• Shrike 10 Fiber outperformed competitors, scoring more than ten points higher
• F10 drone earns Pentagon contract due to eliminating Chinese components effectively
• Fiber-optic guidance makes the Shrike drone resilient to electronic warfare

The Pentagon has recently selected two Ukrainian drone manufacturers as finalists in its $1.1 billion Drone Dominance program.

For years, Chinese manufacturer DJI has dominated the global small drone market, but the US military is now actively seeking alternatives that do not rely on components sourced from China.

“This is not a consumer device company that takes privacy very seriously,” Gamero-Garrido says. Since people use smartphones far more than Alexa or a Kindle, he says an Amazon smartphone today would “significantly increase the scale of the potential privacy harms.”

Gamero-Garrido thinks Amazon could use Transformer as a data-gathering tool to glean how people use its devices, build its advertising network, and compete with the likes of Alphabet and Meta, which are facing regulatory scrutiny in the European Union and California.

One way it could do this is through the Fire TV approach. This is Amazon’s TV streaming platform integrated into a third-party TV (or via a dongle); while you may not have bought a Fire TV-powered TV from Amazon, the data collected by the operating system is still owned by the company.

“Whether they end up succeeding with this phone supplement device, or whether they eventually use a similar model where they install their operating system on other phones or ”light” phones that are built by third parties, it has the same effect,” he says. “Ultimately, what Amazon is doing is centralizing all the network traffic through its own infrastructure so it can improve its advertising business.”

If Amazon can detect when a person is sick from the sound of their voice, then it can recommend that you buy specific cold medicine from Amazon Health—that’s a real patent Amazon owns. If this is now powered on a device you carry everywhere, Gamero-Garrido says it can listen to more of your conversations and serve you better ads.

Even with its past regressions, customers have shown a general acceptance of Amazon’s hardware, says Kassem Fawaz, an associate professor at the University of Wisconsin, Madison, who researches security and privacy in consumer devices.

“I think when it comes to products, unfortunately, consumers value utility and price over privacy,” Fawaz wrote in an email to WIRED.

The accelerant here could be Amazon’s Devices & Services lead, Panos Panay, who joined the company in 2023. Panay famously helped turn Microsoft’s Surface line of computers into an aspirational hardware brand through his “pumped” and emotionally charged keynotes.

Panay has already brought that kind of energy to a few Amazon hardware announcements, like the Kindle Scribe Colorsoft, though he has not matched the success of Surface. If Amazon is truly making a smartphone, it will need to generate a lot of passion to entice customers.

“If someone can do it, it’s going to be Panos,” Jeronimo says. “For that, I have total confidence. He is the right person for these kinds of initiatives.”

The White House’s new policy framework for regulating generative artificial intelligence, released Friday, covers many areas, but one thing is clear: President Donald Trump wants the federal government to set the rules. And those rules appear to fall far short of what consumer and privacy advocates argue is necessary. 

The generative AI revolution has been underway for years, and US legislation is slow to catch up. This is despite the growing awareness of AI’s harms and challenges: chatbots’ dangerous impacts on mental health and child development, the widespread legal wrangling over the copyright protections, the dangerous spread of deepfakes and AI-powered scams, to name a few. 

Sen. Marsha Blackburn introduced the new policy package, called The Trump America AI Act, in Congress on Thursday. The Tennessee Republican’s bill is an attempt to codify a vision based on Trump’s 2025 AI Action Plan, while delving into more legal specifics and providing guidance on implementing new laws (or changing existing ones). 

Trump has maintained that the federal government should be responsible for regulating the AI industry — and that requiring AI companies to comply with 50 different sets of state laws would prevent the US from “winning” the global AI race. However, a proposal to temporarily ban states from regulating AI failed back in July, when it was removed at the last minute from the massive budget bill, known as the “One Big Beautiful Bill Act.” 

Now, the White House is doubling down on its claim to be in charge, with a few exceptions. The plan addresses some of the biggest concerns people have about AI: job loss, copyright chaos for creators, rapidly expanding infrastructure such as data centers and the protection of vulnerable groups like children. But critics say it doesn’t go far enough to regulate the fast-growing AI industry. 

“It is light on protection and heavy on promotion of dangerous AI systems,” Alan Butler, president and executive director of the Electronic Privacy Information Center, said in a statement. “The American people deserve better, and Congress should do better than this.”

The White House’s new proposed AI laws

The White House’s 2026 AI proposal says Congress should not create a new governing body to oversee AI rules, but should let existing agencies and subject-matter experts regulate as they see fit. 

Protecting children: This is one area where the federal government won’t prevent states from creating laws. And many state governments are already leading the charge, especially in regulating romantic or companion chatbots. 

The plan highlights protecting kids from AI-powered deepfakes, a huge issue highlighted in AI creating child sexual abuse material. Shielding young people from the ill effects of AI is an ongoing battle, with several high-profile cases of teenagers using AI for self-harm and suicide.

Blackburn’s policy plan includes general language related to kids’ online safety. Existing bills like the Kids Online Safety Act and the Children’s Online Privacy Protection Rule are, theoretically, designed to protect kids, but advocates and tech experts say they could create a chilling effect on free speech and lead to censorship. 

Though Trump’s AI framework addresses censorship, it’s limited to preventing AI companies from including ideological or partisan bias in their products. Trump has previously railed against what he calls “woke” AI, a term the president and his allies have used to attack concepts like diversity, equity and inclusion.

Job loss: It’s not just translators and data entry folks who are worried about losing their jobs to AI — legacy tech workers like coders and engineers are, too. There have been a lot of concerns about AI disrupting the workforce, with retail giants like Amazon laying off thousands of employees in the name of AI efficiency. The White House says it should use “nonregulatory” methods to focus on youth development and AI workforce training.

Infrastructure: In line with Trump’s previous AI Action Plan, the framework calls for states and local governments to streamline data center construction and operation. These facilities are increasingly controversial, with nearby residents reporting environmental damage and strain on their existing electrical grids, creating higher electric bills. 

Several big tech companies recently agreed to foot the bill for any higher electricity costs, but there’s no way to enforce the voluntary pledge.

Copyright: Whether the use of copyrighted materials in AI training is fair use or copyright infringement is one of the biggest legal issues of the AI age. The plan reiterates the administration’s position that AI companies are covered by fair use — meaning they wouldn’t have to obtain permission or pay for copyrighted content when creating their models. 

But, given the ever-growing number of lawsuits asking the judiciary the same question, the federal government should allow those cases to play out. So far, limited cases with Anthropic and Meta have carved out narrow victories for tech companies, not authors.

The framework document hints that the federal government could become a future licensing partner for AI companies, stating that it should “provide resources to make federal datasets accessible to industry and academia in AI-ready formats for use in training AI models and systems.”

(Disclosure: Ziff Davis, CNET’s parent company, in 2025 filed a lawsuit against OpenAI, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.) 

Does the White House plan do enough?

Tech industry groups praised the administration’s proposals, while consumer advocacy groups offered skepticism at best. 

In a statement backing the plan, the Consumer Technology Association supported a single set of rules for the entire country. 

“AI can and will make us better, and we agree that children need special protection, First Amendment rights are paramount, harmful deep fakes should be regulated, and Congress should not act to restrict AI platforms from relying on fair use protection,” the tech industry trade group said.

But according to Samir Jain, vice president of policy at the Center for Democracy and Technology, the government’s playbook is rife with internal contradictions. While it calls for the federal government to preempt state rules and laws on AI development, it also says the federal government shouldn’t undermine state authority. 

“The White House’s high-level AI framework contains some sound statements of principles, but its usefulness to lawmakers is limited by its internal contradictions and failure to grapple with key tensions between various approaches to important topics like kids’ online safety,” Jain said in a statement.

Ben Winters, director of AI and data privacy at the Consumer Federation of America, said the proposal prioritizes Big Tech over consumers.

“It’s encouraging to see some stated desires to protect people from AI-generated scams and data abuse of minors, but it’s not enough,” Winters said in a statement. “We need to see money where their mouth is on the protections — more money for consumer protection agencies at both the federal and state levels. So far, they’ve done nothing but cut and hamstring them.”

This is BIG, and we’re not just talking about the box it all arrives in.

Boutique label Arrow Video has announced a Complete Series set collecting all five seasons/42 episodes of streaming giant Netflix’s nostalgic horror show, Stranger Things. It will be available in four different versions–Special or Deluxe, 4K or HD Blu-ray–all available on July 28, just 13 days after the tenth anniversary of the premiere.

A phenomenal worldwide hit, Duffer brothers Matt and Ross’ Stranger Things broke viewership records across all seasons, beloved for its quirky characters, heart, humor, dark frights, and perhaps most of all its painstaking recreation of ‘80s middle America. The Bros. are indisputably old-school, stating, “We always dreamed that Stranger Things could be owned in its entirety, not just as a collector’s set, but as a way to preserve the show for decades to come.”

This is huge news for a couple of reasons. Only Seasons One and Two ever had a physical media release, in 2017 and 2018 respectively, as Target exclusives. Both fetched hundreds of dollars on eBay, even though the first season notoriously disappointed many with its lack of high dynamic range and only Dolby Digital 5.1 audio. Season Two did a bit better, with DTS-HD Master Audio and HDR10, but that was the last disc offering… until now.

Moreover, this drop suggests a continued willingness by the streamer to offer some of its most coveted properties on disc. It comes on the heels of the revelation that Criterion Collection editions of KPop Demon Hunters and Guillermo del Toro’s Frankenstein are on the way, recent Oscar winners both. Theoretically, this strategy would take a nibble out of The Big Red N’s core business, but perhaps this is a major-league endorsement that digital delivery and physical media can happily co-exist?

The 25 discs themselves appear to be identical between the Special and Deluxe editions. The production values of the show were always cinema-quality, and the 4K will be presented in Dolby Vision at its proper 2:1 aspect ratio with the original DTS-HD Master Audio 5.1 surround and two-channel stereo for all episodes plus an upgrade to Dolby Atmos for the final two seasons. The on-disc bonus content includes:

• Interviews with the cast and crew 
• Behind-the-scenes featurettes 
• Set tours 
• Bloopers

The Special Edition arrives with a booklet, whereas the Deluxe Edition exclusively packs quite a bit more, including several in-universe souvenirs:

• Palace Arcade alloy-zinc token-style coin
• Self-adhesive Hellfire Club patch
• Hellfire Club 20-sided die
• Enhanced packaging including wraparound box artwork by Juan Ramos
• Twenty-five artcards representing all five seasons
• Five double-sided posters featuring original artwork by Kyle Lambert
• Reversible sleeve inserts featuring new artwork by Juan Ramos and original artwork by Kyle Lambert
• Double-sided foldout map of fictional Hawkins, Indiana
• A 148-page perfect-bound artbook with design sketches, concept art, storyboards and new writing on the making of the series from the Duffer Brothers, Shawn Levy, Andrew Stanton, Kyle Dixon and more

Price & Availability

All editions are available for pre-order now at Amazon and Arrow Video, but won’t start shipping until July 28, 2026.

Related Reading:

Large language models are running into limits in domains that require an understanding of the physical world — from robotics to autonomous driving to manufacturing. That constraint is pushing investors toward world models, with AMI Labs raising a $1.03 billion seed round shortly after World Labs secured $1 billion.

Large language models (LLMs) excel at processing abstract knowledge through next-token prediction, but they fundamentally lack grounding in physical causality. They cannot reliably predict the physical consequences of real-world actions. 

AI researchers and thought leaders are increasingly vocal about these limitations as the industry tries to push AI out of web browsers and into physical spaces. In an interview with podcaster Dwarkesh Patel, Turing Award recipient Richard Sutton warned that LLMs just mimic what people say instead of modeling the world, which limits their capacity to learn from experience and adjust themselves to changes in the world.

This is why models based on LLMs, including vision-language models (VLMs), can show brittle behavior and break with very small changes to their inputs. 

Google DeepMind CEO Demis Hassabis echoed this sentiment in another interview, pointing out that today’s AI models suffer from “jagged intelligence.” They can solve complex math olympiads but fail at basic physics because they are missing critical capabilities regarding real-world dynamics. 

To solve this problem, researchers are shifting focus to building world models that act as internal simulators, allowing AI systems to safely test hypotheses before taking physical action. However, “world models” is an umbrella term that encompasses several distinct architectural approaches. 

That has produced three distinct architectural approaches, each with different tradeoffs.

JEPA: built for real-time

The first main approach focuses on learning latent representations instead of trying to predict the dynamics of the world at the pixel level. Endorsed by AMI Labs, this method is heavily based on the Joint Embedding Predictive Architecture (JEPA). 

JEPA models try to mimic how humans understand the world. When we observe the world, we do not memorize every single pixel or irrelevant detail in a scene. For example, if you watch a car driving down a street, you track its trajectory and speed; you do not calculate the exact reflection of light on every single leaf of the trees in the background. 

JEPA models reproduce this human cognitive shortcut. Instead of forcing the neural network to predict exactly what the next frame of a video will look like, the model learns a smaller set of abstract, or “latent,” features. It discards the irrelevant details and focuses entirely on the core rules of how elements in the scene interact. This makes the model robust against background noise and small changes that break other models.

This architecture is highly compute and memory efficient. By ignoring irrelevant details, it requires much fewer training examples and runs with significantly lower latency. These characteristics make it suitable for applications where efficiency and real-time inference are non-negotiable, such as robotics, self-driving cars, and high-stakes enterprise workflows. 

For example, AMI is partnering with healthcare company Nabla to use this architecture to simulate operational complexity and reduce cognitive load in fast-paced healthcare settings. 

Yann LeCun, a pioneer of the JEPA architecture and co-founder of AMI, explained that world models based on JEPA are designed to be “controllable in the sense that you can give them goals, and by construction, the only thing they can do is accomplish those goals” in an interview with Newsweek.

Gaussian splats: built for space

A second approach leans on generative models to build complete spatial environments from scratch. Adopted by companies like World Labs, this method takes an initial prompt (it could be an image or a textual description) and uses a generative model to create a 3D Gaussian splat. A Gaussian splat is a technique for representing 3D scenes using millions of tiny, mathematical particles that define geometry and lighting. Unlike flat video generation, these 3D representations can be imported directly into standard physics and 3D engines, such as Unreal Engine, where users and other AI agents can freely navigate and interact with them from any angle.

The primary benefit here is a drastic reduction in the time and one-time generation cost required to create complex interactive 3D environments. It addresses the exact problem outlined by World Labs founder Fei-Fei Li, who noted that LLMs are ultimately like “wordsmiths in the dark,” possessing flowery language but lacking spatial intelligence and physical experience. World Labs’ Marble model gives AI that missing spatial awareness. 

While this approach is not designed for split-second, real-time execution, it has massive potential for spatial computing, interactive entertainment, industrial design, and building static training environments for robotics. The enterprise value is evident in Autodesk’s heavy backing of World Labs to integrate these models into their industrial design applications.

End-to-end generation: built for scale

The third approach uses an end-to-end generative model to process prompts and user actions, continuously generating the scene, physical dynamics, and reactions on the fly. Rather than exporting a static 3D file to an external physics engine, the model itself acts as the engine. It ingests an initial prompt alongside a continuous stream of user actions, and it generates the subsequent frames of the environment in real-time, calculating physics, lighting, and object reactions natively. 

DeepMind’s Genie 3 and Nvidia’s Cosmos fall into this category. These models provide a highly simple interface for generating infinite interactive experiences and massive volumes of synthetic data. DeepMind demonstrated this natively with Genie 3, showcasing how the model maintains strict object permanence and consistent physics at 24 frames per second without relying on a separate memory module.

This approach translates directly into heavy-duty synthetic data factories. Nvidia Cosmos uses this architecture to scale synthetic data and physical AI reasoning, allowing autonomous vehicle and robotics developers to synthesize rare, dangerous edge-case conditions without the cost or risk of physical testing. Waymo (a fellow Alphabet subsidiary) built its world model on top of Genie 3, adapting it for training its self-driving cars.

The downside to this end-to-end generative method is the great compute cost required to continuously render physics and pixels simultaneously. Still, the investment is necessary to achieve the vision laid out by Hassabis, who argues that a deep, internal understanding of physical causality is required because current AI is missing critical capabilities to operate safely in the real world.

What comes next: hybrid architectures

LLMs will continue to serve as the reasoning and communication interface, but world models are positioning themselves as foundational infrastructure for physical and spatial data pipelines. As the underlying models mature, we are seeing the emergence of hybrid architectures that draw on the strengths of each approach. 

For example, cybersecurity startup DeepTempo recently developed LogLM, a model that integrates elements from LLMs and JEPA to detect anomalies and cyber threats from security and network logs. 

Microsoft announced on Friday a series of changes focused on improving the quality of its Windows 11 operating system, which notably includes dialing back the number of entry points to its AI assistant, Copilot.

The company said it will reduce Copilot AI integrations in some apps, starting with Photos, Widgets, Notepad, and its Snipping Tool.

Under the heading of “integrating AI where it’s most meaningful,” Pavan Davuluri, EVP of Windows and Devices, wrote on the company’s blog that Microsoft is becoming more intentional about “how and where Copilot integrates across Windows.” Its goal, he explained, is to focus on AI experiences that are “genuinely useful.”

This “less-is-more” approach to integrating AI into existing platforms may reflect the growing consumer pushback against AI bloat. While many people today understand AI to be a useful tool, there are also concerns around trust and safety. For instance, a Pew Research study published this month noted that half of U.S. adults are now more concerned than excited about AI as of June 2025, up from 37% in 2021.

This is not the first time Microsoft has rethought its Copilot integrations. Earlier this month, the news site Windows Central said the company’s plan to ship Copilot-branded AI features across Windows 11 had been quietly shelved. This, the site said, included some system-level integrations within the Settings app, File Explorer, and elsewhere.

Before this, Microsoft had delayed the launch of its AI-powered memory feature, Windows Recall for Copilot + PCs, for over a year as it tried to address users’ privacy concerns. The Recall feature launched last April, but security vulnerabilities are still being discovered.

It’s clear that user feedback is influencing Microsoft’s moves around AI on Windows. Davuluri wrote that he and his team have spent the past several months listening to the community about how they’d like to see Windows improved.

The Copilot rollback is just one of the changes being made.

The company said it’s also introducing the ability to move the taskbar to the top or sides of the screen, giving users more control over system updates, speeding up File Explorer, improving the Widgets experience, updating the Feedback Hub, and making it easier to navigate its Windows Insider Program — a community that offers feedback about Windows’ future.

The dream of fully powering everything from aircraft to cars on just the power generated from solar panels attached to the machine remains a tempting one, but always seems to require some serious engineering including putting the machine on a crash diet. The quadcopter that [Luke Maximo Bell] tried to fly off just solar power is a good case in point, as the first attempt crashed after three minutes and wrecked its solar panels. Now he’s back with a second attempt that ought to stay airborne for as long as the sun is shining.

Among the flaws with the first prototype were poor support for the very thin and fragile PV panels, requiring much better support on the carbon fiber frame of the drone. To support the very large solar array, the first drone’s arms were made to be very long, but this interfered with maneuvering, so the second version got trimmed down and the array raised above the frame. This saved 70 grams of weight from the shortened tubs, which could then be added to the new panel supports.

After an initial test flight resulted in a crash when the PV output dropped, the need for a small battery buffer was clear, so this was added, along with a reduction of the array to 4×7 panels to get the same 20V as the battery. The array also had to be reinforced, as the thin array was very wobbly in addition to making it impossible to fly with any significant wind.

During the subsequent five hours long test flight it was clear that the resulting PV-powered drone was at the limits of its performance, with even some mild cloud cover forcing the battery to provide backup power.

For the test location a tree-sheltered site far away from windy Cape Town was also selected to provide the best possible shot, as keeping position with this drone was very hard. With the low weight and the big surface area of the solar panel array catching any little bit of wind, the GPS-based position keeping was essential. Unfortunately a few hours into the test this feature failed.

Manual position keeping is definitely possible, but [Luke] had to constantly counteract the drone wanting to drift off somewhere else. Ultimately the test flight ended when it was still very much a sunny South African summer’s day, due to the current provided by the array no longer keeping up with the power demands of the motors.

What this perhaps demonstrates best is that if you want to use PV solar power for your flying drone – especially with a significant payload – it’s probably best to use it for recharging while idle, or to extend the battery life by an appreciable amount. That said, props to [Luke] for persevering and making it work in the end.