Artificial Intelligence

You probably know Midjourney for generating AI artwork from text prompts. Now, the company is making a move so unexpected that it describes it as "a little weird and a little crazy." Through a new division called Midjourney Medical, it is developing a full-body ultrasound scanner that it claims can map the inside of your body in under 60 seconds.

The Midjourney Scanner works by submerging users in water at about 2 inches per second, using a ring of ultrasonic sensors. The current prototype uses 40 imaging modules, each containing 358,000 ultrasonic elements, that transmit sound waves through the body and record the returning echoes.

Since sound travels differently through fat, muscle, bone, and organs, the system reconstructs those differences into detailed 3D images. Midjourney describes the experience as being surrounded by half a million tiny dolphins using echolocation.

Anthropic's CEO, Dario Amodei, has sat down with Bloomberg for a lengthy conversation about AI and related topics.

One of the topics touched on during the conversation was Anthropic's decision to sever China's access to its Claude models, which it announced in February 2026. Amodei stated during the interview that Anthropic's decision to limit the use of its models in China cost hundreds of millions of dollars, which at the time represented a significant portion of the company's revenue.

As for what Anthropic has officially stated about the matter, in its February 26 statement titled "Statement from Dario Amodei on our discussions with the Department of War," Amodei explains that it stopped several firms linked to the Chinese Communist Party from using Claude.

NVIDIA has unveiled a framework that enables AI coding agents to train robots to perform high-precision physical tasks without human supervision. Developed alongside Carnegie Mellon University and UC Berkeley at NVIDIA's Generalist Embodied Agent Research lab, the system, called ENPIRE, closes the loop between writing robot training code, testing it on real hardware, and refining it until it learns the desired behavior.

The demo that caught everyone's attention shows a robot arm selecting a graphics card, passing it to a second arm, and carefully seating it into a PCIe slot on a motherboard. The test included a few other tasks, such as sorting metal pins into a box, cutting zip ties with real cutters, and the classic Push-T benchmark.

Across these contact-heavy tasks, the system achieved a 99% success rate under a pass@8 metric, which allows up to eight attempts per subtask with each retry informed by the previous failure.

Anthropic CEO Dario Amodei has sat down for an interview with Bloomberg Originals, where he discussed several topics surrounding AI and AI development, along with geopolitical impacts and Anthropic's internal governance system intended to maintain checks and balances during the quest to ethically develop artificial intelligence.

Amodei was asked, given the power of artificial intelligence and how Anthropic is one of the leading companies in the space with Claude, Claude Code, and its now-banned Mythos-class AI models, why the US government wouldn't take over the company to ensure its tools are used safely. Amodei responded by saying that almost every major technology was built by or with the involvement of the government, and that AI is the first revolutionary-level technology that is predominantly being built by the private sector, and that isn't something Amodei agrees with, nor would it be what he would have chosen if he were given a choice.

Although Amodei believes the government shouldn't completely take over Anthropic, he is scared of both companies having "this technology" and the government having it. So, the middle ground is public governance tactics, which Amodei says is slowly being introduced to Anthropic. For example, Amodei says Anthropic's Long-Term Benefit Trust can appoint and remove a majority of board members and, through that authority, can even remove him as CEO. The idea behind this structure is to prevent the company from being blown ethically off course through a cabal of upper leadership.

A US-based AI startup is making bold claims against NVIDIA's most advanced hardware. Tensordyne has announced the successful tape-out of its Napier chip, a 3nm AI accelerator built on TSMC's process in collaboration with Broadcom and HPE's Juniper Networks. The company is already reporting over $200 million in projected system demand.

The Napier chip packs 138 billion transistors, 2.1 petaflops of Dense FP8 compute, 144GB of HBM3E memory, 256MB of SRAM, and runs at 300W TDP. The architectural hook is a proprietary logarithmic mathematics method that replaces numerous multiplication operations with simpler addition-based computation. Because adders are smaller and more power-efficient than multipliers, Tensordyne claims this frees up significantly more silicon area for SRAM, which it says gives Napier five times as much on-chip SRAM as NVIDIA's Blackwell.

The full rack configuration, called the TDN72, houses 288 Napier chips across four pods of 72 chips each. The complete rack delivers 608 petaflops of FP8 compute, 42TB of HBM3E memory, and operates within a 120kW power envelope, all while being fully air-cooled.

Microsoft CEO Satya Nadella has penned a lengthy X post detailing Microsoft's position on AI models and what companies of the future should value most in AI.

Interestingly, Nadella argues that Microsoft is pushing for a multi-model strategy, and that its Foundry supports different models, agent frameworks, and other Frontier Tuning. Nadella outlines how businesses should operate in an increasingly AI-driven economy and says a business shouldn't adopt a single AI model and build its entire business around it. Instead, Nadella argues that companies should build their own learning systems on top of multiple models, combining employees' expertise with AI capabilities.

Nadella describes this setup as a balancing act between "human capital" and "token capital," with human capital being knowledge, judgment, relationships, ingenuity, and pattern recognition, while token capital represents the AI capabilities a company can develop using its internal knowledge, workflows, and accumulated experience.

Anthropic, the creator of Claude, Claude Code, and the now-banned Fable 5 model, is facing a class-action lawsuit that alleges it misled subscribers about plan usage limits.

The suit specifically states that subscribers of the two highest-paid tiers, Max 5x and Max 20x, have usage limits below what the company advertises. The suit, filed by Karl Khan in the US District Court for the Northern District of California, claims Claude's actual session limits for the Max 20x plan are about "just six to eight times the usage of Pro," with the Max 5x plan being "just three-and-a-half times the usage of Pro.

Khan writes in the lawsuit that, after upgrading through Anthropic's subscription tiers,, he noticed a discrepancy between what Anthropic advertised and what he was able to achieve with his weekly data allocation. The filing goes on to argue that Anthropic has intentionally made information around its usage limits ambiguous, with the suit claiming Anthropic's website, "is a black box, without any meaningful description of how usage is calculated." It adds that Anthropic doesn't clearly define what constitutes a single session.

A new report has shed light on the escalating situation between Anthropic, the US government, and the company's latest AI model, which was only available for 3 days before being forcibly taken offline by the Trump administration.

Anthropic released Fable 5 and Claude Mythos 5 on Tuesday, which are the company's first "Mythos-class" models, seemingly classified as bleeding-edge AI. These Mythos-class models are straight-jacketed versions of the Mythos Preview that Anthropic showcased in April, and announced was too dangerous to be released to the public due to its power to identify cybersecurity vulnerabilities. Essentially, Fable 5 and Claude Mythos 5 are versions of Mythos Preview with safety rails designed to prevent the model from being used in nefarious ways, such as generating malicious code.

These models with the built-in safety rails were banned by the US government over national security concerns, as authorities stated they had learned the Mythos-class models had been jailbroken. Reports state that Amazon, along with five other companies, contacted the White House and raised potential security concerns about Anthropic's Mythos-class models; as a result, the White House issued an export ban preventing non-US nationals from using them. Since that task was beyond Anthropic's capabilities, the AI company disabled the models for everyone.

The US government has banned the use of Anthropic's latest AI models, Fable and Mythos 5, forcing the company to disable access effective immediately.

Only a matter of days after Anthropic released Fable, the company's first publicly released Mythos-level model, which is a cut-down version of the original Mythos model that Anthropic said couldn't be released to the public due to the number of cybersecurity vulnerabilities it detected, which could be used nefariously if placed in the wrong hands. Fable is a somewhat cut-down version of that original model, including safety guardrails that prevent users from using the model on topics and along specific lines of conversation.

Commerce Secretary Howard Lutnick on Friday sent a letter to Anthropic CEO Dario Amodei informing him that Mythos 5 and Fable 5 models will now be subject to export controls to any location outside of the US, and to all foreign persons within the US. Essentially, the US government has banned the use of Mythos 5 and Fable 5 models by anyone who isn't an American living in the US.

Anthropic released Claude Fable 5, the company's first publicly released Mythos-level AI model, and while it is state-of-the-art, Microsoft is reportedly running it through its legal department before authorizing its use in its products.

The news comes from The Verge's Tom Warren, who writes that he has heard from unnamed sources that Fable isn't available in the model picker for Microsoft employees using GitHub Copilot. Notably, every other Anthropic model is available in the model picker, as each complies with Microsoft's Zero Data Retention (ZDR) rules. However, Fable is different: the model's sophistication levels require data retention to operate, which violates Microsoft's internal policies.

Since Fable requires some data retention, Anthropic stores data such as prompts and the AI's responses. Anthropic deletes this data after 30 days, but some prompts and outputs can be stored by Anthropic for up to two years depending on how badly either has violated Anthropic's usage policies. So, to remedy this, Microsoft has handed over Fable to its legal department to evaluate the model, and see how it can use it safely.

Vibe coding is becoming ever-more prevalent, as you may have noticed, and now AI has crept into graphics driver maintenance in Linux.

Phoronix.com reports (via Tom's Hardware) that a fresh effort to tidy up code for the AMD R600 Linux GPU drivers witnessed a bunch of commits (from Gert Wollny) assisted by GitHub Copilot.

This fact was clearly highlighted, with Wollny stating: "This series does a lot of refactoring to make the sfn shader compiler code a bit cleaner. The refactoring was done with the help of Copilot (auto mode)."

Anthropic has officially launched Claude Fable 5, a Mythos-class AI model that now available to the public, and to celebrate the launch Anthropic has also reset current and weekly session limits.

The release marks a major milestone for the company, offering users unprecedented access to a powerful system that balances performance with safety. Anthropic explains in its blog and a series of videos that after it conducted its preview of the Mythos model and discovered thousands of cybersecurity vulnerabilities, which meant the Mythos model could be used for nefarious practices if used by the wrong hands. So, Anthropic created a new model based on the Mythos model, but implemented safeguards.

Fable 5's capabilities surpass those of previous models, excelling in complex reasoning, natural language understanding, and multi-step problem solving. According to Anthropic, the model has been made safe for general use while retaining the advanced features of its Mythos sibling. The company also writes that if users begin asking questions in the following areas Fable will detect the line of questioning and prevent any misuse of its capabilities. Those areas are: cybersecurity, biology, chemistry, and distillation.

NVIDIA has announced the 40-year PC model of opening apps, clicking, typing, and manually interacting with your PC is now over. The new era of PC interaction is here, and it's simply a combination of RTX Spark and Microsoft Windows.

NVIDIA explained at GTC Taipei that AI agents will reinvent the PC after 40 years of keyboard and mouse computing. Huang goes on to say that users traditionally open apps, click around interfaces, type commands, and manually complete tasks, but in this new era of agentic computing, "AI is the UX," and instead of working within apps users will now be able to communicate an objective to an agent and watch as that agent carries out that task autonomously.

The new hardware that will usher in this new era of PCs is RTX Spark, NVIDIA's new super-chip purposely designed for Microsoft Windows. Huang and NVIDIA have described this as "the new PC" and "the personal AI computer". NVIDIA says RTX Spark is a culmination of many years of development as it brings together CUDA, RTX, FP4, TensorRT, OptiX, Reflex, G-Sync, and many other NVIDIA-created software.

NVIDIA has unveiled the RTX Spark Superchip, which features the NVIDIA Blackwell RTX GPU with 6,144 CUDA cores and fifth-generation Tensor Cores. Furthermore, it's connected via NVLink-C2C (chip-to-chip) to a 20-core NVIDIA Grace CPU.

NVIDIA confirmed that RTX Spark was born out of a collaboration with MediaTek, a market leader in Arm-based system-on-chip designs, and that it is specifically designed for personal AI agents.

So, what can RTX Spark actually do? Well, according to Team Green RTX Spark lets creators, AI developers, and gamers render ultra-large 90GB+ 3D scenes, and edit 12K 4:2:2 video. Additionally, RTX Spark users will be able to generate 4K AI videos, run 120-billion-parameter LLMs, and, this is the big one, enjoy increased support for 1-million-token context conversations.

NVIDIA has announced that OpenShell, the company's secure, open-source runtime for autonomous AI agents, is officially coming to leading operating systems, including Microsoft Windows, making these agents accessible to billions of users.

NVIDIA announced during a press briefing ahead of its official keynote at GTC Taipei that OpenShell is coming to leading operating systems, such as Windows, Red Hat OpenShift, Canonical Ubuntu, and other major developer, desktop, and enterprise platforms. The idea behind OpenShell is to provide a safe and secure framework for users to run AI and AI agents at scale.

According to NVIDIA, these recently announced integrations with leading operating systems, such as OpenShift, will mean autonomous agent systems will reach "more than 90% of US Fortune 500 companies". During the Q&A section of the briefing, I asked if NVIDIA's OpenShell platform was a response to the widespread security concerns surrounding OpenClaw.

At Computex 2026, NVIDIA announced the new DGX Station for Windows, a desk-based supercomputer powered by the NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip. Designed for powerful AI agents and heavy-duty enterprise AI workloads, it brings NVIDIA's Grace Blackwell-class AI infrastructure to the Windows ecosystem, enabling trillion-parameter AI.

At the heart of the system, which looks like a workstation, is the NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip. This pairs a powerful 72-core NVIDIA Grace CPU with up to 748GB of memory and can connect to an NVIDIA Blackwell GPU, such as the NVIDIA RTX PRO 6000 Blackwell Workstation Edition, via the NVLink-C2C interconnect for best-in-class communication and performance.

On the network side, you've also got impressive bandwidth, with up to 800 GB/s via the ConnectX-8 SuperNIC for fast network transfers and the ability to connect multiple DGX Station systems. NVIDIA notes that it has developed DGX Station in collaboration with Microsoft to run agents at scale for engineering, design, and more, within the secure, managed open-source NVIDIA OpenShell platform.

NVIDIA CEO Jensen Huang is taking the stage at the Taipei Music Center in one hour for a major keynote focused on the future of AI, accelerated computing, GPUs, and groundbreaking technology. We are on the ground at the GTC Taipei / Computex Taipei 2026 event covering it live, bringing you the key announcements, major talking points, and important updates as they happen.

We've worked hard over the last couple of months to develop a new live blogging news system that delivers updates instantly in a modern, interactive experience. We're debuting it today right here for the NVIDIA keynote. There's no need to refresh the page - updates will stream in and appear automatically. You can even see when an editor is typing an update. We hope you enjoy it! Jensen is due to hit the stage at 11am Taipei time.

Apple's entry into the smart glasses market has hit another speed bump. According to Bloomberg's Mark Gurman, Apple's smart glasses, internally codenamed N50, have been pushed back to a late 2027 launch. The glasses were originally set to be revealed by the end of 2026, with shipments beginning in early 2027, but development delays have pushed that timeline back by roughly a year.

The delay means Apple will enter a market in which Meta has been building momentum for several years. Meta's Ray-Ban smart glasses, which the company recently expanded with two new AI models for prescription wearers, have seen sales triple in 2025 over the previous year. Apple is clearly taking its time to ensure it enters the space with a product worth buying rather than rushing a half-finished wearable to market.

Gurman reports that Apple CEO Tim Cook views the N50 glasses as his top priority before handing the reins to incoming CEO John Ternus on September 1. Ternus has been leading Apple's Vision Products Group for the past two years, making him a natural fit to carry the project forward.

ByteDance is developing its own CPUs to address growing AI infrastructure demands amid rising chip prices and supply shortages. Essentially, ByteDance, the owner of TikTok, has seen how much money hardware companies are making and now wants a slice of the AI pie, and/or it doesn't want to keep relying on third parties to provide it with the necessary hardware.

The company, best known for TikTok, is reportedly designing custom silicon to power expanding AI operations. Sources from Reuters state the chips will be used for AI inference tasks, with production potentially starting this year. The company is said to be in talks with Samsung and TSMC for manufacturing, with at least 100,000 units expected in 2026.

If accurate, the move signals a major pivot into hardware, positioning ByteDance alongside other tech firms like Google and Meta, which have also begun developing in-house AI accelerators intended to assist the demand for as many AI-capable chips as possible.

YouTube is now automatically labeling AI-generated videos, even when creators don't manually disclose their use of artificial intelligence. This marks a significant shift in how the platform handles synthetic content, particularly AI-generated content intended to fool viewers.

The new system will flag videos with "significant photorealistic AI use" and apply an "altered or synthetic content" label. Creators are still required to manually disclose AI use, but YouTube will now step in if they don't.

The label appears just below the video player for long-form content and as a small, subtle overlay on Shorts. For unrealistic or animated videos, which means content that isn't obviously attempting to fool anyone, AI disclosures are in the description of the video.