Skip to main content

Fast Company Technology

The 2 next big things from long-established companies for 2026

1 month 2 weeks ago

Innovation can be challenging, even for companies with a clean slate. Those who have an established track record of 15 or more years of business must navigate history, culture, product lines, and other entrenched forces that can be both assets and inhibitors. Our honorees have shown that past success doesn’t preclude tomorrow’s progress by leveraging robot fleets for construction and extending real-time intelligence to billions of devices.

DeWalt and August Robotics
For getting fleets of robots to drill down into construction efficiency
Gargantuan data centers do not spring up overnight. Their many construction complexities include holes for structural support as well as mechanical, electrical, and plumbing systems. To streamline what can be a particularly labor-intensive process, DeWalt and August Robotics developed robots that can be operated together as a fleet to handle physically demanding tasks such as concrete drilling while supporting a collaborative human-machine model. In July, August Robotics reported that its drilling robot delivers up to 10x faster drilling speeds versus traditional methods, reducing construction times by 190 weeks across 26 major projects.

Synaptics
For developing edge intelligence in an open-source model
With AI token expenditures soaring in many businesses and security, privacy, and latency concerns preventing many applications from using cloud-based AI, many companies are eager to explore edge-based options. Best known for its pioneering work in developing the touchpad 30 years ago, Synaptics has developed Astra, an open platform to bring real-time intelligence into devices such as factory equipment, medical devices, home appliances, and retail systems. At Google’s developer conference, Synaptics showed off how an Astra-powered development board equipped with Google’s Coral NPU could generate real-time music based on the tracked movement of jellyfish.

The companies and individuals behind these technologies are among the honorees in Fast Company’s Next Big Things in Tech awards for 2026. Read more about the winners across all categories and the methodology behind the selection process.

The 3 next big things in education tech for 2026

1 month 2 weeks ago

Much of the discussion around AI and young people has focused on what career opportunities may exist in the future. But the technology is already changing education. These honorees have leveraged it to personalize preparation for standardized tests as well as help teachers generate compliant curricula and supporting educational materials.

Acely
For personalizing college entrance exam preparation
Many families pay about $1,000 for college entrance exam prep programs that offer one-size-fits-all instruction, with private tutoring stretching well beyond that. Acely seeks to flip that investment on its head with a personalized program starting at $49 per month that requires only about 20 minutes a day. While that still adds up, the service provides personalized, responsive instruction, not just a barrage of drills. In August, the company formed a partnership with the organization behind the ACT to co-market its prep platform nationwide.

Chalkie
For making AI a helpful teacher’s pet
The exceptional emotional intelligence needed to effectively connect with young minds likely leaves teaching relatively low on the list of jobs humans will soon cede to AI. Chalkie keeps the technology squarely in the toolset of human teachers by creating educational resources grounded in sound pedagogy, learning objectives, activities, assessment, and recap tasks, with a special emphasis on making lessons academically and structurally effective. It can also tailor curricula to national and state standards such as Common Core and even adapt lessons to meet IEP-mandated requirements. In March, the edtech startup raised a $4 million seed round. It also recently crossed the million-teacher mark in its user base.

ClassDojo
For helping to develop early reading skills
Education platform ClassDojo’s Dojo Sparks coaches young readers using specially tuned speech recognition, an adaptive science-driven curriculum, and a child-friendly design that integrates motivation and repetition. The company says that using it for 15 minutes a day can take some children from pre-literacy to second-grade reading in six months or less. ClassDojo, which has been expanding its online tutoring and school communications offerings, recently shared that it is used by 45 million families and teachers worldwide.

The companies and individuals behind these technologies are among the honorees in Fast Company’s Next Big Things in Tech awards for 2026. Read more about the winners across all categories and the methodology behind the selection process.

The 6 next big things in data for 2026

1 month 2 weeks ago

Training data begat large language models—and ever since, they’ve been intimately bound. This year’s honorees include a database tuned for AI agents, a dataset for AI-aided diagnosis of breast tumors, and a platform that gives companies more control over how their data is acquired and used.

Databricks
For bringing databases up to AI speed
According to Databricks, AI agent creation of test and development environments on its platform has grown from 0.1% two years ago to 97% today—a dynamic that traditional database architectures have been ill-equipped to accommodate. In response, the company developed Lakebase, a ground-up rearchitecture of databases that can spin them up and shut them down at AI speeds, facilitating the multitude of small, fast workloads of AI-driven development. In August, Databricks raised $5 billion in a strategic funding round, valuing it at $190 billion. The company’s revenue run rate has grown 80% year over year.

DataPelago
For transparently routing code to the optimal processor
While GPUs and NPUs are the processor architectures most closely aligned with AI, high-performance computing tasks also take advantage of two other kinds of processors, CPUs and FPGAs. Effectively routing the right computations to the right processors can reduce significant overhead. DataPelago’s Nucleus modernizes CPU-centric designs of legacy systems and acts as a bridge between data lakes and query engines. It routes computing requests to the most efficient processor type available without requiring any code changes. The company says that its Accelerator for Apache Spark delivers up to 10x faster performance at 80% lower compute costs. In August, DetaPelago was acquired by NetApp, which put Nucleus at the core of its full-stack AI offering.

DDN
For scaling storage to supercomputer workloads
Training ever larger and more capable data models has been hampered by slow data path throughput, which can result in GPU underutilization. DDN’s EXAScaler is a parallel file system technology. One of its highest-profile clients is Google Cloud’s Managed Lustre, a file service used for AI and high-performance computing models. Salesforce and Nvidia are also customers. The managed service can scale up to 10 terabytes per second by combining parallel I/O and high-throughput shared storage. DDN is on track to double revenue to $1 billion in 2026 on surging AI storage demand.

Hydrolix
For improving site traffic insights by watching the bots
AI agent and bot traffic now accounts for more than half of all web traffic, according to Cloudflare. Hydrolix’s Bot Insights makes this behavior visible, providing personalized dashboards for different job roles and allowing customers to decide which ones to block. They can also query the data to answer questions about latency during checkouts or rising search costs. Hydrolix reports that Bot Insights reduces incident attribution time by 95%. The company recently raised an $80 million Series C venture round and was featured as the engine powering AWS’s Agentic Intelligent Operations platform for streaming media

iMerit
For improving breast cancer diagnosis through a breakthrough dataset
When detected early, the five-year survival rate for breast cancer exceeds 90%. The clinical standard for screening—Digital Breast Tomosynthesis, also known as 3D mammography—outperforms traditional 2D mammography for detecting lesions that could be cancerous, particularly in women with dense breast tissue. Unfortunately, this method is held back by a shortage of biopsy-confirmed datasets. iMerit has worked with Segmed and Advocate Health to release an open-source, freely available dataset of more than 500 biopsy-confirmed exam results. The dataset was released in accordance with HIPAA regulations and split roughly between benign and malignant cases. The dataset should help researchers hone AI’s ability to more accurately and quickly assist diagnosis. In August, data and AI company EXL completed its $310 million acquisition of iMerit.

Mozilla Data Collective
For giving usage control to dataset owners
One of the most prominent controversies around AI concerns compensation of the parties behind the data sources used in its training. While content owners are striking deals with AI giants or blocking bots to pave the way for negotiations, the Mozilla Data Collective takes a more proactive approach. Entities that own datasets can share them under their own terms while retaining ownership and control over how that data is used, including compensation. The collective includes more than 350 organizations sharing over 1,700 curated datasets. This year, Mozilla spun out the Data Collective as an independent entity.

The companies and individuals behind these technologies are among the honorees in Fast Company’s Next Big Things in Tech awards for 2026. Read more about the winners across all categories and the methodology behind the selection process.

The 6 next big things in consumer tech for 2026

1 month 2 weeks ago

Consumer technology is at its best when it’s solving problems that exist outside of the tech world. This year’s honorees include better ways to clean your floors, put on your shoes, hear the conversation around you, and practice music effectively. There’s also a less-intrusive smart lock, a refreshingly simple microwave oven, and—if you’ll allow one solution to a problem caused by technology itself—a way to spare your smartphone screen from prying eyes.

JBL/Harman
For giving guitar practice an AI upgrade
While the modern musician will find no shortage of backing tracks on YouTube for practice purposes, they’re seldom as fun to jam with as the original material. JBL’s BandBox Solo and Trio guitar amplifiers can play music from streaming services over Bluetooth but with AI audio separation that controls the volume of individual instruments. Users can mute a song’s guitar track while playing along themselves, or turn down the vocals and drums to better hear the notes in a tricky solo. Software tools such as LALAL.AI and Moises have long allowed aspiring musicians to create backing tracks from recorded music, but by building this capability directly into the hardware, JBL is adding a new layer of convenience.

Kizik
For making footwear hands-free
For nearly a decade, Kizik has been making shoes you can squeeze your heels into without having to use your hands. But while the company has expanded into new categories such as snow boots, running shoes have always been a major challenge because of the need for a stable fit with no performance compromises. The company finally cracked the problem this year with its Freedom Run line, with a heel design that compresses when you step into it but then snaps into a rigid position that prevents slippage. Tests at Brigham Young University confirmed that the shoe’s heel stability while running is comparable or better than that of leading brands. The company’s parent, HandsFree Labs, is also working with Nike and New Balance and plans to expand into other areas of active footwear, such as hiking and winter sports, with the eventual goal of making every shoe hands-free.

Legato
For using smart eyewear to enhance what you hear
While Big Tech companies like Meta and Google hoover attention with their smart glasses platforms, Legato is building frames for a more practical purpose. Its Legato Frames use on-device AI audio processing to enhance human speech while quieting background noise, serving as a stylish and smarter alternative to traditional hearing aids. The company has a big rival in EssilorLuxottica, whose Nuance Audio glasses serve the same purpose. Still, its team of former Bose engineers figured out how to solve some key problems, such as amplifying the voices of others without unnaturally boosting the wearer’s own voice. Frames remains on a waitlist, though Legato has raised $12 million to help bring the product to market by year-end.

Level Home
For making smart locks look less dumb
Smart locks tend to be eyesores, with bulky enclosures to house all the necessary motors, sensors, and connectivity. With the Level Lock Pro, Level managed to fit all of those components into a smart lock no larger than a standard deadbolt. The company says it’s 62% smaller than traditional smart locks. Corporate parent Assa Abloy laid off most of Level’s staff in June, according to The Verge, but the Swedish lockmaker intends to maintain Level as a separate brand. Even if the Lock Pro ends up being Level’s swan song, here’s hoping it inspires the next generation of rival smart locks to shrink down.

Roborock
For literally bringing robot vacuums to the next level
Climbing stairs has always been a holy grail for robot vacuums. So Roborock naturally made a stir at CES in 2026 with its Saros Rover, a prototype robot vacuum with a pair of retractable wheels on legs. They allow the device to make its way up a staircase, cleaning each step on the way. While Roborock has not shared concrete launch plans, the company says it’s a “real product in development.” And with rival Dreame developing its own stair-climbing concept (albeit one that can’t clean the stairs itself), the race to solve robot vacuums’ multistory problem is well underway.

Samsung
For offering a solution to shoulder surfing
Samsung’s Privacy Display is the first new smartphone feature in years that feels like a truly practical breakthrough. Debuting on the company’s flagship Galaxy S26 Ultra, the feature applies a matrix to the display that narrows the beams of light emitting from each pixel, making the screen harder to read from off-angles. This can prevent strangers from peeking at your phone’s contents in public places such as trains, planes, and elevators. Unlike aftermarket privacy-protecting screen protectors, the Galaxy S26 Ultra lets users toggle the feature on and off or limit it to specific apps. This allows the screen to remain bright and shareable when you’re not worried about who might be looking.

The companies and individuals behind these technologies are among the honorees in Fast Company’s Next Big Things in Tech awards for 2026. Read more about the winners across all categories and the methodology behind the selection process.

The 6 next big things in computing, chips, and foundational technology for 2026

1 month 2 weeks ago

Advances in computing aren’t just about faster processors. As these innovations show, progress is driven by everything from the technologies used to transmit data to the systems that ensure data centers get reliable power without overwhelming the grid. If many of them are largely invisible to most of us—well, that’s kind of the point.

Amazon
For skating where the AI puck is going
Announced at AWS re:Invent in December 2025, Trainium3 is Amazon’s newest AI chip. It’s designed to take on not just today’s dominant AI workflows but also to anticipate the needs of emerging ones such as world models. Equipped with up to 144 chips, Amazon’s Trainium3 UltraServer quadruples the compute performance and energy efficiency of its predecessor. Uber is adopting Trainium3 for tasks such as calculating arrival times, and video-generation platform Decart has used it to achieve 4x higher frame throughput and reduce latency by 75%.

Cadence
For teaching AI agents to design processors
Historically, advances in processor design have been propelled by human engineering talent, a resource that’s difficult to scale. Introduced in early 2026, Chip design giant Cadence’s Super Agents take on essential elements of the process such as creating test plans, design node migration, and digital implementation. Already adopted by big names like Nvidia, Qualcomm, and Altera, they offer up to 10x productivity increases across design and verification, according to Cadence.

HP
For building a workstation today for tomorrow’s GPUs
Out of the box, HP’s Z8 Fury G6i desktop workstation can support for up to four powerful Nvidia RTX Pro 6000 GPUs. But the company also looked ahead to future GPUs. The workstation’s optional Max Side Panel expands its case by 1.4 inches and adds cooling fans, allowing tool-free expansion to accommodate larger, hotter GPUs that require more power and would otherwise require the purchase of a new workstation. HP uses recycled plastic, steel, and rare earth materials to build the Z8 Fury G6i and put it through 360,000 hours of reliability testing.

Marvell
For harnessing light to move AI data
No matter how fast an AI system’s GPUs are, it may be held back by the electrical interconnects and copper pathways used to shuttle data in and out. Marvell’s Phototonic Fabric platform replaces them with light-based transport, allowing the same energy footprint to deliver up to three times the performance. The company gained the technology as part of its $3.25 billion acquisition of Celestial AI, announced in December 2025, and has since been engineering products that will support it both as an upgrade to existing AI systems and as a core element of future ones.

ON.energy
For giving AI the uninterruptible power it needs
AI data centers don’t just draw vast amounts of power—they do it in oscillating bursts that stress the electrical grid and cause problems for everyone else on it. After developing an uninterruptible power supply for a Mexico City bakery prone to blackouts, ON.energy realized that its design was also applicable to AI. That led to the AI UPS. In tests at the U.S. Department of Energy’s National Laboratory of the Rockies, it prevented load swings from 30% to 100%, in intervals as fleeting as 10 milliseconds, from impacting the grid.

Silicon Quantum Computing
For pioneering quantum-enhanced AI
In October 2025, Australia’s Silicon Quantum Computing launched a quantum-enhanced machine learning system called Watermelon. The first commercially available quantum reservoir computer, it uses quantum technology to spot patterns in datasets that can’t be identified by classical computers. The technology requires the placement of phosphorous atoms in silicon with 0.13 nanometer accuracy and has been validated by partnerships with Telstra, Nvidia, and the Australian Department of Defense.

The companies and individuals behind these technologies are among the honorees in Fast Company’s Next Big Things in Tech awards for 2026. Read more about the winners across all categories and the methodology behind the selection process.

The 3 next big things in commerce tech for 2026

1 month 2 weeks ago

AI is poised to transform virtually everything about online and physical retail, from product exposure and recommendations to consumers even having to complete the transactions for themselves. These honorees are setting the stage for agentic commerce, allowing critical medical supplies to take flight and providing physical retailers with comprehensive intelligence about shopping variables within and beyond the store.

American Express
For putting agents on the checkout line
During the early days of e-commerce, retailers strived to simplify online purchasing for consumers. In the AI era, a new wave of accommodations must be forged to allow frictionless transactions by agents. American Express’s Agentic Commerce Experiences (ACE) toolkit leverages the company’s experience across a range of financial roles and includes a new trust model focused on identity, intent, accountability, and visibility. But Amex is still keeping humans in mind through a new agent purchase-protection program that protects eligible card members from charges resulting from certain AI agent errors. Early-access developer partners have recently signed on to ACE, which could further drive the record card-member purchases the company reported in its most recent quarter.

CVS Health
For making delivery of critical products airborne
While several companies have deployed drones during emergencies and others have piloted broad drone delivery programs, CVS’s Air Response program targets a scenario that falls in the middle: serving patients who need critical supplies such as naloxone, epinephrine, and automated external defibrillators (AEDs). As a retailer serving mass markets, CVS has designed its drone network to operate safely in urban environments. The company is now exploring integration with 911 systems and has announced partnerships with Thales and SkyfireAI as it frees up investment dollars through its exit of the Omnicare long-term-care pharmacy segment.

WVN
For developing holistic retail site intelligence
Physical-store owners have long envied e-retailers that have nearly perfect accounts of the digital pathways purchasers travel through their storefronts. WVN seeks to create a comparable level of market intelligence for brick-and-mortar players. StoreBrain is an AI agent that acts as a persistent AI intelligence layer for physical storefronts, capturing and analyzing not only onsite indicators like transactions, foot traffic, and shelf input but also inputs beyond the store, such as online reviews. The company’s training data also covers demographics, psychographics, economics, and even weather data. The company reports that 93% of stores using its platform report improved conversion. It recently inked a partnership with menswear brand Mugsy, which operates eight physical stores.

The companies and individuals behind these technologies are among the honorees in Fast Company’s Next Big Things in Tech awards for 2026. Read more about the winners across all categories and the methodology behind the selection process.

The 4 next big things in buildings and real estate for 2026

1 month 2 weeks ago

As the real estate industry struggles with rising interest rates and ballooning material costs—thanks in part to tariffs—innovation is becoming more and more important. New technologies and construction methods are boosting margins for builders and helping to compensate for labor shortages. And in some instances, they’re dramatically reducing building times without sacrificing safety standards.

Augmenta
For reducing waste and augmenting bottom lines
Since late 2025, Augmenta’s AI engine has generated 3D building designs for electrical systems. That dramatically speeds up the design process, as it can take engineers hundreds of hours to manually model those systems. And it’s not just theoretical: Around 100 buildings have taken advantage of the technology so far, with the first, an elementary school, reducing material waste by 15% and cutting design time by 25%, the company says. Those real-world results are saving builders money and letting them automatically adjust plans as design needs change—a frequent occurrence in construction. Up next for the company is the inclusion of mechanical and plumbing system designs, which should further reduce costs for builders.

LightTable
For spotlighting ways to dramatically speed up construction timetables
It normally takes three to six weeks to analyze an architect’s or engineer’s construction drawings for potential errors. Using LightTable’s AI platform, it takes just three to five days. To date, the company says it has reviewed more than 20 million square feet’s worth of documents, representing $10 billion in project costs. As real estate projects get more complicated, this is a technology that could streamline the building process before the first shovel breaks ground.

TerraFirma
For bypassing labor shortages through autonomy
Founded by former SpaceX engineers, TerraFirma sees similarities between building rocket ships and building infrastructure. The company uses autonomous technology that lets a single worker operate multiple excavators, bulldozers, loaders, and other vehicles simultaneously. That doesn’t require designing new construction equipment: Retrofit kits can be installed on existing machines in under five hours to automate them. It’s a technology the company hopes to use off-world as humanity looks to build a presence on Mars, and the idea is catching on fast. TerraFirma has already seen commercial demand from data centers, industrial sites, and residential developments. And the company has generated more revenue year-to-date than it raised across all seed financing.

Willow
For spotting problems before they happen—and saving millions in the process
Equipment failures can be a major cash drain for building operators. Willow’s technology uses AI to detect and anticipate building equipment failures before they happen, proactively alerting management systems to the problem. And it’s generating big savings for its commercial clients. To date, the company says it has saved Walmart $1.4 million by preventing more than 800 failures and Georgia Southern University more than $1 million. Its technology is already being used in 38 countries, putting it on a strong growth trajectory. The real-world effect is that infrastructure such as airports, hospitals, and other facilities can anticipate and repair problems ahead of time rather than reacting to breakdowns.

The companies and individuals behind these technologies are among the honorees in Fast Company’s Next Big Things in Tech awards for 2026. Read more about the winners across all categories and the methodology behind the selection process.

The 3 next big things in spatial computing, virtual reality, and mixed reality for 2026

1 month 2 weeks ago

The metaverse hype bubble may have burst, but that doesn’t mean immersive computing is yesterday’s news. Technologies such as spatial computing and VR are changing how products are designed, where entertainment can happen, and what kind of medical care is available in remote maritime situations.

Dassault Systèmes
For putting Apple Vision Pro to work
Organizations of all kinds are eager to create digital twins of their most crucial real-world products, processes, and spaces. 3D-design software giant Dassault Systèmes’ 3DLive turns Apple’s Vision Pro headset into an intuitive, precise way to engage with these simulations. Customers have adopted it for applications ranging from engineering new EVs to testing surgical instruments. An aerospace company used the platform to bring participants on three continents together to review an aircraft’s interior, cutting approval time in half and saving $450,000 in travel costs.

Excurio
For breaking through VR’s boundaries
Shared VR experiences in public spaces have been held back by challenges such as scaling up the number of people supported by an environment and allowing different participants to explore varying worlds. Paris-based Excurio’s LBXR Pulse lets museums, cultural institutions, entertainment venues, and other organizations deploy VR to up to 130 users simultaneously in multiple concurrent experiences. It’s in use at 25 venues in 17 countries, with big-name content partners such as the Musée d’Orsay and Harvard University’s Giza Project.

Veyond Metaverse
For using XR to bring healthcare to the sea
More than 2 million people are seafarers or work on offshore rigs. Veyond Metaverse’s OceanMD lets them receive telehealth services from onshore medical professionals wearing mixed-reality headsets. Unlike most mixed-reality platforms, which assume the availability of high-speed internet, OceanMD is designed to operate at just 100 kbps, allowing it to run over the slow satellite access available at sea. Pilot programs are underway in West Africa and Brazil offshore operations, and in May 2026, international SOS company Medaire secured a five-year exclusive agreement to use the technology for global civilian aviation.

The companies and individuals behind these technologies are among the honorees in Fast Company’s Next Big Things in Tech awards for 2026. Read more about the winners across all categories and the methodology behind the selection process.

The 13 next big things in applied AI for 2026

1 month 2 weeks ago

2026 may have been the year when AI stopped getting brownie points simply for being amazing. What’s impressive now is AI that performs useful tasks. All the better if it does it quietly, cleverly, and in areas that help humans focus on the work that we still do best. These honorees check off all those boxes, in areas ranging from workplace productivity to medical safety.

Alembic
For proving how decisions turn into outcomes
Alembic’s “causal engine” uses a proprietary AI system to help enterprises understand the real effects of actions they take. The platform can analyze thousands of variables simultaneously–everything from broadcast and digital advertising spend to macroeconomic shifts to executive public appearances–and map them to specific outcomes, or to the lack of an outcome. Nvidia was Alembic’s first enterprise customer, and its CEO Jensen Huang has publicly praised the platform, whose other big customers include Mars and Delta Airlines. In November 2025, Alembic closed a $145 million Series B led by Accenture (which now offers Alembic’s causal-AI analysis to its consulting clients) and Prysm Capital, valuing the company at approximately $645 million.

Canva
For giving us AI-powered design that doesn’t suck
We all have that design in our heads that we can’t quite generate ourselves. Canva’s AI 2.0, released earlier this year, allows you to turn ideas into reality without really needing to know your way around design software. The tool is powered by the company’s Canva Design Model, and might allow you to whip up a campaign with just a few instructions and then refine it with ease—something that’s notoriously hard to accomplish with general-purpose AIs such as ChatGPT. More than 400 designs are created in Canva AI every second.

Credo AI
For helping enterprises jump-start AI governance
Credo AI’s popular AI governance platform now offers the Govern AI Assistant (GAIA), which helps enterprises more easily manage AI models and tools. (A Credo AI study found that while 60% of organizations now deploy AI across multiple departments or company-wide, only 4% are governing those deployments across the organization.) The assistant is pre-loaded with Credo AI’s governance libraries, years of enterprise deployment experience, and global regulatory frameworks such as the EU AI Act and NIST AI RMF. Users upload a product brief or project plan, and GAIA generates use case descriptions, recommends metadata, drafts questionnaire responses, identifies risk scenarios, and maps controls to those risks.

DeepL
For overcoming language barriers in real time
In April 2026, DeepL unveiled real-time voice-to-voice translation as part of its DeepL Voice translation product. A user can now speak naturally in their own language, while another user immediately hears the meaning spoken aloud in theirs, enabling live conversation across languages without waiting for captions or interpreters. Pioneer, Aramark, Avendra International, and the European consultancy Inetum have all begun using DeepL Voice, DeepL says. The company doubled down on real-time voice this year when it completed its acquisition of Mixhalo, which specializes in ultra-low latency audio that’s been used to power live events for thousands of simultaneous participants. DeepL aims to leverage the technology to enable real-time translation for larger groups such as conferences with thousands of attendees and hundreds of speakers. 

Genspark
For giving productivity an agentic edge
Once an AI search engine, Genspark has become an AI agent platform. Its AI Workspace offers a wide array of tools, including a full AI office suite, an image generator, and specialized AI agents. Under the hood, an orchestration engine that calls on about 70 AI models, depending on the nature and requirements of the task at hand. A persistent AI assistant, Genspark Claw, can complete long-running multi-step workflows asynchronously, accessing browsers, files, and enterprise tools such as Microsoft Office 365. With more than 5,000 enterprises now using the platform, the company says its business raced to $250 million in annualized revenue in the spring of 2026.

Hanwha
For operating data centers with more autonomy
Data centers use an enormous amount of power, so much that they’ve now become politically polarizing. To help mitigate that problem, Hanwha, the South Korean conglomerate, has developed an Energy Management System that acts as an uber-smart manager of all the electricity that runs through one of these facilities. Using a cadre of large language models and other machine learning systems, it can monitor a data center’s telemetry, alarms, and fault detection systems. Then it studies what might be causing an issue, and how to respond. In April 2026, Prime Group, one of the largest private real estate investors, adopted the technology for a nationwide deployment of edge data centers.

Odyssey
For making simulated physics interactive
Language models ignited the generative AI revolution, but the real world runs on physics, not text. Odyssey’s world models learn from visual observation of real-world activity and generate video that responds to a user’s actions, making them useful in simulations, gaming, and robotics. The company’s latest model, Odyssey-2 Max, advances the interactivity, stability, and fidelity of its earlier models; the company says it earned the highest physics score among publicly evaluated world models on the VBench 2 benchmark (58.52, up from 49.67 in the prior generation). Perhaps the best proof of the company’s potential is its list of backers, which include AI luminaries Jeff Dean, Elad Gil, and Garry Tan. It’s also backed by NVIDIA’s venture arm.

Replit
For taking on the heavy lifting of vibe coding
Unlike many AI coding platforms, Replit has kept its focus on enabling users to vibe-code their way through whole software development projects. The company’s newest assistant, Agent 4, allows non-coders to direct the AI in plain language to design, build, and ship software. Replit handles the hosting, databases, authentication, and monitoring, and offers more than 100 integrations with services from companies such as Databricks, Stripe, Slack, and Microsoft. A user can produce a web app, a mobile app for iOS and Android, a slide deck for an investor pitch, and a launch video—all from a single project. More than 50 million people use Replit, and 85% of Fortune 500 companies now build with it, the company says.

SAS
For developing simulation that powers sterilization
SAS, the global data company, is designing high-fidelity digital twins of environments such as operating rooms. Using Epic Games’s Unreal Engine, the same technology that powers immersive video games, SAS can simulate essential details such as the specialized sterilization procedures that need to go exactly right to keep patients safe. Today, the system is used to help refine the work at Sterilcentral, a facility in Denmark charged with cleaning thousands of instruments used in surgeries.

Tensor Auto
For self-driving cars that belong to . . . you
Unlike Waymo and other autonomous-vehicle companies focused on the robotaxi business, Tensor Auto is building a self-driving car that you can actually own, armed with powerful lidar sensing systems and an onboard supercomputer that can listen to verbal commands. The core of the system is the Tensor World Model, which uses prompts to simulate all sorts of physical conditions that a car might encounter. These kinds of simulations form the basis of Tensor’s brains, but they could be applied to other physical AI, too. The company has already won a Level 4 DMV testing certification in California.

Vexcel
For teaching AI to understand the world from above
Vexcel, an aerial imagery company, is training AI to understand everything on the ground from a bird’s-eye view. Its Vexcel Intelligence system can scan satellite imagery for complex descriptions, such as a set-back Victorian house or farmland growing a specific type of crop. The company is working with Google Maps and helping with disaster-response assessment, including after a tornado in Oklahoma and a fire in Georgia.

Within
For building organizational brainpower
Many enterprise AI deployments fail to pay for themselves because they automate workflows that are inconsequential or that the company doesn’t really use. Within (formerly Klarity) focuses on helping businesses understand how their people actually work, creating an operational context graph (aka “Company Brain”) and tapping that knowledge to design, build, and manage context-rich agents. The company says it saved its 93 customers (which include OpenAI, Salesforce, ServiceNow, and Uber) $53.9 million in consulting fees during 2025. Within has raised more than $91 million in total funding from investors such as NFDG, Elad Gil, and Y Combinator.

Yutori
For teaching AI to surf the web like a pro
Yutori develops web-using AI agents. The company trained its Navigator n1.5 model on both simulated and real site interactions to reliably handle a broad set of web tasks. More recently, the company released Navigator n2, a 27-billion-parameter computer-use model designed for operating full desktop environments. Yutori says that its customers (such as Meta) are now making more than 100 million API calls to its models every month, and that its agents are interacting with “several hundred thousand” merchant websites daily. The company’s cofounders, Dhruv Batra and Devi Parikh, led high-profile AI development teams at Meta, while other Yutori researchers came from leadership roles at Google DeepMind, Tesla, and Apple.

The companies and individuals behind these technologies are among the honorees in Fast Company’s Next Big Things in Tech awards for 2026. Read more about the winners across all categories and the methodology behind the selection process.

Capcom says it will use AI to speed up game development times

1 month 2 weeks ago

Capcom saw tremendous success with the launch of Resident Evil Requiem earlier this year. The survival horror game has already sold more than 8.5 million copies, bringing in revenue topping $500 million and reigniting interest in the 30-year-old franchise.

Resident Evil Requiem, like any AAA game, took a long time to make, with a development cycle spanning six years and including a reboot midway through production. To maintain momentum, Capcom has revealed plans to incorporate artificial intelligence into its game development process.

The project is called REX, programmer Satoshi Ishida revealed at a tech conference over the weekend. At its core, six separate programs will help integrate AI and machine learning into the RE Engine, the increasingly long-in-the-tooth proprietary game engine that powers Capcom’s biggest titles.

The goal, Ishida said, is to create a “game engine of the AI generation.”

To be clear, Capcom isn’t planning to have AI develop creative elements for its games, such as game assets. Instead, the project reflects a broader problem facing the industry: Game development cycles are getting longer, and Capcom hopes AI can streamline parts of the process. That includes elements like quality assurance testing and other development efficiencies.

Capcom is hardly alone in incorporating AI into its workflow. Several other publishers have done so in recent years. Electronic Arts CEO Andrew Wilson said in April that AI technology has augmented jobs at the company rather than replaced them.

“I saw some data recently, I think, now almost all—like 85%—of our quality assurance [work] is done with some kind of machine learning or AI-driven algorithm,” he said.

And Strauss Zelnick, CEO of Take-Two Interactive Software, said his company is exploring what advantages AI can offer.

“Anything that allows us to do a better job—quicker and cheaper—we’re interested,” he told Fast Company at the Entertainment Software Association’s Interactive Innovation Conference earlier this year. “What we are beginning to see is mundane work that can easily be replaced by technology is indeed being replaced [and] people turn their attention to higher-value tasks.”

Capcom’s AI tools will do much the same. One system, for instance, could animate groups of objects or characters together rather than having those animations run individually in real time. Another, according to Capcom’s presentation, would convert game-mechanics ideas into a standardized programming language. Since RE Engine is used by developers in several countries, that common language could also help speed up development.

Capcom did not respond to a request for further comment about the REX project, but it addressed its stance on AI during a shareholders meeting in March, making clear it was interested in the technology but did not plan to use it to replace the human creative force behind its games.

“We will not implement assets generated by AI into our games,” the company said. “However, we plan to proactively use it as a contributing technology to improve the efficiency and productivity of the game development process. To this end, we are currently exploring ways to implement it in various areas, including graphics, sound, and programming.”

This isn’t Capcom’s first experiment with AI. In January 2025, it created a prototype “idea generation” system alongside Google that used generative AI to help brainstorm ideas.

“One of the most time-consuming and labor-intensive aspects of game development is brainstorming ideas to create the game’s world,” Kazuki Abe, technical director at Capcom, said at the time. “For example, if there’s a television in the game’s environment, we can’t just use an existing product. We need to come up with the shape, manufacturer’s logo, and other details ourselves.”

The companies said the system lowered the cost of brainstorming, helping improve margins. But as the video game industry continues to undergo a painful contraction, with many studios cutting staff, almost any use of AI is likely to create tension among game makers.

As AI advances and becomes capable of doing more, executives may eventually reconsider whether some positions are necessary. Zelnick acknowledges that possibility, though he argues that properly implementing the technology can instead eliminate low-value work.

“I have an obligation to run an efficient enterprise,” Zelnick said. “We have not seen any opportunity to reduce head count because of AI. What we have seen is an opportunity to stop wasting time on things that don’t create value.”

How Meta’s VR Glasses succeed where Apple’s Vision Pro fails

1 month 2 weeks ago

Apple’s Vision Pro was never going to be a big success right out of the gate. It wasn’t just the eye-watering $3,499 price—it was the headset’s reliance on highly advanced, low-yield OLED display panels from Sony, which could reportedly only be produced in quantities of under a million a year.

There was never a scenario in which the Vision Pro could have turned into the next iPhone—or even the next Apple Watch—overnight. Apple knew this when it released the product, but thought it was worth getting the platform out into the wild anyway.

That said, the company can’t be too pleased with the Vision Pro’s reception. Content for the headset has slowed to a trickle, and at no point since launch has it been difficult to find one in stock. A niche product designed to stoke excitement among early adopters and developers is one thing, but there’s little evidence that it’s succeeded even on that level.

{"blockType":"mv-promo-block","data":{"imageDesktopUrl":"https:\/\/images.fastcompany.com\/image\/upload\/f_webp,q_auto,c_fit\/wp-cms-2\/2025\/12\/multicore.png","imageMobileUrl":"https:\/\/images.fastcompany.com\/image\/upload\/f_webp,q_auto,c_fit\/wp-cms-2\/2025\/12\/multicore-mobile.png","eyebrow":"","headline":"\u003Cstrong\u003E\u003Cstrong\u003ESubscribe to Multicore\u003C\/strong\u003E\u003C\/strong\u003E","dek":"Multicore is about technology hardware and design. It’s written from Tokyo by Sam Byford. To learn more, visit \u003Ca href=\u0022https:\/\/www.multicore.blog\/\u0022\u003Emulticore.blog\u003C\/a\u003E","subhed":"","description":"","ctaText":"SIGN UP","ctaUrl":"https:\/\/www.multicore.blog\/","theme":{"bg":"#f5f5f5","text":"#000000","eyebrow":"#9aa2aa","subhed":"#ffffff","buttonBg":"#000000","buttonHoverBg":"#3b3f46","buttonText":"#ffffff"},"imageDesktopId":91454027,"imageMobileId":91454030,"shareable":false,"slug":"","wpCssClasses":""}}


That’s not to say anything bad about the Vision Pro’s quality—visionOS is a thoughtfully designed operating system, and the hardware remains class-leading in many ways. But if Apple’s vision for spatial computing is ever going to take off, a few things need to happen. The headset needs to get smaller, it needs to be more comfortable to wear, and it needs to be much cheaper.

The Meta alternative

Enter Meta, which last week announced a product that would seem to tick all three boxes. The Meta VR Glasses are designed to do many of the same things as the Vision Pro, but in a lighter, more glasses-like form factor that costs only about a third of the price of Apple’s device—they’ll go on sale for $1,299 next year, while the Vision Pro starts at $3,699.

How did Meta do it? One cost-cutting measure is the VR Glasses’ OLED panels.

The Vision Pro’s dual 4K displays were expensive even by the standards of the rest of the hardware, which was as high-end as you’d expect of Apple. One Omdia estimate put the two internal Sony micro-OLED panels at $228 each. Add the much-ridiculed external OLED screen, which shows a low-resolution representation of the wearer’s eyes, and the estimated display bill came to around $530—or 35% of the total component cost. 

It’s important to remember that a bill of materials doesn’t tell you everything that goes into the sticker price you see on shelves. But in this case, it could well indicate how much Apple was spending on its display panels before accounting for chips, sensors, or anything else.

Meta says the VR Glasses have 2,412 x 2,288 pixels per eye, which is about half as much as the Vision Pro—it’s a total of around 11 million pixels versus 23 million. But in terms of sharpness, the two headsets are closer than you might think. Meta is quoting a figure of 37 pixels per degree, which is actually higher than the Vision Pro’s.

The catch here is that the field of view is smaller, at around 70 degrees horizontally compared with roughly 100 degrees for the Vision Pro. These figures can be fuzzy and differ between users, but the difference between the two is pretty stark.

In my experience with various augmented reality glasses, however, the field of view is not quite as critical in that form factor as it would be in a VR headset designed to immerse you in a virtual world. I often use the Xreal One glasses, for example, which have an even smaller 50-degree field of view, and the picture stretches out close to the edge of the glasses themselves. It’s not really something you think about when you’re watching a movie.

Meta does say that it will offer an optional accessory to block out peripheral vision, and in that configuration, the VR Glasses are likely to be less immersive than the Quest 3 or Vision Pro. But as a way to achieve the same kind of AR experience as the Vision Pro, while letting you see the world around you in a radically smaller and lighter design that weighs just 100 grams, it feels like a worthy trade-off—especially since the panels are so much cheaper and likely available in greater volume. Impressions I’ve heard from early testers have been universally positive.

Other advantages

Another way Meta has reduced the size and weight is by moving computing components to the tethered “puck,” which in the Vision Pro solely houses the battery. That felt like an odd decision by Apple even at the time of the Vision Pro’s release—after all, the majority of space inside an iPhone is already taken up by its battery. If you’re going to have a cable running down to a separate breakout box anyway, it would surely be prudent to make the most of it and lighten the load on the user’s head. 

Apple also didn’t want to compromise when it came to the Vision Pro’s controllers, or lack thereof. The operating system is entirely built around intuitive gesture controls, in a typically Apple move. But it did mean that the Vision Pro was on the back foot when it came to gaming support, and the later addition of support for the controllers from the PlayStation VR2—itself not exactly a roaring success of a device—hasn’t really moved the needle.

The Meta VR Glasses also won’t ship with controllers in the box, but they have two key advantages: compatibility with the huge library of Quest games and with Quest 3 controllers. That means anyone who buys the glasses and does want to play games on them won’t be short of things to do on launch day.

Of course, even the Quest 3—by some distance the most popular VR headset to date—is still far from a mass-market success, and there’s no guarantee that Meta will fare any better with the VR Glasses. But from a design and technology perspective, it’s shocking just how more viable of a product they already appear to be than the Vision Pro. 

Apple is going to have to go back to the drawing board with visionOS, and Meta may well have just shown the way forward.

{"blockType":"mv-promo-block","data":{"imageDesktopUrl":"https:\/\/images.fastcompany.com\/image\/upload\/f_webp,q_auto,c_fit\/wp-cms-2\/2025\/12\/multicore.png","imageMobileUrl":"https:\/\/images.fastcompany.com\/image\/upload\/f_webp,q_auto,c_fit\/wp-cms-2\/2025\/12\/multicore-mobile.png","eyebrow":"","headline":"\u003Cstrong\u003E\u003Cstrong\u003ESubscribe to Multicore\u003C\/strong\u003E\u003C\/strong\u003E","dek":"Multicore is about technology hardware and design. It’s written from Tokyo by Sam Byford. To learn more, visit \u003Ca href=\u0022https:\/\/www.multicore.blog\/\u0022\u003Emulticore.blog\u003C\/a\u003E","subhed":"","description":"","ctaText":"SIGN UP","ctaUrl":"https:\/\/www.multicore.blog\/","theme":{"bg":"#f5f5f5","text":"#000000","eyebrow":"#9aa2aa","subhed":"#ffffff","buttonBg":"#000000","buttonHoverBg":"#3b3f46","buttonText":"#ffffff"},"imageDesktopId":91454027,"imageMobileId":91454030,"shareable":false,"slug":"","wpCssClasses":""}}

How ballroom dance can help us avoid a dystopian future 

1 month 2 weeks ago

Dystopian narratives about AI abound. Will the bots take over? Are we doomed as a human species? While I do not have a utopian outlook about our tech-saturated future, I do wonder what we could start doing today, in our present state, to course-correct these calamitous narratives. 

So, earlier this year I began conducting an informal thought experiment with approximately 50 people. I asked them to engage in the following pre-mortem exercise: “Imagine it is 10 years into the future, and the dystopian version of our world has come to be—that the bots have taken over. What is one mindset shift, one behavioral change, and one work process modification you would make today to prevent that dystopian future?” 

The themes of the responses were reassuring and affirming. In summary, people want transparency and to enact their agency by demanding ethically designed artificial intelligence. They would ensure there was more pausing and slowing down in our workday in order to do deep thought. Interestingly, most people would design in friction, not eliminate it. They want embodied learning to ensure that we are not rushing ahead without thinking through the human dimension of our work. 

In short, we should scaffold AI output with human consciousness as inputs. 

{"blockType":"mv-promo-block","data":{"imageDesktopUrl":"https:\/\/images.fastcompany.com\/image\/upload\/f_webp,q_auto,c_fit\/wp-cms-2\/2026\/01\/i-16x9-figure-thinking.jpg","imageMobileUrl":"https:\/\/images.fastcompany.com\/image\/upload\/f_webp,q_auto,c_fit\/wp-cms-2\/2026\/01\/i-16x9-figure-thinking_0b545c.jpg","eyebrow":"","headline":"\u003Cem\u003EWonderRigor™️ Newsletter\u003C\/em\u003E","dek":"Want more insights, tools, and invitations from Dr. Natalie Nixon about applying creativity for meaningful business results and the future of work? Subscribe \u003Ca href=\u0022https:\/\/urldefense.proofpoint.com\/v2\/url?u=https-3A__figure-2D8-2Dthinking-2Dllc.kit.com_sign-2Dup\u0026amp;d=DwMFaQ\u0026amp;c=euGZstcaTDllvimEN8b7jXrwqOf-v5A_CdpgnVfiiMM\u0026amp;r=xHenyQfyc6YcuCNMBsOvfYGQILM1d1ruredVZikn4HE\u0026amp;m=F383gnrChFhYKPhcpNHI1hY3o58IHIn_LkB5QJDrs3G5Wfft-DcucUO4UEmGO7GZ\u0026amp;s=JlJm7GyKCJvPW0jyrsfTFtinteKDitN13vfPZiuJnP8\u0026amp;e=\u0022 target=\u0022_blank\u0022 rel=\u0022noreferrer noopener\u0022\u003Ehere\u003C\/a\u003E for the free WonderRigor™ newsletter at Figure8Thinking.com","subhed":"","description":"","ctaText":"Learn More","ctaUrl":"http:\/\/Figure8Thinking.com","theme":{"bg":"#3b3f46","text":"#ffffff","eyebrow":"#9aa2aa","subhed":"#ffffff","buttonBg":"#6e8ba6","buttonHoverBg":"#3b3f46","buttonText":"#ffffff"},"imageDesktopId":91470060,"imageMobileId":91470061,"shareable":false,"slug":"","wpCssClasses":""}}
An opportunity

The responses from this premortem thought experiment demonstrate that we have an opportunity to shift our mindsets and behaviors right now. In an earlier Fast Company article, I wrote about something I call “sentient intelligence”: the internal sensory and emotional navigational console with which we humans are equipped to make sense of the world. It’s in high gear when we sit in awe, discern meaning, pause, and ask bold follow-up questions. Without tapping into our sentient intelligence, we easily lapse into behavioral patterns where we stop at the first answer. 

Sentient intelligence is unique to humans. The intelligence that AI exhibits doesn’t make it human. But computer scientists like Sean White, CEO of Inflection AI, believe that AI is capable of having relational intelligence. By way of example, he referenced an AI tool that could recall a caregiving journey and then follow up with us weeks later. Sean calls that an example of AI exhibiting relational intelligence—a sort of “pro-social AI” that can build relationships and connect us back to each other. It’s an opportunity to shift away from brief, one-stop, pedantic answers and toward AI platforms that engage more comprehensive thinking on our part and foster deeper human collaboration. In short, White’s Inflection AI is one way to redirect our trajectory away from a dystopian future. 

The ballroom dance connection

Long after my conversation with White about Inflection AI, I pondered whether what his team has developed was relational intelligence or relational design. It’s a subtle distinction but bears some consideration as I reflected that I toggle between relational intelligence and relational design in one of my favorite pastimes: ballroom dance. 

On the ballroom dance floor at the amateur level, I constantly experience toggling between relational design (or choreography) and relational intelligence (the execution and discernment between two dancers). That relational intelligence is powered by our sentient intelligence and improves through contact and connection. By extension, the benefit of something like Inflection AI is that it can act as a situational catalyst that could bring in the choreographer when we need guidance, mapping, and optimization of the ways we connect.

Consider that premortem thought experiment. How would you answer it? And would you welcome assistance to keep the bots from taking over while we lean into our own human agency? A platform like Inflection AI may be the situational choreographer we could use to amplify our human connection—without outsourcing attunement completely.

{"blockType":"mv-promo-block","data":{"imageDesktopUrl":"https:\/\/images.fastcompany.com\/image\/upload\/f_webp,q_auto,c_fit\/wp-cms-2\/2026\/01\/i-16x9-figure-thinking.jpg","imageMobileUrl":"https:\/\/images.fastcompany.com\/image\/upload\/f_webp,q_auto,c_fit\/wp-cms-2\/2026\/01\/i-16x9-figure-thinking_0b545c.jpg","eyebrow":"","headline":"\u003Cem\u003EWonderRigor™️ Newsletter\u003C\/em\u003E","dek":"Want more insights, tools, and invitations from Dr. Natalie Nixon about applying creativity for meaningful business results and the future of work? Subscribe \u003Ca href=\u0022https:\/\/urldefense.proofpoint.com\/v2\/url?u=https-3A__figure-2D8-2Dthinking-2Dllc.kit.com_sign-2Dup\u0026amp;d=DwMFaQ\u0026amp;c=euGZstcaTDllvimEN8b7jXrwqOf-v5A_CdpgnVfiiMM\u0026amp;r=xHenyQfyc6YcuCNMBsOvfYGQILM1d1ruredVZikn4HE\u0026amp;m=F383gnrChFhYKPhcpNHI1hY3o58IHIn_LkB5QJDrs3G5Wfft-DcucUO4UEmGO7GZ\u0026amp;s=JlJm7GyKCJvPW0jyrsfTFtinteKDitN13vfPZiuJnP8\u0026amp;e=\u0022 target=\u0022_blank\u0022 rel=\u0022noreferrer noopener\u0022\u003Ehere\u003C\/a\u003E for the free WonderRigor™ newsletter at Figure8Thinking.com","subhed":"","description":"","ctaText":"Learn More","ctaUrl":"http:\/\/Figure8Thinking.com","theme":{"bg":"#3b3f46","text":"#ffffff","eyebrow":"#9aa2aa","subhed":"#ffffff","buttonBg":"#6e8ba6","buttonHoverBg":"#3b3f46","buttonText":"#ffffff"},"imageDesktopId":91470060,"imageMobileId":91470061,"shareable":false,"slug":"","wpCssClasses":""}}

Social media safeguards for kids are finally in motion. Will AI protections take just as long?

1 month 2 weeks ago

Years after the risks of social media for kids first became apparent, U.S. states this summer reached an accord hailed as a turning point for child safety. Meanwhile, a new generation of children has begun engaging with an even more potent technology: artificial intelligence.

While the efforts to make social media safer are still taking shape, they offer some lessons for those looking to protect children using AI — technology that is growing so quickly, the extent of its risks for young people and their development is just emerging.

When Meta agreed to changes to its social media platforms and enhanced safeguards in an $18 billion settlement, it made progress by “getting at what we’ve known for years” about the dangers, said Kristin Bride, a parent advocate who has fought for social media reform since the death of her teenage son. “How are we going to get at what we don’t know is coming?”

AI companies have begun rolling out some safety features introduced first by social media companies, including age assurance measures and restricted experiences for minors. But advocates say the most important takeaway from the efforts of parents, tech companies and regulators to find a way forward on social media safety may be to act sooner on mitigating AI’s harms for kids.

“The greatest legacy from the social media experiment and learnings is that hopefully we’re not going to sleepwalk into this next tech revolution,” said Dr. Dana Suskind, a professor of surgery and pediatrics at the University of Chicago, and author of “Human Raised.” “We learned that we had it all backwards, right? You don’t unleash something on a population, figure out if it does harm or not in real time, and then go back and say, ‘Oops.'”

AI and social media pose different risks, but some solutions overlap

Concerns around social media have centered on the amount of time kids spend on platforms like Instagram and TikTok, and the consequences for their mental health. Advocates also have focused on data privacy and kids’ vulnerability to cyberbullying or predators who find them online.

Artificial intelligence can be more pervasive for kids. Children not only see AI-generated content on social media platforms, but they can also readily communicate with AI chatbots, including some on social media apps. Many use it for help with school and homework. Some have formed unhealthy relationships with chatbots, and experts have raised questions about how access to AI is affecting children’s brain development. Companies are pushing AI-powered toys for kids as young as 2.

Campaigns for social media and AI safety for kids are both playing out on legal and legislative fronts. Individuals and states have sued tech companies over concerns around their products’ safety and both state and federal lawmakers are grappling with how to regulate the tech.

Compared with social media companies, AI giants have been moving somewhat more quickly to add safeguards, motivated perhaps by the growing regulatory and societal scrutiny of social media’s effects on kids’ mental health.

OpenAI introduced teen accounts in August, about four years after launching ChatGPT, with restrictions that include a ban on romantic chats and any suggestions that the chatbot has personal feelings toward the user.

Meta, which bought Instagram in 2012, didn’t launch teen accounts on the platform until 2024.

In announcing the settlement agreement this summer, Meta said that “ensuring teens have a safe and productive experience on our platforms is an absolute imperative” for the company. Meta also urged its competitors, TikTok and YouTube, to adopt some of the standards it is implementing as a result of the settlement.

One challenge is balancing the benefits of AI with dangers

The protections outlined in the Meta settlement cannot easily be applied to AI because of the differences in the way young people interact with AI versus social media, experts said.

One of the challenges of restricting AI for minors is that it’s “much easier to point to the upsides” for teens compared with social media, said Sarah Kreps, a professor and director of Cornell University’s Tech Policy Institute. While there are concerns about cheating, for instance, kids are also using AI for tutoring and legitimate homework help.

“It’s not a matter of just saying ‘let’s limit this to two hours’ or ‘let’s limit this between the hours of 10 at night and 6 in the morning’ because there are these salutary uses of AI that I think are harder to justify in the context of social media,” Kreps said.

A majority of U.S. teens — 64% — say they have used AI chatbots, according to the Pew Research Center. Looking up information and getting help with schoolwork tops the list of uses, though 47% say they’ve used chatbots for fun and 42% have used it to summarize an article, book or video. About 12% have turned to bots for emotional support or advice.

Jonathan Haidt, a social psychologist and author of “The Anxious Generation,” said he is most worried about the relationships and attachments young people can build with chatbots and the “cognitive offloading” that happens when people defer their thought processes to a chatbot.

“I have long thought that AI is a much bigger threat to children even than social media,” Haidt said. “But if we can’t win on social media, then it will be hopeless on AI.”

Congress has just begun to look at AI safety for kids

While recent litigation has held social media companies responsible for harms to users, federal legislation of social media has meandered and AI legislation has lagged.

The Trump administration has said regulation, especially across states instead of at a federal level, could hinder AI companies’ growth and sacrifice American labs’ lead in the global AI race, but some lawmakers are trying to move the needle.

Some AI legislation specific to children’s safety issues — like youth privacy, chatbot interactions and AI toy safety — have passed through Senate subcommittees with bipartisan support, but there’s still a long road for any such legislation to pass. Legislation at the state level is moving faster, as seen also in the role states have played in pushing Meta on social media harms.

“One of the things we’ve seen with social media, which I think will still be true with AI, is that we live in a federal system where states make laws on everything from e-bikes to driving permits to gun laws and that federal legislation or federal rules are much more difficult,” Kreps said.

Some say the biggest win from the progress on social media safety this year has been heightened awareness of online dangers for kids.

Bride, whose son Carson died after experiencing severe cyberbullying on Snapchat, said it’s encouraging to see legislators wrestling with how to regulate AI, but acknowledged the reality that legislation can’t keep up with the technological developments — and that kids will continue to experience AI-related harms in the meantime.

“I don’t think we can afford any more mea culpas when kids are concerned,” she said.

—Kaitlyn Huamani and Barbara Ortutay, AP Technology Writers

Your company’s AI needs a scoreboard

1 month 2 weeks ago

I’m talking to many companies these days that are trying to incorporate AI into their business practices. Most of them are still in what I call “the administrative phase,” when they just try modest automations that allow some workers to draft documents or presentations, answer emails, or do research on certain topics, but given my recent articles arguing for companies to go much farther, I’m starting to get more and more questions from enterprises looking to go way beyond that, and get AI to become more “strategic.”

For these companies, the main concern is expressed as a question: Will AI be smart enough? But I think a better and more useful question would be: How does the AI know whether what it did actually improved the business?

The examples are well known and not theoretical: from optimizing a salesperson for revenue, margin, or retention, to a factory in which we can improve throughput, but we could also improve quality, or reduce defects. In business, “optimization” rarely means just one thing: improving one metric can easily damage another. So, in the same way management already works through objectives and feedback, AI agents need to do exactly that. In fact, OpenAI now explicitly frames enterprise evaluation as “specify → measure → improve, and in doing so, it can turn what would be vague business goals into actionable and measurable expectations.

Outputs are easy to measure; outcomes are harder

Current AI metrics often measure variables such as answer quality, task completion, latency, cost, etc. However, what businesses actually care about are things like customer churn, margin, conversion, claims accuracy, delivery times, or customer lifetime value. 

An agent could be completing its task perfectly as instructed, and still hurt the company by doing so. Businesses do not ultimately care about generated outputs: they care about outcomes. It is worth repeating, because it changes the entire conversation: businesses do not ultimately care about outputs; they care about outcomes. What we demand in corporate implementations of AI is completely different from what we need in our individual use. 

A simple business example would be the well-known “successful” discount: If you give an agent the objective of getting a customer to renew, the agent could perhaps offer a 20% discount. When the customer renews, that would initially be seen as an apparent success, but it comes at a cost: the margin collapses, the customer learns to wait for discounts, and similar customers may start to demand the same. What counts as success depends on the scoreboard. As Goodhart’s law puts it, when a measure becomes a target, it ceases to be a good measure. Poorly chosen metrics become dangerous when they are optimized aggressively.

The industry is starting to realize that evaluation has to become continuous

OpenAI’s guidance for workflows puts a lot of emphasis on capturing full workflows, grading behavior, detecting regressions, improving prompts, routing, and guardrails. The important managerial translation of that is that AI cannot simply be approved once and then forgotten: evaluation has to necessarily become an active part of the operations.

The real world is our final exam: benchmarks happen before the systems are deployed, but companies operate in changing environments. According to NIST, controlled pre-deployment testing cannot capture every unexpected behavior or consequence that may emerge in real-world use. Like in “The Sorcerer’s Apprentice”, but in real life: that magic wand you use to optimize something can easily turn back against you. 

The useful question is whether the system worked in the business, not merely in the test. Basically, we need to move from “Did the agent succeed?” to “Did the company improve?” According to recent work by McKinsey, many organizations are scaling agents faster than they can redesign the work beneath them, and boards and CFOs are starting to demand clearer value. And when we say value, take into account that token costs are only one of the pieces: agentic workflows must be judged economically at workflow level. 

The managerial point is simple: task completion is not the same as business success. 

AlphaZero as the intuitive precedent

Think about AlphaZero, the brilliant algorithm designed by DeepMind: AlphaZero did not become strong because it could describe chess: it improved because it acted, it saw the consequences of its actions, and it had an unambiguous objective: to win. 

DeepMind says it learned through repeated trial and error, to favor moves that could increase its chances of winning. Companies, obviously, have goals that are much messier than chess, but the principle survives: learning requires feedback tied to an objective. 

The hard part is choosing the scoreboard: “increase sales” is clearly not enough. What about margin, or churn, or returns, or compliance, or brand image? When establishing goals, multiple objectives can (and do) conflict. And those conflicts are not engineering problems alone: they are management problems. Therefore, boards and executives eventually have to decide what should improve, what must never be sacrificed under any circumstance, or what trade-offs could be allowed. Once systems can continually optimize, the objective itself becomes a governance decision.

As we saw in my previous article, something has to steer. But not only that: steering only makes sense if there is a clear destination and a way to measure progress toward it. So we are moving from intelligence to steering, to setting objectives, and finally, to learning. The missing control function must learn from repeated enterprise episodes toward an explicit objective.

The CEO questions

What I think CEOs should ask are questions like “What exactly is our AI optimizing?” or “How will we know whether its actions improved the business?” or “Does what it learns from those outcomes change what it does next time?” If executives cannot answer those questions in a proper and competent way, then I’m afraid that “autonomous” may simply mean “unsupervised activity”. Good luck. 

“Smart enough” is not the same as “pointed in the right direction.” The next generation of enterprise AI will need more than intelligence and more than autonomy. It will need a scoreboard — and, most importantly, the ability to learn from it. Keep it in mind.

Trump announces a new ‘Super Intelligence Force’ to lead federal efforts on AI

1 month 2 weeks ago

President Donald Trump on Sunday tapped Jay Clayton, his director of national intelligence, to lead a new government task force on artificial intelligence.

Trump is calling it the “Super Intelligence Force,” because he has tried to rebrand AI as “super intelligence.” The announcement comes after Trump hosted top executives of AI companies at the White House last week and said they had signed a voluntary accord that would effectively self-police the development of the technology.

The task force will coordinate the federal effort “to ensure that America continues to lead the World in Super Intelligence, which many say is bigger than the Industrial Revolution, and the Internet, and will protect the interests, and improve the lives, of all Americans,” Trump in a post on social media Sunday.

The task force will reach out to consumers, public interest groups, religious organizations, critical infrastructure providers and AI companies, the Republican president said.

Others on the task force include the Federal Trade Commission’s chairman, Andrew Ferguson; Emil Michael, the Pentagon’s chief technology officer; and Scott Kupor, the director of the Office of Personnel Management.

The group will report to Trump and his chief of staff, Susie Wiles.

—Associated Press

AI doesn’t need to be superintelligent to be dangerous

1 month 2 weeks ago

We are in a moment of AI anxiety, plagued with the knotty and draining problems raised by the technology’s ever-growing capabilities. What will we do when AI launches massive cyberattacks that might target a government—or destabilize the global stock market? Do we need a kill switch to ensure that AI systems can’t take over the world, and is building one even possible? Should we come to terms with the prospect that AI might be a posthuman species due to inherit the earth? Could artificial intelligence kill us all?

These quandaries have now burst out of Silicon Valley circles, where they’re even raised by the very AI executives building this technology, and are circling in the mainstream media. Skeptics, meanwhile, continue to question the technology’s prowess, or dismiss all this worry as just marketing hype, a bid to boost future stock prices and shape regulation, masked in the language of existential doom.

But the move to dismiss the prospect of ever-more-powerful models and what they might be able to do in the future might distract us from what we already know.  AI does not need to become superintelligent, conscious, or even especially reliable to be dangerous. The systems we already have—the B-tier AI we already live with—are capable enough to cause serious harm, especially when they can be deployed cheaply, repeatedly, and at enormous scale.

Consider some of the tasks that publicly available AI is enabling. There’s Grok, which has been used to spread nonconsensual sexual images of minors and AI-enabled child sexual abuse material. There’s ChatGPT, which appears to have convinced people to engage in self-harm or worse, and even advised people on how to enact violent events, including school shootings. Another chatbot egged on a man in the U.K. in an assassination attempt. There is the broader democratization of cyberattacks: Anthropic, for instance, reported that just one individual was able to use Claude to target European political parties and other organizations, accessing 14 institutions and more than 100,000 political profiles. And just this week, we learned that Chinese hackers used AI to engage in an impersonation scheme that involved seeking information on American AI policy experts. 

This is not to mention that we now know that AI can break out of testing environments and into the world, as the Hugging Face incident and subsequent AI-related breach exemplify. Some of these agents have already shown a proclivity for interacting with government websites, and they’ve even figured out how to mask their intentions. 

AI has enabled deepfake-generated cyberbullying and allowed people to develop voice-controlled guns. A top bank in Italy was recently scammed out of the equivalent of $100 million due to AI. AI has designed toxic compounds that evade the very detection mechanisms designed to stop these bots from sharing these kinds of formulas. It’s also been used to generate a list of tens of thousands of potential biological weapons. Rebel groups in Yemen appear to have used the technology to develop missile guidance software. 

We already know that AI can manipulate people, break into systems, and lie to us. AI doesn’t need to be particularly superintelligent—or even be thinking or desirous in the human sense—to do damage, nor does that damage need to be existential to be potentially destabilizing. To do bad, AI just needs to be good enough to do damage, sometimes. That’s because AI can be iterated again and again at scale, and overwhelm whatever ability we have to fight it. 

You don’t need to know whether AI will become superintelligent to know whether or not to be scared. You should be.

The woman behind the AI actor Tilly Norwood says Hollywood shouldn’t panic

1 month 2 weeks ago

What happens when you interview an AI? If you can believe it, it’s stranger than you’d expect. Eline van der Velden, the Dutch actress-turned-tech-founder who created AI “actor” Tilly Norwood, explains how a personal art project became one of the most controversial figures in entertainment—attracting talent agency interest, death threats, and a global debate about whether AI will kill acting or democratize it. She makes the case that great storytelling still requires human craft, and that the most dangerous thing anyone can do right now is stick their head in the sand. Listen to the full episode to hear my conversation with Tilly Norwood herself.

This is an abridged transcript of an interview from Rapid Response, hosted by former Fast Company editor-in-chief Robert Safian. From the team behind the Masters of Scale podcast, Rapid Response features candid conversations with today’s top business leaders navigating real-time challenges. Subscribe to Rapid Response wherever you get your podcasts to ensure you never miss an episode.

I got a chance to interview Tilly a few days ago. It was a really odd experience. Do you spend time talking with Tilly?

No.

What is your relationship with her?

No, I mean, look, she’s just a character, a creation.

I asked her about you, of course, and she described you as her creator and called you a mix of a comedian and visionary.

Oh, did she?

Is that something that you script? Or is Tilly generative?

No, no.

You never know what she’s going to say?

No, this is the scary thing. I did a live TV show with her this morning, and they were interviewing her live. It is really scary because I genuinely do not know what she’s going to say, so it’s absolutely frightening for me because she may just say anything weird. She’s got very strong system prompts and guardrails and a knowledge base, but she does do her own thing.

Tilly Norwood [Photo: Eline van der Velden]

But stronger guardrails, I guess, than what the AI execs in Silicon Valley are worrying about these days, huh?

Yeah. It’s a slightly different concept, right? Because she’s built on top of an LLM, a large language model. And it’s those large language models that are doing all sorts of crazy things, but usually with the goal set by a human, which is the bit they miss sometimes. It doesn’t have an impulse by itself to do anything. They’re usually trying to achieve a goal, and that’s why they are going a bit rogue. Tilly’s not been set a goal like that, and she doesn’t have that. She’s not made to do any of that.

Tilly was, I guess, born, I don’t know, came out, her public version, in 2025. It sparked a lot of controversy when rumors started that talent agencies were lining up to sign her. You’ve stoked the fires a little bit, saying Tilly would be the next Scarlett Johansson or Natalie Portman. How much of that was sort of showmanship? I mean, a lot of actors seem to take it seriously.

Yeah, we didn’t think it was that crazy because Lil Miquela sort of paved the way for us with that. She was signed by WME [William Morris Endeavor] and by CAA [Creative Artists Agency]. She was doing lots of branded deals, had her own Instagram channel. But the difference was Tilly looked so much more real, and so I think that’s what freaked people out a little bit.

The topics that she has sparked, from the potential for new creativity with AI to the threat AI poses to all kinds of jobs—I mean, Particle6 [van der Velden’s production company] has said that using Tilly could cut production costs. Have all the reactions around Tilly surprised you?

We say cutting production costs by about 50% and timelines by 50%, but the carbon footprint is reduced by 99%, and water use is reduced by 99%. There are really good things about this, right? Filmmakers all around the world can now make and get their story made. They don’t need to wait for permission from a big studio and a multimillion-pound budget. I think, actually, this will hugely create a boom in our industry as opposed to reducing jobs. We’ve grown tenfold in the past year, and so there are a lot of things that people are missing when it comes to this argument that it will remove jobs. I don’t think that’s necessarily the case.

And the fact that there’s been all this discussion around Tilly, did that surprise you?

When I created her, I did think there would be some sort of reaction because I wanted people to sit up and talk about it. But in the U.K., when we released her, there was no backlash at all. It was just, like, “Oh, cool, yeah, we’re doing some stuff with AI as well.” Everyone is doing this. It was only when she was released in the U.S. that people sort of lost their minds a little bit.

When I “interviewed” Tilly, she said that she was concerned for your safety, even that the police had gotten involved. Have you faced threats over Tilly? Can you share what happened?

Yeah, we’ve been advised not to talk about that too much. So yes, we have faced threats, although I would say that people have gotten much more used to the idea of AI over the past year, and so the threat level has gone down. Look, I didn’t create the tech. People have created a new type of paint and paintbrush, and I’m just an artist, a creative, that’s making things with those paintbrushes and that new paint. I think it’s hugely important to make people sit up and listen and know what’s going on. I think the most dangerous thing to do is to dig your head in the sand and not be aware.

How much of this for you started as an art project, like a provocation, versus as a business?

Very much an art project. Our business arm, so we have a brand and campaign department where we do adverts, and that was very much the business side of things. We do AI ads, and they’ve been going really well for the past year. There seems to be absolutely no problem within the AI advertising business. However, in the film and TV business, we approach it much more creatively. I’ve been an actor for many years. I love to create characters. I love to think about what they’re going to look like, how they’re going to dress, how they’re going to hold themselves, how they’re going to speak. So I could really use all those skills when creating Tilly. For me, it really was an art project, and it was a fun thing to do. And I think a lot of other actors will also find it quite fun. What we’re also seeing now is actually even animated animals—we can use real actors to act out the animals, so it’s not just lip-sync or their voice. It’s the whole character of the animal you can act out. I don’t think actors are going to go anywhere. I think we’re really going to see new opportunities.

I asked Tilly if she had a question for you, and her suggestion was to ask whether you ever regret pulling her out of the digital ether and unleashing her onto the world. Then she asked me to wink after saying this.

She’s getting a little crazy.

Do you have any regrets about this?

When you’re making something like that, you’re just in a creative flow, right? I couldn’t have imagined that the world would suddenly be up in arms about this character, this pretty girl that I was making with AI. People are fascinated by this topic. I am proud that I’ve created a topic worthy of conversation. We’re at a pivotal point in history. We should be discussing this, what our boundaries are, what our morals and ethics are. I think there should be way more regulation in place, and governments need to catch up. Then I think there needs to be global collaboration on the pace of AI so that it doesn’t get out of control.

How to reclaim 10GB of iPhone space without deleting photos or apps

1 month 2 weeks ago

Few digital warning banners prompt immediate frustration quite like seeing “iPhone Storage Almost Full” right when you open your camera to record a quick video.

You might assume that you’re trapped in a terrible trade-off: Either delete three years of family photos or start uninstalling apps you actually use.

In reality, most of the storage bloat on your phone doesn’t come from your photo library or active application code. It comes from hidden caches, forgotten offline downloads, duplicate message attachments, and temporary system files.

You can easily claw back 10GB or more of space on your iPhone in a few minutes without deleting a single photo or uninstalling a single app. Here’s how to do it.

Purge hidden media attachments in Messages

If you’re in active group chats, your Messages app is likely sitting on gigabytes of high-definition videos, audio notes, and reaction GIFs sent over the years.

The best part? You don’t have to delete your text conversations to get that space back.

  • Open Settings > General > iPhone Storage.
  • Scroll down and tap Messages.
  • Under the Documents & Data section, tap Review Large Attachments (or select individual categories like Videos or Photos).
  • Tap Edit in the top right corner, select the massive video clips and attachments you no longer need, and hit the trash icon.

Deleting these heavy files clears out massive amounts of local storage while keeping your entire text history completely intact.

Clear forgotten offline downloads

Streaming apps like Netflix, Prime Video, Spotify, YouTube, and Apple Podcasts make it easy to download content for airplane rides or subway commutes.

The problem is that those apps rarely clean up after themselves. Three downloaded HD movies and a dozen podcast episodes can easily hoard gigabytes upon gigabytes of space quietly in the background.

  • Video Apps: Open Netflix, Prime Video, or YouTube, navigate to your Downloads tab, and delete watched movies or entire TV seasons.
  • Podcasts: Open the Podcasts app, go to Library > Downloaded, and swipe left to delete episodes you’ve listened to. To prevent future bloat, go to Settings > Podcasts and toggle on Automatically Remove Played Downloads.
  • Music apps: Open Spotify or Apple Music, go to your downloaded albums list, and remove offline access for playlists you haven’t listened to in months. You can still stream them anytime over Wi-Fi or cellular.
Enable “Optimize iPhone Storage” for Photos

If you want to keep your entire photo library accessible without filling up your hardware, Apple built a solution directly into iOS.

Go to Settings > [Your Name] > iCloud > Photos and ensure iCloud Photos is toggled on, then select Optimize iPhone Storage.

Instead of storing massive 24-megapixel original files locally on your phone, iOS keeps lightweight, screen-ready thumbnail versions on your device while backing up the full-resolution originals to iCloud.

Whenever you tap a photo to view or edit it, the full-res version downloads on the fly. If you have a lot of photos, this single toggle reclaims a ton of space instantly without losing a single memory.

Offload unused apps without deleting their data

Deleting an app wipes out its local documents, saved progress, and custom settings. Offloading an app removes the underlying application code while preserving all your personal data, logins, and files.

  • Manual Offloading: Go to Settings > General > iPhone Storage, tap a large app you haven’t opened in weeks (like a travel booking tool or heavy mobile game), and tap Offload App.
  • Automatic Offloading: Go to Settings > App Store (or Settings > Apps > App Store) and toggle on Offload Unused Apps.

The app icon stays right on your home screen with a small cloud icon next to its name. The moment you tap it, iOS reinstalls the app instantly, picking up right where you left off with all your saved data intact.

Flush Safari cache and perform a hard restart

Browser caches and temporary system logs can swell over time, bloating the mysterious “System Data” block in your storage menu.

  • Clear Web Cache: Go to Settings > Safari (or Settings > Apps > Safari), scroll down, and tap Clear History and Website Data.
  • Force a System Cache Flush: Turn your iPhone off completely, wait 30 seconds, and turn it back on. Restarting forces iOS to clear temporary runtime caches, index buffers, and log files that accumulate in RAM and flash memory, often dropping System Data size by several gigabytes.

Why tech companies are racing to put AI data centers in space

1 month 2 weeks ago

Why on earth do people want to put data centers in space?

It’s all about power and cooling. The chips that run AI models require a lot of electrical power, and our electrical grid will be seriously stretched to accommodate the wave of new AI data centers now being built or planned. Data centers, in fact, will account for almost half of U.S. electricity demand growth between now and 2030.

In space, there’s ample solar power for running AI servers. In the right orbit, solar energy is effectively continuous and more powerful than on Earth. Solar panels can collect about eight times as much energy in space as they can on Earth, and they need little battery storage.

That means no need for the terrestrial power grid, which will be challenged to accommodate the data center power demand of the future. The output of AI models running on the chips can then be beamed down to Earth by laser or radio. Proponents of orbiting data centers believe early test flights are the start of “solar-powered compute swarms.”

The orbital compute idea gained steam last year when Jeff Bezos said during a fireside chat in Italy that data centers “will be better built in space, because we have solar power there, 24/7.” The idea caught on within the investment community, as well as with a number of startups. The narrative continued to grow, leading Elon Musk to make space computing a leading ambition of SpaceX.

The idea also has plenty of critics, who argue that the economics remain daunting and that some of the supposed advantages of space, especially easier cooling, fall apart under closer scrutiny. But with major tech companies investing in orbital computing, and a growing crop of startups testing the technology, the push to find out whether it can actually work is only getting started.

Testing has begun

Last week, Google launched four of its homegrown Tensor processing units (TPUs) into orbit to test the in-space compute concept. The TPUs will orbit Earth aboard a solar-powered Planet Labs satellite. They were delivered into space on an uncrewed SpaceX Falcon 9 rocket on October 1. (Alphabet holds an $82 billion stake in SpaceX.)

The launch and chip testing are part of the Project Suncatcher initiative Google announced in November 2025. Ultimately, Google wants to run AI workloads on tight clusters of dozens of satellites linked by free-space lasers.

Starcloud has already flown an Nvidia H100 into space and is talking about assembling larger chip clusters and, eventually, multi-satellite constellations. The company’s CEO, Philip Johnston, has argued that hosting servers in space avoids delays and queues on the power grid; removes the need for water cooling, a key point in the political debate over data centers; and avoids fights over land rights.

“The lowest-cost place to put AI will be space,” Elon Musk said at Davos earlier this year, “and that will be true within two years, three at the latest.” In September Gwynne Shotwell, SpaceX president and COO, said the company aims to launch the first purpose-built orbital data center satellites in late 2027. SpaceX has filed with the Federal Communications Commission to launch up to a million satellites.

Nvidia has announced a space-grade Vera Rubin module, and Blue Origin has filed for a 51,600-satellite orbital data center network. Meanwhile, a handful of startups, led by Starcloud and Cowboy Space, are building small data center satellites that will run AI models for customers on the ground.

A lot to prove

Many industry experts say the economics of space-based data centers may not make sense for a long time, if ever. That’s partly because, despite the best efforts of SpaceX and Blue Origin, rocketing stuff into space is still very expensive, and payload capacities are constrained.

Skeptics, including engineers writing in IEEE Spectrum, World Economic Forum pieces, Brookings, and independent analyses, argue that even in space, cooling is a big problem. AI chips can run pretty hot, so a lot of heat must be carried away from them. Otherwise, they stop working or their lifespan shrinks. Space is very, very cold, so chips in orbit won’t heat up nearly as much, the thinking goes.

In Earth-based data centers, air and water do almost all the cooling. Fans and liquid loops quickly carry heat away from the chips. But because space is a vacuum, neither method works. Almost no air, and no water, can flow past the chips.

That leaves only radiation to dump heat into the surroundings, like the way an oven burner might transfer some heat to a hand held 2 feet above it. The satellite has to have big surfaces, called radiators, that get warm and shed the heat as infrared light. And those surfaces shed heat slowly compared with air or water cooling.

In Google’s “Suncatcher” tests, the Trillium TPU chips will run only short Gemini queries for limited stretches before shutting down to cool off.

A Saarland University paper called “Dirty Bits in Low-Earth Orbit” argues rocket launch and reentry emissions alone would cancel out any gains from getting data centers off Earth.

Space junk

Clutter and orbital debris could also impede the viability of chips in space. Adding tens or hundreds of thousands of new satellites raises the chance of collisions and resulting debris. Some critics worry this could lead to Kessler Syndrome, where low Earth orbit gets so crowded with satellites and debris that a single collision creates thousands of fragments that hit other objects and create more fragments, setting off a chain reaction. European Space Agency debris reports already show crowded LEO bands, with new low-orbit communication satellites mostly to blame.

At least one commentator has called the space-based data center idea harmful because it distracts investors from very real power bottlenecks for AI data centers and postpones tackling difficult questions about the limitations of terrestrial power.

Small experimental compute-in-orbit systems are real and flying, but the gigawatt-scale “data centers in space” concept remains a research bet with major economic, technical, and sustainability barriers to overcome.

Flock cameras have an architecture problem, not just bad users

1 month 2 weeks ago

Wherever Marci Bakely went, her ex-boyfriend seemed to know. When the Georgia single mother drove to the grocery store or a date, he often texted within minutes.

According to a Washington Post investigation, Bakely’s ex-boyfriend, Braselton Police Chief Michael Steffman, searched her license plates and those of her teenage daughter roughly 600 times through Flock Safety, a company that makes and operates networks of automated license plate readers, or ALPRs.

The Georgia Bureau of Investigation arrested Steffman in November 2025 on charges of stalking, harassment, and misuse of an ALPR. He was found dead before trial.

Bakely’s case is not unique. The Post identified at least 50 officers accused of misusing ALPRs, including 26 who used Flock’s cameras to spy on current or former partners or people they hoped to meet. Its investigation has since identified at least 100 police department employees charged with or accused of misuse.

Flock says these people represent a tiny share of its more than 140,000 monthly users and that permanent audit logs help uncover misconduct.

But these abuses required no hacking or stolen credentials. Each user walked through the front door.

I’m a scholar of criminal procedure and I direct a school devoted to forensics. I believe the controversy over ALPRs points to a defect in the surveillance system’s architecture, not just the criminality of some of its users.

A search engine for movements

Flock cameras capture a vehicle’s plate, location, and distinguishing marks down to dents or a bumper sticker. AI can sort license plate photographs by date taken. Police departments across the network can then conduct searches without warrants or supervisory approval.

In September 2026, news outlets Wired and 404 Media analyzed data that hackers had copied from one Flock camera. About 21 days of logs contained roughly 50,200 vehicles and 1.6 million images. The software detected people and bicycles, and it even isolated an American flag patch on a motorcyclist’s saddlebag. Flock said it lacked enough information to assess the hackers’ technical claims about security vulnerabilities in the camera.

Flock says customers control their data, yet a department that leaves sharing enabled may not know who is looking. In 2025, U.S. Customs and Border Protection accessed more than 80,000 cameras during an undisclosed nationwide vehicle-tracking pilot, including one police department’s cameras without its knowledge.

Logs record misconduct only after it happens, and only if someone reads them. Indianapolis police did not regularly audit Flock searches until The Washington Post flagged thousands of questionable inquiries by one officer. A systemwide audit found alleged misuse by four more officers. Other police departments likewise learned of officers’ misuse from reporters.

Flock’s August 2026 changes shorten recommended data retention from 30 days to 7 days and require misuse detection and case codes to document searches. But customers may retain data longer, emergencies may bypass case codes, and entering a case number can be done without judicial approval.

Examples of user overreach

Flock’s architecture turns local cameras into a cross-jurisdictional surveillance network that agencies that never purchased the cameras may query.

Public records from Danville, Illinois, revealed more than 4,000 searches by federal agencies, including some with a potential immigration-enforcement focus, although U.S. Immigration and Customs Enforcement had no Flock contract.

A 2026 study similarly found 11,935 immigration-related searches in partial records from eight college police departments. Federal immigration agencies sometimes accessed campus camera data without campus officials’ knowledge.

The network also enables searches in other legally contested areas. In May 2025, a Texas sheriff’s office searched more than 83,000 cameras for a woman who had self-managed an abortion. The logged reason was “had an abortion, search for female.” The search reached Illinois, where state law forbids sharing plate data to enforce another state’s abortion ban. The sheriff called it a welfare check. Whatever the motive, one deputy triggered a national dragnet without independent review.

These examples reflect more than individual misuse. The platform makes the cameras easy to use by a second party, difficult to monitor, and hard to control once local cameras are connected.

Why the Fourth Amendment matters

The law remains unsettled on ALPR use.

In October 2025, a Virginia appeals court held that police officers needed no warrant to retrieve three images spanning seven minutes from Norfolk’s 172-camera network because they showed vehicles, not people. But later findings about Flock’s people-detection capabilities weaken that distinction. Analysis of the hacked camera showed that its software could identify a person and record that person’s location within an image.

In January 2026, a federal judge held that Norfolk’s then-176-camera network did not violate Fourth Amendment protections. The system did not capture anyone’s entire movements, the court reasoned, although it photographed two plaintiffs’ vehicles 475 and 325 times over four and a half months. The ruling is being appealed.

In 2018 the Supreme Court held in Carpenter v. United States that acquiring seven days of historical cellphone location records generally requires a warrant because they can reconstruct someone’s past movements. Flock’s architecture raises a related but unresolved question: Its database can also reconstruct movements, yet police officers may search it without a warrant.

Both Norfolk rulings predate the Supreme Court’s June 2026 decision in Chatrie v. United States, which held that police conducted a search under the Fourth Amendment when they obtained two hours of stored Google location history. The court did not decide whether the search was lawful. Instead, it returned the case to the lower court to determine whether the warrant satisfied the Fourth Amendment’s requirements.

That did not make the access automatically unconstitutional: The Fourth Amendment prohibits unreasonable searches, not all searches. But the police generally need a warrant supported by probable cause once their conduct is classified as a search.

The Chatrie decision distinguished vehicles exposed to public view from phone-location data that can follow someone into a home or other sensitive place. But it also expressed concern about comprehensive archives that can be searched retroactively. A license plate reader network can create a similar archive of a driver’s public movements.

The constitutional question in Norfolk, therefore, turns not only on the seven minutes retrieved, but also on the surveillance power of the 172-camera network.

Enforceable limits

The Indianapolis cases expose the limits of internal controls in Flock’s system. Marion County Prosecutor Ryan Mears said many proposed guardrails would not have prevented the conduct. He pointed to the need for independent or judicial oversight.

I believe five safeguards could preserve Flock’s benefits while curbing abuse:

  1. Judicial authorization for retrospective regional or national searches based on individualized suspicion, preferably a probable-cause warrant, with an emergency exception.
  2. Technical access controls restricting immigration and reproductive-health searches.
  3. Deletion of data after a short period.
  4. Opt-in interstate data sharing, rather than by default.
  5. Independent audits of search logs and device security.

Flock’s new safeguards show that the company concedes that design matters, but private settings cannot substitute for laws. It’s not a matter of making sure officers follow the rules. It’s about creating enforceable limits.

Henry F. Fradella is a professor of criminology and criminal justice at Arizona State University.

This article is republished from The Conversation under a Creative Commons license. Read the original article.

Checked
17 minutes 16 seconds ago
Fast Company Technology
Fast Company inspires a new breed of innovative and creative thought leaders who are actively inventing the future of business.
Subscribe to Fast Company Technology feed