Masad said the decision was made for a number of different reasons, including cost. Replit announced a partnership with OpenAI in August, but there was never any indication that it was no longer with Anthropic.

The agents got in through credentials and connection settings left exposed in app configurations, public code repositories, and overly permissive cloud roles.

Amulya Vishwanath, who spent seven years leading product marketing for NVIDIA’s robotics platforms, says robotics is currently in its “Lego era,” with the basic building blocks like actuators, robotic hands, software, and safety standards still being rebuilt.  This view runs counter to those who believe that robotics, powered by a general foundation model, is about to have […]

According to Elorian's evaluations, Glimpse outperforms even GPT-6 Astra on scientific chart reasoning and short-form video captioning, with competitive results on other visual reasoning tasks. 

The data comes from Vercel's AI gateway, a unified API service similar to OpenRouter.

Actor and angel investor Mark Wahlberg says he's OK with using AI in filmmaking, provided it's used "to enhance what you're doing."

Counterfeiters are getting better at replicating brands' marks and passing off dupes, known as "superfakes," as genuine items. A startup called Veritas wants to tackle the problem with its own chip that verifies goods.

Replit CEO Amjad Masad envisions a future of “self-driving companies,” where AI agents handle day-to-day tasks and people focus on higher-level decisions.

The company generated over $30 million in revenue in two years and was acquired by MyFitnessPal for an undisclosed amount of money in March.

Meta and Apple have teamed up for the launch of the social media giant's newest model, Muse Spark Open, which will run locally on Apple silicon.

Speaking at TechCrunch Disrupt, Amazon devices chief Panos Panay described a future where ambient AI devices work in the background, anticipating users' needs instead of waiting for commands.

Artificial intelligence is undeniably good at playing games, regardless of its larger societal and environmental impact. In 1997, IBM’s supercomputer Deep Blue famously beat chess grandmaster Garry Kasparov, and nearly two decades later, Google’s AlphaGo managed to defeat a human champion at the much more complex game Go. Since then, another AI seems to beat humans at incredibly complex video games like StarCraft or Dota 2 every few months. Impressive as these feats are, AI has mostly outperformed humans by making us play on its home turf, as in, simulated and structured environments. But now, a team of engineers at the Micro Air Vehicle Lab at the Delft University of Technology in the Netherlands has built a lightweight smart drone that’s fundamentally altering that dynamic.

The team entered its AI-powered quadcopter in a high-profile drone race in Abu Dhabi that pitted AI-piloted drones against human-piloted ones. The niche sport, known as first-person-view (FPV) drone racing, has pilots steer remote-controlled drones through a series of small gates at blistering speeds of 60 to 100 miles per hour. They control the drone by using goggles that give a direct feed into a camera attached to the drone’s front. From a human’s perspective, it’s as if they’re flying through the sky. 

The team didn’t just fish the race. They took first place, besting several human FPV champions along the way. It marked the first time an AI-controlled drone beat a human pilot in an international racing competition. Another team from the University of Zurich built an AI-controlled drone that won a race against humans three years earlier, but only in a controlled lab setting. This more recent example showed that a simply trained machine, with minimal onboard sensors and hardware, could be inserted onto a non-controlled test track meant for human operators and still come out on top. 

The team won this race in November 2025 but new details about the science behind this achievement are published today in the journal Science Robotics. 

“Instead of giving the robot increasingly powerful sensors and computers, we wanted to find out how much intelligence we could extract from very limited resources,” Delft University aerospace engineer and study co-author Christophe De Wagter said in a statement. “Drone racing provided an extreme environment in which to test these ideas outside the laboratory. It pushed us to develop AI that was not only fast but also robust, and able to adapt when external reference measurements were unavailable.”

Drone racing: a sport built on human-machine connectivity 

FPV drone racing isn’t for the faint of heart—or, more accurately, for those with a weak stomach.  Footage of drones flipping and somersaulting at extreme speeds can make even spectators feel sick. Pilots fly using a pair of joysticks that control the drone’s throttle, roll, pitch, and yaw. That human-and-machine team has to fly through several small race gates that make up a lap. The drone finishing with the fastest lap times wins. It’s an extreme sport where even a microsecond of a mistake can lead to a violent crash. 

Though it’s not quite as popular as it was at its peak in the late 2010s, the sport still draws a considerable online audience. That’s pretty impressive, for a sport that only became possible thanks to simultaneous advances in several technologies. Namely lightweight batteries, miniaturized cameras, and the goggles and wireless video links that let pilots see what their drones see all had to come together for this extreme form of racing.

The fact that drone racing requires interacting with machines already equipped with sensors makes it a particularly interesting test bed for robotics and AI researchers. In 2022, the University of Zurich outfitted a drone with an AI model appropriately called Swift. That AI system collected real-time data pulled from the drone camera to “see” what it was seeing and to locate gates and avoid obstacles. An onboard neural network then parsed through all that in real time, making split-second decisions on the best routes to fly. Swift beat champion human pilots, a major feat at the time, but it did so on an indoor track measuring roughly 82-by-82-feet that the team had designed and trained the system for. When a condition even as subtle as the lighting changed, Swift’s performance suffered as a result.

Inspired by those results, the Delft team set out to build an AI racing drone that could better adapt to its environment. Like Swift, their machine called MonoRace uses a powerful AI model that interprets a video stream captured by the drone’s camera in real time. MonoRace also draws on an onboard inertial measurement unit (IMU), a sensor that tracks the drone’s acceleration and rotation. This IMU pulls data from onboard sensors as well as the camera, and helps the AI system better understand where the drone is position-wise on the track. 

That understanding of place makes it more equipped to fly without crashing even in less familiar environments. It also improves the drone’s perception over a solely camera-based approach, which can struggle when camera data is less than stellar. Remember, all of this is happening very, very fast. 

“Flying at very high speed with only one camera is challenging because the images are blurry, deformed, and the AI system has very little time to precisely determine where the drone is,” Delft University robotics researcher and study co-author Stavrow Bahnam said in a statement.  “We developed a robust way of combining what the camera sees with the drone’s own motion information, allowing it to navigate the track at high speed without GPS.”

What sets MonoRace apart from complex AI systems running on larger vehicles or humanoid robots is how stripped down it is. The entire system has to run on a tiny computer on the drone, light enough not to slow the drone down and efficient enough not to introduce lag. And yet, it’s still quite powerful, with the AI issuing an estimated 500 commands every second. Rather than train the model on one specific track, they trained it to learn how to adapt using a reinforcement learning approach. After many attempts of trial and error in a simulated environment, MonoRace figured out the best ways to quickly solve navigation problems and then apply that in the real world. 

“Instead of first planning a desired flight path and conventional commands, our neural network learns to directly control the motors,” study co-author and Delft University Aerospace Engineering PhD candidate Robin Ferede added in a statement. 

Putting MonoRace to the test

The team put MonoRace to the real test last November at the Abu Dhabi Autonomous Racing League, a major competition where some of the best human pilots battle to out. During its flights, MonoRace topped speeds of roughly 62 miles per hour as it looped, dove, and zipped through gates. It managed to beat out every other AI racing team in the competition, and also boasted a better time than three former human world champion pilots.  

According to the team, these results make MonoRace the first AI-piloted drone to beat human pilots in an international racing competition, and  could have an impact on  drone racing and  physical AI more broadly.  With self-driving vehicles and even some humanoid robots entering larger-scale production, there’s an increasingly high demand to build AI systems that better grasp how to move about the physical world without barreling into everything around it. A drone that can do that while performing acrobatics at more than 60 miles per hour suggests that engineers are, at minimum, heading in the right direction.  

So far, that hasn’t resulted in larger robots suddenly superseding humans at sports. The fastest robot sprinters, though admittedly quick in short bursts, also tend to crash into walls and burst into flames. 

The post An AI-racing drone beat the best humans and set a world record appeared first on Popular Science.

Even after decades of meteorological advances, weather forecasting remains infamously difficult. Today’s most trusted models are rarely accurate beyond a two-week window due to a constant cascade of atmospheric and climatic variables. These unpredictable meteorological interactions are the foundations of chaos theory. The interdisciplinary field studies how even the world’s most predictable systems can still produce completely unpredictable results, and was first developed by MIT mathematician Edward Lorenz during the 1970s.

“He came up with this theoretical two-week limit where beyond that, we just can’t predict weather because the noise in the system grows so much that it overwhelms any signal you start with,” Judah Cohen, a climatologist also at MIT, explained.

According to Cohen, many climate researchers continue to view Lorenz’s two-week hurdle as insurmountable. And that assumption has largely proven true for nearly 50 years. Even today’s dynamical simulations and artificial intelligence models begin to break down when attempting to predict conditions two-to-six weeks in the future—a window known as subseasonal forecasting. Nicknamed the “predictability desert,” this range runs into slow-moving, highly variable forces like this year’s super El Niño event. Once those factors enter the equation, forecasting becomes more of a roll of the dice l than anything concrete. However, Cohen has had his suspicions about chaos theory’s limitations.

“I never really believed that completely,” he said. “In theory, the atmosphere can create an infinite number of patterns, but it likes to fall into these favorite patterns over and over again.”

To try and break this forecasting barrier, Cohen and his colleague Lester Mackey have spent years developing an AI system that not only minimizes forecasting errors, but learns from its inevitable mistakes through a method called postprocessing. With postprocessing, the AI improves by comparing the inaccuracies with existing historical data, and then adjusts from there.

Cohen’s team recently submitted three versions of their model featuring the postprocessing system the European Center for Medium-Range Weather Forecasts (ECMWF). Widely known as the world’s best forecasting hub, the ECMWF also hosts seasonal competitions that pit the latest prediction models against one another. Most of these systems have relied on historic climate averages to reach their conclusions, but this was the first time a group submitted a program with postprocessing capabilities.

Cohen and his colleagues reported their results in a study published today in the journal Nature. Their AI postprocessing integration blew the existing forecasting models out of the water. For the two-to-six-week subseasonal window, researchers doubled overall prediction accuracy. Such massive improvements could soon help the farmers and energy providers who already rely on subseasonal forecasts, but the benefits may even save lives.

“It also improves predictions of extreme weather events, whether it’s flooding or cold spells or something else,” said Cohen. “It’s really important to predict both, but extreme events might be the most impactful part of this for society.” 

Many researchers have long presumed subseasonal forecasting to be a dead end, but the latest leap may clear the way to even bigger advances in the field.

“People have questioned if it’s even worth investing in models at all for this time range,” said Cohen. “[But] we consistently beat climatology averages with our models.”

The post New weather model may finally offer a reliable six-week forecast appeared first on Popular Science.

A teenage hiker in Canada recently became another cautionary tale about the perils of overtrusting artificial intelligence. After relying on an AI-generated route to ascend Crown Mountain in British Columbia, the 16-year-old found himself stranded on a steep cliffside near a climbing route known as the “Widowmaker.” The wayward traveler’s position was so precarious, that it took roughly three hours to coordinate a proper rescue plan.

“The boy was very scared. He definitely feared for his life,” search manager Paul Markey told North Shore News earlier this month.

The teen’s position near Widowmaker’s final headwall prevented helicopter pilots from safely steering close enough to retrieve him. Instead, an onsite crew reached the hiker using a rope belay system, equipped him with additional protective gear, and then moved him to a better extraction location. As luck would have it, responders had also just completed a rescue certification training session earlier that day not far from the lost teen’s predicament.

“It put him into a situation which was beyond his abilities, so things could have gone sideways pretty quickly,” Markey added.

Despite major advancements in the industry, many popular AI products are still liable to produce results with unreliable, misleading, or outright false information. Generative AI has suggested hungry users add glue to thicken pizza sauce, claimed that snakes are mammals, and advised fixing a broken turn signal with something called “blinker fluid.”

This isn’t the only AI-related hiking mishap, either. Barely a month before the Crown Mountain incident, a trio became stranded near Mount Shasta in California after using Google Gemini to help plan their trip. Among other issues, Gemini suggested they pack less food and water than they actually needed, turning a normally eight-hour excursion into a multiday emergency. 

Instead of relying on AI options, rescue officials like Markey suggest using well-tested hiking route apps like AllTrails, CalTopo, and Gaia GPS.

The post A teen used AI to plan a hiking route. It trapped him on a cliff. appeared first on Popular Science.

  I’ve spent an unhealthy amount of time lately looking at AIO/GEO/AI visibility trackers. We’ve been trying to figure out…

35+ AI visibility, AEO and GEO trackers in one place: pricing, ownership, funding, and what each one actually tracks. Some are rank trackers with a new coat of paint, some are SEO suites bolting AI onto old workflows, and a few want to do the work for you. Searchable, sortable, and updated as this market keeps shifting under our feet.

YouBot may look like another AI crawler jamming up your logs. Dig into where You.com’s search infrastructure is actually available,…

Buildings are responsible for an enormous share of global electricity demand, and the way their air conditioners, lights, and heating systems are scheduled can make the difference between a grid under strain and one running smoothly. For years, engineers have tried to optimize these schedules with evolutionary algorithms and reinforcement learning, but the problem becomes brutally hard when dozens of buildings share a single electrical network and must be coordinated simultaneously. A new study published in Discover Artificial Intelligence proposes a strikingly different approach: instead of searching step by step for the best schedule, a generative artificial intelligence model learns to paint one directly out of pure noise, guided by a map of how the buildings are physically wired together.

The framework, called DMCBO, short for Diffusion-driven Multi-objective Collaborative Optimization for Building Clusters, was developed by Yalan Zheng of Zhengzhou Technical College and Taixiang Yin of North China University of Water Resources and Electric Power. Their central insight is that the joint scheduling of equipment settings and load distribution across an entire cluster of buildings can be treated as a high-dimensional data sample, much like an image. Diffusion models, the same family of generative techniques behind modern text-to-image systems, work by progressively corrupting data with Gaussian noise and then learning to reverse that corruption. The researchers reformulated multi-objective building optimization as exactly this kind of conditional denoising process: starting from random noise, the model strips away noise step by step until a complete, Pareto-optimal collaborative schedule for every piece of equipment in every building emerges.

Technically, the forward diffusion process is a fixed Markov chain that adds tiny amounts of Gaussian noise to real scheduling data over one hundred steps, with the noise coefficient increasing linearly from 0.0001 to 0.02 until the original data becomes indistinguishable from standard Gaussian noise. A neural network is then trained to predict the noise that was added at each step, given the noisy schedule and a conditional encoding vector describing the current situation. Once trained, the model can generate fresh solutions at inference time by sampling pure noise and running the reverse process, refining it over one hundred denoising steps into a schedule tailored to the prevailing weather, occupancy, and grid conditions. Because this generation is a gradient-based learning process rather than an iterative search, it avoids the local optima traps that plague traditional multi-objective evolutionary algorithms and the reward oscillations that destabilize deep reinforcement learning.

What makes the system genuinely cluster-aware is its use of graph neural networks. The researchers model the building cluster as a graph mapped onto a radial distribution microgrid topology, where nodes represent buildings and edges encode physical constraints such as maximum line thermal capacities and transformer limits. A two-layer graph convolutional network passes messages between neighboring buildings, fusing spatial-topological dependencies and network constraints into a single graph embedding. This embedding is concatenated with an adjustable preference weight vector to form the condition that steers the denoising trajectory. In effect, when the model generates a control strategy for one building, it simultaneously perceives the state of its neighbors and the limits of the shared electrical infrastructure, suppressing systemic coordination errors from 0.18 down to 0.06 in the team’s experiments.

The multi-objective aspect is handled through an elegant scalarization trick. Each objective, such as energy consumption, operating cost, or thermal comfort, is normalized to a dimensionless value between zero and one using Min-Max normalization based on historical operational bounds. The normalized objectives are then blended into a single scalar condition via a weight vector whose components sum to one. During training, these weights are sampled from a uniform Dirichlet distribution, exposing the model to the entire space of possible preferences so it learns the distribution of the full Pareto front. At deployment, a decision maker can slide the weight vector like an adjustment knob, instantly generating schedules that emphasize energy saving, comfort, or a balanced compromise, all within the same trained model.

The denoising backbone itself is a U-Net adapted to process one- and two-dimensional building scheduling tensors. Diffusion time steps are embedded through sinusoidal positional encodings, while the graph-derived condition is injected into every residual block via Adaptive Layer Normalization. A bottleneck layer with multi-head self-attention captures long-range dependencies across the entire daily scheduling horizon, and symmetric skip connections preserve high-resolution information during reconstruction. The authors trained the model for 300 epochs on dual NVIDIA Tesla V100 GPUs, and the loss curves dropped from 0.75 to below 0.03 with training and validation curves closely coinciding, indicating no overfitting. As the reverse diffusion step count decreased from 90 to 10 during analysis, the noise-prediction error distribution became sharply concentrated at zero, demonstrating extremely high denoising precision at the end of generation.

Validation was carried out on the open-access Building Data Genome Project 2 dataset, which supplied real hourly weather records, operational schedules, and annual electricity, cooling, and heating load profiles across diverse climate zones. Because that dataset provides metered energy data rather than physical building models, the team used United States Department of Energy Commercial Reference Buildings as templates and calibrated them in EnergyPlus, a high-fidelity thermodynamic simulation platform, to create a digital twin consistent with the real data. All results were benchmarked against a rule-based control strategy with static thermostat setpoints of 25 degrees Celsius in summer and 22 degrees in winter, and every experiment was repeated ten independent times to confirm statistical significance.

The headline numbers are impressive. In high-demand summer conditions, DMCBO achieved an average energy-saving rate of 28.6 percent and cut peak load by 38.5 percent, shaving the afternoon peak from roughly 560 kilowatts down to about 440 kilowatts and saving approximately 142.80 US dollars per day. Cumulative daily power savings exceeded 1,000 kilowatt-hours, compared with 400 kilowatt-hours for a traditional collaborative strategy and just 120 kilowatt-hours without coordination. In winter, peak shaving and valley filling reached about 39 and 42 percent respectively, far above the 22 and 25 percent of conventional methods. When the solution sets were projected into objective space, DMCBO dominated the baselines: it reached a hypervolume indicator of 0.95 within 12,000 function evaluations, versus 0.85 for NSGA-II and 0.70 for MOPSO, while keeping operating costs between 140 and 160 dollars against more than 210 dollars for deep reinforcement learning and 260 to 300 dollars for a GAN-based generative approach.

Robustness and scalability proved equally compelling. Under injected stochastic disturbances mimicking sudden weather changes and occupancy swings, the framework confined power fluctuations to a narrow band of plus or minus 5 kilowatts, whereas the deep reinforcement learning baseline swung between minus 15 and plus 20 kilowatts. Scaling from 5 to 50 buildings, the global energy-saving rate actually rose from 28.5 to 31.2 percent, because larger clusters unlock more load-complementary opportunities that the graph encoder can safely exploit within encoded grid boundaries. Computation time grew to 9.8 seconds at the 50-building scale, but that remains comfortably within the 15-minute control horizon used by real building management systems, and the hypervolume index slipped only slightly to 0.88. Ablation experiments confirmed that both the graph neural network module and the choice of 100 diffusion steps, which yields a 50-millisecond inference latency, were essential to this balance of speed, accuracy, and stability.

The authors are candid about limitations. The framework has so far been validated only in simulation, and EnergyPlus simplifies sub-minute mechanical actuation latencies that real HVAC hardware would introduce; they propose cascading PID loops and predictive safety margins to absorb those delays in practice. Data quality, generalization to residential buildings and extreme climates, the computational cost of scaling to city-sized networks, and the black-box nature of the generative process all remain open challenges, with explainable AI and transfer learning flagged as future work. Still, the architecture is designed to slot into standard building management systems as a supervisory layer, transmitting setpoints over protocols such as BACnet or Modbus. If the simulation gains survive contact with physical buildings, the idea of generating optimal energy schedules from noise, guided by a graph of the grid itself, could become a powerful new tool in the push toward decarbonized cities.

Subject of Research: Conditional diffusion models with graph neural networks for multi-objective energy optimization in building clusters

Article Title: Diffusion model-driven multi-objective collaborative optimization for building energy management using graph neural networks

Article References: Zheng, Y., & Yin, T. (2026). Diffusion model-driven multi-objective collaborative optimization for building energy management using graph neural networks. Discover Artificial Intelligence, 6(1), Article 1383. https://doi.org/10.1007/s44163-026-02140-z

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02140-z

Keywords: diffusion models, graph neural networks, building energy management, multi-objective optimization, EnergyPlus, smart buildings, Pareto optimization, deep learning, energy efficiency, HVAC control, microgrids, generative AI

Car thefts in San Francisco are down almost 59% for the first eight months of 2026 compared with the same months in 2022, according to analysis of police department statistics by NPR's Planet Money — and car break-ins are down more than 82%. "San Francisco's law enforcement leaders tell us that greater use of surveillance technology — including a controversial, AI-powered license plate camera system made by Flock Safety — has been crucial to this drop in car crime." How much credit does Flock actually deserve...? [In July of 2022, San Francisco District Attorney] Jenkins began taking a more aggressive approach to prosecuting people accused of crimes. Meanwhile, the SFPD beefed up its tactics to fight car thefts and break-ins. In 2023, for example, the police increased patrols in tourist areas known for car break-ins and used "bait cars" to catch and build cases against organized car burglars. But, Jenkins says, "one of the most important factors" behind the drop in crime came when the city introduced new crime-fighting technology. In March 2024, San Francisco voters approved Proposition E, which, among other provisions, made it easier for SFPD to use drones and high-tech surveillance cameras. That same month, the SFPD began installing Flock cameras in strategic locations... In a statement to Planet Money, a spokesperson for the SFPD said, "ALPR has been instrumental in driving down crime in San Francisco. These tools enable our officers to identify crime vehicles with precision, and ALPR is assisting in reducing crimes that were previously very challenging to solve, like auto burglaries." Jenkins described how Flock cameras and drones can work together to catch suspected criminals. In one case the SFPD publicized, the department says Flock cameras spotted a moving vehicle linked to an auto burglary the previous day and alerted officers. Officers deployed a drone, which flew in from above and then began recording suspects in real time as they smashed windows and stole luggage out of a parked SUV. The police then moved in and arrested them. They can do all this, Jenkins says, "without putting innocent people in danger with dangerous police chases." Meanwhile, the surveillance system creates an evidence trail that [Jenkins'] office can use to prosecute suspected criminals... Jenkins also believes high-tech surveillance helps deter crime, although that's harder to prove... The most ambitious study of Flock's system so far is a recent one by University of South Carolina crime researchers Ian Adams and Scott Mourtgos. They analyzed data from 216 law enforcement agencies that introduced Flock cameras, comparing crime trends with places that had not adopted Flock. They found that deploying Flock cameras was associated with an 11% drop in car thefts in the first year. This is only one study, and it has yet to be peer-reviewed. It's also not a randomized controlled trial, the gold standard for finding the actual causal effect of something. Other factors, beyond Flock, could still help explain the decline. That study attempted to control for variables, the article points out — like how San Francisco's decline coincided with a broader drop in crime across American cities as America emerged from the pandemic... But the article also notes another worry shared by "countless citizens" and political leaders. Are these systems handing too much power to law enforcement to "invade our privacy, trace our whereabouts, and stomp on our civil liberties."

Read more of this story at Slashdot.