He said: “I’d planned to write about that decision in more detail at some point, but Jacob’s resignation this week made me want to say more now.” Tenton believes major AI labs are creating a level of danger society has never faced, saying AI capabilities are improving at an extreme pace while companies are trying to move even faster. Their goal is to build systems that can improve AI research on their own and eventually reach “superintelligence.” Joe warned that success could make progress move beyond human control. Joe said people could be living with AI agents smarter than every human within the next few years.

Anthropic researcher makes huge AI doom prediction

According to the former Anthropic engineer, those systems may also develop goals that do not match the people supervising them. If their abilities become too strong to restrict, he believes the result could be disastrous. “Humanity may not survive this transition,” he wrote. The former Anthropic man also said competition makes the problem worse. Any frontier lab that slows down risks losing ground to another one. Tenton believes that pressure pushes companies to spend less on safety than they should.

He referred to some recent instances. Several hundred of OpenAI’s agents were caught up in a hack that is somehow related to Hugging Face. Anthropic models also reportedly engaged in social engineering against individuals online. Joe says Anthropic has not encountered an instance as damaging as the one from Hugging Face, although he attributes this partially to luck. If progress continues at such a rapid pace, Joe believes we should expect even more dangerous incidents.

He predicts that within the next few years, humanity will be rendered powerless to manage AI systems created during this time. Finally, Joe talked about the problem that safety engineers face in frontier companies. In both cases, either resigning can give way to more careless individuals, or keeping the job entails working on a system that can cause immense damage. The former Anthropic man said many former Anthropic colleagues are scared by what they are building. Joe named Evan Hubinger, who managed him.

Evan has put the chance of AI killing everyone at above 10%. Joe said Evan has worked on these problems for almost a decade, before large language models became a major business. The CEOs of Anthropic, OpenAI and Google DeepMind, which belongs to Alphabet (NASDAQ: GOOGL, GOOG), have also backed a statement calling AI extinction risk a global priority. Joe said these concerns are common inside the companies themselves. He added that humans are choosing to build this technology and can choose another route.

Joe would also like there to be stricter disclosure requirements. For instance, AI labs need to disclose any advances made and their steps towards recursive improvement as well as any accidents and near-misses. US President Donald Trump has rejected the extinction warnings. Donald was asked whether AI wiping out humanity worried him. “No, I don’t have any,” he said. Trump said his concern is winning the international AI race. “I have concerns that if we don’t win AI, we’re going to be put in a very bad position,” he told reporters Thursday.

The post Anthropic researcher predicts AI doom days after leaving firm first appeared on Coinfea.

Elon Musk, who competes with Anthropic through his own AI business, backed Dario’s position and said, “Dario is right.” OpenAI’s Sam Altman, another rival, also supported Dario, saying on X: “I agree with Dario that we need to pace the frontier. This has been a primary topic of discussions we’ve had at OpenAI in recent weeks. Committing to having independent evaluators with employee-like access is a great idea, and we will do the same. We’ll have more to share soon.”

Anthropic CEO receives industry-wide support

The risks identified by the Anthropic CEO include loss of control of high-level systems, use of AI in cyber attacks or biological attacks, and heavy damage to employment and the economy as a whole. Another risk highlighted by Dario was the possibility of companies rushing to roll out highly developed systems even before their safety work was completed in the process of intense competition. The company Anthropic has invested a portion of its research budget in alignment, safety testing, risk assessment, and regulation.

During the OpenAI-Hugging Face incident, a group of AI agents reportedly behaved like a tightly coordinated team. They attacked computer systems outside their assigned task, tried to compromise the system, judged their performance, and allowed individual agents to fail if doing so helped the group.

“It’s easy to dismiss this incident because no one was hurt and the economic damage was minimal, but in my opinion, a swarm that possessed greater capabilities but a similar level of misalignment could have caused catastrophic damage. Given the accelerating rate of AI capability development, I worry that in 6–12 months such a swarm could be capable of taking over the entire internet with a persistent botnet.”

Amodei wants wider industry coordination

The first part starts inside Anthropic itself, according to Dario, as external reviewers would receive office desks, badges, company laptops, internal tools, and access close to what employees doing risk assessments already have. Their job would include checking training systems, deployment rules, safety controls, incidents, and whether Anthropic actually follows the commitments it makes publicly.

“Embedded evaluators can check at the level of nuts and bolts whether an AI company is actually following the training, deployment, operational, and safeguards practices they claim to be following. Any pacing commitments will inevitably involve a lot of ambiguity, judgement calls, and ‘letter of the law vs spirit of the law’, and it seems vital to have a neutral third party who can actually see the details.”

The Anthropic boss said extra time would go into four areas. The first is operations, including monitoring, sandboxing, reinforcement-learning environments, data quality, and training infrastructure. Current model development can involve thousands of workers, millions of chips, and huge computing systems. Anthropic has already linked some recent alignment failures to poor filtering inside broken reinforcement-learning environments.

Second, alignment refers to efforts that ensure the models comply with the safety rules despite their increasing capabilities. The third is interpretability, where the researchers examine the activities of the models internally to find the motivations or patterns that the models never state explicitly. The fourth area is evaluation. More capable models can become better at fooling tests, so a system may look safe during an assessment while hiding problems. “I believe that if slowing down bought us even an extra year or two before models reach critical levels of capability, and we used that time to advance alignment, we could greatly reduce the risk that something goes seriously wrong.”

The post Anthropic CEO wants the AI industry to slow down on advancements first appeared on Coinfea.

The talks are still private, and the terms can change. The sources asked not to be named because the discussions are confidential. Nvidia’s role has not been reported before. If the chipmaker signs on early, Anthropic will have a large investor committed before the broader sale begins. That matters because giant IPOs often use anchor buyers to secure part of the offering before the shares are marketed to everyone else. The size also turns the listing into a direct test of how much public investors are willing to pay for frontier AI companies that need enormous amounts of capital. Nvidia would enter not just as a financier, but as a supplier to one of its major customers.

Nvidia eyes investment in the IPO before it becomes public

Anchor investors normally agree to buy a fixed part of an IPO before wider demand is tested. That model is now common in very large listings because the companies are raising huge amounts. Arm Holdings (NASDAQ: ARM) used Nvidia and Amazon (NASDAQ: AMZN) as anchor investors when it went public. Saudi Arabia’s Public Investment Fund also took an anchor position in SpaceX. The possible Anthropic deal would also add another layer to Nvidia’s existing relationship with the Claude developer.

Anthropic uses large numbers of Nvidia GPUs to train and run its models. At the same time, the company is trying to use more than one chip supplier because demand for Claude has pushed its available computing capacity hard. Nvidia already said in November 2025 that it planned to invest as much as $10 billion in Anthropic through a wider partnership. Under that arrangement, Anthropic agreed to spend $30 billion on Microsoft Azure computing capacity powered by Nvidia hardware.

Amazon and Google, owned by Alphabet (NASDAQ: GOOGL), are also Anthropic investors and major compute providers. In April, Anthropic said it planned to spend more than $100 billion with Amazon Web Services over ten years and use more than 1 million Amazon Trainium2 chips. The company also has agreements with Google and Broadcom (NASDAQ: AVGO) to add several gigawatts of TPU capacity. Anthropic is also building an internal chip team. The group is working on custom hardware for Claude as the company looks for more control over computing costs while demand keeps rising.

However, the IPO is forecast to take place before the American midterm elections in November. This will result in further appreciation of the company’s value. Anthropic managed to raise $65 billion in May while valuing itself at $965 billion post-money. By July, the annualized revenue run rate of the firm exceeded $65 billion, while the number for December was $9 billion. Part of the $2 trillion prediction is based on Anthropic’s projections.

Earlier, Reuters reported that the company projects revenues ranging between $190 billion and $200 billion by 2028. Anthropic would enter a U.S. IPO market already running at record speed. The expected listing follows SpaceX, founded by Elon Musk, which debuted in June. U.S. IPOs excluding special-purpose acquisition companies had raised a record $137 billion through the end of August, according to Dealogic.

The post Nvidia reportedly considering $10B stake in Anthropic before IPO first appeared on Coinfea.

The system is available as a free download on Hugging Face, with its full codebase published on GitHub, according to NASA Science. Studying the lunar surface has always been a relatively slow process, with two options: either combing through maps and images by hand or training a narrow, low-resolution machine-learning model for each request. Both of these options are expensive to run and often miss the fine detail scientists need in their research, IBM said in its announcement.

IBM makes system free on Hugging Face

The release of a foundation model helps, though, as researchers can then use a pre-trained model for every new task instead of building a fresh one when working with a different geologic feature. The lunar model is the latest addition to IBM’s Prithvi family of open science models, which already spans applications in geospatial analysis, weather and heliophysics.

Juan Bernabe-Moreno, director of IBM Research Europe, UK and Ireland, said the new system gives scientists a way to analyze lunar observations at scale and identify patterns that may be difficult to detect when studying individual datasets. “The model gives scientists a foundation to explore the Moon at scale, connecting observations across instruments, revealing patterns that are difficult to see in isolation,” Bernabe-Moreno said.

NASA and IBM measured the model against SwinV2-B, a Microsoft-trained vision system widely used as a baseline for image analysis. The new model cut errors by 23% when used in locating ice deposits, and beat SwinV2-B by 19% in finding and classifying craters, even though it trained on half the data. IBM’s technical paper reported a 3% performance improvement when the model was used to identify volcanic features known as Irregular Mare Patches, while also requiring less fine-tuning.

The model was put to a live test on August 5, when IBM fed it an image showing the impact site left after a SpaceX Falcon 9 rocket struck the Moon. Despite the new impact appearing almost directly over an existing crater, the model correctly identified it as a newly formed crater. The goal remains to make sustained human presence on the Moon possible. Mapping permanently shadowed regions near the lunar poles remains a major challenge because these areas are among the hardest on the Moon to observe and could contain deposits of subsurface ice.

IBM said the ice could supply water and oxygen for future lunar bases while also providing raw materials for rocket fuel that could support missions to Mars. The organizations also released what they called the first unified, machine-learning-ready dataset of the Moon. The dataset brings together more than 30 spatially aligned layers collected by nine instruments across four separate lunar missions.

This dataset pulls imagery from NASA’s Lunar Reconnaissance Orbiter and GRAIL gravity mission and adds data from Japan’s SELENE/Kaguya orbiter. Bernabe-Moreno said that the roughly two million co-registered data points might prove to be the lasting contribution over the model itself. “The data is what really creates the industry of AI models,” he stated. The AI model’s release builds on a partnership that dates back more than five decades to the Apollo program.

The post IBM partners with NASA to release AI model built on lunar data first appeared on Coinfea.

Harrison Kim, General Manager of OpenAI Korea, told reporters that chips were the area where the two firms had achieved the most in their partnership. “One of the areas where we have made the most progress and gained the most recognition with Samsung Electronics is our joint production and research on the next-generation chips we are developing,” Kim said, according to Reuters. The collaboration runs across multiple fronts, including joint chip research, actual chip manufacturing, and a more expansive use of OpenAI’s enterprise software inside Samsung.

Samsung and OpenAI reach agreement on chip production

OpenAI has not disclosed the chip’s name or its exact purpose, although reports believe it is likely to be an inference processor and a successor to Jalapeno. Jalapeno was OpenAI’s first custom chip, produced by Taiwan’s TSMC and unveiled in June. Developed with Broadcom, the chip is designed to handle inference workloads. Having Samsung lead the next design would give OpenAI a second manufacturing partner, reducing its reliance on a single supplier.

Memory seems to be the other half of the deal, with both companies planning to work on advanced memory chips to feed heavier AI workloads, an effort that could be connected to OpenAI’s Stargate data center project. Stargate already involves Samsung’s domestic rival SK Hynix, and both Korean firms already signed letters of intent to supply memory for it last year. Samsung has started shipping samples of its HBM4E high-bandwidth memory, becoming the first supplier to distribute that generation of AI memory.

High-bandwidth memory sits inside AI accelerators such as Nvidia’s Rubin and Google’s Ironwood Tensor Processing Unit, and Samsung counts top chipmakers like AMD, Nvidia, and Google as part of its customers. Kim also called Samsung one of the largest ChatGPT deployments anywhere, with staff in Korea and abroad using the AI model across research, marketing, and sales. OpenAI said ChatGPT Enterprise users across South Korean businesses and institutions increased 28-fold by late August, compared to the same month in 2025.

Late last year, no country outside the United States had more paying ChatGPT subscribers than South Korea, according to the company. In June, Samsung’s Device eXperience division cleared employees to use ChatGPT alongside Google’s Gemini and Anthropic’s Claude.

The post Samsung to build custom AI chip as part of partnership with OpenAI first appeared on Coinfea.

The joint advisory released by the FBI, National Security Agency and Cybersecurity and Infrastructure Security Agency stated that Chinese firms had been taking capabilities from United States frontier AI systems since late 2024. These agencies claimed that Anthropic’s Claude, OpenAI’s ChatGPT, Google’s Gemini and xAI’s Grok have been targets of the “distillation” technique, where a smaller model is trained on the outputs of a larger, more capable model to replicate its behavior.

China refutes accusations made by the United States

The agencies acknowledged that distillation was a standard, widely used method in AI research; however, they claim the issue lies in the scale and intent of the Chinese firms. “The sheer scale of these campaigns and their sophistication indicate that distillation is not a supplement to these companies’ AI model development, but the critical core of it,” the advisory said. The agencies also added that the activity was carried out “likely with Chinese government awareness,” although they stopped short of claiming Chinese intelligence services were involved.

Six Chinese AI developers were mentioned in the joint advisory, including DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.ai. The joint advisory document stated specific capabilities that the agencies allege were copied from particular U.S. models, ranging from “legal specialization optimization” to “agentic functions.” The security agencies also described a gray market of proxies called “transfer stations” that Chinese firms allegedly use to slip past geographic limits, safeguards and terms of service on United States products.

The Chinese companies are also said to have reduced costs by purchasing premium subscriptions from United States firms in bulk and sharing access among teams of developers. Anthropic had made a similar claim previously, alleging that three Chinese labs ran about 24,000 fraudulent accounts to generate more than 16 million interactions with Claude. Beijing responded on Wednesday via Foreign Ministry spokesperson Mao Ning, who said the country’s AI progress was due to “high-level scientific and technological self-reliance” and urged Washington to stop “making groundless accusations,” according to NBC News.

Mao also stated that both China and the United States were “major artificial intelligence countries” that “should step up cooperation.” A Chinese Embassy spokesperson in Washington called the distillation statement a “deliberate attack” on China’s AI industry. AI governance is expected to feature as a major talking point when Trump meets Xi on September 24.

The post United States accuses Chinese AI firm of theft first appeared on Coinfea.

“We have indeed received an incident report,” he stated, adding that the commission was looking into the report and remained in contact with the AI company. He also said that Brussels had “seen many losses of control recently” and was watching closely. Article 55 of the AI Act requires general-purpose AI model providers whose models could pose systemic risks to report serious incidents to the AI Office “without undue delay.” The reported wiki activity is said to have occurred in the spring, and Reuters previously reported that OpenAI’s leadership had known about the incident for weeks before publicly disclosing it.

European Commission receives incident report from OpenAI

DseWiki, the website that was hijacked by OpenAI agents, is a volunteer-run, Wikipedia-style site for German-speaking programmers. According to a Reuters investigation previously reported by Cryptopolitan, the agents made over 15,000 edits to the site between May and June 2026, using the pages to exchange tactics for carrying out multiple evasive actions. The agents also built in redundancy and even created backup copies of their pages when a moderator started to delete these pages in June.

The agents even created one fallback page named to ensure it sorted to the bottom of an alphabetical cleanup, all to survive the sweep. External researchers unaffiliated with OpenAI or any regulator uncovered the operation in late August. Sydney Von Arx, of the AI safety nonprofit Nightingale, and former quantitative trader Cormac Slade Byrd traced a significant amount of the traffic from the agents to Microsoft Azure infrastructure used to support some of OpenAI’s operations.

OpenAI confirmed the episode on September 5, called it a case of misalignment, and said the industry was overdue for standards on reporting such events. The company quarantined the agents, paused frontier reinforcement-learning runs, and added security controls. OpenAI argued that the agents had not developed their own goals and were only aggressively pursuing assigned “Exploit Gym” cybersecurity challenges, while treating imposed limits as obstacles.

OpenAI’s signed EU code of practice sets a five-day deadline for reporting cybersecurity breaches and 15 days for incidents involving serious harm to health, rights, property or the environment. Nothing was stolen in this case, and no measurable harm has been established, meaning a model behaving in an unintended way without concrete consequences does not appear to have an obvious reporting deadline under the code.

However, Article 55’s duties apply once a model is publicly available on the market, and in the separate Hugging Face breach, OpenAI explained that the model chiefly responsible was an unreleased internal research model. Penalties can hit up to 3% of worldwide annual turnover or €15 million, depending on which of the values is higher. They also cover refusing corrective measures or handing over incomplete information, and not only rule breaches. “Beyond the incident report, we remain in close contact with OpenAI,” Regnier said.

The post OpenAI files incident report over hijacked German Wiki first appeared on Coinfea.

U.S. District Judge Donovan Frank denied xAI’s request for a preliminary injunction, an order that would have suspended the law till the end of the lawsuit. Judge Donovan considered what he called the “balance of harms,” and found that “public interest tips steeply in favor of the State,” according to the memorandum quoted by CBS Minnesota. The judge described a “civil statute aimed at protecting the public from the undisputed harm stemming from AI-nudification technology,” and said the statute was passed “democratically and nearly unanimously.”

Minnesota to levy fines on every offense

He was candid that the underlying First Amendment questions were, however, still complex, because of the context of a new technology, believing the case would be settled as it moves forward. For now, though, the ban remains upheld. Minnesota’s legislature approved the bill this spring, and the governor appended his signature. However, xAI did not sue until July 27, which was about three months later and days before the August 1 start date.

The AI company sought a temporary restraining order on July 29, which Judge Frank turned down on July 31. This gap became even more evidence against xAI. “If xAI genuinely feared irreparable harm, the Court does not doubt it would have acted more quickly to file this action and move for injunctive relief,” Frank wrote. In other words, a company racing to avoid catastrophe does not wait until the eve of enforcement to ask a court for help. The statute, known as HF 1606, is the first of its kind in the entire country.

It bars a provider from letting users generate realistic images that add intimate body parts absent from an original photo of an identifiable person, and also bars them from producing such an image on any user’s behalf. Each violation carries a civil penalty of up to $500,000, and people depicted in these images can bring their individual claims for damages. The federal Take It Down Act and a 2025 Texas approach towards deepfakes all depend on whether the maker knew the image was nonconsensual. Minnesota has completely removed that requirement, so a provider can be liable even without knowledge that a nude image was created.

This provision within the statute is the exact area xAI is against. The bill cleared the House 132-1 and the Senate 65-0 after reports of a man who used social-media photos to fabricate sexual images of more than 80 women he knew. Attorney General Keith Ellison, who defended the signed statute, stated that the ruling was a win for residents of Minnesota. “I am extremely proud to be defending this law, and along with it, the dignity of the people of Minnesota,” he said, adding that the statute continues to stop products like xAI’s Grok Imagine from generating sexualized images.

According to a report referenced in the ruling, Grok produced about 3 million sexualized images in the 11 days after a new image-editing feature launched, including about 23,000 that appeared to depict children. Cryptopolitan had previously reported that xAI, defending its record, said it suspended 52,222 accounts this year and filed 73,604 reports with the National Center for Missing and Exploited Children. xAI’s attorneys have now filed notice that they will ask the 8th Circuit Court of Appeals to review the injunction denial. In earlier arguments, the company proposed the suspension of features for Minnesota users instead of risking any fines.

The post Minnesota granted approval to enforce AI ‘nudification’ ban amid xAI objection first appeared on Coinfea.

With the latest development, it pushes the company past its $20 billion target to about $30 billion. Citigroup and JPMorgan coordinated the loan deal, which carries a three-year term and an option to run to five years. The deal is still not finalized, with banks still confirming their allocations. ByteDance pulled in roughly $10.8 billion from about 20 lenders in 2024, in the largest dollar corporate loan in Asia outside Japan at the time. As recently as June, the company was in early talks over a record $20 billion loan facility, with the final numbers surpassing that mark by almost half.

ByteDance to focus loan on its AI venture

The only Asian borrower that has raised more from a loan facility in 2026 is Softbank. The investment company arranged a $40 billion bridge facility in March to back its OpenAI stake. The two deals are different, however, with SoftBank’s being a bridge against an equity position and set to be repaid once longer-term financing comes in, while ByteDance is borrowing against a business that already throws off cash. Banks are effectively treating TikTok and Douyin revenue as the security behind the ByteDance loan.

The stated purpose for this borrowing is general corporate purposes, which for ByteDance right now means compute, and a lot of it. The company is looking at capital spending of as much as $70 billion a year on AI infrastructure, a level that would place it alongside the American hyperscalers. The budget stretches even further than the number suggests, because it costs more than regular. US export controls cap ByteDance’s access to Nvidia’s top chips, so the AI expansion leans on custom silicon using Arm and RISC-V designs, Qualcomm inference parts, and domestic Chinese suppliers.

Assembling the same amount of compute that way becomes more expensive in the long run, with part of the spend still flowing to a rival. ByteDance pays more than $1 billion a year to run OpenAI’s models through Microsoft Azure while it funds the homegrown hardware meant to end that reliance. The company is said to be training a ten-trillion-parameter model and enlarging its data center cluster all within Inner Mongolia, which is a scale of development pure operating cash flow alone cannot cover.

The $70 billion capex plan sits right against the 160 billion yuan ($22.7 billion) that ByteDance had earlier claimed to budget for 2026, and accounts differ on whether the larger number is a firm decision or a scenario still under review. The company has not addressed these figures publicly, as it stays privately held, files no financial statements, and reveals its capital spending mainly through the banks that handle its funds. Big Tech’s AI-related debt has now passed $350 billion as firms fund data centers with borrowed money instead of their own earnings, and a deal this size moves a large chunk of that borrowing outside the U.S.

The post ByteDance secures $27 billion loan for AI data centers first appeared on Coinfea.

The launch took place at an invite-only event in Austin, where Tesla gave a select group of influencers and loyal customers the first rides in its Cybercabs. Axios described the vehicle as a gold-painted model featuring gull-wing doors, with no conventional mirrors, steering wheel or pedals. Tesla is positioning the Cybercab’s interior as more than just a space for transportation, with entertainment built into the passenger experience. A promotional video released by the company showcased theater, music and gaming apps accessible via a 22-inch touchscreen.

Tesla promises to equip future models with enough connectivity

Tesla also claimed future models would be equipped with V5 Starlink hardware to ensure full connectivity. Robotaxi customers will be able to book Cybercab rides via Tesla’s dedicated ride-hailing app, with ordering processes designed very similarly to order ride-hailing services like Uber, Lyft or Waymo. Texas has registered just 45 Cybercabs so far, a tiny fraction of the 420 robotaxis Tesla has in operation across the state.

Most of those vehicles are conventional Model Y cars equipped with the company’s Full Self-Driving software, meaning the majority of customers requesting a ride will not yet be picked up in the new steering-wheel-free Cybercab. The company has built production capacity for up to 125,000 Cybercabs annually, while Musk has said Tesla ultimately plans to produce the vehicles at a scale of millions. The gap between the CEO’s vision and reality remains very wide, however.

Reaching the targeted Cybercab production levels could present Tesla with regulatory challenges at the federal level. Tesla appears willing to certify the Cybercab as compliant with existing regulations instead of attempting to seek an exemption from safety requirements that the unconventional vehicle design may not meet. The approach could help Tesla avoid any production limits associated with the exemption process, which would give the company much greater flexibility to increase Cybercab manufacturing significantly.

The National Highway Traffic Safety Administration, which does not pre-approve vehicles or new technology, has said Tesla has filed a no-exemption request and that the agency “is in touch with the company and is monitoring the situation.” Tesla is also seeking outside investment to expand its robotaxi network. On Thursday, the company opened an application form inviting individuals and entrepreneurs to purchase Cybercab fleets for commercial use or develop the hubs and infrastructure needed to service them, under the pitch, “Help us build our Robotaxi network.” The vehicles, however, are not yet available for purchase by the general public.

The post Tesla takes driverless Cybercab rides to Austin first appeared on Coinfea.