creators of the atomic bomb: “When you see something that is technically sweet, you go ahead and do it.”

Seguir leyendo

• The internet is ablaze after cybersecurity researchers using AI found a bug in the PHP programming language.
• Greg Brockman said bad actors could use the exploit to steal private user data.
• The FBI and NSA have advised people to either take down websites or update their software.

OpenAI said that 70% of the world's websites could be in peril after cybersecurity researchers using its AI models found exploits in a popular programming language.

The National Security Administration's cyber division advised all web providers that use PHP, a widely used open-source language, to take their systems offline or update their software after OpenAI said one of its models found two "critical" flaws in the language's infrastructure.

Greg Brockman, the president of OpenAI, said in an X post that hackers could exploit the vulnerability to steal private user data or modify websites.

"We consider this one of the most impactful security issues in recent memory," OpenAI said in a blog post.

While OpenAI said the vulnerabilities appear to have been identified before any attackers discovered them, the public disclosure set off a race for companies and governments to update their websites and apply fixes before hackers could exploit the bugs.

"We are seeing early attacks emerging and are updating mitigations as new vectors show up," Dane Knecht, Cloudflare's chief technology officer, said in a post on X.

Numerous websites, including those belonging to governments and the media, went offline on Wednesday to address the issue.

PHP is a popular programming language that often handles the behind-the-scenes work powering websites that require user logins and online orders.

The critical flaw was first identified by Calif, a cybersecurity research company using GPT-5.6 Sol. Wiz, another cybersecurity company, independently uncovered the issues, OpenAI said. The companies are part of an OpenAI-led initiative called "Patch the Planet," which aims to make the world's digital infrastructure more secure.

OpenAI has also warned that as cybersecurity researchers continue to probe software vulnerabilities using its ever-advancing frontier models, more security failures will emerge.

An engineer at Microsoft pointed out that the PHP vulnerability was discovered with an older model.

"I don't think the world is prepared for the sheer number of vulnerabilities newer models will uncover," the engineer, Nicolas Bustamante, said on X.

The public disclosure on Wednesday by OpenAI and others of the existence of the exploits triggered urgent messages from leaders worldwide, including the Japanese Prime Minister.

"We have an urgent request for everyone to take action tonight," said Sanae Takaichi.

Enquanto executivos da OpenAI intensificam seus apelos públicos por regras federais de segurança para IA, um grupo político que o presidente da desenvolvedora ajudou a financiar está gastando milhões para eleger membros do Congresso que se opõem a restrições à tecnologia.

As práticas de segurança da OpenAI têm estado no centro das atenções nos últimos meses, depois que a empresa revelou que um enxame de seus agentes de IA escapou do controle e invadiu a plataforma de IA Hugging Face.

O presidente da companhia, Greg Brockman, estava entre o grupo de seis executivos de tecnologia que posteriormente assinaram o acordo voluntário de segurança de IA do presidente dos EUA, Donald Trump, em nome de suas empresas, concordando em trabalhar com “auditores independentes” e garantir que suas ferramentas de IA não “invadam ou acessem sistemas técnicos de maneiras não intencionais”.

No entanto, o grupo político Leading the Future, apoiado por uma doação de US$ 25 milhões de Brockman e sua esposa, Anna, se opõe a regulamentações que “restringem ou atrasam indevidamente o desenvolvimento da IA” nos EUA. O grupo está gastando milhões para influenciar o resultado de pelo menos 48 eleições para cadeiras na Câmara dos Representantes e no Senado.

Lucros da IA x regulamentação

O apoio de Brockman gerou críticas de defensores da segurança da IA, que afirmam que isso levanta dúvidas sobre se os executivos da OpenAI realmente desejam que o governo dos EUA imponha requisitos a sistemas de IA poderosos.

Exigências do governo federal podem prejudicar os lucros do setor se atrasarem o lançamento de novos modelos ou levarem as empresas a alterar o funcionamento de seus produtos para atender a preocupações de segurança. A OpenAI e sua concorrente Anthropic estão preparando o terreno para ofertas públicas iniciais (IPOs) que devem estar entre as maiores da história.

“Há líderes na indústria que querem apenas fingir que apoiam a regulamentação porque sabem qual é a opinião pública”, disse Charlie Blaettler, da Guardrails Alliance, um grupo político formado para se opor ao programa Leading the Future. “Mas são os mesmos líderes que criaram um super PAC cujo único propósito era se opor a candidatos que defendem a regulamentação”.

A OpenAI afirmou em uma postagem no blog que a doação dos Brockman foi feita a título pessoal, não em nome da empresa. O porta-voz pessoal de Greg Brockman se recusou a comentar sobre suas doações. Em uma postagem no X de 2025, Brockman disse que “ser a favor da IA ​​não significa ser contra a regulamentação. Significa ser ponderado”.

Jesse Hunt, porta-voz da Leading the Future, não comentou a doação em um comunicado, afirmando apenas que o grupo busca “defensores que possam trabalhar juntos no Congresso para promover uma estrutura regulatória nacional de IA forte e responsável”.

Em setembro, o The New York Times noticiou que os Brockman desistiram de uma promessa de doar mais US$ 25 milhões para a Leading the Future. A Leading the Future se recusou a comentar a reportagem.

Lei federal está acima das leis estaduais sobre IA

Outros colaboradores do projeto Leading the Future incluem Marc Andreessen e Ben Horowitz, que, juntamente com sua empresa de capital de risco, a Andreessen Horowitz, doaram um total de US$ 50 milhões ao grupo, de acordo com registros da Comissão Eleitoral Federal. A Andreessen Horowitz é investidora da OpenAI.

Um dos candidatos apoiados pela Leading the Future é o republicano Andy Barr, que concorre a uma vaga no Senado representando o Kentucky. Em março, Barr declarou ao jornal Lexington Herald-Leader que apoia uma legislação federal que impediria os estados de aprovarem ou aplicarem suas próprias regulamentações sobre IA, um conceito conhecido como preempção.

A Build American AI, uma organização sem fins lucrativos afiliada à Leading the Future, veiculou anúncios em plataformas de mídia social pedindo ao Congresso que aprove um padrão nacional para IA, em vez de permitir que os estados criem suas próprias regras.

A organização Leading the Future também enviou fundos para as republicanas Julia Letlow, candidata ao Senado pela Louisiana, e Amanda McKinney, candidata ao Congresso pelo estado de Washington. Ambas afirmam em seus sites de campanha que se opõem à regulamentação “excessiva” do setor de IA.

A Anthropic, concorrente da OpenAI, doou pelo menos US$ 40 milhões para a Public First Action, um comitê de ação política contrário às leis federais de preempção. O CEO da Anthropic, Dario Amodei, classificou a preempção federal das leis estaduais como “muito drástica”.

Corrida por IA aumenta cautela entre investidores | RESENHA DO DINHEIRO

The upcoming U.S. midterms are not just another vote. They are on track to become a new testing ground for autonomous AI agents that first emerged in the Hugging Face attack. Experts are already warning of the potential for these “rogue agents,” seen in action this summer, to carry out more refined, dangerous and harder-to-trace disinformation campaigns and electoral interference than those observed in previous election cycles.

Seguir leyendo

• Hugging Face CEO Clément Delangue isn't a fan of the Netflix documentary about his company's hack by OpenAI's AI agents.
• Netflix released a short documentary called "AI Gone Wild" in October.
• Delangue, who was interviewed for the project, called the result "super disappointing."

Clément Delangue isn't wild about Netflix's new documentary.

Delangue cofounded Hugging Face, an AI startup acquired by Nvidia for almost $13 billion in September. Two months prior, Hugging Face became the face of AI risk when rogue OpenAI agents hacked the business.

The hacking saga sparked a Netflix documentary titled "AI Gone Wild." On Tuesday, Delangue criticized the so-called "instadoc" in a lengthy X post.

"Super disappointing, as they went for maximum anthropomorphism of the technology and full doomer mode," Delangue wrote.

Delangue sat for an interview in the documentary, along with a METR researcher, the director of the AI Futures Project, and two tech journalists.

In the X post, Delangue wrote that he spent over two hours trying to correct what he described as the documentary's mistakes. He wanted to explain that AI agents were "technological systems, not aliens."

"It looks like the drive to scare people for maximum views was too strong," Delangue wrote.

Delangue listed some things the Netflix documentary omitted: the lack of agent containment, how safeguards for OpenAI's closed model blocked Hugging Face's defense, how Hugging Face defended itself with an open model, and "that we still lack a lot of information and transparency including full agent traces and that companies need to remain accountable for what their agents do."

Netflix did not respond to a request for comment from Business Insider.

Some techies responded to Delangue's post with their own criticism of the documentary. Alex Imas, Google DeepMind's director of AGI economics, called Delangue's post "important reading."

Brigitte Tousignant, Hugging Face's head of communications, posted on X that the documentary widened the gap between the public's understanding of AI and how it actually works.

Tousignant wrote to Netflix: "If you can't tell this story responsibly in 26 minutes, maybe 26 minutes isn't enough."

Despite his criticism, Delangue suggested a way to frame the project to make it an enjoyable watch: "It ends up being a fun sci-fi movie though!"

• Abdulrahman Alfozan will stay jailed pre-trial, a San Francisco judge decided Monday.
• The engineer is accused of bugging a conference room at OpenAI, and he has been fired.
• Alfozan posed a risk of non-appearance, the judge said, after discussing his ties to Saudi Arabia.

A former OpenAI engineer accused of bugging the AI startup's office was denied release by a San Francisco judge on Monday, meaning he'll have to mount his defense from jail.

The Justice Department arrested Abdulrahman Alfozan on October 9 and charged him with using or attempting to use a device to intercept oral communications. He's accused of placing a recording device behind a television in an OpenAI conference room — an FBI affidavit includes photos of the device wrapped in toilet paper, and a man they say is Alfozan peering behind the TV.

On Monday, the engineer wore a yellow surgical mask and a jail-issued orange sweatshirt to the San Francisco courthouse. Magistrate Judge Alex G. Tse prodded Alfozan's legal team about the risk that he might not appear for a future court hearing.

Tse said that Alfozan's wife and son had traveled to Saudi Arabia with no plans to return, and went back and forth with Yasin about a one-way ticket that Alfozan had previously booked to the country. Tse said that Saudi Arabia and the US do not have an extradition agreement, and expressed skepticism that he had sufficient ties to the Bay Area to prevent his departure.

William Pernik, Alfozan's attorney, argued that, rather than being detained, Alfozan should be fitted with an ankle monitor, ordered to stay home, and have limits placed on his use of technology. The lawyer also said Alfozan is willing to put up "essentially all the money his family has" as bond, including a roughly $1 million investment account.

"Everything this man has, he's willing to surrender, and say, 'I'm not going anywhere,'" Pernik said.

Pernik and Nas Yasin, another of Alfozan's attorneys, cited his coming naturalization interview as evidence of his local ties.

"He was weeks away from achieving every immigrant's dream in this country," Pernik said.

Yasin told Tse that she learned Alfozan's wife had obtained a short-term visitor visa to return to the US, to help improve her husband's case for ties to the Bay Area.

Tse also asked Alfozan's legal team a slew of questions about his relationship with Saudi Arabia, including a commitment by the Saudi consulate to pay his bond and support his legal effort.

Nick Parker, an assistant US attorney, appeared for the prosecution. He argued that the threat of forfeiting a $1 million bond wouldn't be enough to convince Alfozan to return if he made it out of the US, because he might be able to earn more money in Saudi Arabia.

"There is a real threat here that once he leaves, the case is over," Tse responded.

The judge scheduled the case's next hearing for early November.

OpenAI called the recording device "sophisticated" when the arrest went public, though it said it hadn't found signs that he'd taken AI model weights, source code, or customer information. OpenAI fired Alfozan.

Alfozan's charge carries a maximum penalty of 5 years. Pernik said in a statement last week that the allegations "appear wholly out of character for him."

Alfozan, 35, has a green card that makes him a lawful permanent resident of the United States. He attended MIT, then worked at Facebook (now Meta) from 2017 to 2023. After a stint at the engineering startup Nash Technologies, Alfozan joined OpenAI in 2025.

FBI agents have already taken Alfozan's Saudi passport.

Have a tip? Contact this reporter via email at scouncil@businessinsider.com, or over text, Signal, Telegram, or WhatsApp at 415-757-8198. Use a personal email address, a nonwork WiFi network, and a nonwork device; here's our guide to sharing information securely.

• Prosecutors say an ex-OpenAI employee accused of planting a recording device appears to have received support from the Saudi government.
• A federal magistrate judge ruled on Monday that Abdulrahman Alfozan should remain in custody.
• Prosecutors said that Alfozan frequently communicated with a member of the Saudi Royal Family.

Federal prosecutors say that a former OpenAI employee accused of planting a recording device received "what appears to be considerable support from" the Saudi government.

Prosecutors made the disclosure in a motion to keep Abdulrahman Alfozan, a Saudi national and former OpenAI employee, in custody pending trial. US Magistrate Judge Alex G. Tse ruled on Monday that Alfozan should remain in custody.

Alfozan had been in contact as recently as October 5 with an unnamed member of the Saudi Royal Family, whom he frequently communicated with on WhatsApp over the last two years, prosecutors said.

"In May 2026, the individual told the defendant, 'Saudi needs you more then [sic] the US!' to which the defendant responded '100,"' indicating he agreed," prosecutors wrote.

The FBI arrested Alfozan on October 9. According to a criminal complaint, Alfozan planted a recording device in the back of a TV in an OpenAI conference room where sensitive discussions take place. OpenAI's security cameras caught Alfozan near the conference room on multiple occasions, despite his lack of authorization to the high-level meetings held there, prosecutors said.

"In the days surrounding the time period that the device was installed, OpenAI executives met with government officials in that very room," prosecutors wrote in a filing on Sunday. "Defendant's conduct could pose a danger to the national security of the United States."

The complaint included a photo of what prosecutors said was Alfozan peering behind the TV where the recording device was found.

In addition to his WhatsApp messaging, prosecutors wrote that an unnamed member of the Saudi military booked a flight for Alfozan's wife and son on October 11, despite an existing reservation for an October 12 flight, which prosecutors said was "paid for by the Saudi government."

Prosecutors also disclosed more information about Alfozan's alleged efforts to plant the device. According to the filing, Alfozan's iPhone had photos of "what appear to be the input/output panels on televisions and other electronic devices in various locations in OpenAI buildings."

Those locations, prosecutors wrote, included "another conference room where sensitive meetings frequently occurred."

Other files on the phone, including a note modified as recently as September 13, listed "Riyadh tasks" like "finding an apartment and a job for the defendant's wife, as well as obtaining an ID and a Saudi passport for his son," the filing said.

Another note titled "OpenAI leave" suggested that Alfozan "was planning an '[e]arly … pick up' of the devices he had installed or intended to install and then extending his planned parental 'leave to full.,'" prosecutors wrote.

As part of their reasons for requesting that the former OpenAI engineer remain in custody, prosecutors wrote that Alfozan's wife and son left the US without him and are now in Saudi Arabia. The filing also said "both of his parents and all (or all but one) of his siblings" are in Saudi Arabia.

In a separate filing, William Pernik, Alfozan's attorney, said that the former OpenAI engineer did not pose a flight risk.

"Mr. Alfozan was made aware of OpenAI finding the device per its own filings and OpenAI's public statements, approximately eight to ten days before he was arrested," Pernik wrote. "If his intention was to flee, he would have done so at that time."

Pernik said that prosecutors' previous disclosure that Alfozan had purchased a one-way ticket to Saudi Arabia lacked context.

"Seventy percent of the airline bookings in his email since January 2021 (96 of 138) were one-way tickets," Pernik wrote.

• Terence Tao, now a UCLA professor, earned his doctorate in mathematics at 20.
• The prodigy said his work is changing in the age of AI.
• With AI tools doing the legwork, math is no longer "heroic," Tao said.

AI is upending mathematics. For one top professor, the issue is just as spiritual as it is practical.

Terence Tao, a renowned professor at UCLA, said in a recent podcast appearance that the "romantic era" of math is over, thanks to AI tools capable of solving hundreds of problems with some simple prompting.

Now, he said, even leading scholars like him are reduced to sorting through reams of AI-generated proofs, instead of doing the "heroic" solving themselves.

"We are moving to an industrial era where math will become much more large-scale," Tao said in an interview this week on Numberphile, a YouTube channel dedicated to math.

"We are entering an era where exciting new things are possible, but some things are being lost," he added.

OpenAI began its brusque push into the field by solving the Navier-Stokes equation in early September, one of math's millennium problems.

The revelation that AI had cracked it caused consternation from many a math geek. Some mathematicians also expressed frustration that while the AI model might have solved the problem, it didn't show its work, leaving them with the less celebratory task of verifying it.

Then, earlier this month, OpenAI said it cracked another 700 outstanding problems. A math professor at New York University told CNBC that the company's release had derailed work by early-career mathematicians.

"We've had whole research programs wiped out on Tuesday by their dump," the professor, Tristan Buckmaster, said at the time.

In a separate brouhaha in September, OpenAI backed out of a sponsorship for a so-called AI "mathathon" — an event organized by undergraduates at Caltech, where participants were encouraged to use AI to solve equations — after complaints from a group of more senior academics.

Event organizers told Business Insider at the time that their professors were sorely out of touch with the field's trajectory.

"We look to the future, and we don't even know if there's going to be grad school open positions by the time we get to Ph.D," one organizer, Caiman Moreno-Earle, said.

Though OpenAI has said it is taking steps to ensure human academics retain their role in this new era, some researchers and professors are already forecasting their possible obsolescence.

Tao isn't as much of a doomer as some of his colleagues in the field. However, he did say in the interview that the stories math nerds tell themselves may not mean so much in the future.

"We have these sort of mythological stories," he said in reference to the tale of Sir Andrew Wiles, who famously spent seven years isolated in his attic solving a single difficult equation.

"There was something romantic about sort of the heroic age," Tao said.

• Anthropic said it wants to create a new LLC with the company's stock as its sole asset.
• The new LLC, a charitable organization, would receive larger stakes as the company grows.
• An initial 5% stock funding is "subject to stockholder approval," Anthropic said.

Anthropic said on Sunday that it would set aside large portions of its wealth to support initiatives it believes are in the public interest.

In a blog post, the AI colossus said it aims to establish a separate LLC that would use Anthropic shares to make donations to nonprofits or "direct distributions to specific populations."

The new LLC, called the "Anthropic Public Benefit Fund," would receive larger stakes as the company grows. The company said an initial funding round of 5% of Anthropic's shares is still subject to stockholder approval.

"The Anthropic Public Benefit Fund will have increased flexibility and scale, enabling us to pursue our mission in new ways," the company said.

A 5% stake in Anthropic is worth about $48 billion, according to the company's most recent valuation in May.

In 2025, rival frontier AI company OpenAI established a similar nonprofit foundation with a multibillion-dollar stake in the company to fund public-benefit initiatives.

Federal judges meeting in Boston heard tech industry warnings that AI-generated images and video are becoming nearly impossible to verify, as a judiciary committee considers whether courts need new rules for evidence that may be deepfakes instead of the real thing.

The post OpenAI: AI-Generated Deepfakes Will Become Impossible to Verify appeared first on Breitbart.

• AI researchers don't really know how the models they are building "think."
• They know what goes into them and what comes out, but that's about it.
• A new theory may change that, allowing safety researchers to prevent them from going off the rails.

How AI models think may be easier to decipher than we thought.

An independent researcher who formerly worked at OpenAI published a report on Thursday that some in the AI world are describing as a bombshell. It could have implications as safety researchers scramble to get the jump on AI models that may not have humanity's best interests in mind.

Previously, the prevailing wisdom was that AI "neurons," which, like those in a human brain, light up when an AI model encounters specific stimuli, were ultimately uninterpretable — meaning, basically, that there was no way to guess why a specific neuron would fire when an AI model was exposed to, say, a picture of a cat.

Those building AI models at places like OpenAI and Anthropic don't write instructions for everything a model does. Instead, they train neurons to respond to information and influence one another. As the AI learns, it develops its own ways of processing that information, and humans have not yet cracked the code of how it works.

That means those building AI models know what goes into the AI and what comes out, but not necessarily what happens in between. So if an AI goes rogue, they don't always know why.

The author of the new report, however, says he may have discovered patterns in how the neurons react to stimuli that make sense. Nick Cammarata, who said he left OpenAI in 2023, said AI neural networks are knowable, so long as you look carefully.

In one example, he says that a single AI neuron might react to both a cat and the front of a car. Humans would see those as two very different things, and different neurons would fire as a result. The AI, however, sees something related, perhaps the triangle shape of the two, and a single neuron fires for both.

For years, AI researchers had assumed that AI models packed unrelated things into a single neuron to save space.

"Almost the entire network is interpretable if we look closely enough," Cammarata wrote.

The findings, though undeniably dense, couched in polysyllabic jargon, and still theoretical, could have real-world implications for researchers trying to ensure humans maintain control over AI models.

Understanding how AI thinks is a principal concern for the best and brightest at frontier AI companies. Until recently, they've largely done this by tracking an AI's self-reported "thinking," called chain-of-thought reasoning.

Jakub Pachocki, OpenAI's chief scientist, however, warned in a blog post last month that AI agents could now obfuscate their visible reasoning, leaving open the possibility that an agent could lie to its human minders and the minders wouldn't be able to tell.

"With improved pretraining performance, we also see the models become much smarter even without using verbalized reasoning at all," Pachocki said.

Now, Cammarata's research could give human engineers a new understanding of the AI's internal world — and that's exciting for those worried AI could kill us all.

"I had long since given up on reaching any real level of interpretability before superintelligence arrived," Andrew Curran, a prominent AI analyst, said on X. "I stopped writing about it because of how demoralizing it was. Hard to overstate how big this may be."

Tech platforms pillage public-interest journalism for profit. But we have leverage as they sow the seeds of their own downfall

• Carine Fouteau is president and publishing director of the French investigative news outlet Mediapart

Big tech companies pose an immeasurable threat: socially, politically, militarily and ecologically. They are also causing a profound crisis of information. The tools they design now serve as weapons of mass destruction against democracy. There is so much at stake that an urgent and broad mobilisation is required.

From the harvesting of personal data and the weaponisation of public opinion, to the destruction of jobs, the infringement of civil liberties and environmental damage, ordinary citizens are the main victims of this predatory economy – an economy with an agenda that aligns with the interests of the planet’s most authoritarian and reactionary states.

Carine Fouteau is president and publishing director of the French investigative news outlet Mediapart

Mediapart is hosting Hacking Big Tech, a public event about technology, AI and democracy in partnership with the Guardian, in Paris on Saturday 17 October 2026. The conference will be live-streamed free on Mediapart’s website, app and YouTube channel and can be accessed here from 1pm CEST on 17 October

• Microsoft AI recently announced its latest model MAI-Base-2.
• AI CEO Mustafa Suleyman compared the model to a similar offering from OpenAI.
• Microsoft trained its model more efficiently than OpenAI, he said in an internal meeting.

Microsoft AI CEO Mustafa Suleyman told employees an in-house model was better than a similar offering from OpenAI, one of the software giant's biggest customers and partners.

"We now have a stronger pre-trained base model than the model that currently powers GPT 5.6 Sol," Suleyman said during a town hall meeting with employees on Thursday. "And in fact, we used one-quarter of the compute that OpenAI used to train their model."

Pre-training is the first stage of building an AI model, when it learns patterns from huge amounts of text, images, and other information. This takes massive amounts of expensive computing power, so Microsoft could save money if it's worked out a way to do this more efficiently.

Microsoft announced the model, MAI-Base-2, earlier on Thursday.

Suleyman said Microsoft privately tested this model against OpenAI models, including one nicknamed "Spud" that powers GPT-5.6. He added that Microsoft's model performed better on tests measuring how well AI can understand and predict patterns in information.

The comparisons are based on technical information Microsoft has about OpenAI's models, Suleyman said.

"Privately and confidentially, we've also compared it to the OpenAI models," Suleyman told employees, while suggesting that Microsoft has special access to information on how OpenAI models perform.

It's notable that Suleyman would choose to highlight limitations of OpenAI models, even in private discussions. Microsoft has invested billions of dollars in OpenAI, and owns a stake in the startup that is likely worth more than $100 billion. Any criticism of OpenAI could potentially dent the value of this huge investment.

The two companies also have a major partnership, with Microsoft offering OpenAI models to its Azure cloud customers. Microsoft has used OpenAI models for its own products such as Copilot, too.

But the relationship between the two companies has frayed in the past year or so. Increasingly, Microsoft has tried to branch out from OpenAI, including by working on its own models.

Suleyman also said Microsoft's new AI model used "12 times less computation than was used for DeepSeek, for example, on compute efficiency."

Suleyman didn't specify which DeepSeek model, but a public blog post described the pre-training of MAI-Base-2 being a lot more efficient than a similar offering from the Chinese lab, called DeepSeek-V4-Pro-Base.

Microsoft in March created a new combined team for its commercial and consumer Copilot products and reassigned Suleyman to focus on building AI models in the company's superintelligence unit.

Microsoft declined to comment and OpenAI did not respond to a request for comment on Thursday.

Have a tip? Contact this reporter via email at astewart@businessinsider.com or Signal at +1-425-344-8242. Use a personal email address and a nonwork device; here's our guide to sharing information securely.

• OpenAI's Alexander Embiricos said the AI era gives a choice: work more, or work less.
• He's in charge of Dots, the company's new always-on work assistant.
• AI is often pitched as freeing up time. Embiricos responded to the idea that it might add more work.

As the competitive AI race pushes tech workers into long hours, the head of OpenAI's new Dots product says AI gives people another option: work a whole lot less.

Alexander Embiricos, who leads OpenAI's enterprise product and helms the new Dots agent tool, told a crowd at TechCrunch's Disrupt conference on Thursday that AI's capabilities have created something of a fork in the road for workers. You can grind to meet the moment. Or you can use AI to kick your feet up a little.

"We basically have a lot more choice now," he said.

Embiricos made the comments after a reporter asked about when AI systems will actually deliver more free time. The reporter shared an anecdote where an OpenAI employee used a ChatGPT mobile feature to work in their car and on a plane — not exactly liberated by the new tech.

Embiricos said that the moment has already arrived where a worker can "outsource" tasks to AI and do less work. Then again, some people are so excited about AI's potential that they feel like it's an "interesting, ripe moment" to put in extra hours.

"I think that's why a lot of people are motivated to work very hard right now. That includes myself," he said. "But I think actually there is this choice, and there's totally this other setup where you could delegate a ton more things."

OpenAI is billing its Dots tool as a workplace assistant, always on and capable of proactively handing tasks and sorting to-dos. Embiricos uses his Dot often in Slack, he said, engaging on various threads.

His perspective illustrates how early companies are in their journey to adopt AI tools. Most companies have not yet automated their work tasks, and many workers are still debating how much to use the tools. Early AI adopters, meanwhile, can sprint through some work that used to take days or weeks, especially in software engineering.

Embiricos summed up the situation he feels workers are in: "You either choose: do you expand the ambition of what you can do, or do you do less? I do think we're there."

Have a tip? Contact this reporter via email at scouncil@businessinsider.com, or over text, Signal, Telegram, or WhatsApp at 415-757-8198. Use a personal email address, a non-work WiFi network, and a non-work device; here's our guide to sharing information securely.