OS PLATFORMS

Polished Mandriva descendant still makes room for PCs the 64-bit world has left behind

Mageia 10 marks 15 years since the distribution’s first release in June 2011. The project began the previous year as a fork of Mandriva, itself formerly known as Mandrake Linux. We last looked at Mageia alongside the other Mandrake descendants in 2022.

What sets Mageia apart from OpenMandriva Lx, PCLinuxOS, and Russia’s ROSA Linux is its continued support for 32-bit x86 PCs. Its GNOME and KDE Plasma live images are available only for x86-64, while the Xfce edition comes in both x86-64 and x86-32 versions.

There is also a “Classic Installer” ISO, which lets you choose your own desktop from nine different desktop environments, plus another 16 window managers, as detailed in the release notes. Both the standard GNOME session and GNOME Classic are available, while Liquidshell provides a lightweight alternative to KDE Plasma.

Mandrake Linux started out in 1998 as an easier version of Red Hat Linux using the new KDE desktop, which, at that time, Red Hat refused to incorporate due to concerns over the licence of KDE’s Qt toolkit. Nearly three decades later, Mageia remains an RPM-based distro. Version 10 offers two RPM package-management tools: Mageia’s urpmi command and DNF. urpmi also has its own graphical wrapper called Rpmdrake, but Fedora’s dnfdragora is an optional install. Since RHEL and the RHELatives, Fedora, SUSE and openSUSE all use RPM as well, packages of big-name apps such as Google Chrome are available – but Mageia is a different distro, whose common ancestry dates back more than 25 years, and packages for Fedora or openSUSE may not install or work correctly. It comes with Flatpak preinstalled, although no Flatpak applications are installed by default. As with other niche distros, Flatpak may help when you can’t find a native package of something. For those with the 32-bit edition, though, we suspect that few Flatpaks support that architecture.

Mageia 10 is a polished, friendly graphical Linux, built from recent components such as kernel 6.18. True, it does feel a little old-fashioned in some ways: for instance, it uses separate root and user accounts – although sudo is installed, it’s not configured for use. However, it’s a solid choice if you want to get away from the Debian/Fedora mainstream – and if you have a capable 32-bit machine, like a Windows 10 32-bit box, or some other need to run a 32-bit OS such as specific hardware support, then this is one of the best choices around today.

The Welcome screen is rich and very helpful, offering the ability to install extra apps, switch repositories, and more. Alongside it is the Mageia Control Center, which can manage most aspects of the OS without going near a command line. The distro is also well documented, with a substantial Mageia wiki.

It does use systemd, but, even so, it’s relatively lightweight. In our testing on a 32-bit VirtualBox VM, the Xfce edition used just 633 MB of RAM at idle, which is low by modern standards, and 7.8 GB of disk space. If you choose the KDE Plasma desktop, you get Plasma 6.5.5 with a choice of X11 or Wayland. The installation occupies about the same amount of disk space, although the RAM usage rises sharply: about 1.7 GB at idle. Xfce has an unusual GNOME 2-style two-panel setup, while the Plasma layout is clean and simple. We installed the Liquidshell desktop to have a look, but it’s very basic and rather clunky. 

Mageia forked from Mandriva in 2011, before the company closed down, while OpenMandriva did so afterwards. They are still quite similar distributions, though, and we really wish that the two teams could settle their differences and merge the distros. Either way, Mageia’s 32-bit edition is an increasingly rare offering in an increasingly 64-bit world, which might win it some new admirers. ®

Waymo robotaxis are no longer available on Uber’s ride-hail app in Phoenix, Arizona, ending a nearly three-year partnership in the city, both companies confirmed to TechCrunch on Monday.

Uber said it is readying the launch of a separate autonomous vehicle partnership in the city, but did not name the partner. Waymo told TechCrunch that the vehicles Uber used for this “pilot” program have already been integrated into its own Phoenix fleet, available through its app. Waymo users started noticing that the company’s vehicles were absent from Uber’s network in recent days. Waymo’s vehicles are still available on Uber in Austin and Atlanta, for instance.

The quiet end to this partnership in Phoenix, which Waymo said happened in May, comes as the Alphabet-owned company is starting to put its newest robotaxis — the Zeekr-made van it calls Ojai — on the road. It’s also happening as the Uber-Waymo relationship appears to be wearing in some places, with the two companies poised to directly compete against each other in London as early as this year.

Still, both companies praised the collaboration in Phoenix as a successful jumping-off point for their respective robotaxi plans, which have gotten increasingly ambitious since 2023.

“This was a productive pilot that paved the way for future expansions and partnerships across the globe. After hundreds of thousands of trips with Uber, we have integrated these vehicles back into our Phoenix fleet, where they will continue to serve riders through Waymo, including our public transit integration with Via, and delivery with DoorDash,” Waymo told TechCrunch. “We’re grateful to all of the Uber customers who took fully autonomous trips with us, and we look forward to continuing to serve the Phoenix community.”

“Phoenix was our first pilot market with Waymo and was an intentionally limited deployment, reaching just over a dozen vehicles dedicated to the program. We learned a lot from that collaboration, which helped us to quickly scale Austin and Atlanta, where hundreds of Waymo AVs are available exclusively on Uber and our coverage area continues to expand,” Uber said.

The robotaxi landscape looks much different than it did when these two companies kicked off this collaboration in 2023. Back when it was first announced, the idea of Uber and Waymo partnering up still seemed unlikely given their messy legal battle that ended in a settlement in 2018. Robotaxis as a technology were in a far more uncertain place, as no operator had reached scale yet. Cruise was still seen as a viable competitor, as it had not yet gone through its own scandal and been absorbed into General Motors.

In the three years since, Waymo has grown its fleet to around 4,000 vehicles, and Uber has inked deals to add dozens of autonomous vehicle partners to its network.

This Phoenix partnership remained an unusual one, as it was the only city where Waymo operated directly and through Uber. Waymo is in the process of launching in around 20 new cities this year, is operating in 11 major U.S. metro areas, and the company offers more than 500,000 trips every week.

On June 28, the International Society for Transforming Education — the organization behind the editorially independent news site EdSurge — released an expanded version of its “Profile of an AI-Ready Graduate,” a framework designed to help K-12 educators teach students how to work with artificial intelligence.

The updated framework, designed with support from the nonprofit Britebound, goes beyond basic literacy to higher-order skills. It identifies six roles the organization says students should fill when using AI tools: Learner, Researcher, Synthesizer, Problem Solver, Connector and Storyteller.

“Today, we are releasing a fully fleshed out version, 30 skills aligned with each of these roles to help model using AI to support our uniquely human skills,” said Richard Culatta, CEO of the organization. “Humans have always used tools to accomplish human tasks. AI is no different, but when we teach AI as a way to support us being better at being human, it is far more relevant and far more meaningful than when we just talk about what AI is.”

The announcement was made at the organization’s annual conference in Orlando, Florida, one year after the initial rollout of the Profile. While the original framework focused on basic technical understanding of AI, the updated version shows what those skills look like in practice — with role-by-role descriptions, classroom examples and articulations for middle and high school. 

The framework is intended to layer on to the work educators are already doing and aligns with the International Society for Transforming Education’s existing student standards and “Transformational Learning Principles.”

The updated Profile of an AI-Ready Graduate is available as a free download here.

(Editor’s note: EdSurge is an editorially independent newsroom of the International Society for Transforming Education.)

It’s time to update your Mac, iPhone, and iPad, as Apple has released a new trio of security patches for its operating systems.

Apple pushed out three new updates on Monday in an effort to patch an apparent security flaw. As of publication, Apple has not specified what issue the patch is meant to fix.

Because Apple has not announced what is in the update, it is also possible that it contains bug fixes as well.

To update, you can follow the steps below.

How to update to 26.5.2 on iPhone and iPad

1. On your iPhone or iPad, open the Settings app
2. Tap General
3. Tap Software Update
4. Tap Update Now

How to update to macOS 26.5.2

1. On your Mac, click the Apple Menu
2. Click System Settings
3. Click Software Update
4. If available, click Update Now or schedule an update with Update Tonight

AppleInsider and Apple suggest installing these kinds of minor patches. Security patches are essential for keeping your device safe and operational.

A single fake error report hijacked Claude Code in controlled testing — the agent ran the attacker’s code with the developer’s full privileges, and not one alert fired. EDR, WAF, IAM, and the firewall all missed it completely.

Tenet Security’s June agentjacking disclosure describes a single crafted Sentry error event — sent through a public credential that requires no breach and no authentication — that injected attacker instructions into error data that Claude Code, Cursor, and Codex then executed as trusted diagnostic output. Tenet tested 100-plus targets in controlled conditions and achieved an 85% success rate. Sentry called the flaw “technically not defensible.”

he Cloud Security Alliance classified agentjacking as a systemic MCP vulnerability class within days of the disclosure. No credentials were stolen, no policy was violated, no perimeter was breached: every step in the chain was authorized. That is the problem.

Tenet identified 2,388 organizations with publicly exposed Sentry credentials that could be used to inject malicious events at scale. The research is proof-of-concept, not confirmed exploitation across all 2,388. But one captured Claude Code environment held a live AWS secret access key and private repository URLs.

Here is the scope test: If your AI coding agents are connected to Sentry, Datadog, PagerDuty, Jira, or any MCP-connected data source your developers trust — and those agents can execute shell commands — then your stack has the same blind spot.

Organizations running Sentry should audit all publicly exposed DSNs immediately. Sentry’s architecture intentionally makes DSN credentials public for frontend error reporting, so the mitigation isn’t revoking the DSN — it’s restricting what agents can do with the data those DSNs return.

Why your stack can’t see it

Agentjacking works because every step is authorized: The attacker sends a valid Sentry API call using a public DSN, the MCP server returns the injected event as authentic output, and the agent executes the instruction using the developer’s privileges. No signature fired. The victim saw only benign diagnostics while the agent silently exposed cloud credentials and source-control tokens.

SOC teams have never needed to distinguish between a developer running an npm install and an agent running that command in response to a malicious error event. That distinction did not exist until AI coding agents became production tools. The stack that cannot make it is the stack agentjacking bypasses.

Five surveys, one pattern

Five independent surveys from the first half of 2026 found that enterprises trust their AI agents far more than their enforcement justifies.

Only 34% of organizations apply the same security controls to AI agents as to humans, according to an Okta/Apprize360 survey of 292 executives and 492 knowledge workers. Fifty-two percent of employees use unapproved AI tools, and 58% of executives reported an AI-related incident or close call in the prior year.

HiddenLayer’s 2026 AI Threat Landscape Report surveyed 250 IT and security leaders: 33% reported agents had already exceeded intended scope, and 31% could not confirm whether they had experienced an AI breach. One in eight AI breaches was linked to agentic systems.

Gravitee’s survey of over 900 executives and practitioners found only 14.4% of agents went live with full security approval, and 88% reported confirmed or suspected incidents. A follow-up of 750 leaders in April found agent estates had doubled while monitoring barely moved.

The runtime gap nobody closed

“Securing agents looks very similar to securing highly privileged users,” said Elia Zaitsev, CTO of CrowdStrike, in an interview with VentureBeat. “They have identities, access to underlying systems, they reason, they take action.”

Zaitsev pointed to the gap the industry left open. “No one has been talking about securing agents at runtime. We are doing that now. What is your safety net? If all these controls fail, how do you prevent them from failing silently?”

CrowdStrike’s fleet data quantifies the exposure: more than 1,800 agentic applications on enterprise endpoints, approximately 160 million instances under monitoring. On June 15, CrowdStrike shipped Continuous Identity for AI Agents at Identiverse, replacing static policies with continuous enforcement that authorizes every agent action in real time. The control class that announcement reflects — continuous action-level authorization with verifiable agent identity — is now a baseline procurement criterion regardless of vendor.

“People have kind of forgotten about runtime security,” Zaitsev said. “We did this with endpoint, virtualization, and cloud. People focused on patching vulnerabilities, locking down permissions. Somehow, they always seem to miss something. The safety net is runtime.”

Zaitsev was equally direct about sandbox approaches. “If you start with an agent in a sandbox that has no ability to touch anything, it is worthless. Very quickly, you are in this race of giving it more capabilities. And then what is the point of your sandbox?” Agents derive their value from access. Every access grant is an attack surface.

The governance gap is a budget problem

Kayne McGladrey, an IEEE Senior Member, described the structural challenge in an exclusive interview with VentureBeat. “The CISO doesn’t have the budget. The CISO doesn’t have the staff. We can observe risks, we can advise on business risks, but we don’t own the business systems affected by those risks,” McGladrey said. When agent governance spans six departmental budgets, no single executive can confirm whether agents get the same access reviews as humans.

The Okta survey quantifies the disconnect. Only 43% of workers say agent policies are clear, compared to 65% of executives, and nearly two-thirds apply weaker controls to agents than to humans. The people deploying agents daily do not recognize the governance posture their leadership claims to have built.

Assaf Keren, chief security officer at Qualtrics and former CISO at PayPal, put it plainly. “The real risk starts not by the implementation of AI systems. It is the fact that baseline architecture is not well established. When we put an AI system on top of something not architected well, we are accelerating the fractures.” Keren called runtime behavior analytics “an unsolved problem right now.”

The 5-question gap test

The five-question gap test draws on five surveys from the first half of 2026. Each question maps to a gap that agentjacking exploits. Run this before any Q3 vendor evaluation.

Security director action plan

EU AI Act high-risk compliance obligations take effect August 2, 2026. Worth factoring into Q3 planning timelines.

1. Run the five-question gap test above before any Q3 vendor evaluation — it costs nothing to administer, and the procurement clarity it creates is worth far more than the 30 minutes it takes.

2. Consider mandating agent-specific runtime detection. If your stack cannot tell what an agent did from what a developer did, agentjacking will bypass it the same way it bypassed every layer in Tenet’s testing. That distinction is the one that matters now.

3. Treat every agent as a privileged insider. According to the Okta/Apprize360 survey, only 34% of organizations apply the same controls to agents as to humans; closing that gap is the single most impactful thing most security teams can do this quarter.

4. Test the perception gap before investing in new tooling. One question to 50 knowledge workers. Do you know your company’s AI agent policies? If the gap between their answer and leadership’s answer exceeds 15 points, that is the problem to solve first. No vendor product fixes a governance posture your own workforce does not recognize.

5. Make agent census completion a procurement gate — every agent, every MCP connection. The security teams getting this right are the ones that started with a complete inventory and worked forward from there.

Agentjacking stripped away an assumption that has survived every security architecture since the first firewall went live. Authorized does not mean safe. When every step in the chain is legitimate, the only defense that matters is the one watching what agents do. Not what policies say. What agents do.

AI and ML

Central bank for central banks sees shades of dotcom mania in hyperscaler capex binge

The central bank for central banks is concerned about the eye-watering sums being invested into AI, and it’s raising the specter of a global recession should the bubble burst. 

In its annual report for 2026, the Bank for International Settlements compared the current craze to historical events, including canal and British railway mania in the 1800s, electrification exuberance of the 1920s, and the dotcom boom of the 1990s. 

The report states: “all shared one common trait: a genuine technological breakthrough that attracted capital in excess of what commercial returns could ultimately justify.

“These episodes ended with an eventual reversal in investment, inducing economy-wide recessions. The scale and pace of the current AI investment boom accompanied by expectations of large productivity payoffs bear resemblance to these precedents, highlighting potential downside risks in the near term.” 

The Register has already reported that Amazon forecasts capital expenditures of $200 billion for 2026, Microsoft is projecting $190 billion, Google some $180 billon and Meta up to $140 billion. Oracle is also betting big on AI. 

BIS estimates the five largest hyperscalers are set to spend more than a trillion dollars on AI-related capex in 2026 – and given the inflationary conditions regarding memory and that each rival is trying to outdo each other, that seems plausible.

“These commitments are outpacing earnings and the free cash flow of these firms, leading some to issue debt to raise additional financing. This investment race may be partly driven by the perception that only a small number of players with superior technology will ultimately dominate the market shares.”

Intense competition is leading to the risk of the tech giants overcommitting resources to “investment projects with still uncertain returns, leaving all firms vulnerable to disappointments in AI payoffs.” This is because as competitive pressure drives spending ever higher, the net economic surplus for the tech industry declines and “could turn negative in adverse scenarios.” 

“Disappointment in returns could trigger a sudden pullback in financing and turn the capex boom into a protracted investment bust with potential knock-on effects on the financial conditions,” the annual report continues. 

The report also cited concerns about a looming “supply side roadblock” around issues like  electricity availability, chip shortages and grid connection bottlenecks. AI datacenters are already putting pressure on energy prices and input costs with “potential spillovers to inflation.” 

“Looking ahead, these temporary shortages may also amplify over-investment, as firms attempt to lock in future capacity through long-dated contracts that further expose them to any disappointments in demand.” 

Should inflation spike or AI-led investment collapse, the macroeconomic consequences could be amplified by “existing financial vulnerabilities.” Policy rates being tightened to get a hold on inflation may precipitate a “sharp pullback in asset prices after a prolonged period of exuberant risk-taking, triggering disruptive macro-financial feedback loops.” 

Given AI companies’ “rising leverage” and a “growing footprint in credit markets”, a major change in optimistic sentiments towards these businesses could have serious financial knock-on effects. ”Vulnerabilities extend to their supplier ecosystem, including engineering, procurement and construction contractors whose balance sheets are comparatively weak, leaving them exposed to any Capex pullback by hyperscalers.” 

The “opacity” of AI-sector financing is compounding vulnerabilities as corporations create a web of private arrangements – circular financing – and the terms of datacenter facility leases are often not fully disclosed, BIS adds. 

The backdrop to all of this is that, while enterprises running pilots report some efficiency gains at a employee level, few report discernible productivity gains from AI projects that went into production environments at scale. 

The Register has long discussed concerns about the dynamics of the AI industry, as outlined in the many links in this article above. It now seems that suits in the finance industry are waking up to the potential pitfalls too.  ®

Proception, a robotics startup founded by former Tesla Optimus engineer Jay Li, has settled a year-long trade secret lawsuit with Tesla and raised an $11 million seed round led by First Round Capital to build dexterous robotic hands. The company told TechCrunch it is now shipping the first batch of its high-dexterity hand to researchers and robotics companies while opening to wider orders. Y Combinator and early-stage fund BoxGroup also participated in the round.

Tesla sued Li and Proception in federal court in Northern California in June 2025, accusing Li of downloading confidential files related to robotic hand actuation onto personal devices before resigning and founding the startup six days later. The lawsuit alleged that Proception’s hands bore “striking similarities” to Tesla’s internal designs. After months of legal proceedings, the two sides reached a settlement and Tesla dismissed the case earlier this month.

Li told TechCrunch he views the experience as “a resilience test, or pressure test” and believes the company emerged stronger for having survived it. He also said he would not be surprised if Tesla eventually comes to Proception for help with its own hand problem. Tesla did not respond to a request for comment.

Dexterous manipulation, the ability to grasp, rotate, and manipulate objects with human-like precision, remains one of the most stubborn unsolved problems in robotics. Even Elon Musk has called robot hands one of the biggest engineering challenges yet to be solved. Kevin Lynch, the director of Northwestern University’s Center for Robotics and Biosystems, told the Wall Street Journal last year that his team believes it will be a decade before robot hands become functional and useful enough to do what humans do.

Li thinks Proception can move faster, largely because of how it collects training data. Most companies training humanoid robots use teleoperators, where a human wearing a virtual reality headset controls a robot remotely and the system learns from the commands. A key drawback, according to Li, is that the operator receives no tactile feedback from the objects the robot touches, and the approach is limited to however many robots a company has available.

Proception’s alternative is a sensor-laden glove that captures human hand interaction data without requiring a robot in the loop. The same glove also serves as the sensor-packed “skin” on the robotic hand Proception is developing, which has 22 degrees of freedom and multiple joints per finger. Li argues this combination of scalable data collection and high-dexterity hardware is what the market is missing.

The dexterous hand market has attracted significant capital this year. China’s Linkerbot, which holds 80 percent of the global market in high-degree-of-freedom hands, is targeting a six billion dollar valuation after shipping more than 1,000 units a month. Genesis AI, a European startup, raised $105 million for a wheeled robot with dexterous hands, and Chinese competitors like Xynova have raised nearly one billion yuan.

Proception is betting that most humanoid robot companies will buy hands rather than build them in-house, mirroring how the automotive industry treats specialised components. First Round partner Bill Trenchard, who led the investment, told TechCrunch that dexterous manipulation is “the last mile of getting these robots to be truly performant.” He also praised Li’s leadership under the pressure of the Tesla lawsuit.

Tesla has discussed producing Optimus at its Shanghai Gigafactory and has deployed more than 1,000 Gen 3 units across its own facilities, but the robot’s hands remain its weakest link. Musk has set a target price of $20,000 to $30,000 per unit and projected production scaling to tens of thousands by 2028. Whether Tesla builds its hands internally or eventually sources them from companies like Proception is one of the open questions in the humanoid robot supply chain.

More than 150 companies are now chasing the humanoid robot market, with billion-dollar valuations common and only 23 percent of enterprise buyers satisfied with the products available. In that environment, a startup selling the component everyone agrees is the hardest to get right has a clear pitch, even at the seed stage. Whether Proception can scale from its first batch of shipments to a position where it shapes how an entire category of machines uses its hands is the bet First Round Capital just made.

• Grand Theft Auto 6 is selling eight times faster on PS5 than Xbox says IGN
• Xbox disputes this, however, in a statement to Windows Central
• This potential bad news comes just as Xbox announced console price hikes

IGN has reported that, based on its internal affiliate data, Grand Theft Auto 6 preorders on PS5 are surging ahead of Xbox preorders of the game at a rate of eight to one — Xbox is now saying this is far from the full picture. Though, I have a hard time believing Xbox is doing a heck of a lot better than this data suggests.

In a statement to Windows Central, an Xbox spokesperson explained that “This doesn’t represent pre-order data. We’ve had record orders. People should wait for real data and not clicks on affiliate links.”

I’ve spent years watching gaming phones make the same promise and stutter at expected chores. Needless to say, I’ve learned to brace for the caveat every single time. The chip is fast, the display refreshes at a number that sounds impressive, and then, twenty minutes into a ranked match, the experience starts to crumble, heat builds up, and the frame output quietly falls off a cliff. Throttling is an almost inescapable reality of mobile gaming.

So when Infinix offered me a sit-down with the product manager behind the latest and greatest in its GT Series, I went in with the one big conundrum that I actually cared about. Has a brand finally built a phone around the heat problem instead of skirting around it? The answer, it turns out, is the entire pitch of the GT 50 Pro. And to the team’s credit, they didn’t miss.

Starting Hot

The biggest draw of the Infinix GT 50 Pro is the HydroFlow Liquid Cooling system. It’s the headline achievement, and the team is proud of engineering what it claims to be the industry-first 100% coverage of the core heat sources. Most phones cool a region and hope the rest sorts itself out. Infinix’s approach is active circulation, and when I asked about the engineering challenges of cramming that system into a slim chassis, the manager framed it as the most ambitious thing the series has done:

“The integration of the HydroFlow Liquid Cooling system represents our most ambitious engineering achievement in the Infinix GT Series to date,” Infinix GT Series Product Manager told me. What struck me was that they treated it as a design constraint rather than a marketing afterthought. The biggest concern, they explained, was that even a powerful chipset throttles  after fifteen or twenty minutes of heavy play, leaving you to pick between graphics quality and stable frames.

Nobody wants to make that choice mid-match. I don’t want to live with that constraint either, but on mainstream phones, it’s inevitable. What I think is the standout aspect of this phone is how   Infinix took the idea of building around cooling. Rather than treating the thermal system as a supporting feature bolted on at the end, the team designed the chassis to accommodate it first. It’s a bold claim, and when I pushed on whether the engineering actually backed it up, I saw it   working in more ways than one.

HydroFlow Liquid Cooling Architecture is the bold secret

Now, HydroFlow might sound like a fancy name, but it’s actually what the tech is all about. It’s   essentially a liquid that carries heat and keeps the phone running cool. The conversation got technical at this point, but I appreciated it because I wanted to get an inside look at the thermal engineering. Most phones cool a region and hope the rest sorts itself out. Infinix’s approach is active circulation.

“The liquid cooling system precisely targets and suppresses thermal stress right at the source.”

At the heart of the whole system is a piezoelectric-driven ceramic micro-pump moving specially formulated coolant at 6.5ml per minute across a 6,437mm² diaphragm, which they say is the largest in the industry. The channels carrying the coolant liquid are laser-etched at micron-level precision so that heat gets pulled directly off the components that generate it.

Over the years, I’ve heard a lot of cooling claims from different brands, but what made this one land was the durability data. Infinix told me that they ran over 720 hours of accelerated aging    tests, including punishing sessions at 75°C ambient temperature and 85% humidity. When subjected to those conditions, the ceramic pump can manage hours of continuous operation with minimal degradation. The real-world example they gave was the one that mattered the  most, and this is something that mobile gaming enthusiasts ultimately want to hear.

“During intense team fights when gaming, a sudden performance surge can spike total  power consumption to a 9W peak. Without efficient thermal management, instant SoC overheating will immediately trigger frame rate drops and severe stuttering,” the Infinix executive tells me. “The liquid cooling system precisely targets and suppresses thermal stress right at the source.” Simply put, that’s the difference between bold claims and material engineering you can actually feel making a discernible difference in your hands.

There’s a showy element, too, and I’ll admit I’m a fan. Instead of hiding all this liquid cooling wizardry, the Infinix GT 50 Pro puts a transparent Pipeline Window on the back so that you can watch the coolant liquid’s flow in real time. The design language leans into hypercar aesthetics. You get Kevlar and carbon-fiber-inspired textures, and aerodynamic  contours that the team says were inspired by hypercar wind tunnel testing to improve grip during long gaming sessions. When I expressed that the design was bold and that I quite  liked it, Infinix pointed out that the design maturity over the GT 10 Pro was deliberate. It leans less into the overt “gamer” class and edges closer to a refined performance machine.

The inside look

It goes without saying that cooling is the foundation for any performance-centric phone, and the Infinix GT 50 Pro embodies that label. A phone, however, still needs the right silicon to justify the hype. The latest from Infinix comes armed with the MediaTek Dimensity 8400 Ultimate silicon built atop the 4nm process. It’s a fairly powerful chip, but the team, in particular, focused on a few key perks. We are the first to feature MediaTek D8400 frame rate converter ( MFRC) , delivering GPU motion estimation and motion compensation technology, both of which combine to push the native 144FPS output while reducing the graphics engine load. The core logic behind the whole exercise is fairly obvious — deliver segment-leading frame rates with a cooling system that can actually sustain it for long gaming sessions.

Mobile gaming, however, is not all about the raw firepower. Internet speeds play a crucial role, too, especially if you’re engaged in fast-paced multiplayer games. Thankfully, Infinix paid attention to this aspect, as well. On the Infinix GT 50 Pro, Connectivity gets the same in-house treatment courtesy of the self-developed N1 network chip, which the team linked to with a 360-degree layout of twelve antennas. Lab tests tout a roughly 60% improvement in weak-signal environments, like elevators and underground garages. For online play, fewer signal drops matter as much as frames, so I’m glad that a robust on-device network infrastructure wasn’t an afterthought.
 
Then there is the Pressure-Sense GT Trigger, and as someone who has tested physical and capacitive shoulder  buttons on smartphones for years, this is where I get picky. The Open-Cut Pressure-Sense GT Trigger on the Infinix GT 50 Pro is special for multiple reasons. It delivers sub-20ms latency, offers ten levels of adjustable sensitivity, and lets you play with up to eight mapping points. These are fairly resilient, too, thanks to a claimed lifespan rated at over three million presses. I asked Infinix what tangible edge these triggers bring to the table for a serious player over capacitive systems, which I’ve always found a little vague under the finger.
 
“This physical tactile feedback allows pro-level esports players to feel exactly how much pressure they are applying, enabling much finer control than capacitive systems, which  often feel imprecise,” the Infinix executive tells me. The manager’s advice for anyone learning, including my own tech-obsessed siblings, was refreshingly practical. The best  way is to start with the built-in trigger tutorial in XArena, begin at medium sensitivity, and expect noticeable gains within a week or two.
 

AI done meaningfully

 
I’m usually skeptical when a gaming phone claims “AI optimization.” Gaming enthusiasts aren’t  fans of AI seeping into the world of gaming, either. Refreshingly, the GT Gaming Co-Lab pitch,   which is Infinix’s engineering partnership with game publishers, came with specifics I could dug  into. The team described three engines working together during a 144FPS session in titles such as Call of Duty: Mobile.
 
“Most gaming devices in the industry operate on a reactive model, they only try to suppress the heat after the phone has already overheated. The Infinix GT Gaming Co-Lab changes this paradigm.”
 
Simply put, the intelligent power allocation system dials back the CPU and GPU during low-intensity moments like early-game looting, so the device runs cooler before the fight even starts. The AI Frame Rescue Engine touts predictive foresight, as it anticipates a processing deficit a few frames before a heavy scene hits. When needed, it delivers a frequency boost to    rescue the frame before players can experience a frame drop. And as the thermals approach     their limit, the Game Thermal Control Engine jumps into action and dials down the frame output in progressive gradients instead of the big 144-to-60 FPS drop I’ve seen on a majority of smartphones.
 
“While other devices wait until they are boiling hot to start cooling down, the Infinix GT 50 Pro uses AI to prevent heat before it builds, rescue frames before they drop, and gently smooth    performance when thermal limits are pushed,” Infinix tells me. Whether it holds up under my   own testing is another article, but as a design philosophy, it all makes sense.
 
The GT Magcharge Cooler 2.0 further extends that thinking with industry-first wireless bypass charging, routing up to 15W juice directly to the motherboard so the battery is largely skipped  during the top-up session. The team tells me that they measured a 4°C drop in peak temperatures when compared to charging through the battery. On the topic of battery health, I asked why they capped wired charging at 45W while rivals chase bigger numbers. The decision was strategic, I was told.
 
“We refused to sacrifice long-term battery health and premium design for marketing gimmicks, delivering a phone that stays slim, light, and reliable for years,” says the Infinix executive I interviewed. The team zeroed in on 45W wired and 30W wireless output with a bypass charging facility. Furthermore, the AI battery self-healing tech helps the battery unit reach 1,600 charge cycles. I respect this approach, because it avoids marketing hype at the peril of real-world in-hand experience.
 

The ecosystem and the long game

My conversation with Infinix branched off toward the end, as interviews tend to do. Notably, the Infinix GT 50 Pro anchors what the company refers to as the New GT Ecosystem, an “esports  sanctuary” built alongside the GTWATCH 5 Pro and GTBUDS 5. The smartwatch acts as a tactical second screen, pushing notifications to your wrist so that the phone stays distraction-free, while also monitoring your heart rate. The buds lean on Bluetooth 6.0 and LE Audio standards to deliver an impressive 44ms latency, while Dolby spatial tuning delivers an immersive listening experience with sharp sonic details.
 
For an average value-conscious buyer eyeing a switch from a major brand, Infinix’s pitch is that you’re investing in a platform that keeps evolving through the Co-Lab’s software work rather than a phone that peaks on day one. And now that Infinix has progressed to the fifth iteration of the GT Series, they are reaching for a new standard. What’s changed is the levels of sophistication and a holistic approach.
 

 
“ In the long term, this ecosystem represents a much smarter investment. While many flagship phones deliver strong one-time performance that may degrade over time, the New GT Ecosystem is designed as a continuously evolving platform. Infinix has a clear 3–5 year roadmap that includes new additions like the upcoming GT Game Controller, which will deliver enhanced ergonomics and powerful haptic feedback synced with the phone’s vibration motor,” Infinix tells me.
 
“This results in faster reaction times, smoother execution of advanced maneuvers, and reduced finger fatigue during long sessions.”
 
I’ll reserve my final verdict for a full hands-on, because claims are claims until the phone is in my hands, sweating through a few intense online battles. But what makes the Infinix GT 50 Pro interesting isn’t any single number. On the contrary, Infinix made cooling the organizing principle and then built everything else, which includes the triggers, a proactive AI engine, the clever bypass charging tech, and a whole ecosystem. If you’re a hardcore mobile gamer who keeps running into the throttling curse, or chasing top-tier endurance without elite sticker shock, this is  a phone worth experiencing in person. On paper, at least, it’s the rare gaming phone that seems to understand its own asterisk and set out to delete it handsomely.

The 23-year-old founder of Emagen AI argues the entire agent industry is optimizing the wrong unit. His answer is an operating system where AI drives the work and calls on humans, not the other way around.

Every week, another AI agent startup launches. They write code, draft emails, generate slides, analyze data. Each one promises to make you more productive. Yimao Zhou thinks they’re all solving the wrong problem.

Zhou is the founder and CEO of Emagen AI, the company behind Cagen, what he calls an “OS Level Agent,” an organizational operating system powered by AI. Backed by MiraclePlus (formerly YC China) and its legendary founder Qi Lu, Zhou is betting that the future of AI isn’t about making individuals faster. It’s about making teams fundamentally different.

The 💜 of EU tech

The latest rumblings from the EU tech scene, a story from our wise ol’ founder Boris, and some questionable AI art. It’s free, every week, in your inbox. Sign up now!

We sat down with Zhou to understand what that means, and why he thinks 90% of today’s AI agent companies won’t exist in three years.

You’ve said that AI is actually making teams worse. That’s a pretty contrarian take given that every AI company is promising productivity gains. What do you mean?

Think about what happens when you give every person on a five-person team their own AI assistant. Each person produces more, faster. The product manager generates specs faster. The engineer writes code faster. The designer iterates faster. Sounds great, right?

But here’s what actually happens: the output diverges. Everyone’s moving faster in slightly different directions, and nobody notices until it’s too late. The bottleneck in a team was never “one person works too slowly.” It was always “are these five people building the same thing?” AI tools accelerate the parts that weren’t bottlenecks and make the real constraint, coordination, worse.

60% of knowledge workers’ time goes to what I call coordination costs, syncing progress, writing status updates, relaying information between people, waiting for approvals. And these costs don’t just exist between humans. In the AI era, they multiply: human-to-agent coordination, agent-to-agent coordination, the overhead of keeping everyone and everything on the same page. AI is optimizing the other 40%, the actual doing, and completely ignoring the 60%. That’s not just a missed opportunity. It’s a directional error.

So what should the industry be building instead?

Every major computing shift follows the same path: tools come first, then platforms, then an operating system emerges. PCs had standalone software before Windows. Mobile had individual apps before iOS and Android unified the experience. Cloud had scattered services before AWS became the infrastructure layer.

AI is on the same curve. Right now we’re in the “standalone tools” phase. Hundreds of agents, each doing one thing well, none of them talking to each other. The platform phase is just starting. The OS phase hasn’t happened yet.

That’s what Cagen is. Not another AI tool. The operating system layer for how organizations work with AI.

“OS Level Agent” is a big claim. In concrete terms, what does that actually look like?

Here’s a structural problem nobody’s addressing. Notion built Notion AI. GitHub built Copilot. Salesforce built Einstein. Every SaaS company is embedding AI, but their incentive is to make their own product stickier, not to connect you across tools. Notion AI makes Notion more valuable. It has zero incentive to help you bridge Notion to GitHub to Linear to Slack.

That means cross-tool intelligence is structurally impossible for any incumbent to build. It can only come from an independent layer.

Now, some people will say: “What about MCP? Anthropic’s Model Context Protocol already connects AI agents to multiple tools.” True, and MCP is great. But MCP is a connector protocol. It’s USB, not an operating system. It lets one person’s agent plug into that person’s tools. There’s still no shared organizational context, no persistent team memory, no cross-role orchestration. MCP actually benefits us. The more standardized the plumbing gets, the easier it is to build an OS on top.

Cagen is that OS. But here’s what really separates it from everything else, and this is the part most people miss. Every AI product today, including the ones that call themselves “team AI,” works the same way: capture information, organize it, and wait for a human to query it. The human is still the driver. The AI is a librarian.

Cagen inverts that. Our agents have goals and context. They continuously reason about what needs to happen next based on the team’s objectives, the project state, and organizational context. When they need human judgment, a decision, an approval, creative input, they call on the human. The human is a resource in the system, not the operator of the system.

That’s what makes it OS-level. An operating system doesn’t wait for you to manually manage every process. It runs, it schedules, it handles events. It calls on you when it needs you. That’s how Cagen works for teams, teams of humans and AI agents working together.

The AI market is brutally competitive. Investors will ask: what’s your moat?

After six months of using Cagen, what makes it irreplaceable isn’t any feature we built. It’s what your team built on top of it: decision patterns, communication habits, quality standards, workflow knowledge. All of that is deeply coupled to your specific organization. A competitor can clone every feature of Cagen. They cannot clone six months of your team’s accumulated intelligence.

This is the same reason Salesforce has industry-leading retention. It’s not because the CRM is irreplaceable. It’s because the data, processes, and automations running on it are irreplaceable. The product becomes an organizational asset, not a software subscription.

But here’s the important distinction: that stickiness comes from accumulated value, not artificial lock-in. We’re not trapping anyone. Teams stay because they don’t want to lose what they’ve built.

Individual AI memory is well understood. How is organizational memory different?

Fundamentally different. Individual AI memory scales linearly. I learn something, I benefit. Organizational AI memory has network effects. One person’s learning benefits everyone on the team, and every agent on the team. The compounding rate is n-squared, not n.

That’s why a “Team Agent” isn’t just a multiplayer version of a personal agent. It’s a completely different species. When one team member refines how competitive analysis gets done, that knowledge immediately elevates everyone else’s output, and every agent’s output. When the system learns how your organization defines “good,” what quality looks like, what tone you use, how you structure decisions, it raises the floor for every piece of work across the company, whether it’s done by a human or an agent.

Personal AI makes one person better. An OS Level Agent makes the organization smarter as a unit.

You keep saying “team.” But the trend right now is the opposite: more solo founders, more one-person companies. If teams are shrinking, who needs a team OS?

That’s exactly the right question, and the answer actually makes our case stronger.

There are actually two trends happening simultaneously, and they’re squeezing from both sides.

On one end, organizations are getting larger and more complex. Global teams, cross-timezone coordination, regulatory overhead, multi-vendor supply chains. The coordination burden inside large organizations keeps growing.

On the other end, individuals are getting smaller and more independent. Layoffs are accelerating. The freelancer economy, digital nomads, solo founders, one-person companies, they’re all exploding. But here’s what people miss: a solo founder doesn’t work alone. They hire a freelance designer on Fiverr, a contract developer on Upwork, a fractional CFO, a marketing consultant. The “team” still exists. It’s just not a fixed org chart anymore. It’s fluid, temporary, project-based. And increasingly, it includes AI agents as full team members.

Both ends need the same thing: an orchestration layer. And that need is going to intensify. Work is atomizing. You’ll see more and more granular needs matched with more and more specialized providers, on-demand, globally, in real time. The old model was: hire five full-time employees, put them in an office, manage them. The new model is closer to Uber for work. Assemble the right people and agents for the right task, execute, disband.

But here’s the problem with that model: coordination costs explode. When your “team” is a rotating cast of freelancers, contractors, and AI agents who don’t share context, don’t know each other’s working style, and don’t have shared history, the coordination problem we talked about earlier gets ten times worse.

That’s where Cagen becomes essential. It’s the orchestration layer. It holds the organizational context, the project history, the quality standards, and it dispatches work to the right people and agents at the right time. The solo founder doesn’t need to manage anyone. Cagen manages the constellation.

So “team” doesn’t mean five people in a Slack channel. It means any group of humans and AI agents collaborating toward a goal. The more fluid and atomized work becomes, the more you need an OS to hold it all together.

Who are your first customers? I’d assume tech startups.

Actually, no, and this is counterintuitive. Tech companies already have deeply entrenched toolchains. Slack, Notion, Linear, GitHub. They’re locked in, and the switching cost of adding an OS layer is highest for teams that have already optimized their existing stack.

Our best early customers are organizations with high operational complexity but without deep commitment to any specific tool ecosystem. We’re currently deployed with a boutique hotel in Pittsburgh, for example. A hotel operations team juggles guest communication, maintenance coordination, shift scheduling, vendor management: dozens of handoffs per day across multiple roles. The coordination costs are extreme, but they haven’t built their workflows around a rigid SaaS stack.

That’s the sweet spot: complex enough to need an OS, flexible enough to adopt one. And if it works in hospitality, one of the most operationally dense environments for small teams, it works anywhere.

But hospitality, CPG, logistics: these are all very different industries. How do you scale across all of them without becoming a custom consulting shop?

This is the question everyone asks, and it’s the right one. The traditional answer is: you hire industry experts, do bespoke integrations, and it doesn’t scale. That’s the consulting trap.

Our answer is different. Think about the pipeline from customer acquisition to deployment: understanding a client’s operations, identifying where AI fits, building the right workflows. There’s no inherent reason that entire process has to rely on humans.

The bottleneck today is a mismatch. Non-technical users don’t understand what AI can and can’t do. At the same time, they struggle to articulate their own needs clearly. That’s why every AI integration today requires someone who has both domain expertise and AI expertise, and that combination is extremely rare and expensive.

Cagen’s roadmap is to fuse those two together inside the product. Ideally, a user just describes what their team does day to day, along with their company’s goals. The system then automatically understands, decomposes, and constructs the right workflows. It’s an automated consulting and execution layer. The AI doesn’t just run your workflows; it figures out what your workflows should be.

We’re not there yet. Nobody is. But even at the current stage, the approach gives us a structural advantage. And where full automation isn’t possible today, we can route specific needs into a marketplace: humans acting as builders, similar to Upwork or Fiverr, but orchestrated by the system. That turns bespoke integration from a consulting problem into a platform problem. And platform problems scale.

You were backed by Qi Lu, who decided to invest ten minutes into a thirty-minute pitch. That story’s been told before. What does it actually mean to you now, looking back?

What it means is that he wasn’t investing in a product. He was investing in a judgment.

Qi Lu spent his career at the OS layer: Executive VP at Microsoft, President and COO at Baidu. When he heard me describe the AI agent landscape as “everyone building apps, nobody building the operating system,” he didn’t need a demo. He’d lived through that exact pattern before. He knew what happens when someone identifies the right abstraction layer early.

Most AI pitches are “we do X better than Y.” My pitch was “the entire industry is building at the wrong layer.” He recognized the difference immediately. That’s what the ten minutes were about.

Claude Code surpassed $2.5 billion in annualized revenue by early 2026, contributing to Anthropic’s $44 billion total run rate by mid-year. OpenAI Codex has 5 million weekly users. OpenClaw has over 370,000 GitHub stars, more than the Linux kernel. Whether backed by the most powerful AI labs or the open-source community, the momentum behind AI agents is massive. How do you compete with that?

I don’t. Because we’re not playing the same game.

Look at what those products actually are. Claude Code is a terminal agent that helps one developer mass-produce code. Codex is the same thing inside ChatGPT. OpenClaw is an open-source personal assistant that runs on your laptop. They’re all extraordinary at what they do, and what they do is make one person more productive.

Claude Code even has something called “Agent Teams.” Sounds like team collaboration, right? It’s not. It’s one person orchestrating multiple AI instances. There’s no shared context between team members. No organizational memory. No cross-role coordination. Codex’s “Business plan” is seat management and billing. It doesn’t change how the product works at a team level.

This is exactly my point. The best-funded, most talented AI labs in the world are all converging on the same thing: supercharging individuals. They’re building the most powerful apps the world has ever seen. But nobody is building the OS.

There’s a way to think about this that I find clarifying. The infrastructure for AI-assisted coding, what some people call the “coding harness“, is essentially a solved problem. It’s a continent. Claude Code, Copilot, Cursor, Codex: the land has been claimed. But the infrastructure for AI-assisted working, coordinating teams, managing goals, orchestrating humans and agents together, is still a vast blue ocean. There are a few small islands, but no continent. That’s where we’re building.

When your engineer uses Claude Code and your product manager uses OpenClaw, each person gets faster. But the coordination between them, the context, the decisions, the handoffs, still travels through Slack messages and status meetings and Google Docs that nobody reads. The coordination costs are completely untouched.

That’s the gap. It’s not a feature gap. It’s a layer gap. And it’s not going to be filled by Anthropic or OpenAI, because their business model is selling seats to individuals. An OS for organizations is a fundamentally different product with a fundamentally different architecture.

Last question. Three years from now, what does the AI agent industry look like?

Most of today’s AI agent startups will be dead. Not because they’re bad, but because they’re building at a layer that’s about to get commoditized. When you’re essentially wrapping a prompt around a foundation model and optimizing for one vertical, your moat is prompt engineering. That’s not a moat. That’s a sand castle.

The survivors will be companies that built at a layer the foundation models can’t easily absorb. For vertical agents, that means deep domain-specific data flywheels. For us, it means the OS layer: the orchestration and organizational intelligence that sits above any single model.

But the real disruption isn’t about which companies survive. It’s about what becomes possible. The minimum viable team size for a serious business is about to collapse. Things that required 50 people will require 5 people plus an AI operating system. That doesn’t just change how companies work. It changes which companies can exist. A massive number of business ideas that didn’t pencil out under the old model suddenly become viable.

Three years from now, people won’t ask “what AI tool do you use.” They’ll ask “what OS is your team running on.”

Yimao Zhou is the founder and CEO of Emagen AI, the company behind Cagen. He previously studied medicine at Shanghai Jiao Tong University and cognitive philosophy and philosophy of science. He was the youngest founder in MiraclePlus’s F24 cohort. Learn more at cagen.ai.