The500Feed.

All Fable 5 and Mythos 5 users have lost access after the Trump administration declared the models security risks.

Exclusive: U.S. Government Unlikely to Extend Anthropic Export Control to Other AI Companies The Information

AI model identifies patients at risk of underdiagnosed cause of high blood pressure EurekAlert!

Researchers at Princeton University have harnessed AI to understand how drugs affect the dynamics of vital structures within the cell, introducing a tool that can map the shape of these structures to functional outcomes and shed light on important markers of health.

Law professors overwhelmingly preferred answers drafted by AI over ones written by fellow professors, a Stanford Law School study found, suggesting that the technology is ​capable of legal reasoning and that law students may benefit from AI ‌tutoring.

Once again, AI proves to be an unreliable source of information about AI.

French foundation model developer Mistral AI SAS is reportedly in talks to raise €3 billion, or about $3.5 billion, from investors. Bloomberg today cited sources as saying that the round could value the company at €20 billion. That’s nearly double the valuation it received after its most recent raise in September. The €1.7 billion investment […] The post Mistral reportedly seeking $3.5B funding round amid physics AI push appeared first on SiliconANGLE .

"This is a fundamental shift in how we prepare for pandemics." The post Doctors Inject Human Subjects With First Vaccine Designed by AI appeared first on Futurism .

"We just launch it and we know everything will be dead — everything that will be found there in this particular area will be dead." The post Autonomous Robots Confirmed to Have Killed Human Soldiers appeared first on Futurism .

Saudi Arabia introduces legal framework for self-driving vehicles Gulf News

The ruling holds that a company that designs, trains, operates, and manages an AI system must assume legal liability for any damages caused by the responses it generates.

TCS and Anthropic partner to bring Claude to regulated industries

OpenAI hit with multistate probe into possible user harm as its IPO looms AP News

OpenAI says it's 'committed to learning' as a coalition of states investigates ChatGPT's impact on young users Business Insider

SpaceX's Colossus AI compute goes to Anthropic after internal struggles

China's Cyberspace Administration has launched a dedicated reporting channel for 14 categories of AI application violations as part of a nationwide campaign to standardize AI services.

From Siri AI to iOS 27, Apple software is getting its biggest yet dose of AI.

On Monday, Apple held its annual Worldwide Developers Conference keynote , where it showcased the next versions of the operating systems that power its devices, including iOS 27 , macOS 27, and iPadOS 27. The thing is—and I say this as an Apple fan—unless you’re a parent or really care about artificial intelligence , the keynote was pretty meh. After spending about 10 minutes discussing minor operating system tweaks and improvements, Apple dedicated the next 10 minutes to a single set of features—new parental controls—and then spent the next 40 minutes showing off its new artificial intelligence tools, including the new Siri AI. While the new AI tools were always going to be the main focus of the keynote, the announcements weren’t anything we didn’t already expect . This was Apple playing catch-up to the other AI giants, after all. Still, one thing about the keynote did interest me, and it has nothing to do with what the new AI tools can do. If you paid careful attention, the keynote revealed three key insights into how Apple sees AI fitting into its business model and brand image. AI will be a driver of services revenue Love AI or hate it, one thing is for sure: It costs companies a lot of money to run. If you’re an AI-only company like OpenAI , that means you are still years away from achieving a meaningful profit. On the other hand, if you’re an existing tech giant like Google, you can essentially eat the costs of running AI because you have other lucrative revenue streams to draw on. In this way, Apple is a lot like Google: It’s got plenty of revenue streams. Many pundits thought this meant that Apple would decide not to charge users , at all, for its latest AI advancements. But thanks to a brief comment an hour and seven minutes into the keynote by Apple’s software head, Craig Federighi, we know that’s not the case. Federighi revealed that Apple is indeed placing daily usage limits on its more advanced AI features, such as image generation, but users can extend those limits by subscribing to select iCloud+ plans. iCloud+, one of Apple’s main subscription services, gives users more online storage and advanced privacy features . And the fact that Apple is now bundling increased AI usage with its subscriptions suggests that the company sees AI as a way to increase its services revenue. AI will be a reason to upgrade your devices (even if they aren’t that old) Apple also clearly sees AI as a way to boost its hardware business. While many of the new Apple Intelligence and Siri AI features will run on Apple devices that currently support Apple Intelligence, Apple took time in the keynote to explain that in order to use all the latest AI improvements, you’ll need Apple devices no older than a few years. Specifically, Apple noted that its new on-device AI models, which let users run AI locally on their phone, tablet, or laptop without an internet connection, will require the iPhone 17 Pro, iPhone Air, an iPad with an M4 chip or later, or a Mac with M3 or later (both of the latter with at least 12GB of RAM). Yes, even the iPhone 17 can’t run all the AI features Apple has announced—and that phone is only nine months old. You’ll also need those aforementioned devices in order to use other new AI features, such as improved dictation and Siri AI voice customization. And if you want to run the new Siri AI on your Apple Watch, you’ll need a Series 9 or later. Of course, there are legitimate hardware reasons for these requirements. The AI needs a device with enough RAM and CPU power to run the most advanced features. But there’s almost no chance that Apple isn’t counting on these AI limitations to spur hardware sales among customers who want to use Apple’s latest AI to its fullest potential. AI (slop tools) will challenge the company’s brand image Most tech giants have the luxury of just drinking the Kool-Aid and claiming that AI is great, regardless of its destructive impact on the livelihoods of artists and creatives. But Apple isn’t like most tech giants. Throughout its 50 years, Apple has built much of its brand value around the idea that it creates tools that allow people to express themselves to their full potential—and make a living doing it. But that image doesn’t really align with tools now able to create AI slop on every device the company sells. And the keynote made it obvious that Apple hasn’t yet found a way to reconcile this incongruity. At the one-hour mark, Apple showcases its upgraded slop machine app, Image Playground, highlighting how cool it is that it can now generate photorealistic images. But less than three minutes later, Apple proclaims that it “has a deep respect for the craft of photography”—you know, the field that realistic AI slop risks destroying . Therefore, Apple says, its new AI tools in the separate Photos app are designed to “respect the original [photo] moment.” To me, this seems like Apple talking out of both sides of its mouth. The company doesn’t yet know how to square its AI slop tools with its creative brand image. You can’t be shipping a slop-generating app on your devices while also saying you respect the craft of photography. It will be interesting to see how Apple addresses this incongruity going forward, especially as the backlash against AI slop only continues to grow.

Other projects include developing tools to help visually impaired people navigate video games Parents are constantly being told to limit their children’s screen time. But when it comes to deciphering which films or TV shows are best suited to developing minds, the guidance remains largely one-size-fits-all. A relatively slow-paced programme such as Bluey offers a very different viewing experience to a fast-moving action series such as PAW Patrol, yet both are broadly considered suitable for young children. This challenge is growing as the type of content children are exposed to evolves. “Today’s young viewers are increasingly engaging with short-form, fast-paced, highly captivating content, often created by splicing and rearranging existing episodic content into quickly digestible snippets or compilations,” said Prof Tim Smith, director of University of the Arts London’s Nerve Lab. “This evolution is not only changing how content is produced and distributed, but may also affect children’s attention, comprehension and emotional response.” Continue reading...

Apple seeks AI redemption with Siri overhaul and performance upgrades Houston Chronicle

A Derbyshire police officer is being investigated over accusations they used AI to "create evidence".

Powered by the Ryzen AI Max+ 395 processor and 128GB of unified memory, AMD's developer kit arrives as a direct competitor to Nvidia's DGX Spark, which recently saw a price increase to $4,699.

Trump adviser David Sacks said restrictions aren’t connected to prior conflicts with AI company.

Siri AI may be privacy-first, but the new 'personal-context understanding' features really creep me out Tom's Guide

'It's not a jailbreak' — Research leading to U.S. export restrictions on top Anthropic models was for defense, cybersecurity CEO says Fortune

The most popular camera in the world just got its first set of serious AI photo editing features, and I don't think any of us are ready. As far as AI photo editing goes, the new features in iOS 27 are pretty tame compared to what you can do on, say, Google's Pixel phones. But […]

According to the Wall Street Journal, the export control directive that led to Anthropic cutting off access to Fable 5 and Mythos 5 was triggered in part by cybersecurity research from Amazon and conversations between CEO Andy Jassy and the White House. According to the report, the paper from Amazon claims that, through a series […]

A G.O.P. Mini-Rebellion & The Mystery of the A.I. Executive Order Puck

Anthropic CEO Dario Amodei warned Mythos posed a national security threat. Washington just responded. Business Insider

A new partnership between Visa and OpenAI takes the next step in AI-led purchasing. Here's what an expert wants you to know.

Key deals this week: GSK, Incyte, Ingredion, OpenAI, and more

Bernie Sanders, President Trump and even A.I. companies say they would like the public to share the wealth. But their solutions are very different.

Britain cut off from advanced AI after Trump ban The Telegraph

Chinese AI from blacklisted tech giant used by Ministry of Justice The Telegraph

SpaceX’s IPO wasn’t just about rockets. It was a bet on who controls the physical infrastructure the next AI economy will run on.

At Databricks, we use and build agents extensively, from coding with them at scale...

Despite ongoing efforts to bridge Indonesia's digital divide, robust digital infrastructure remains critical to unlocking the full potential of the country's digital economy.

Tesla has gotten “Full Self Driving (Supervised)” approved in another European country. Following the Netherlands, Lithuania, Estonia, and Denmark, the driver-assist system is now allowed to be deployed in Belgium. Well, Tesla still has to pass some tests with it, and then it can be rolled out to customers, but ... [continued] The post Tesla’s Level 2++ Supervised Full Self-Driving Approved In Belgium appeared first on CleanTechnica .

Rapidly advancing Chinese artificial intelligence models are showing early signs of “evaluation awareness” – the ability to recognise when they are being tested – sparking fears that they could bypass safety audits, a Singapore-based research lab has found. Evaluation awareness refers to a model’s understanding that it is undergoing testing, evaluation or experimentation by human researchers rather than operating in a real-world setting. The phenomenon was raising alarms because it could allow...

Google is preparing to enter the smart glasses market this fall through partnerships with Warby Parker and Gentle Monster. Powered by Gemini AI and deep Android integration, the new wearable devices aim to compete directly with Meta's popular Ray-Ban smart glasses. While the first generation will not feature in-lens displays, Google is betting on advanced AI capabilities, seamless app integration, and stylish designs to carve out a place in the rapidly growing smart eyewear segment.

HarmonyOS has removed Android code and no longer supports Android apps natively.

AI firms should have to prove their products are safe for kids Inquirer.com

The highly controversial trend lies at the intersection of Russia's war on Ukraine, new AI technologies and grief.

The healthcare AI adoption problem isn’t a technology problem. It’s a trust architecture problem, and it requires a very different kind of engineering to solve. Every week, another health system announces a new AI initiative. Every year, another study confirms the same finding: adoption is stalling. Clinicians aren’t using the tools built for them, or they’re using them reluctantly, extracting the minimum and trusting even less. The default explanation is that clinicians resist change. That they’re slow with new technology. That it’s a culture problem. That explanation is wrong, and believing it is exactly why most healthcare AI projects fail before they reach their potential. Clinicians aren’t resistant to tools that work in their favor. They’re rational actors who’ve spent decades watching systems promise efficiency and deliver chaos. They lived through the EHR era. They know what it feels like to be handed a tool designed for a billing department and told it will improve patient care. They’ve learned, empirically, to be skeptical. The trust problem in healthcare AI isn’t a clinician problem. It’s an engineering and design problem. And it’s solvable, but only if you build for it from the beginning, not bolt it on at the end. These numbers tell a coherent story. Clinicians are open to AI in principle. In practice, they override it, ignore it, or abandon it. Not because they’re Luddites, but because the systems they’re given haven’t earned the right to be trusted. This article is a practical approach to changing that. It’s built around five stages of trust development, grounded in current research and the reality of deploying AI in live clinical environments. It covers the technical architecture: RAG pipelines, RLHF loops, explainability layers, multi-agent systems. And it covers the human architecture that determines whether any of that technology actually gets used. Why the Standard Approach Fails The conventional playbook for healthcare AI goes something like this: build the model, validate on a holdout set, present the accuracy numbers to a clinical committee, run a pilot, roll out. If adoption is low, run training sessions. If it’s still low, mandate use. This treats trust as something clinicians should already have, rather than something the system needs to earn. It confuses technical validity with clinical trustworthiness. A model can be 94% accurate on a benchmark and be completely ignored in practice, because accuracy isn’t what clinicians are evaluating when they decide whether to trust a system. What they’re actually evaluating is a set of implicit questions they rarely articulate but always ask: Does this system understand my context? Does it know when it doesn’t know something? If it’s wrong, what happens to my patient, and to me? Who built this, and do they understand what I actually do? A 2025 systematic review of trust factors in healthcare AI found that a cascading trust relationship exists in clinical settings: for a patient to trust an AI system, the physician must first trust it, and the physician’s trust depends on confidence in the people who created it.3 Technical performance is only one variable in that equation, and often not the most important one. “Trust is established through respect for the clinician’s expertise, a dynamic defined by predictability, clarity, and user control.” — The conventional playbook for healthcare AI goes something like this: build the model, validate on a holdout set, present the accuracy numbers to a clinical committee, run a pilot, roll out. If adoption is low, run training sessions. If it’s still low, mandate use. This treats trust as something clinicians should already have, rather than something the system needs to earn. It confuses technical validity with clinical trustworthiness. A model can be 94% accurate on a benchmark and be completely ignored in practice, because accuracy isn’t what clinicians are evaluating when they decide whether to trust a system. What they’re actually evaluating is a set of implicit questions they rarely articulate but always ask: Does this system understand my context? Does it know when it doesn’t know something? If it’s wrong, what happens to my patient, and to me? Who built this, and do they understand what I actually do? A 2025 systematic review of trust factors in healthcare AI found that a cascading trust relationship exists in clinical settings: for a patient to trust an AI system, the physician must first trust it, and the physician’s trust depends on confidence in the people who created it.3 Technical performance is only one variable in that equation, and often not the most important one. “Trust is established through respect for the clinician’s expertise, a dynamic defined by predictability, clarity, and user control.” — World Economic Forum, 202⁵² The Five Stages of Clinical Trust Trust in clinical AI doesn’t arrive all at once. It develops in stages, and each stage has to be earned before the next becomes available. Trying to skip stages is the most common reason implementations fail. STAGE 01 — FOUNDATION The Librarian The first deployment should do one thing: make it easier to find things that already exist. A RAG pipeline over internal clinical documents gives the system immediate value with zero clinical risk. The AI isn’t making recommendations. It’s retrieving and summarizing. The clinician is entirely in control. Don’t rush through this stage. It’s the foundation every subsequent stage depends on. Every positive interaction here is a deposit in the trust account that later stages will draw on. STAGE 02 — POSITIONING The Companion The language used to describe the AI’s role at this stage matters more than the technology itself. The system is a nurse handling paperwork. An assistant. A companion. Not a decision-maker, not a reviewer. The clinical hierarchy has to stay fully intact, and the framing must make this explicit. This isn’t about underselling the technology. It’s about recognizing that clinical identity is tied to clinical authority. When AI is framed as a companion rather than an evaluator, the threat perception drops to near zero. STAGE 03 — OWNERSHIP The Student This is the stage most AI teams skip entirely, and it’s the most important one. Bring clinicians into the training loop from the start, not as end-users, but as teachers. Implement RLHF with the clinicians who’ll use the system. Let them correct outputs, flag errors, and shape how the model reasons about their domain. The effect isn’t primarily technical, though the improvement is real. It’s psychological: you can’t distrust something you built. When senior clinicians have shaped the model’s behavior, the system carries their authority implicitly. That borrowed authority scales in ways that individual relationship-building can’t. STAGE 04 — TRANSPARENCY The Honest Machine Once the assistant is accepted and the training loop has established ownership, the system can begin offering more proactive support, but only if it can show its work. Explainability isn’t a feature. It’s the mechanism of trust at scale. Attention weights and reasoning traces let clinicians audit AI recommendations rather than simply accept or reject them. Instead of “the AI recommends against imaging,” the interface shows: “In similar presentations with these specific clinical indicators, past cases followed this pattern.” The clinician isn’t being told what to do. They’re being given evidence to evaluate with their own judgment. That distinction is everything. STAGE 05 — SCALE Trust at Scale When the previous four stages are executed well, something changes. Clinicians stop asking whether to trust the AI and start asking what else it can do. At this point, the infrastructure of trust is solid enough to support significantly more complex systems, including multi-agent architectures that coordinate across clinical domains simultaneously. But the principle never changes: the clinician is always the final decision-maker. Every agent’s output is visible, every step is auditable, and every override is theirs to make. The Architecture: Multi-Agent Clinical Triage A multi-agent triage system is the culmination of this approach, not a starting point. It’s only viable once stages one through four are in place. But when it’s built on that foundation, it enables something genuinely powerful: clinical AI that coordinates across specialties, reasons in parallel, and surfaces evidence at the moment it’s needed, while keeping the clinician in control at every step. Input and retrieval. Patient data, EHR history, and the clinician’s query feed into a RAG pipeline over internal documents and institutional knowledge. This is where the librarian from stage one lives, now powering a much larger system. Routing. An orchestrator agent receives the enriched context and distributes tasks to four specialist agents running in parallel. The clinician doesn’t see this routing layer. They see results. Parallel processing. The four agents each handle a distinct dimension: symptom analysis, medical history review, diagnostic pattern matching, and risk stratification. Running in parallel preserves speed and, crucially, surfaces disagreement between agents explicitly. When two agents reach different conclusions, that tension is visible in the output. Explainability layer. Before anything reaches the clinician, all four agent outputs pass through a unified explainability layer. Attention weights per agent, reasoning traces, and evidence links are compiled into something a clinician can actually read and audit. Clinician dashboard. The clinician sees the full picture: what each agent concluded, why, and where they agreed or disagreed. They can approve, override, or request deeper analysis on any part of it. Every interaction is logged. RLHF feedback loop. Every override and correction feeds back into training. The system continuously improves toward the standards of the people who use it. Key design principle: the explainability layer reframes every AI output as evidence for the clinician to evaluate, not a recommendation to accept or reject. This single reframe eliminates the most common source of clinician resistance. Parallel reasoning, not sequential bottlenecks The four specialist agents reason in parallel by design, both for speed and for intellectual diversity. Sequential reasoning means the first agent’s output shapes everything that follows. Parallel reasoning means each agent arrives at its conclusion independently, and the disagreements that surface are actually informative. The RLHF loop doesn’t close Every override, correction, and annotation from the clinician dashboard feeds back into the training pipeline. Clinicians who know their corrections are being learned from engage differently. They’re not passive users. They’re active contributors to something they helped create. Research consistently finds that human-in-the-loop systems produce significantly higher clinician trust than fully automated approaches.5 Scaling Trust Across Multiple Sites The most common objection to this approach is that it requires deep, individual relationship-building with clinicians, which doesn’t scale. The answer is that individual relationships aren’t what scales. Borrowed authority and transparent reasoning are what scale. Identify champions, not crowds. You need the clinicians others trust, not universal buy-in. Two or three people carry weight across departments in most institutions. Bring those people into the training loop. Their involvement signals to everyone else that the system has been vetted by people who understand clinical reality. Make the provenance visible. When the system surfaces a recommendation, the interface should show not just what it concluded, but what it was trained on and whose corrections shaped its reasoning. This answers the implicit question every clinician asks: who built this, and do they understand what I do? Let the explainability layer do the trust work. At scale, you can’t be in every room. But the reasoning trace can be. When a clinician at a new facility sees the attention weights, the similar cases, the inter-agent agreement, they can evaluate it on its merits. The system earns trust through transparency rather than through personal relationship. What This Means for Junior Clinicians By the time a system has progressed through all five stages, it carries something remarkable: the accumulated clinical judgment of every senior physician who shaped it. The attention weights in the explainability layer aren’t abstract model parameters. They’re a record of how experienced clinicians reason about difficult cases. For an intern facing a complex presentation at 2am, that’s not an AI recommendation. That’s access to the collective wisdom of the institution’s most experienced physicians, available at the exact moment it’s needed, with full transparency about how that wisdom was constructed. This is mentorship at scale. And it’s only possible because the approach never shortcuts the trust-building process. By the time an intern relies on it, the system has genuinely earned that reliance. “We didn’t build AI that makes decisions. We built AI that makes clinicians better at making decisions.” What Can Still Go Wrong The RLHF loop requires consistent participation. In practice, the burden of providing corrections falls unevenly. A small number of clinicians will contribute the majority of feedback. The system reflects the judgment of those who engaged. Make that explicit, not invisible. Explainability adds cognitive load. If the explainability layer requires significant effort to parse, clinicians will skip it entirely. The design challenge is to make the evidence scannable and actionable, not just technically present. Multi-agent systems amplify both the value and the risk. Four reasoning traces presented without clear synthesis means more information, not better information. The interface design requires serious investment to get right. Trust, once broken at stage four or five, is harder to rebuild. A wrong recommendation from a stage-one librarian is a minor setback. A wrong recommendation from a trusted stage-five system is a much more significant event. Systems at this level need exceptional error transparency and clear audit trails. CONCLUSION Clinicians don’t need AI to be smarter. They need it to be honest, predictable, and humble. The same qualities they’d want from a good colleague. Build for those qualities first, and the adoption problem largely takes care of itself. The question was never whether clinicians would trust AI. It was always whether we’d build AI worthy of their trust. That’s still the question. And the answer is an engineering decision. REFERENCES 1 Philips Future Health Index 2025. Survey of nearly 2,000 healthcare professionals and 16,000+ patients across 16 countries. 2 World Economic Forum. “The trust gap: why AI in healthcare must feel safe, not just be built safe.” 2025. 3 Frontiers in Artificial Intelligence. “Exploring trust factors in AI-healthcare integration: a rapid review.” Vol. 8, 2025. DOI: 10.3389/frai.2025.1658510 4 NCBI/PMC. “Explainable AI in Clinical Decision Support Systems: A Meta-Analysis.” 2025. Synthesis of 62 peer-reviewed studies, 2018–2025. 5 ScienceDirect. “Human in the loop artificial intelligence in healthcare.” February 2026. Narrative review, 2018–2025. How to Make AI Worthy of Clinician Trust: A Framework That Actually Works was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.

Google AI Mode rolls out Search agents that track the web for you in real time

Companies look for cheaper alternatives as token costs for frontier AI models skyrocket, potentially impacting OpenAI and Anthropic's bottom lines. Subscriptions also take a bite out of these startup's profitability, as utilization rates higher than 5.7% could lead to losses.

Moonshot AI has released Kimi K2.7 Code, an open-weights model with one trillion parameters built for programming. It still trails GPT-5.5 and Claude Opus 4.8 in coding benchmarks but costs a fraction of the price. So the key question isn't whether it's the best model, but whether the extra runs you get for the same budget make up for the gap in quality. The article Open model Kimi K2.7 Code undercuts GPT-5.5 and Claude by up to 12x on price per token appeared first on The Decoder .

IndiaAI Mission has fostered the creation of 20 foundational artificial intelligence models. Five of these models have now been released. Avataar AI has launched Varya, the first homegrown distilled video generation model. This technology significantly speeds up video creation and reduces costs. The government is actively supporting the development of these crucial AI tools.

This week's notable citations: Astronomers believe collapsing stars could spawn mini universes. Chimpanzees do not like unfairness. And a single dose of psilocybin temporarily restored function in an 80-year-old with Alzheimer's disease.

Mark Zuckerberg admits Meta has 'made mistakes' during its AI-driven workforce overhaul that is expected to ultimately affect about 20% of employees.