News

When you open a forecasting app, you simply select a location, choose an hour and read the answer.

Behind that screen, however, lies one of the hardest prediction problems in science.

The atmosphere is a three-dimensional, constantly changing system.

Before a model can predict what it will do next, it must first answer a surprisingly difficult question: what exactly is the atmosphere doing right now?

No single instrument provides a complete picture of the weather. Weather radars detect precipitation particles. Satellites observe clouds, moisture and temperatures from space. Ground stations measure conditions at individual locations. Weather balloons and aircraft provide information from different altitudes.

Each source has its own resolution, update frequency, coverage gaps and measurement errors. Building the current state of the atmosphere is therefore like reconstructing traffic across an entire city using cameras that point in different directions, refresh at different times and leave many streets completely invisible.

Meteorologists call the process of combining observations with a previous forecast “data assimilation”. It produces an estimate of the current atmospheric state, known as an analysis, from which the next forecast begins.

As ECMWF explains, forecast quality depends crucially on the accuracy of this analysis. Even an excellent model will struggle if its starting picture is incomplete, inaccurate or already out of date.

One common short-term forecasting technique is optical flow. It analyses consecutive radar images, estimates how precipitation is moving and extrapolates that motion into the future. This can work well when a rain band is moving steadily. The problem is that storms are not static shapes sliding across a map. They form, intensify, weaken, split and merge. A new convective cell can appear where no rain existed 20 minutes earlier.

Imagine a storm approaching a mountain. A simple extrapolation may predict that it will continue moving at the same speed. In reality, the terrain can alter airflow, slow the system and contribute to heavier rain or hail. The challenge is therefore not only predicting where existing precipitation will move. A useful model must also predict how the weather system itself will change.

Another issue is that large-scale models are good at describing weather fronts, pressure systems and atmospheric circulation across countries and continents. But many events that affect people and businesses happen on a much smaller scale. A thunderstorm may affect one part of a city while leaving another almost dry.

A cloud front can sharply reduce output at one solar farm but miss another facility 20 kilometers away. Wind conditions can differ substantially between the ground, a rooftop and the altitude at which a drone operates. Those they can’t predict.

Weather is also highly sensitive to its initial state. Small uncertainties in temperature, humidity or wind can grow as a forecast extends further into the future. This is the real meaning of the “butterfly effect”. It does not mean that one faulty sensor automatically destroys a global forecast. It means that we can never measure the atmosphere perfectly, and uncertainty in its initial state grows over time.

That is why modern forecasting systems increasingly produce ensembles: multiple plausible forecasts rather than one supposedly certain answer. For many decisions, knowing that there is a 30% probability of a severe storm is more useful than receiving a confident prediction that later turns out to be wrong.

Latency is one of the least discussed sources of forecast error. Consider a model that produces an excellent forecast based on atmospheric conditions measured two hours ago. If a thunderstorm formed 20 minutes ago, the forecast may already be operationally useless, regardless of how sophisticated the model is.

The World Meteorological Organization defines nowcasting as detailed local forecasting from the present up to six hours ahead. For real-world applications, however, nowcasting is more than a forecast horizon. It requires a continuous pipeline that collects new observations, checks their quality, synchronizes them in time, runs the model and delivers an updated result within minutes.

At the same time, real-time weather forecasting does not mean literally zero delay. What system does is keeps listening to the atmosphere and updates quickly enough to influence the decision being made.

And that means a lot for some industries. A consumer deciding whether to take an umbrella may tolerate an imperfect hourly forecast. Many operational systems cannot. A delivery drone needs to know whether heavy precipitation, icing or dangerous wind will cross its route before it reaches the area. An airport may need to adjust runway operations as a thunderstorm develops nearby.

A solar operator needs to anticipate a fast-moving cloud front before generation suddenly falls. Logistics and mobility platforms can reroute vehicles before flooding or severe rain disrupts a road network.

The same principle applies to outdoor events, emergency services, agriculture and energy trading. These users do not simply need “tomorrow’s weather”. They need to know what is changing, where it is changing and whether there is still time to act.

AI weather models can generate forecasts much faster and with less computing power than many traditional numerical models. Some already outperform physics-based systems on particular variables and forecast horizons. But AI does not eliminate the observation problem.

Most global AI forecasting models still begin with an atmospheric analysis produced by traditional numerical weather prediction infrastructure. If that analysis is several hours old, fast inference cannot make the initial state current again.

AI models can also underestimate the intensity of rare events because extreme examples are underrepresented in their training data and because some training methods favor smoother, more statistically likely outcomes. Recent research has found that leading AI models can produce larger errors than physics-based systems when forecasting record-breaking heat, cold and wind events.

This is why the future is unlikely to be a simple contest between AI and physics. ECMWF already runs its operational AI forecasting system alongside its traditional physics-based model.

The strongest architecture will combine global models, high-frequency local observations, AI-based nowcasting, physical constraints and probabilistic estimates of uncertainty.

We've reviewed, rated, and ranked the best cloud backup platforms.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

• Researchers found banking malware requests VPN permissions to block Google Play and Play Protect on infected Android phones
• ToxicPanda 2.0 bypasses security to install hidden payloads, targeting 349 banking and crypto apps across 16 countries
• The malware can even seize shell-level control of a device

Security researchers have flagged a new twist in Android banking malware: a trojan that turns your phone's own VPN feature against you.

A report from mobile security firm Zimperium details how ToxicPanda 2.0 abuses VPN permissions to shut down Google's built-in protections before it strikes.

The tactic is effective because it hides in plain sight. Plenty of legitimate apps ask for VPN access, and even the best VPN apps rely on the same underlying permission to route your traffic. ToxicPanda copies that request, but uses the access to cut your phone off from Google Play instead.

Once Google Play Protect can no longer reach the device, the malware has a clear runway to install its payload and begin harvesting your data.

ToxicPanda operates as a "dropper," an app that smuggles in a second, hidden program. According to Zimperium researchers, the malware first shows a fake installation screen and prompts the victim to grant VPN permissions. That request looks routine, so many users tap "allow" without a second thought.

Granting it lets ToxicPanda create a local network interface that sits between the phone and the internet, giving the malware control over all traffic passing through the device. It immediately uses that control to block communication with Google Play and Google Play Services.

Cutting off this connection lets the malware interfere with app verifications, updates, and Play Protect, the security layer that would normally flag or remove a harmful app. With Google effectively blindfolded, ToxicPanda decrypts a payload hidden in its own files, installs it, and then asks for Accessibility Service permissions to dig deeper.

This release is a major escalation from the previous ToxicPanda iteration.

As Zimperium says, ToxicPanda 2.0 now supports 167 remote commands. It can also overlay fake login screens on 349 banking, e-wallet, and crypto apps across 16 countries, up from just 16 apps previously. A separate module harvests PINs from more than 140 financial apps using invisible overlays that capture your taps.

The trojan can also spoof your Android lock screen to steal your PIN, pattern, or password, and it abuses Android's Wireless Debugging (ADB) feature to gain shell-level access and grant itself permissions without prompts.

ToxicPanda first appeared in 2024, targeting European banks. Other similar malware such as Rokarolla, has also hit hundreds of banking and crypto apps in recent months.

The strongest defense is to keep the malware off your phone in the first place.

Only install apps from the official Google Play Store, and avoid sideloading APK files from links, ads, or third-party sites. Be aware that dangerous VPN links have even slipped into official app stores, so treat any surprise VPN prompt with suspicion.

A legitimate VPN remains a valuable privacy tool, but this campaign is a reminder that the permission itself is powerful, so grant it only to apps you genuinely trust.

It took 127 blood-splattered minutes for 2024’s outrageously violent flick Deadpool & Wolverine to reach a kill count of 178. In my two-hour preview of Insomniac’s Marvel’s Wolverine, I’m pretty sure it took even less time to match it.

The game starts with a delirious murder spree against the forces of human supremacist Bolivar Trask on the Asian island of Talenbang, who has taken Team X’s leader Nathaniel Essex captive.

I’m soon slicing arms off in brutal finishing moves and pummeling their insides into paste with barrage attacks. When I get to staggering foes into the blast radius of a nearby grenade, turning them into a mushy pile of goo, it’s abundantly clear this isn’t your friendly neighbourhood Wolverine.

It’s an impressively percussive, weighty combat system. With Wolverine’s adamantium skeleton in mind, I’d describe it as heavy metal combat.

At first, soldiers are summarily dispatched with a chain of light attacks and a heavy attack finisher. Later, in Canada, shield-bearing Reavers demand an initial kick before a five-knuckled skewer to the sternum. And the Hand, red-gowned ninjas recently spotted in Spider-Man: Brand New Day, come with longer blades of their own, sometimes left lodged in Wolverine’s body to impede him until it either falls out or he rips it out to furiously stab an enemy.

The environment plays an important role too, whether you’re knocking hapless enemies off a rooftop or slamming them into a wall to leave them vulnerable to a barrage attack sure to relieve them of their entrails. Those enemies often come prepared, with mutant power inhibitors locking Logan’s regeneration unless you target the machines first — either by stealth or all-out assault.

Wolverine’s slower speed meant I took a few more hits than I needed when I couldn’t cancel attacks to dodge an incoming blow or use a Hurricane move to clear my vicinity. But once I’d adjusted to the pace (and learned how crucial the leap attack is for staying in combat), I was stringing together brutal extended combos to the point that Devil May Cry’s style meter wouldn’t have gone amiss.

In my one-man blood donation campaign, I’m building up my rage meter. As it ascends through tiers, you first unlock the Last Stand recovery mechanic, next enhanced finishing moves, then, lastly, the ability to trigger all-out carnage.

All of a sudden, environments fade away into monochrome irrelevance. The bodies of enemies pulse bright and ripe for shredding. Instakills become available with one squeeze of the trigger buttons.

It’s a thrill of delirious violence, and delete-on-sight kills comparable to Ninja Gaiden 4’s immensely satisfying Berserk mechanic last year.

Jess Reiner-Reed, Senior Project Director, tells me that turning players into frenzied berserkers was always the plan, but balancing that aggression with Wolverine’s rapid healing ability required some mechanics to shapeshift in a manner worthy of Mystique.

“When people think of Wolverine, they say, ‘How are you going to make a video game about a guy that doesn't die?’ We had some versions that didn't work very well — versions where we were incentivizing the player to hide and have their health regenerate, like an FPS mechanic. But then we really landed on this version where we're incentivizing you to be aggressive.”

After years producing family-friendly fare in the Spider-Man series and Ratchet & Clank, the exact degree of this aggression and its gratuitous violence surprised many. Reiner-Reed gleefully reminds me that Insomniac isn’t new to this after the M-rated Resistance series.

“In Spider-Man, we always had to make sure Spider-Man doesn't kill people, because that's not Peter Parker; he webs guys to the side of the building. In Marvel's Wolverine, we get to kill people in the face.”

Focusing up close rather than tearing through Manhattan at 60mph meant Insomniac can lavish attention on details that would pass in a blur for Spider-Man. The game’s richly detailed levels look fantastic, with a focus on satisfyingly physics-driven destructible objects.

One moment in the demo particularly looked like it could have come straight out of Ratchet & Clank, and not for the better. I was pitted against the Proto Sentinel, a giant armoured robot boss rooted in place and limited to pummelling fists, laser-eyes and missile launchers. The somewhat formulaic boss battle was bested by dodge rolls and sprinting from clearly telegraphed blast zones. Despite Team X’s occasional interventions to enliven the fight, it felt like a dated chore compared to more recent similar battles like Clive’s fights against Final Fantasy XVI’s Eikons.

I found the Nightmare Doors sequences similarly deflating. These are portals into Logan’s Canadian wilderness cabin, where he unlocks forgotten memories by navigating a timed obstacle course, scrambling up tree trunks and leaping from eerily suspended shipping containers. The two examples I saw felt like perfunctory diversions compared to the far superior storyline chases.

Wolverine can throw hands with the best of them, but does the unrelenting focus on its excellent close quarters have legs? With no weapon mixups, ranged abilities, or a wide suite of movement abilities evident so far, the variety must come from elsewhere.

Reiner-Reed reassures me that variety will be found through build variations created through the player’s chosen special moves and custom augmentations to shape their build, whether they want to focus on stealth or accelerate into a raging rampage.

Moment to moment, it certainly has the X-Factor you’d expect from the X-Men’s headline act; I catch my pulse racing at the end of my demo, and the Team X team-up attacks with the gigantic Sabertooth and Mystique have me itching to see what punchy combinations await.

But in the space of two hours, I was craving more movement options and could already feel the leap attacks become a crutch.

Insomniac has far more to offer than just Deadpool & Wolverine’s glib humour and sadistic mayhem, as much as Reiner-Reed relishes team meetings to discuss blood physics. Off duty in a trademark flannel shirt and a cowboy hat, as Wolverine, I got to visit Lowtown for some R&R in the lawless slums of Madripoor.

The sequence provided ample opportunities for Liam McIntyre to show his range between macho posturing with Sabretooth and self-effacingly confronting his inability to find a suitable gift for Jean Grey (despite the advantages he gets from sniffing out her trail).

But it was a simple yet artfully composed game of Never Have I Ever at the bar with Mystique that impressed me most. Logan confronts his misdeeds and inner demons through the bottom of a glass. As I choose to drink and not drink at some of Mystique’s revealingly probing interrogations, I can safely say this: never have I ever been so struck by just waiting to see if Logan’s hesitant outstretched arm raised the cup to his lips.

“I love Logan as a character and always have because I found him one of the most emotionally complex characters,” says Liam McIntyre, the gravel-voiced Aussie actor who’s demonstrating Logan’s unfettered rage and nuanced regret in even these quieter moments.

“He doesn't want to go and high-five people in New York City; that's not his deal. He wants to stay off and be remote and do what he has to do and not make a big deal out of it [...] He feels the most grounded of the superheroes to me.

McIntyre does such a fine job making Wolverine a character you want to spend time with; I was almost disappointed when The Hand’s intrusion brought this lull to an abrupt close. Although crossing blades with them on an exhilarating rooftop chase that closed out the demo soon banished any momentary disappointment. The single-minded narrative momentum I saw here can remove any fears Spider-Man’s open world would be missed.

Both McIntyre and Reiner-Reed point to the game's ending as the moment they're most looking forward to seeing players’ reactions. If it can top what I saw in the first act’s (unfortunately embargoed) rollercoaster climax, I’m eager to see how Marvel’s Wolverine can top it when it arrives on PlayStation 5 on September 15.

• Anti-AI sentiment and data center opposition is rising, and communities are costing tech firms billions with their opposition
• Trump says communities opposed to data centers are "making a mistake", and calls fossil fuel burning generators "beautiful"
• Texas Gov. Greg Abbott says that AI data centers "basically dug their own grave", adding that "they got the backlash they deserve."

The US government has gone all in on AI, so it’s no surprise that President Trump doesn’t fully understand why so many Americans are opposed to their presence within their communities.

During an interview with his ally Michael Cohen, Trump said “Communities that don't take a data center, they're making a mistake,” adding that their presence adds “tremendous amounts of jobs and money” to communities.

Texas Governor Greg Abbott recently offered his own take on data centers, saying that “They basically dug their own grave for the problem that's been caused for them, and that's why they got the backlash they deserve.”

AI data centers are requesting an unprecedented amount of energy from US electrical grids, prompting Trump to not only lift clean air restrictions on data centers, but also push forward one of the biggest repeals of environmental protections.

As a result, data centers are being built with enormous gas turbine plants that release pollutants and cause health issues for local residents. In his interview, Trump said, “They're making their own power plants. They're building the most beautiful, you've never seen power plants like this.”

Across the political spectrum in the US, there has been a singular, clear message. AI is making people very concerned. Outside of the environment, the main concern comes in the form of job replacement. The majority of Americans in a recent poll said the technology will lead to fewer jobs.

Young people are especially distrustful of those running AI companies, again arguing that the technology would lead to fewer jobs. Almost three quarters of Americans believe that the pace of AI is moving too quickly - and they may have a point. Regulations and governance on AI are quickly falling behind.

Representatives of red and blue states have been banding together to push for more restrictions on AI development and moratoriums on data center construction to give the US time to put together a plan on how best to not only build out AI capacity, but also create an AI ecosystem that benefits all Americans, not just those at the top.

But how does Trump’s argument that data centers are creating “tremendous amounts of jobs and money” hold up under scrutiny?

The first thing to note is that while data centers do create short-term jobs while under construction, they only require a skeleton staff once operational. The companies hired for data center construction are often larger firms with the capability to handle large infrastructure projects who often aren’t local companies at all.

Where data centers do create jobs is further down the digital pipeline as the AI capacity is leveraged by existing industries. While some industries are seeing a rise in AI-related jobs creation, 2026 has been the worst year for AI layoffs with more than 176,000 people laid off for AI-related reasons.

On the money side of the equation, Meta has recently revealed that its 4,000-acre data center campus has led to local teachers receiving bonuses of as much as $50,000 due to tax revenues gained from the site.

But Meta received $3.3 billion in tax exemptions for the site. This is a common incentive to attract investment, but many states are quickly realizing the exemptions they have granted for AI data centers are quickly becoming unsustainable.

Ohio projected a revenue loss of $136 million when it granted tax exemptions for new data centers constructed in the state, but Ohio Governor Mike DeWine recently revealed that number has reached $1.6 billion. Now, states are quickly repealing or amending their tax exemptions.

Until the Trump administration recognizes that the American public wants AI to benefit everyone in a sustainable way, rather than just AI bosses and the super rich, they’ll likely continue facing widespread opposition to their AI agenda.

There's a long list of things to buy this school season, especially since my firstborn is off to kindergarten this year, and at the top of my list is Apple's best-selling AirTag.

• Shop more daily deals at Amazon

Apple's AirTag is a Bluetooth tracking device that works with your iPhone to locate an item that it's attached to. In my case, I'm attaching the AirTag to my child's backpack for peace of mind.

I bought the newly released AirTag for full price, but you can currently get the 2nd Generation Apple AirTag 4 Pack on sale for $79.99 (originally $99). That's a new record-low price, and each AirTag comes out to just $20, which is an absolute steal for what you get in return.

AirTag 4 Pack (2nd generation): was $99 now $79.99
You can get the best-selling Apple AirTag 4 Pack for $79.99, which is the lowest price we've ever seen. A great way to track a child's backpack or luggage, the AirTag can be attached to anything you don't want to lose, and your iPhone can locate it.View Deal

The 2026 Apple AirTag launched in January and is one of Amazon's top three best-selling gadgets. The AirTag isn't just for backpacks; you can attach the tracker to your keys, wallet, or luggage, or anything else you don't want to lose, and your iPhone will locate it. The 2nd Generation model offers a 50% longer Precision Finding range and a louder speaker with a higher-pitched chime. You can also now share your AirTag location with family and friends on shared items or airline partners for lost luggage.

Keep in mind, this is a limited-time offer and the lowest price we've ever seen for the 4 Pack Apple AirTag model. If you're not interested in a 4 Pack, you can grab a single Apple AirTag for $24.