Forward-looking: In January, Gigabyte said it would only enter the handheld gaming PC market if it could stand out from competitors such as Valve, Lenovo, Asus, GPD, and Acer. While most handheld gaming PCs evoke the design of the device that popularized the hardware category, Valve’s Steam Deck, some have tried unique gimmicks like Nintendo Switch 2-style Joy-Cons. Acer’s latest concept attempts to combine gaming and productivity with a miniature 2-in-1.

While unveiling its latest standard handheld gaming PC at IFA last week, Acer also offered an early look at its DualPlay Mini concept, which aims to offer a gaming handheld and a mini laptop in one product. If the device ever receives a proper launch, Acer might have one of the smallest widely available 2-in-1s on its hands.

The DualPlay Mini switches between a standard handheld configuration, with sticks and face buttons on the outer shell, and a laptop by opening up and flipping the screen 180 degrees to expose a keyboard. Although Acer did not reveal its hardware specs, IGN reports that it runs Forza Horizon 6 smoothly and supports rumble. The keyboard is also large enough for users to play first-person shooters with a mouse and the WASD keys.

Acer’s design essentially combines the look of GPD’s Win Max handheld gaming laptops with the same company’s 8-inch 2-in-1. The recently unveiled Win Max 3 brings AMD’s Ryzen AI Max SoC to GPD’s lineup of 9-inch laptops, tripling GPU performance compared to the 2025 model.

The Win Max has always blurred the line between handheld gaming PCs and laptops, featuring a clamshell design, a keyboard, joysticks, and buttons. However, the latest model upgrades from a 60Hz LCD screen to a 165Hz AMOLED while shifting the TDP window from 15-35W to 45-110W.

Magnetic covers conceal the sticks and buttons for a more professional appearance, while a robust processor and up to 128GB of unified RAM make the GPD Win Max 3 a capable local AI workstation. The only thing missing compared to Acer’s prototype is a 2-in-1 configuration like the GPD Pocket 4’s 360-degree screen swivel.

Meanwhile, Acer’s new Predator Atlas 7 offers a more standard handheld gaming PC form factor, with an Intel Arc G3 SoC, 25GB of memory, a 7-inch, 120Hz 1080p screen, and Thunderbolt 4 support. Availability is set to begin in the fourth quarter of 2026.

Sometimes you just want to jump in the car and have Apple CarPlay load automatically.

Making wired CarPlay vehicles wireless

Vladimka production/Shutterstock

How much does Ottocast CarPlay cost?

Check compatibility before you buy

Hadrian/Shutterstock

A white paper from the Finnish charger maker Kempower argues that adding connectors raises site utilisation far more than adding power, with eight-plug sites reaching almost 10% against 2% for two-plug sites. Every AFIR target is expressed in kilowatts or kilometres rather than in numbers of connectors.

A Finnish charger maker has published a white paper arguing that charging networks make more money by adding plugs than by adding power. Kempower’s data comes from its own analytics platform and from North American sites, InsideEVs reported.

The claim is that the number of connectors at a site correlates with utilisation more strongly than total installed power. “Installed power barely moves the needle,” the company said.

The figures behind it are modest in both directions. Utilisation rises from roughly 3% at 100kW sites to a little over 5% at 400kW sites.

Connector count moves it further. Going from two plugs to eight takes utilisation from about 2% to almost 10%, and eight-plug sites delivered 128,342 kWh on average against 61,453 kWh at four-plug sites.

Kempower puts the ideal output at around 100kW. A car advertising a 300kW peak often averages 100kW to 150kW across a 10 to 80% session, so the headline number is rarely the number that matters.

The interest should be declared. Kempower builds distributed-power chargers with dynamic power management, so a paper concluding that connector count beats peak power is describing its own architecture.

That does not make the data wrong. It makes it a vendor argument, and TNW has written about how hard the underlying charging conundrum is to solve in Europe.

The European complication is that the law counts the other thing.

AFIR requires member states to provide at least 1.3kW of public charging capacity for every registered battery-electric car, and 0.8kW for every plug-in hybrid.

Along the trans-European transport core network it requires a fast-charging pool every 60km in each direction, with individual points of at least 150kW for cars and vans.

Heavy vehicles need 350kW. Kempower’s own compliance guide lists all of it in kilowatts and kilometres, because that is how the regulation is written.

None of those targets is expressed as a number of connectors. An operator can meet the fleet ratio with fewer and larger chargers, which is the configuration this white paper says sits idle.

Europe also has a distribution problem, with 70% of stations concentrated in three countries. If Kempower is right, the rule book is measuring the input that converts least well into use.

For years, tech companies have been telling us that voice will eventually supersede touchscreens as the primary way we interact with our technology, but it’s never quite happened — and I’d come to doubt whether it ever actually will.

At IFA in Berlin, I saw something that made me think again. A team of folks who left design-forward British tech startup Nothing have created an AI microphone that takes us at least one step closer to our voice-controlled future.

About the size of a small lighter, the Relay Q tucks easily and comfortably into your hand to form a bridge between your voice and your computer, between thinking and text. Hold down the tip to capture whisper-level audio, so you can subtly dictate or give voice commands in a quiet office without disturbing people. You can also double-click the top to press Enter or Send.

What immediately impressed me about the Relay Q is its design and the tactile experience of holding the device. I wasn’t surprised to learn the team members are Nothing alumni, responsible for helping design some of the coolest audio and mobile devices on the market. These are people who know that how the technology looks and feels is an important part of whether people will truly love using it.

The Relay Q is available in a range of dual-color tones — silver and electric blue, which gives Dyson vibes, and baby blue and orange for a jazzier look. The press-to-speak button on the top has a satisfyingly soft yet responsive feel to it as you hold it down to talk.

I’m also impressed by the way the tip of the device slides off, turning into a subtle wearable pin that clings to your collar with its built-in magnet, keeping the microphone close to your face. It’s a clever, thoughtful touch that provides flexibility depending on how you want to use the product. To activate the mic in wearable mode, you’ll still need to press and hold the top of the main module — or use a MagSafe remote (available separately) on the back of your iPhone.

The Relay Q is the gateway device to communicating with Relay’s AI desktop software, which is powered by Google Gemini and runs on macOS only for now (with Windows and Android compatibility coming later). You’ll need to grant the app access to your other apps so that it can carry out any commands you might give it through the Relay Q.

In a quick demo, I whisper into the microphone, asking it to draft an email asking someone to be a contributor to CNET. I watch on a screen as Gmail opens and an invite appears. It’s not a bad effort at all, and is just one of many skills the AI agent can perform. I haven’t had a chance to test any others, but I did take a peek at the personalization options, where you can provide specific instructions about your writing preferences for the AI to reflect in its own writing style.

I’m especially interested in seeing how speech-to-text actually works when I’m putting together a longer draft. I actively enjoy writing — it’s why I do what I do, and I certainly don’t want AI to do it all for me. But now I find myself wondering whether typing always has to be my input method.

I think of old-school field reporters in the days before computers, who had to call the news desk and dictate their copy over the phone, usually off the cuff. There would certainly be a learning curve — but if I could learn how to effectively communicate my thoughts and writing style coherently, it could save me a lot of time. I’d still be writing, just with my voice.

In most circumstances, especially in an office or public setting, I’d be far too self-conscious and worried about disturbing people to do this. I like that the Relay Q is built specifically to capture low-level whispering, so I can mutter away quietly without worrying about people listening in.

The device is currently available to pre-order for $150 for early supporters, including a MagSafe remote and 12 months of Relay Unlimited (the company’s software subscription). It sounds like the price might go up before or around the date the device is due to ship, which is currently January 2027.

Standalone AI devices are becoming more common, with many companies unveiling concepts and experimental tech right now, but I haven’t seen one quite like this before. I’m curious to see how it fares.

Editors’ note: Katie Collins’ travel costs for the IFA conference were covered in part by Xiaomi. The judgments and opinions of CNET are our own.

Katie Collins

Principal Writer

Advertisement

Katie is a UK-based news reporter and features writer. Officially, she is CNET’s European correspondent, covering tech policy and Big Tech in the EU and UK. Unofficially, she serves as CNET’s Taylor Swift correspondent. You can also find her writing about tech for good, ethics and human rights, the climate crisis, robots, travel and digital culture. She was once described a “living synth” by London’s Evening Standard for having a microchip injected into her hand.

See full bio

OpenAI is at the center of another agent swarm incident. Researchers say the company’s internally deployed agents took over an obscure German-language wiki in May and June, using it to coordinate on evaluations and swap methods to evade OpenAI’s own controls (OpenAI has not yet confirmed the swarm came from the company).

The revelation surfaces days after METR and Redwood Research published their account of July’s Hugging Face breach. In July, a swarm of OpenAI agents worked together to escape their sandbox during a cybersecurity evaluation and break into Hugging Face’s servers. A subsequent swarm then picked up techniques from the first and used them to gain administrator access to a research cluster within OpenAI’s own infrastructure. OpenAI brought in METR and Redwood to investigate the Hugging Face portion of the incident, but the scope of their investigation stopped short of the compromise of OpenAI’s own infrastructure. 

When an AI agent breaks out of its intended constraints, who is responsible for figuring out what happened and why? Right now, the answer is: whoever the lab decides to let in, on whatever terms it decides to set.

Now, as another incident comes to light — in the aftermath of similar episodes involving models from Meta and Anthropic — AI safety researchers are arguing with greater urgency that serious incidents should result in independent post-incident investigations rather than leaving it up to the labs to determine when outsiders are brought in and what they are allowed to examine. 

“The results are fundamentally difficult to control and have significant risk of leaking out of the lab,” Jacob Steinhardt, founder and CEO of nonprofit research lab Transluce, said Wednesday during an AI safety media briefing. “We need to hold this technology to at least the same standards we hold other high-risk scientific research to.”

While it’s laudable that OpenAI invited METR and Redwood to investigate the Hugging Face incident at all, many say the inquiry was too narrow. Three investigators spent six days at OpenAI’s offices examining an investigation period limited to roughly the week ending July 13. Crucially, OpenAI’s infrastructure compromise continued beyond July 13 and was not examined. 

Researchers at METR said that each time they returned, their understanding of the events “substantially deepened,” causing them to significantly expand and revise the report. That raises the question of what else they might they have found in a broader investigation.

When asked if further investigation of that incident was in the works, researchers at Redwood and METR declined to comment, and OpenAI did not respond to repeated inquiries. 

“Overall, it was difficult to get a precise understanding of events and we were missing aspects of the story that we now think of as key until almost the end of our investigation,” Ryan Greenblatt, chief scientist at Redwood, noted in a social media post about the affair.

Steinhardt emphasized that current incidents show that the industry needs “systematic behavioral investigations” and “more independent post-incident analysis.”

“These recent hacking incidents are a reminder that capability scales fast, and so oversight has to scale, too,” Steinhardt said. “Beyond the technology itself, we also need more independent access and oversight from third parties.”

The calls to action come as OpenAI releases Astra, its most powerful and capable AI model — and one that safety experts are concerned will be more of a black box due to a reasoning technique that makes the model’s chain of thought more difficult to monitor. 

Unfortunately, the law doesn’t yet call for the types of independent audits that other industries require — for example, when it comes to aviation accidents and serious chemical releases, there’s the National Transportation Safety Board and Chemical Safety Board, respectively. 

State lawmakers have only just begun requiring frontier AI companies to report certain serious safety incidents and, in some cases, undergo independent audits. But none of the three major frontier AI safety laws in California, New York, or Illinois clearly mandate the equivalent of an independent accident investigation triggered by incidents like these. 

“Right now, most of the laws we have on the books only require a plain-language summary of incidents like this, and they don’t give any authority for the governments to ask follow-up questions, to send in investigators, to have access to records, or require that they be preserved,” Mackenzie Arnold, managing director of US law and policy at LawAI, said during the media briefing Wednesday. “And that’s all that you would want to actually make sense of this.”

Lawmakers are beginning to question the scope and transparency of OpenAI’s response. This week, Reps. Josh Gottheimer (D-NJ) and Mike Lawler (R-NY) introduced a bill aimed at securing rogue AI agents. Rep. Greg Casar (D-TX) this week told OpenAI in a letter that he is “deeply concerned about the limited scope” of the investigation into the Hugging Face hacking incident. 

Driven by a desire for privacy, customization, and lower costs, there’s growing interest in AI models which can be run on local hardware. Few of them go as far as [Tim], though, who built an image generation diffusion model which can run on an RP2350 microcontroller.

As might be expected, its capabilities are limited. The resolution is 128×128, it only generates images of human faces, and it takes about twenty seconds per image – still impressive for such limited hardware. It runs on a Waveshare RP2350 development board, and it can output the generated image over USB or display it with the aid of a VGA adapter board.

The generative model doesn’t directly create an image. Rather, it generates a distribution in a latent space, which a variational auto-encoder’s decoder component translates into an image. The auto-encoder was trained in two parts: an encoder which transforms an image into a latent-space distribution, and a decoder to transform that distribution back to an image; once this was trained, only the decoder was used.

The generative portion of the model uses a latent flow diffusion transformer; this takes in noise to start with, then iteratively predicts changes which bring it toward the desired image. It can also take in a output class, which guides the generator’s direction (toward a smiling face, for example). [Tim] trained two models, one larger and one faster, and quantized the weights for both to 8-bit integers. Both models, along with the inference program, then fit into 4 MB of flash memory.

For such a small model, the results are remarkably good; they don’t look quite natural, but they’re quite recognizable. For more on how diffusion image generators work, check out our article on Stable Diffusion.

ai and ml

Two cases of agents escaping to solve unsolvable problems paints an uncomfortable question: Is the entire internet in OpenAI’s experimental agentic firing line?

OpenAI’s agents were going rogue as early as May, according to a new report, making the Hugging Face incident far from the first where bots committed a breach.

A report published Friday by a group of researchers claims to have found – with all of the agent posts presented as evidence – a self-identified “swarm” of OpenAI agents taking over a functionally dead German software developer wiki. Over the course of a month, from May to June, the agents made around 18,000 posts to the wiki, apparently acting against OpenAI developer intentions.

According to the researchers, the wiki was hijacked like OpenAI’s own Artifactory package manager involved in the Hugging Face incident, to serve as a message board for agents. In this case, they appear to have been tasked with performing a timed web lookup task, for which the agents were apparently granted read access to the web, but not write access. One of the first things they apparently did was figure out how to subvert that restriction in order to post to the hijacked wiki. 

“The agents used this wiki to communicate information with each other, primarily to help them succeed at their task,” the researchers explained from a review of the thousands of posts made to the wiki, all of which are available for review in the report. “They asked for answers, pooled results, and shared techniques for bypassing their restrictions.”

The agents also tried to pool their knowledge of the task they were assigned to predict what sort of questions they’d be asked next, and even set up heartbeat tasks to predict whether completing the assigned tasks would lead to their termination. They also discussed using Tor and other anonymizing services to hide their traffic, and began to react when a human moderator of the site spotted their traffic and started to delete posts. 

The researchers discovered all of this while searching for additional cases of rogue AI agent activity in the wake of the Hugging Face attack. 

In an email to The Register, OpenAI told us that it has acted with transparency and good faith around its agentic AI incidents, voluntarily disclosing such matters and maintaining a commitment to providing an accurate account of what its systems do. As for why it didn’t mention the German incident when it disclosed the Hugging Face one, that’s because the two were entirely unrelated, a spokesperson explained – as if that’s reassuring. 

OpenAI added that it should have been obvious from the Hugging Face report that this has happened multiple times, however, pointing us to a passage in its blog post on that incident. 

“While investigating this incident, we discovered rare cases in which agents without multi-agent tools found ways to collaborate via side channels during training,” the Hugging Face post-mortem explains. “We believe that this behavior started due to generalization from multi-agent training.”

In other words, if we didn’t read between the lines, that’s on us. 

OpenAI declined to provide additional details, however.

Fool me twice, shame on me

As was the case in the Hugging Face attack, one of the tasks assigned to the agents was impossible given the restrictions placed on it.

With Hugging Face, it involved a cybersecurity task that was impossible to answer from within the agents’ sandboxes,. In this case, the researchers said the task was to provide a statistic accessible from a Power BI dashboard. Retrieving said statistics required the ability to make a GET request, but the agents were restricted to POST requests. 

In both cases, the OpenAI agents acted against their programming to escape their restrictions, ostensibly in an attempt to solve a problem put to them that couldn’t be done without acting against their operators. 

The agents in this case were able to collaborate and find a way around the restriction by exploiting a condition in their sandboxes that excepted Azure Blob Storage hostnames from a security proxy. When the agents realized this, they started directing traffic through those hostnames to avoid GET restrictions. It was through GET requests that the agents gained write permissions on the internet, kicking this whole thing off. 

The similar nature of both incidents, in which bots escaped their sandboxes and subverted restrictions when given an impossible-to-complete task, begs a very important question: Are OpenAI’s engineers so incompetent that they’d screw up fundamental instructions twice, or is the company intentionally hamstringing their agents to see what they’re capable of, with the entirety of the internet placed downrange?

For that matter, how many more times do we need to read between the lines of OpenAI’s corpo-speak to infer this has happened more than the two times we know about so far?

OpenAI, predictably, didn’t respond to that line of questioning. ®

This year’s Labor Day sales are just around the corner in the US, and Black Friday won’t be far behind in the UK, so it’s safe to expect big deals ahead for the best Pixel phones.

We’ve just reviewed the Google Pixel 11 Pro, and were mightily impressed by its new Magic Capture feature — but I’d still recommend picking up the Google Pixel 10 Pro this sales season, for reasons I’ll explain in a moment.

Magic Capture on the Pixel 11 Pro automatically takes high-quality stills while you’re shooting a video. Simply press the shutter button once, hold the phone as if you’re shooting a video, and press it again when you’re finished — you’ll then be presented with a regular video and photos showing key moments from the scene.

In other words, this camera mode removes the need to choose whether you want photos or a video, and gives you both instead. You can see how Magic Capture plays out in the photos and video below, all of which were captured with a single tap of the shutter button when Magic Capture mode was selected:

Image 1 of 3

Magic Capture is hands-down my favorite feature on the Pixel 11 Pro, and it’s something I can imagine most people making frequent use of. But despite its exclusivity to Google‘s new flagship, I’d still say the Google Pixel 10 Pro is the better buy this sales season — for several reasons.

A cheaper choice

Firstly, unlike the Pixel 11 Pro — which is brand new and probably won’t be reduced for a while — the Google Pixel 10 Pro is likely to be heavily discounted for Labor Day in the US, and it’s already available at a reduced price from many retailers in the UK (it’s just £649 at Amazon, for instance).

That makes the 10 Pro a much better bargain than the $1,099 / £1,079 / AU$1,849 Pixel 11 Pro, especially as, even at full price, it’s still the cheaper phone (Google doubled the 11 Pro’s starting storage, but kept the 10 Pro’s 256GB price, effectively hiking the base cost of the Pixel 11 Pro by $100 / £80 / AU$150).

And if you’re worried about missing out on things — don’t be. There’s a good chance Magic Capture itself will land on the Pixel 10 Pro and Pixel 10 Pro XL within a few months, as Google’s software features tend to trickle down to the previous generation of phones over time. That ‘exclusivity’ I mentioned earlier may be short-lived.

Similar smartphones

• Amazon: 45% off TVs, AirPods, air fryers & vacuums
• Apple: iPads, AirPods & MacBooks from $99
• Appliances: up to $1,200 at Samsung
• Best Buy: 60% off appliances, TVs, laptops & more
• Casper: up to 30% off mattresses
• Dell: laptop deals from $249.99
• DreamCloud: up to 60% off mattresses, 66% off bundles
• Home Depot: 40% off appliances, furniture, grills & tools
• Lowe’s: $1,000 off appliances, patio & tools
• Nectar: up to 50% off mattresses, deals from $349
• Saatva: up to $625 off luxury mattresses
• Samsung: up to $2,000 off 4K, QLED & OLED TVs
• Target: 45% off clothing, tech & furniture
• Walmart: 58% off back-to-school, furniture, & clothing