Category: Technology

  • I looked ridiculous wearing this solar powered hat that couldn’t even charge my phone

    I looked ridiculous wearing this solar powered hat that couldn’t even charge my phone

     

     

    Solar hat being used by a man with cable connected to power bank.

     

    Me wearing the hat with the cable connected to a power bank.

     

    Richard Baguely/Zooey Liao/CNET

    Like many middle-aged men, I have a large solar-powered area on the top of my head: it’s called a bald spot. If I work in the yard for a long time, it turns very red and glows at night. So I wear a hat, but all that beautiful solar energy bouncing off my dome goes to waste — or it did until I tried the $129 hat EcoFlow Power Hat (currently $99 on sale)a new sun hat/solar power source combination.

     

    It has eight solar panels that power two USB ports (one USB-A and one USB-C) in a small box with a small LED under the back edge. While it adequately protected my bald spot, it failed to charge my phone to any significant extent. Plus, it’s just an ugly hat.

    Pros

    • Covers my head, adequately doing the job of a hat
    • Technically it can charge a phone slowly, so I guess it works

    Cons

    • Very low charging rate with just 5 watts of power
    • Need a cable running across your back
    • The design will only appeal to Wicked cosplayers (it’s ugly)

    Don’t miss any of our unbiased technical content and lab-based analysis. To add CNET as Google’s preferred font in Chrome.

    Design: The least attractive hat I’ve ever worn

     

    Richard Baguely/Zooey Liao/CNET

     

    There are eight small solar panels on the edge, but unless the sun hits them correctly, you won’t get much power from them.

     

    Richard Baguely/Zooey Liao/CNET

    Let me be clear: this is a really ugly hat. Sunhats aren’t usually elegant, but this thing looks like a melted flower pot or a giant version of one of those horrible flower pot holders made out of vinyl records. I’m not a fashionista, but the Oasis reunion made bucket hats cool again and hats with wider brims like the Tilley TS1 Protect yourself from the sun without looking like you’ve been hit, Wile E. Coyote style, by a falling satellite dish. I’m a big fan of practicality over style, but there are limits, especially since it’s not very practical.

    Not easy to clean or use

    All the electronics in the hat mean you can’t wash it. When my baseball cap gets dirty, sweaty, and covered in dirt from hard work, I can throw it in the washing machine. Try that with the EcoFlow Power Hat and you’ll destroy it. You also need a cable connecting the hat to the phone, which isn’t very practical when you’re trying to pull weeds. I ended up running a cable through the back of my shirt, which just made the whole thing look even sillier.

    Charging: Can’t charge anything worth using

     

    Electric hat with garden tools.

     

    A six-hour gardening session barely moved the needle on charging my devices.

     

    Richard Baguely/Zooey Liao/CNET

    The problem is how solar energy works. Sunlight hits a solar panel, which converts it into electrical energy. This is called the photoelectric effect, first explained by Albert Einstein, who won the Nobel Prize for discovering it in 1921. The amount of energy depends on the size of the panel and the amount of light that hits it. This, in turn, depends on how the panel is positioned in relation to the sun.

    That’s why your roof’s solar panels are angled south to face the sun. The EcoFlow Power hat features eight small solar panels positioned around its large brim. This means that most do not receive much or any direct light unless the sun is directly overhead. The sun has an annoying habit of rising and falling, so most of the time you won’t get a consistent amount of direct sunlight.

    What does all this mean in practical terms? That means this sun hat doesn’t work very well. I tested the hat in full midday sun in my backyard near Boston and found that, at best, it generated about 5 watts of power; that’s not much. It generated 5 volts at the USB port, but the current flow never exceeded 1 amp, meaning it still only generated about 5 watts of power. In the morning or afternoon, when the sun was at a certain angle, the current generated dropped to less than 0.3 amps, approximately 2 watts. EcoFlow claims it can generate up to 12 watts, but I’ve never gotten anything close to that.

    100 hours to charge an iPhone 17 Pro

    To test this further, I conducted an arduous gardening session. Well, okay, I put it on a pole, connected it to a rugged portable power bank and watched from the deck to make sure my local groundhog (we call him Wilbur the Pig Whistle) did not steal it. After six hours of this hard work, the charge level increased by 9%, representing around 225 mAh of stored charge. This is about 5% of the capacity of a so this hat would take about 100 hours to charge the phone. That’s assuming the phone isn’t currently working.

     

    Power Hat charging iPhone 17 Pro

     

    The Power Hat needs 100 hours to charge an iPhone 17 Pro if the phone is turned off.

     

    Richard Baguely/Zooey Liao/CNET

    Specifications

    • Two sizes: Medium-Large (56-58 centimeters) and Large-XL (59-61 centimeters)
    • Dual USB-A and USB-C ports to charge two devices simultaneously.
    • It has eight small solar panels on the edge.
    • SPF 50+, sun and UV protection
    • Claims 24% thermal to electricity conversion and up to 12 watts of power; tested to only 5 watts
    • IP65 waterproof and dustproof rating

    Buying Advice: Skip, there are better hats and better solar charging options out there

    Many of the products I test end up being a case of a great idea but poor execution. The practicalities of manufacturing products and real-world engineering compromises often detract from the product’s intended purpose. However, for the full price of $129 or $99 on sale if you buy direct from EcoFlowEcoFlow Power Hat is one of the rare exceptions that is a bad idea with poor execution.

    It’s ugly and has solar panels that are too small to be effective for, well, anything. Maybe it would be better if they went with a flat lid designor a Sahara or hiking hat design with the solar panels on the top or neck flap. It is also an inferior solution to simply getting a fast and compact portable chargerlike the InfinityLab InstaGo 5000 or Anker 523 PowerCore Slim 10K PD. Another alternative is to use a larger portable solar panelwith which you can pair a power station or clip to your backpack while walking to charge your devices, like the Bluetti 2 speakerphone.

    All of these options will provide more power than the EcoFlow Solar Hat and will do it better and faster. Therefore, keep solar panels away from your head and stick them on your roof.

    avots

  • Ready to ditch your Windows PC? I Found a Powerful Mini PC Optimized for Linux

    Ready to ditch your Windows PC? I Found a Powerful Mini PC Optimized for Linux

     

     

    focahero.jpg

     

     

    Key findings from ZDNET

    • This powerful PC is available now starting at $1,230
    • NX Gen3 PC was able to run gpt-oss:20b LLM without issue
    • This PC with Kubuntu is probably more powerful than your current PC.

     

    Follow ZDNET: Add us as a preferred source on Google.


    Whenever I get the chance to test a Linux PC, I’ll be happy to do so. It’s not just because I can experiment with how third parties approach the operating system, but also because it means there are more possibilities for consumers to purchase Linux systems.

    If you’ve heard of Kubuntu, you know it’s a version of Ubuntu that focuses on the KDE Plasma desktop – a shiny desktop interface. Kubuntu Focus, on the other hand, is a company dedicated to selling laptops and desktops equipped with the Kubuntu operating system.

    Also: The best Linux laptops

    After receiving the Kubuntu Focus NX Gen3unpacking and configuring, I turned on the power and watched Kubuntu load. Within seconds, I was greeted by the onboarding wizard. In stark contrast to any Windows PC I’ve tested, it took me less than two minutes to get the NX Gen3 up and running.

    My experience

    Reviewing a Linux laptop or PC is always a breath of fresh air for me because I know the operating system very well (so I know exactly what to expect) and it’s almost always a positive experience from start to finish.

    The NX Gen3 was no exception.

    The model I received housed an Intel Ultra 7 255H CPU/GPU, Arc T140 8 Xe, with 32 GB of RAM and Intel Mesa graphics, which costs around $1,230. You can spec a system with up to 96GB of RAM, a 4TB drive, which increases the price to $2,075.

    Kubuntu Focus NX Gen3.

    There are plenty of ventilation openings to keep this PC cool.

    Jack Wallen/ZDNET

    Of course, it’s KDE Plasma, so the desktop is glorious. And yes, the default is a dark theme, which I customized right away. Once I got that out of the way, I ran the available updates (which took three minutes total) and started doing the one thing I always do when testing a new PC: pushing it to the limit.

    As? Nowadays, this is very easy. I installed Ollama, pulled LLM llama3.2 and ran a query. I got used to slow responses to Ollama queries on review machines, and at first I thought the NX Gen3 would give me the same results. After a few seconds, however, Ollama spat out the answer to “What is Linux?” much faster than I thought. In fact, I’ve never seen a small form factor PC respond so quickly to a local AI query.

    Impressive.

    Also: How do I feed my files to local AI for better, more relevant answers

    I then decided to download a larger LLM (gpt-oss:12b, which is 65 GB). Of course, the pull took forever (because of its size, which doesn’t reflect NX Gen3. Once complete, I ran the same query (“What is Linux?”), which was significantly faster than llama3.2 LLM. While gpt-oss:20 responded to the query, I started using other applications, and to my surprise, nothing was harmed by Ollama punching the CPU. The NX Gen3 didn’t even blink.

    In addition to the AI ​​portion of my test, I performed real everyday tasks (installing apps, opening and using apps, moving windows, running updates, browsing… you know the drill) and I wasn’t at all surprised as the NX Gen3’s performance was absolutely stellar. I’d put this little machine at about the same level as my System76 Thelio (which is the most powerful desktop PC I’ve ever used).

    I would go further and say that the Kubuntu Focus NX Gen3 PC is powerful enough for pretty much everything you need for average use. No, you’re not going to mine massive data or launch a spacecraft to Mars with this computer, but for the average human, the NX Gen3 offers well-above-average performance.

    Also: How to Run DeepSeek AI Locally to Protect Your Privacy – 2 Easy Ways

    Even with various KDE Plasma window effects running (like Wobbly Windows), I didn’t see any lag, tearing, or pixelation.

    No matter what I did with the NX Gen3 PC, I was impressed. This little machine is a powerhouse and would serve anyone well. And yes, KDE Plasma is a suitable desktop environment for those new to Linux. So whether you’re new to the open source operating system or have been using it for years, you can bet that the Kubuntu Focus NX Gen3 PC is up to whatever task you want to tackle.

    I didn’t bother running benchmarking tests on this machine because I tend to prefer my reviews to be of value to those who would actually buy such a machine to use every day. Benchmarking is great, but these numbers don’t tend to mean much to the average user.

    ZDNT Buying Advice

    If you’re in the market for a new PC and want to finally jump on the Linux bandwagon, the Kubuntu Focus NX Gen3 It’s not only a great place to start, but a great place to stay. This small size computer seriously impressed me. It’s powerful, quiet, easy to set up and easy to use.

    Also: How to install Steam on any Ubuntu-based Linux distribution so you can play a world of games

    If you’re looking at the end of life for your Windows 10 PC and are looking for something better than what’s on the shelves at big box stores, head over to Kubuntu Focus official website and buy an NX Gen3 PC. You won’t regret it.

     

    Kubuntu Focus NX Gen3 technical specifications

    I will list the basic specifications here:

    • CPU/GPU – Intel Ultra 5 225H CPU/GPU Arch T130 Xe
    • RAM – 16GB
    • Storage – 500GB PCIe 4
    • Disk encryption – optional
    • Warranty – 1 year limited
    • Shipping – 3-6 days free

    avots

  • Lifetime access to AdGuard is just $16 today

    Lifetime access to AdGuard is just $16 today

     

    DR: Get one AdGuard family plan lifetime subscription for just $15.97 with code FAMPLAN (MSRP $169.99) until 11:59pm PT on October 12th and protect 9 devices from ads, trackers, and online threats.


    Tired of dodging pop-ups, dealing with slow-loading websites, and feeling like your every move online can be tracked? You are not alone. THE the internet should be a safe, peaceful and private space. The AdGuard Family Plan can help you smooth things out, digitally speaking.

    Just $15.97 with promo code FAMPLANyou’ll get lifetime protection for up to nine devices. Whether on your phone, tablet or laptop, AdGuard blocks intrusive ads, prevents trackers from spying on your data and protects your family from potential malicious websites.

    Ads aren’t just annoying; they slow down your browsing, consume your data, and sometimes even carry malware. And if you have kids online, you want to make sure they’re safe from inappropriate content. The AdGuard Family Plan does all this and more. With built-in parental controlsyou can filter harmful websites and ensure your family’s online experience is safe and enjoyable.

    With coverage for up to 9 devices, the AdGuard Family Plan is ideal for homes where everyone is connected. And because it’s a one-time payment for lifetime access, you’ll never have to worry about recurring fees or annoying subscription renewals.

    So whether you’re protecting your own devices or those of your entire family, AdGuard facilitates a smoother online experience. Say goodbye to annoying ads and hello to a safer, faster Internet.

    Mashable Offers

    Don’t miss the opportunity to save a lifetime of protection from the AdGuard Family Plan although it only costs $15.97 (MSRP $169.99) with code FAMPLAN until October 12th at 11:59 pm PT.

    StackSocial prices subject to change.

    avots

  • What I Learned Testing the Starlink Mini Deep in the Washington Wilderness

    What I Learned Testing the Starlink Mini Deep in the Washington Wilderness

     

    On almost every hike or camping trip I take in Washington state, there’s a moment on the drive when service drops, music cuts out in the car and texts stop coming in. That’s when I start hoping I remembered to download the trail map.

    This disconnection is usually a welcome reprieve. I go off into the mountains partly to get a break from my phone, after all. But on this brisk Sunday morning in the North Cascades, that moment never came. I’d brought along a Starlink Mini in my backpack.

    As my friends waited in line for the bathroom at the trailhead, I pulled out the small white square, about the size of a laptop, plugged it into my portable battery and waited for it to communicate with the Starlink satellites zooming invisibly 342 miles above our heads.

    This was a slightly unusual way to use the Starlink Mini. The dish comes with Starlink’s Roam plans, which are designed for “high-speed internet on the go,” according to the company. It uses less than half the power of the full-size Starlink dish and weighs a little over 3 pounds, so it’s feasible to bring it along on a backcountry hike or camping trip. But more common use cases are RVs, vans and boats that aren’t tied to one specific location.

    To test out the Starlink Mini, I also spent three days connected to it at my apartment in Seattle. I did everything I normally would on my regular old cable internet: chatted on Zoom meetings, streamed live sports and completed crosswords online with a friend.

    I’ve written skeptically about Starlink in the past, but as I sat on top of a mountain in the North Cascades watching live NFL games in HD, I couldn’t help but grin. In this article, I’ll cover everything you need to know to get the most out of a Starlink Mini, and give you a full picture of the kind of performance you can expect from it.

     

    man-using-cellphone-next-to-starlink-mini-dish-with-mountains-in-background

     

    It took about ten minutes for the Starlink Mini to find a connection in the mountains.

     

    Joe Supan/CNET

    How to set up the Starlink Mini

    Setting up the Starlink Mini is incredibly simple. Here’s everything you’ll get in the box:

    • Starlink Mini dish (16.92 x 13.14 x 3.11 in)
    • Kickstand
    • Pipe adapter and flat mount
    • Power cable (49.2 feet)
    • Power supply
    • Plug

     

    contents-of-starlink-mini-box-laid-out-on-wooden-table

     

    The Starlink Mini was incredibly simple to set up, with only six items in the box.

     

    Joe Supan/CNET

    Download the Starlink app

    Your first step is to download the Starlink app, which you can access from a QR code on the packaging. This will guide you through the entire setup process, from finding an ideal location to creating a Wi-Fi network.

    Check for obstructions

    After the app is downloaded, it will prompt you to find a good resting place for the dish. Starlink says you need a “clear view of the sky” away from obstructions like tree branches, telephone poles or roofs. You can use the app to check for obstructions by holding your phone up to the sky.

     

    Screenshots-of-Starlink-obstructions-tool-in-app

     

    The Starlink app directs you to hold your phone up to the sky to analyze obstructions in the area before mounting.

     

    I wouldn’t get too bogged down in finding the perfect location. By design, the Starlink Mini probably won’t be staying in one place for too long. If you have a video meeting you want to ensure a stable connection for, you may want to run a quick obstruction check before you hop on. But for most Starlink Mini users, the obstructions surrounding the dish will constantly be in flux.

    Plug in the dish

    Once you’ve found a suitable spot, you plug one end of the power cable into the supply box and the other end into the dish. The power cable runs 49.2 feet, so you have some room to place the dish far away from the nearest electrical outlet.

    Connect to Wi-Fi

    Once it’s plugged in, you’ll be prompted to set up a Wi-Fi connection. Remember to follow best Wi-Fi practices to keep your network secure: Use a strong and unique password, create a guest network for visitors and set a time for automatic software updates.

    It will take a few minutes for Starlink to communicate with satellites in the sky. Don’t panic if you’re offline or have slow speeds to start. It takes up to 30 minutes to optimize its connection and sometimes you won’t see maximum speeds until it’s been in the same spot for a full day.

    Align Starlink

    Once your network is set up, you can fine tune the dish’s connection with Starlink’s satellites by tweaking the alignment. You’ll be prompted to rotate the dish until it’s facing the right direction at the ideal angle.

     

    Screenshots-of-Starlink-alignment-tool-in-app

     

    Starlink’s alignment setup guides you to the optimal angle to mount your dish.

     

    Joe Supan/CNET

    Mounting

    There are many options for mounting your Starlink Mini. Several mounts are available from Starlink and others from third-party manufacturers. The Mini comes with a pipe adapter and flat mount accessory, which is designed to be installed on a pipe or flat surface in a “stationary location.” If you want to use the dish on a vehicle, Starlink sells a mount that’s designed to be removable from a roof rack and another that stays in one place permanently.

    Optimizing your Starlink Mini for travel

    The Starlink Mini is meant for on-the-go connectivity, but this presents a few challenges that you don’t have with a stationary dish. Here’s everything you’ll need to know to get internet on the road — or the trail, sea or campground.

    Powering the Starlink Mini

    Since you may not have access to an electrical outlet, you’ll have to find an alternative power source. Starlink says that the Mini dish requires a 100-watt (20V/5A) power source to operate optimally, or a minimum of 65 watts (12-48V). Absent a generator, you have three options for this kind of power: a vehicle, portable power station or power bank.

    Using a vehicle to power the Starlink Mini is generally the simplest option. You’ll need to purchase an adapter that can be used with an automotive 12-24V auxiliary power outlet, also known as the car cigarette lighter. Starlink sells a Mini Car Adapter for $45, and there are a number of options available from third-party sellers. One popular combination is UGREEN’s 130W USB-C Car Charger paired with a Starlink Mini USB-C to DC Power Cable.

    A portable power station is a good option for camping if you don’t want to worry about draining your vehicle’s battery, but these are more expensive and heavier than most power banks, so they wouldn’t necessarily be a good fit for hiking or backpacking. These generally have AC outlets on them, so you can use the provided power supply without having to purchase a separate adapter.

    If you want something more portable or for more intermittent use, a power bank is your best bet. I used the Anker 737 Power Bank to test the Starlink Mini, which gets you 24,000 Milliamp Hours (mAh) capacity and 140-watt output. (Again, Starlink says the minimum you can use is 65 watts, but I’d recommend going with at least 100 watts for any power bank.)

     

    starlink-mini-connected-to-power-bank-in-grass

     

    A 24,000mAh Anker 737 Power Bank powered my Starlink Mini for about four hours at a time.

     

    Joe Supan/CNET

    The capacity here is the key number. At 24,000 mAh, the Anker 737 is on the higher end of most power banks you’ll see. Starlink says the Mini consumes between 25 and 40 watts on average. Working from home on the Mini, my power draw was just 21 watts on average. That meant my Anker battery could power the Starlink Mini for about 4 hours at a time.

    If you go the power bank route, you’ll need to buy a separate cable that uses a USB-C. There are a bunch of these for about $10 on Amazon. I ended up going with a 10 foot cable that had both a USB-C and cigarette lighter plug. That way, I could use the power bank as my primary source but also use my vehicle in a pinch.

    Trees: Starlink’s Achilles heel

    Starlink isn’t exaggerating when it says you need a clear view of the sky for it to work properly. I tried putting it under a variety of trees around my house, and it failed to establish a connection under any of them. Even where you could see patches of sky through the branches, the Starlink app still returned an “Obstructed” error message.

     

    starlink-mini-on-ground-under-trees

     

    The Starlink Mini failed to establish a connection in wooded areas.

     

    Joe Supan/CNET

    You could potentially move the dish to a clearing, especially if you’re camping near a lake or river, and set up the dish there. I was still getting a pretty strong connection from about 50 feet away from the dish, so you have some leeway in finding an optimal location.

    Troubleshooting common issues

    If you have a clear view of the sky and the Starlink app is still showing no connection to the internet, the first place you should check is the power. The status light on the back of the Starlink Mini should be blinking slowly. If there’s no light at all, it means there’s a problem with the power source.

    If the power is on, but you’re still having trouble establishing a connection with the Starlink Mini, my first recommendation is to wait. When I powered it up on my hike for the first time, I was disappointed at first that I couldn’t get a connection. But after about 10 minutes of the dish searching for satellites, I was online and checking my fantasy football scores.

     

    man-kneeling-next-to-starlink-mini-dish-with-phone-in-hand

     

    It took about 10 minutes for the Mini to establish a connection on my hike.

     

    Joe Supan/CNET

    After 10 minutes, your Starlink Mini should have established some connection to the internet (and should continue to get stronger over the next 24 hours). If you’re still not getting a signal, Starlink recommends unplugging the Mini from the power source and plugging it back in. If that doesn’t work, you can reset it to factory settings, which will require setting up a new Wi-Fi network.

    Starlink in the wild: How did it hold up in the backcountry?

     

    man-walking-in-mountains-with-starlink-mini-dish-on-backpack

     

    To test out the Starlink Mini on the move, I strapped it to my backpack for the hike down from Cutthroat Pass.

     

    Joe Supan/CNET

    To test out the Starlink Mini in the wilderness, I powered it up at various points along the 10-mile hike and took some speed tests. Once I got to the top (and wouldn’t annoy my hiking partners too much), I ran some more real-world tests. The Starlink Mini passed all of them beautifully. I streamed live NFL games, FaceTimed a friend and streamed Game of Thrones in 4K. In every case, the experience felt no different than if I were doing those things on my cable internet connection at home.

    That said, the numbers tell a slightly different story. I took 12 speed tests on the 5-hour hike, and the Starlink Mini returned averages of 127Mbps download speed, 17Mbps upload speed and 46 milliseconds latency. That was shockingly close to what Ookla reports as Starlink’s median US performance: 105Mbps download, 15Mbps upload and 45ms latency. (Disclosure: Ookla is owned by the same company as CNET, Ziff Davis.)

    While those numbers were certainly serviceable for most online tasks, they didn’t come close to my cable internet through Xfinity. Though it’s a slightly unfair comparison. Starlink Mini is meant to keep you connected where broadband infrastructure doesn’t exist, but I think it’s worth putting Starlink’s speeds into context.

    I pay $63 per month through Xfinity and got speeds of 422/175Mbps with 20ms latency when I tested my connection over the same period. That’s less than I would pay for Starlink Roam ($80 per month), and I’d only get 50GB of data each month with Starlink.

    I use about 90GB of data at home in an average week, and I don’t do any particularly bandwidth-intensive activities, just a few Zoom meetings every day, working in Google Docs and streaming TV for an hour or two. With that diet, I would run out of data with Starlink’s 50GB Roam plan in a few days. Starlink charges $1 for every GB you go over your data cap each month. If you want unlimited data for the Starlink Mini, it’ll cost you $195 monthly.

    Starlink’s upload and download speeds were high enough that most small to mid-size households probably wouldn’t experience many limitations. But there is one group that would: online gamers. Starlink’s had an average latency of 45 milliseconds in my tests. I usually say anything below 50ms is good enough for gaming, but there were several spikes above 100ms in my 24 speed tests with Starlink. That’s not a huge deal if you’re just checking emails or streaming Netflix, but it could make an online gaming experience nearly unplayable over the course of a couple of hours.

    Starlink Mini: Key benefits and drawbacks

    Pros

    • Internet isn’t tied to one location
    • Lightweight and portable
    • Plenty of speed for most online activities
    • Simple setup and installation
    • Intuitive and helpful app

    Cons

    • Signal is easily blocked by trees and other obstacles
    • Uses a significant amount of power
    • Latency spikes were common
    • 50GB monthly data cap is extremely low for most people

    Is the Starlink Mini worth it?

    I don’t think it’s an overstatement to say Starlink is a miraculous technology. One user in Alaska described it to me in a previous interview as an absolute game changer where he lives.

    “You bring a Starlink dish out there and plug it in. Two minutes later, you’ve got the whole world again in the palm of your hand,” he said.

    My experience with the Mini was no different. It was absolutely surreal to be on top of a mountain deep in the backcountry watching a live NFL game in crystal clear HD. But there are only a few specific situations in which I’d recommend the Starlink Mini.

    If you’re going to be traveling for a long period of time and won’t have consistent access to an internet connection, the Starlink Mini is essentially your only option for staying connected. But that comes at a significant price. The dish itself retails for $299, though Starlink is currently offering it for free to new customers. And you’ll pay another $165 per month if you want unlimited data, which is essential if you plan on using it like you would a home internet connection.

    If you just want to use it periodically, you might be able to get by with the 50GB plan for $50 a month. Starlink used to let you pause service on Roam plans, so you could activate it only for specific trips. As of August, you now have to pay $5 a month for that stop-restart privilege.

    I could see how having the peace of mind to check email or look up directions while traveling would be worth the cost for some people. And in those situations, the Starlink Mini performs beautifully. Just don’t expect the same performance you’d get at home.

    avots

  • The Fixer’s Dilemma: Chris Lehane and OpenAI’s Mission Impossible

    The Fixer’s Dilemma: Chris Lehane and OpenAI’s Mission Impossible

     

    Chris Lehane is one of the best in the business at making bad news disappear. Al Gore’s press secretary during the Clinton years, Airbnb’s chief crisis manager through every regulatory nightmare from here to Brussels – Lehane knows the drill. Now he’s two years into what may be his most impossible job: As vice president of global policy at OpenAI, his job is to convince the world that OpenAI actually cares about the democratization of artificial intelligence, while the company increasingly behaves like, well, every other tech giant that ever claimed to be different.

    I spent 20 minutes with him on stage at Elevate conference in Toronto earlier this week – 20 minutes to get past the talking points and get into the real contradictions that erode OpenAI’s carefully constructed image. It was neither easy nor completely successful. Lehane is genuinely good at what he does. He’s friendly. He seems reasonable. He admits uncertainty. He even talks about waking up at 3 a.m. worrying about whether it will actually benefit humanity.

    But good intentions don’t mean much when your company is subpoenaing critics, draining water and electricity from economically depressed cities, and bringing dead celebrities back to life to assert its market dominance.

    The company’s Sora problem is really at the root of everything else. The video generation tool was launched last week with copyrighted material apparently embedded into it. It was a bold move for a company that was already being sued by the New York Times, the Toronto Star and half the publishing industry. From a commercial and marketing point of view, it was also brilliant. The invite-only app rose to the top of the App Store as people created digital versions of themselves, said OpenAI CEO Sam Altman; characters like Pikachu and Cartman from “South Park”; and dead celebrities like Tupac Shakur.

    Asked what motivated OpenAI’s decision to launch this newest version of Sora with these characters, Lehane responded that Sora is a “general purpose technology” like the printing press, democratizing creativity for people without talent or resources. Even he – who calls himself a creative zero – can make videos now, he said on stage.

    What he guessed is that OpenAI initially “let” rights holders opt out of having their work used to train Sora, which is not how copyright use normally works. So, after OpenAI realized that people really liked using copyrighted images, it “evolved” into an opt-in model. This is not iteration. This is testing how much you can get away with. (By the way, although the Motion Picture Association made some noise last week about legal threats, OpenAI appears to have largely escaped.)

    Naturally, the situation brings to mind the aggravation of publishers who accuse OpenAI of training their work without sharing the financial spoils. When I pressed Lehane about publishers being priced out of the economy, he invoked fair use, that American legal doctrine that is supposed to balance the rights of the creator with public access to knowledge. He called it the secret weapon of US technological dominance.

    Techcrunch Event

    San Francisco
    |
    October 27-29, 2025

    Perhaps. But I recently interviewed Al Gore – Lehane’s old boss – and realized anyone could just ask ChatGPT about this instead of reading my TechCrunch article. “It’s ‘iterative,’” I said, “but it’s also a substitute.”

    Lehane listened and abandoned his speech. “We’re all going to need to figure this out,” he said. “It’s very simplistic and easy to sit here on stage and say we need to discover new economic revenue models. But I think we will.” (We’re making it up as we go along, so I hear.)

    Then there is the question of infrastructure that no one wants to answer honestly. OpenAI already operates a data center campus in Abilene, Texas, and recently opened a massive data center in Lordstown, Ohio, in partnership with Oracle and SoftBank. Lehane compared the adoption of AI to the advent of electricity – saying that those who accessed it last are still playing catch-up – but OpenAI’s Stargate project appears to be targeting some of those same economically challenged locations to create facilities with their huge appetite for water and electricity.

    Asked during our meeting whether these communities will benefit or just pay the bill, Lehane addressed gigawatts and geopolitics. OpenAI needs about a gigawatt of power per week, he noted. China produced 450 gigawatts last year, in addition to 33 nuclear facilities. If democracies want democratic AI, he said, they have to compete. “The optimist in me says this will modernize our energy systems,” he said, painting a picture of a reindustrialized America with transformed electrical grids.

    It was inspiring, but it wasn’t an answer to whether people in Lordstown and Abilene will see their utility bills increase as OpenAI generates videos of The Notorious BIG. AI that consumes the most energy outside.

    There’s also a human cost, which became clearer the day before our interview, when Zelda Williams took to Instagram to beg strangers to stop sending her AI-generated videos of her late father, Robin Williams. “You are not making art,” she wrote. “You are turning the lives of human beings into disgusting, over-processed hot dogs.”

    When I asked how the company reconciles this kind of intimate harm with its mission, Lehane responded by talking about processes, including responsible design, testing frameworks, and government partnerships. “There’s no manual for these things, right?”

    Lehane showed vulnerability at times, saying he recognized the “huge responsibilities that come with” everything OpenAI does.

    Whether or not these moments were planned for the audience, I believe him. In fact, I left Toronto thinking I had watched a masterclass in political messaging — Lehane threading an impossible needle while dodging questions about company decisions that, as far as I can tell, he doesn’t even agree. Then the news revealed that complicated situation, which was already complicated.

    Nathan Calvin, a lawyer who works on AI policy at the non-profit organization Encode AI, revealed that at the same time I was speaking with Lehane in Toronto, OpenAI had sent a sheriff’s deputy at Calvin’s house in Washington, D.C., during dinner to serve him a subpoena. They wanted your private messages with California lawmakers, college students, and former OpenAI employees.

    Calvin says the move was part of OpenAI’s scare tactics surrounding a new AI regulation, California’s SB 53. He says the company weaponized its ongoing legal battle with Elon Musk as a pretext to target critics, implying that Encode was secretly funded by Musk. Calvin added that he fought OpenAI’s opposition to California’s SB 53, an AI safety bill, and that when he saw OpenAI claim that it “worked to improve the bill,” he “literally laughed out loud.” In a social media kerfuffle, he went on to specifically call Lehane a “master of the dark political arts.”

    In Washington, that might be a compliment. At a company like OpenAI, whose mission is to “build AI that benefits all humanity,” this sounds like an accusation.

    But what matters much more is that even the folks at OpenAI are conflicted about what it’s becoming.

    As my colleague Max reported last week, several current and former employees took to social media following the release of Sora 2, expressing their doubts. Among them was Boaz Barak, an OpenAI researcher and Harvard professor, who wrote about Sora 2 which is “technically incredible, but it’s premature to congratulate ourselves for avoiding the pitfalls of other social media apps and deepfakes.”

    On Friday, Josh Achiam – head of mission alignment at OpenAI – tweeted something even more remarkable about Calvin’s accusation. Prefacing his comments by saying they were “possibly a risk to my entire career,” Achiam went on to write about OpenAI: “We cannot do things that make us a frightening power rather than a virtuous one. We have a duty and a mission to all humanity. The bar for fulfilling that duty is remarkably high.”

    It’s worth pausing to think about this. An OpenAI executive publicly questioning whether his company is becoming “a frightening powerhouse rather than a virtuous one” is not on the same level as a competitor taking shots or a reporter asking questions. This is someone who chose to work at OpenAI, who believes in its mission and who now recognizes a crisis of conscience despite the professional risk.

    It’s a moment of crystallization, the contradictions of which can only intensify as OpenAI moves towards artificial general intelligence. It also made me think that the real question isn’t whether Chris Lehane can sell OpenAI’s mission. It’s whether others – including, critically, the other people who work there – still believe it.

    avots

  • Lenovo Yoga Pro 9i 16 review: A true MacBook Pro rival?

    Lenovo Yoga Pro 9i 16 review: A true MacBook Pro rival?

     

    We shouldn’t expect any Windows laptop with a powerful discrete GPU to truly replicate what the MacBook Pro does. Yes, there are more powerful systems out there, but efficiency is not the goal of these systems. Even with the improvements Nvidia has made to Advanced Optimus (automatic switching between discrete GPUs when needed), battery life suffers, especially when running heavier applications. In a local video playback test, the Yoga Pro 9i 16 lasted about 12 hours. Despite using the same 84-watt-hour battery, this appears to be a slight improvement over last year’s model, although it’s difficult to make a head-to-head comparison. I know battery life decreases quickly under heavy load, as it died in just 45 minutes while running a benchmark. You’ll want to be connected if you’re doing something very serious.

    Regardless of the task, you’ll get twice the battery life in an M4 Max MacBook Pro. Only when we have ARM-based systems with powerful integrated graphics that rival the M4 Pro and M4 Max will there be competition for Apple. The closest thing we’ve seen so far is AMD’s exclusive Ryzen AI Max+ processor, which appeared in the Asus ROG Flow Z13 and used a massive integrated graphics chip to challenge traditional discrete graphics. But we still have a long way to go.

    The only other Windows laptop that could be better is the Asus ProArt P16which I haven’t tested yet. Now it even comes with an RTX 5070 or 5090 option, which could make it significantly more powerful than the Yoga Pro 9i. However, it’s also a much more expensive laptop, configured with a 4K OLED display and only high-end GPUs. The Yoga Pro 9i is also hundreds of dollars cheaper than the Dell 16 Premium when configured similarly.

    The Lenovo Yoga Pro 9i 16 gets a lot of things right—enough to make it worth recommending as a worthy MacBook Pro competitor. Ultimately, it’s the performance, screen, and premium quality that make it a worthy content creation machine, and the Yoga Pro 9i succeeds on all of those fronts, perhaps better than any other Windows machine I’ve tested.

    avots

  • Seattle Space Week Offers Tips for Dreamy Entrepreneurs — and a Preview of a Hypersonic Jet Gun

    Seattle Space Week Offers Tips for Dreamy Entrepreneurs — and a Preview of a Hypersonic Jet Gun

     

    Wave Motion co-founders James Penna and Finn van Donkelaar show off their company’s Jet Gun prototype and test projectiles in the back of a box truck parked in Seattle’s Pioneer Square. (GeekWire Photo/Alan Boyle)

    Most weeklong tech events offer opportunities for entrepreneurs to network and trade tips, serious sessions where CEOs and public officials share their visions, and happy hours where future deals are made. But how many “tech weeks” include a show-and-tell featuring a military-grade Jet Gun?

    This was one of the bonus attractions during Seattle Space Weeka hodgepodge of events catered by Northwest Space and its partners.

    By the time attendees were seated for Monday’s opening session in the Pioneer Building in the heart of Pioneer Square, team members from Wave Motion Launch Corp. parked a box truck outside the building and opened the back to reveal the prototype jet blaster they are testing for the US Army.

    Two of the Everett, Washington-based startup’s co-founders, CEO Finn van Donkelaar and chief operating officer James Penna, climbed into the truck and explained their project to a crowd that gathered on the sidewalk.

    Wave Motion’s Jet Gun uses a concentrated blast of gas and small particles to propel a projectile at hypersonic speeds. Unlike a rocket, which has to carry its own propellant, the Jet Gun would launch the projectile (or, say, a satellite) into the air (or, say, space) from the ground. Wave Motion claims the barrelless blaster has the potential to be up to 100 times more compact than a conventional rocket or cannon of equivalent power.

    The weapon’s prototype is a steampunk-style contraption about 10 feet long.

    “This is a demonstrator in the same way as SpaceX’s Hopper, right?” Penna told the crowd. “It’s a proof of concept [to show] that we can use this propulsion concept to at least make a vehicle go up and down, or wherever it wants to go. …So as we build bigger, more powerful jets, and build multiples of them, we will be able to design the jet.”

    The prototype in the back of the truck is designed to launch a jet at speeds ranging from five to 10 times the speed of sound, Penna said. “In the orbital prototype, it will be a greater-than-Mach 20 stream of hellish steel and fire, extending tens of kilometers through the atmosphere,” he said. “It’s going to be extremely cool.”

    The launch of Wave Motion was founded in 2020 and won a $1.35 million award from the U.S. Navy in 2022 for work on the Jet Gun. Last year, the company received a $1.6 million contract from the Army to support continued development. “Our business strategy is to make the first demonstrations of the principle, because this is the first propulsion system of its kind built in the world,” said Penna.

    It seems reasonable to think that a blaster capable of releasing a hypersonic stream of hellfire and steel could be used as a weapon as well as a propulsion system. Van Donkelaar acknowledged that this may well be something to consider. “We hope to get some military contracts out of this as well,” he said.

    FundingQuest coach Bryan Brewer makes the point during a Seattle Space Week panel. Other speakers include Mehran Mesbahi, executive director of the Joint Center for Aerospace Innovation; Andrew Hanna, senior manager of government programs and business development at Stoke Space; Joe Nguyen, director of the Washington State Department of Commerce; Nicole Brown, Starfish Space senior mission manager; and moderator Sean McClinton, co-founder and director of entrepreneurship at Space Northwest. (GeekWire Photo/Alan Boyle)

    Here are some other highlights from Seattle Space Week:

    • Space Northwest co-founder Sean McClinton said his industry group’s goal was to increase Washington state’s annual economic activity in the space sector from $4.6 billion to $40 billion by 2040, with 90 new space startups created in the next decade (compared to 40 in the last decade).
    • Washington Governor Bob Ferguson marked Seattle Space Week with a proclamation – and Joe Nguyen, director of the Washington State Department of Commerce, said he wanted to pave the way for the space industry. “This is how we can really make the government an ally and not an obstacle to the work that is being done, because this work is very urgent,” Nguyen said. “Other states and other countries are competing for this same industry.”
    • Nguyen said his priorities included streamlining the licensing process for manufacturing facilities and testing sites. “Can we invest in infrastructure and energy so that we have the tools we need to succeed? Can we help foster research partnerships between universities and our companies as well?” he said.
    • The state is not likely to give space companies the same types of tax incentives that Boeing received for airplane manufacturing, Nguyen said. But he said support could come in the form of infrastructure development: “One of the other priorities we have at the Department of Commerce is building clean energy…. We have identified 21 key projects that we believe can come in time for the IRA.” [Inflation Reduction Act funding] and being able to develop seven gigawatts of energy for our communities here, and that has already brought a huge benefit. We can do exactly the same thing for the space economy.”
    • Several business representatives laid out a to-do list to support Washington state’s space industry, ranging from giving a bigger boost to education and workforce development to providing more access to testing facilities for space startups. “What we’re going to do is start developing a database of who has what testing resources and figure out how we can help each other,” said Marcy Mabry, CEO of the Seattle-based company. Space Launch. “So, Joe, we can talk to you.”
    • Nicole Brown, Senior Mission Manager in Tukwila, Washington Star of the Sea Spacesaid accessibility and competition from other technology sectors pose major challenges for recruiting employees. “It’s very expensive here, and when you’re trying to recruit talent, even people who would really love to live here need to think twice,” Brown said. “I think we have a mission that people are very excited about, and we’re moving quickly, and people are happy to come and be creators, but we can’t compete pay-wise with some of the other tech companies here in the region.”

    Seattle Space Week ends today with “Proposals for startups and networking: Seattle Space Week edition” in the Pioneer Building, followed by Space Week Karaoke in Golden Cocks in Pioneiro Square. Registration for pitch session required (but probably not for karaoke).

    avots

  • Marineland says its 30 belugas may have to be euthanized. How would this happen?

    Marineland says its 30 belugas may have to be euthanized. How would this happen?

     

    If you can’t leave them, sell them, or send them to a sanctuary, what do you do?

    Although Marineland has no current plans to euthanize the 30 belugas that swim in what remains of the Niagara Falls amusement park, the company threatened to do so last week in a letter to Canada’s Federal Fisheries Minister Joanne Thompson, obtained by CBC News.

    The park would face the “devastating decision to euthanize” unless the federal government could provide them with financial support, Marineland said in a letter.

    The message came days after the minister denied Marineland’s request for authorization to send the whales to Chimelong Ocean Kingdom, a theme park in China that was interested in purchasing the mammals. The government said it intended to prevent whales from being used for entertainment purposes.

    So far, Marineland has been unable to find a suitable sanctuary or other facility to house the whales. Theme park owners said in a report that a proposed sanctuary in Nova Scotia is heavily polluted and not on track to be developed soon.

    Marine mammal experts say that if the threat of beluga euthanasia became a reality, the option would come with a number of logistical and moral issues.

    WATCH | Marineland’s Request to Export Belugas to China Denied by Feds:

    Marineland’s request to export belugas to China denied by feds

    The Department of Fisheries and Oceans has denied Marineland Canada’s request to export the remaining 30 beluga whales to the Chimelong Ocean Kingdom theme park in China. The decision is in line with regulations to protect marine mammals from exploitation.

    Euthanasia medication is problematic, says vet

    For animals like those at Marineland, who are used to being examined by veterinarians, euthanasia is usually a two-step process, says veterinarian and marine biologist Chris Harvey-Clark.

    First, veterinarians would have to give the whales a sedative to ensure they didn’t feel any pain during the process, and then administer a medication that would end their lives.

    “That’s okind do gold standard method for euthanasia,” said Harvey-Clark, who is also affiliated with the Whale Sanctuary Project in Nova Scotia as a consultant.

    But the drug used to end life tends to be a barbiturate, Harvey-Clark says, which presents other problems when the procedure is performed.it’s done. Drugs can penetrate groundwater if animals are buried or can kill scavengers that make whales’ meal if they were left to decompose.

    And with 30 whales in total, each weighing between one and two tons, that leaves a lot of whale to dispose of.

    Andrew Trites, director of the Marine Mammal Research Unit at the University of British Columbia in Vancouver, adds that whales are, in many cases, considered too large to be cremated, making this method of disposal difficultalso.

    And because they spent their lives in captivityIn life, Trites says whales wouldn’t know how to hunt for food, so releasing them into the wild isn’t an option.

    Ethical issues

    Trites and Harvey-Clark warn that there are also a number of ethical obstacles to potentially euthanizing the whales.

    Euthanasia is a veterinary optionAnimals tend to feel morally good when an animal is suffering and the end of its life is near, Harvey-Clark said.

    In cases where death is not immediatenente, he says the morningevery choice is not that simple.

    “You really need to look at the well-being of the animal – what its life is like and what its future is going to be like, and from that make a, you know, an intelligent decision about whether this animal should be humanely euthanized or whether there is another way,” Harvey-Clark said.

    LISTEN | The fate of Marineland’s beluga whales:

    The current19:17The fate of Marineland’s beluga whales

    Marineland wanted to send its 30 captive beluga whales to an amusement park in China – but the federal government stopped them. And with the once-iconic amusement park closing its doors, the question of what to do with the whales and who should be responsible remains unclear. W

    Furthermore, belugas are highly social and skilled at communicating with each otherallowing them to form close bonds with other whales. Harvey-Clark says the loss of the whales one by one would presumably be something the belugas would notice, which could cause damage before they are ultimately killed.

    “They are certainly aware enough to understand that the environment they are in is changing very quickly, and that would be very stressful,” Harvey-Clark said.

    Trites adds that because animals are often euthanized when they are sick, there may be some experimentation when it comes to the exact dosage the whales should receive in order to euthanize them as well.

    black silhouettes of people in front of a blue tank of water where white beluga whales swim
    Tourists line up at an observation area to see two Marineland attractions, a killer whale calf swimming with its mother and a small pod of beluga whales in Niagara Falls, Ontario. on Wednesday, July 18, 2001. (CP PHOTO/Scott Dunlop) (Scott Dunlop/Canadian Press)

    “None of these [euthanasia processes] It seems like something I or anyone I know would want to be involved with,” Trites said.

    Trites says he hopes the threat never comes to fruition and that the whales can find homes where they can live good lives and contribute to future research on belugas, which face increasing challenges due to climate change.

    “But ultimately, this threat could become reality if no one moves any of their pieces on this chessboard,” he said.

    avots

  • Teens Hack School Cell Phone Bans with Creative Workarounds

    Teens Hack School Cell Phone Bans with Creative Workarounds

     

    NEWNow you can listen to Fox News articles!

    Across the country, schools are cracking down on cell phone use. At least 18 states have implemented doorbell-to-doorbell bans, with New York calling phones “distraction devices.” Teachers are praising the change, saying classes seem more focused. But teenagers? They don’t give up so easily.

    Students are avoiding the bans in the most millennial-inspired way possible, turning Google Docs into digital chat rooms. With their laptops open, it looks like they are working on tasks. In reality, they are typing messages in real time, like a traditional AOL chat room.

    SCHOOL SECURITY TOOLS ARE SPYING ON CHILDREN — EVEN AT HOME

    Sign up for my FREE CyberGuy report
    Receive my best tech tips, urgent security alerts, and exclusive offers straight to your inbox. Plus, you’ll get instant access to my Ultimate Scam Survival Guide – for free when you join my CyberGuy.com/Newsletter

    A girl looks at her laptop.

    Students secretly turn Google Docs into real-time chat rooms. (Kurt “CyberGuy” Knutsson)

    A Creative Workaround for Cell Phone Bans in Schools

    Parents and teachers admit that the alternative solution is smart. One teacher said she respects her students’ determination to stay connected and even acknowledged that the phone ban has improved behavior and concentration in class. Still, she worries that turning Google Docs into chat rooms could open the door to bullying or cheating. Parents are also participating. One parent told CyberGuy that some kids in her district are buying MacBooks just so they can text each other through iMessage. Others, the father added, are relying on email threads or even old post-its to keep the conversation alive.

    A girl writes at a table in front of an open laptop.

    Teenagers proudly share their classroom tricks on TikTok. (Kurt “CyberGuy” Knutsson)

    Reaction on social media

    On TikTok, students proudly display their “secret” Docs conversations. Captions range from playful: “Your cell phone rule would never stop me” to defiant: “We can never silence each other, queens.” The creativity is bringing laughter from older generations who remember the days before smartphones. But the trend also raises debate. Some parents see this as a harmless way for children to adapt, while others fear that it undermines the entire purpose of the ban. Educators are also divided, amused by the ingenuity but frustrated that students still find ways to get off task during classes. Viral clips prove one thing for sure: when it comes to technology, today’s teens will always find a workaround.

    A girl uses a MacBook's trackpad.

    Some kids buy MacBooks to keep texting via iMessage. (Kurt “CyberGuy” Knutsson)

    Tips for Parents for Dealing with Cell Phone Bans at School

    If your child’s school has adopted a phone ban, there are some ways to help your child adapt while keeping communication open and safe:

    • Talk about the rules at home: Explain why schools are implementing these bans and set expectations for how your child should behave with laptops and other devices.
    • Offer secure communication plans: Work with your child and the school to establish how you will contact each other in the event of an emergency. Some districts allow phones in lockers or require phones to remain unplugged in backpacks.
    • Encourage balance: Remind your child that downtime from screens can actually help them focus better in class and relax during the school day.
    • Monitoring alternatives: Keep an eye on how your child uses tools like Google Docs, email, or messaging apps. What starts as a conversation with friends can sometimes turn into bullying or cheating.
    • Be open to comments: Ask your child how the ban is affecting their school day. Their perspective can help you understand where the real challenges and benefits are appearing.

    TEENS AND PHONE USE WHILE DRIVING: WHY THIS DEADLY HABIT PERSISTS

    What does this mean for you

    If you are a parent, this shows how inventive children can be when rules are set. Cell phone bans may reduce scrolling, but students are quickly switching to other tools. They’re chatting through shared Google Docs, buying MacBooks to use iMessage during classes, exchanging notes via email, and even sticking to old-school Post-its to stay in touch. While some of these workarounds seem harmless, they also come with risks, from distractions that take your focus away from learning to new opportunities for bullying or even cheating. For teachers, it’s a reminder that managing distractions in the classroom goes beyond phone policies. Laptops, messaging apps and even simple sticky notes can become backdoors for the same behaviors schools are trying to limit.

    Take my test: How safe is your online security?

    Do you think your devices and data are really protected? Take this quick test to see where your digital habits stand. From passwords to Wi-Fi settings, you’ll get a personalized analysis of what you’re doing right and what needs improvement. Take my test here: CyberGuy.com/Quiz

    CLICK HERE TO GET THE FOX NEWS APP

    Kurt’s Key Takeaways

    Phone bans are reshaping the school day, and educators are already seeing benefits. However, students are proving that they will always find ways to connect, whether through phones, laptops, or even retro workarounds that echo the dawn of the Internet era.

    What do you think? Are these bans helping kids learn better, or are they simply encouraging students to get sneakier with technology? Let us know by writing to us at CyberGuy.com/Contact

    Sign up for my FREE CyberGuy report
    Receive my best tech tips, urgent security alerts, and exclusive offers straight to your inbox. Plus, you’ll get instant access to my Ultimate Scam Survival Guide – for free when you join my CyberGuy.com/Newsletter

    Copyright 2025 CyberGuy.com. All rights reserved.

     

    avots

  • Nvidia researchers boost LLMs’ reasoning skills by making them ‘think’ during pre-training

    Nvidia researchers boost LLMs’ reasoning skills by making them ‘think’ during pre-training

    Nvidia researchers have developed a new technique that flips the script on how large language models (LLMs) learn to reason.

    The method, called reinforcement learning pre-training (RLP), integrates RL into the initial training phase rather than saving it for the end.

    This approach encourages the model to “think for themselves before predicting what comes next, thus teaching independent thinking behavior early in pre-training,” the researchers state in their article.

    By learning to reason in plain text without needing external verifiers, models trained with RLP show significant improvements in learning complex reasoning tasks downstream, suggesting a future of AI that is more capable and adaptable for real-world tasks.

    The typical LLM training cycle

    Typically, large language models are first pre-trained on large amounts of text using a “next token prediction” objective, where they are given a string of text and asked to continually guess what the next word (or token) will be. At this stage, they learn basic grammar, facts, and associations.

    In the later post-training phase, models often learn complex reasoning skills such as chain of thought (CoT), where a model explains its reasoning step by step. This step usually involves supervised fine tuning (SFT) or reinforcement learning from human feedback (RLHF), which require specialized and curated datasets.

    The paper’s authors argue that this sequential process does not correspond to human understanding, which is “not a linear token-by-token process, but rather a parallel integration of input with prior knowledge.” Existing pre-training methods lack this mechanism, hindering a model’s ability to develop deep reasoning from the start.

    How reinforcement learning pre-training works

    RLP reformulates this process by treating CoT generation as an action that the model takes before predicting the next token. At each step, the model first generates an internal report “thought” or chain of reasoning. It then predicts the next word in the text, using the original context augmented with your new thought.

    The model receives a reward based on how much its thinking improved the accuracy of its prediction compared to a baseline that did not generate a thought (pure prediction of the next token). This reward signal is automatically calculated based on the change in probability, eliminating the need for external verifiers or human-labeled data.

    The reward is positive only when the thought generated helps the model better predict the next token. By rewarding thoughts based on their predictive benefit, RLP effectively teaches the model how to think usefully on the same massive, unstructured data sets used for standard pre-training.

    This continuous feedback loop allows the model to learn when a simple predictive guess is sufficient and when it needs to engage in deeper reasoning. As the researchers say, “RLP was designed to shape thinking into basic models, rewarding only those thoughts that measurably help in predicting the next token.”

    This fundamental approach, however, does not make the later stages of fine-tuning obsolete. According to Bryan Catanzaro, vice president of deep learning applied research at Nvidia and co-author of the paper, RLP is designed to complement, not replace, these crucial steps. “RLP is not intended to replace later post-training stages such as supervised fine-tuning or reinforcement learning from human feedback,” Catanzaro told VentureBeat. “These stages remain crucial to refining the model’s behavior… It was actually designed to amplify the effectiveness of these later phases, giving the model a head start.”

    RLP in action

    In experiments with Qwen3-1.7B and Nemotron-Nano-12BThe Nvidia team tested RLP on a set of mathematical and scientific reasoning benchmarks. The results show that RLP-enhanced models consistently outperformed their conventionally trained counterparts, with particularly strong gains on heavy reasoning tasks.

    For a company, this improved reasoning could translate into more reliable results in multi-step workflows like financial analysis or legal document summarization.

    “RLP encourages the model during pre-training to think before predicting, helping the model internalize a more coherent reasoning style,” Catanzaro said. “This could help reduce subtle logic errors, especially in longer workflows.”

    While emphasizing that RLP-trained models will still need the usual safeguards like verification layers, human oversight, and consistency checks, Catanzaro said that “RLP provides a stronger foundation.”

    Importantly, the benefits of RLP increase rather than disappear during subsequent stages of fine-tuning (catastrophic forgetting is a common problem in LLM training, where later stages of training cause the model to forget its previously learned skills and knowledge). The RLP-trained model achieved an overall score 7-8% higher than baseline values ​​after an identical post-training regimen. The researchers conclude that RLP “establishes robust reasoning foundations that are not eliminated by post-alignment but are instead combined post-training.”

    The efficiency of the technique is an important discovery. In the Qwen3-1.7B model, RLP improved performance by 17% over standard continuous pre-training and also outperformed a similar technique called Reinforcement Pre-training via Prefix Matching (RPT) rewards. This advantage held even when the baseline model was trained with 35 times more data to match the computational cost, confirming that the gains come from the method itself, and not just from more processing.

    Furthermore, RLP demonstrates impressive scalability and versatility, successfully extracting a reasoning signal from general-purpose web data – not just select datasets. When applied to the Mamba-Transformer hybrid model Nemotron-Nano-12B RLP achieved a 35% relative improvement over a heavily trained baseline while only using a small fraction of the data.

    While these results point to a more efficient path to building powerful models, Catanzaro frames the innovation as a fundamental change in the learning process itself, rather than an immediate solution to high training costs.

    “This research is interesting because it offers a change in the way models absorb information during pre-training, leading to a smarter learning process.” he explained. “It wouldn’t replace large-scale pre-training, but it would offer another creative method for building the best models possible.”

    A New Foundation for AI Training

    Ultimately, RLP points to a future where pre-training will no longer be a monolithic process of predicting the next token. Instead, the next generation of models could be built on a hybrid of goals, creating AI that learns to think more robustly from day one. Catanzaro offers a powerful analogy to frame this shift:

    “Predicting the next token teaches a model what the world is like; Reinforcement-style goals like RLP can teach you to think about what you’re seeing,” he said. “Combining these two goals can help models develop deeper, more structured thinking much earlier in training… Tools like RLP can build on this foundation, making learning more active, curious, and even more efficient.”

    There is still a lot to learn about the dynamics of reinforcement learning in the pre-training phase, but what seems clear is that “introducing exploration early in training opens up a new axis for scaling – not just in size, but in the way models learn to reason,” said Catanzaro.

    avots