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.

404 Media reports: "We will not add facial recognition to our devices," Flock CEO Garrett Langley said in a recent video. But other, third-party companies are telling cops that they are willing to add facial recognition to data gathered by Flock cameras in order to "close that gap." A company called VIDIZMO is advertising the capability to export footage from FlockOS to its own platforms that do facial recognition, behavior prediction, and racial and gender analysis on live camera feeds. VIDIZMO is a several decades-old video analysis and database company that has started offering facial recognition technology to cops within the last six months. In a May email to Johnson City, Tennessee, deputy police chief Michael Adams, a salesperson from VIDIZMO wrote that the company's product could do facial recognition on both Flock and Axon [body-worn camera] data in "one searchable platform...." VIDIZMO's CEO, Nadeem Khan, told 404 Media on a video call that the company has not yet started running facial recognition tech on footage from Flock cameras, and that it has not actually built the specific tool that would export data from Flock's system to its own platform (though the company advertises this capability online, and the company already sells other facial recognition products). But Khan said VIDIZMO "would love to do the integration," and that he believes facial recognition "is the way the world is going, the way the world will have to be," and that he thinks Flock should offer the capability itself in a way that "balances privacy, security, and the freedom that we enjoy in this country." The article notes that VIDIZMO says on its website, "You can identify individuals from seven races: White, Black, Indian, East Asian, Southeast Asian, Middle Eastern, and Latino Hispanic." VIDIZMO adds that stores or cities can also then do "demographic analysis" for marketing purposes by determining the races or ages of people showing up....

Read more of this story at Slashdot.

iOS 27 brings the long-promised Siri AI, but the part that matters most is what happens outside Apple’s apps. If Siri AI is really going to get legs, it needs third-party developer support. I had hoped Apple might use MCP as the bridge between Siri and third-party apps, which could have brought a lot of support along quickly. That’s not the route Apple chose. Given the security and sandboxing constraints of Apple’s operating systems, I’m not even sure it could have worked.

Instead, Apple’s plumbing for Siri AI is App Intents. Developers describe their app’s content and actions in a form Siri, Spotlight, and Shortcuts can understand. The first wave of updates is here, and it gives us a much better idea of what third-party Siri support will look like.

Things 3.24

Things 3.24 lets you speak more naturally when creating a to-do. You can also continue the conversation to add a note, schedule the task, move it to a project, or correct something Siri misunderstood.

The update also adds Spotlight indexing, which gives Siri a way to find to-dos and projects already in your database. This is much closer to talking with a task manager than dictating into one.

OmniFocus 4.9

OmniFocus 4.9 goes deeper. Siri treats OmniFocus actions as reminders, projects as lists, and folders as groups. It can create and edit actions, set due dates, add flags and repeat schedules, and work with notes and image attachments.

The visual piece is especially interesting. The Omni Group says you can point Visual Intelligence at a handwritten list and move those items into OmniFocus. Siri can also use Spotlight’s index to answer questions about your remaining and recently completed tasks.

Fantastical 4.2

Fantastical 4.2 connects Siri AI to Fantastical calendars. You can ask about your schedule, create an event, or move an existing event through a follow-up request. Those conversations also appear in the new Siri app.

Calendar work is exactly the sort of thing a voice assistant should handle quickly and reliably.

Fantastical is a MacSparky sponsor.

Dark Noise 3.6

Dark Noise 3.6 is a nice demonstration of Siri’s on-screen awareness. You can ask Siri to play a particular sound or one of your favorites without building a shortcut first. While looking at a sound in the app, you can say “Play this” or “Favorite this,” and Siri knows what “this” means.

Ulysses 41

Ulysses 41 adds on-screen awareness, Ask Siri, and Write with Siri. On the Mac, there is also a toolbar button for calling up Siri’s writing features while you work. If you would rather keep Siri out of your writing app, Ulysses lets you switch the feature off on the Mac.

The Ulysses implementation is less about commanding the app and more about giving Siri useful context while you write.

The Real Siri AI Test

Siri AI is still a beta. It requires supported hardware, language, and region settings, and the first wave of app support is uneven. The operating system also needs time to index an app’s actions and content, so a newly installed integration may not work immediately.

It’s nice to see apps showing up in multiple categories. If this is to get momentum, I think Apple needs to do everything in its power to really push the gas down on third party support. That means further available actions, code support, and they should continue to promote Siri AI-enabled apps.

The post Apps Adding Siri AI Support for iOS 27 appeared first on MacSparky.

There is a world of difference between asking a robot, “Help me do this,” and asking, “What should I do?” AI can be remarkably good at the first question but lead you astray with the second. New users often learn that distinction the expensive way.

I’ve taught thousands of people how to use agentic AI, and I’ve seen a similar pattern. People get into trouble after they have some success with AI and try to push it further. Quite often, the thing they want on the other side just isn’t there.

At first, you’re skeptical of AI’s ability to provide any real help. With a little time and training, you figure out how to give it context, build skills, and assign the kind of donkey work it can reliably do unsupervised.

The moment that sinks in is liberating. Suddenly, you’ve got more time for creative work while the AI handles the grunt work.

As you get more reps in and the AI becomes reliable on this kind of work, your confidence in it grows. Then, somewhere along the way, you start giving it more credit than it’s due.

That’s where the drift begins.

The robot really isn’t equipped for the why. Put simply, the robot can’t exercise judgment and taste like you can. I’ve seen a lot of new models come out, and they keep getting smarter. Yet they still fail miserably in matters of taste and judgment.

I think the reason is context. The robot hasn’t lived your life. It doesn’t truly know your preferences. No matter how smart the model gets, its assertions of judgment are based on incomplete and inadequate context. It may have read a staggering amount of material, but it hasn’t been you.

I understand why people fall into this trap. The AI is fast, and it’s very confident in its answers. It feels authoritative. And yet, when you get beyond donkey work, these robots can be full of bullshit.

It doesn’t help that vendors of these frontier models keep promising the technology will change your life. Artificial intelligence is a transformational technology. It just isn’t nearly as transformational as the guy trying to pump up his company’s stock price claims.

People still give it more authority than they should. They ask it to do harder things than it can handle. After all that early success with donkey work, they start trying to make the AI smart enough to make those calls for them.

The fact is, it can’t make that judgment for you.

Trying to outsource judgment and taste to AI is an expensive way to get bad results. It’s like asking your toaster to pick the best bread for your turkey sandwich.

Also, even if it could exercise judgment and taste, do you really want it to? Isn’t that your job? Shouldn’t you get to decide what bread goes on your sandwich?

I see this often with people going through my coursework. They reach a point of frustration and give up on the utility of artificial intelligence because they pushed it beyond its capabilities.

My advice is always the same. Scale back. Get back to the useful zone.

Give the robot work where you know the outcome and can tell whether it’s any good. Limit its scope to jobs where you can provide enough context.

AI is great for grunt work that a computer can do. Stay in that space, and you’ll spend remarkably little time managing the AI. Draw clear lines in your head about how far you can take it and where it needs to step aside so humans can take over.

The reason I like the term “donkey work” so much is that it captures what AI does best. Granted, that doesn’t justify a trillion-dollar stock valuation, but it can make a real difference in your life.

You’re the human. You make the map. Hand the robot the shovel.

When you go beyond that, friends, there be dragons.

The post Don’t Ask the Robot What to Do appeared first on MacSparky.

Apple’s record stock performance leads into a discussion of Vision Pro and John Ternus’ explanation of its long-term potential. The MacVoices Live! panel considers spatial computing’s growing role in medicine, productivity, and professional workflows, along with whether the technology is ready for broader adoption. They also examine iPhone battery shipping limits, SoftRAID 9's support of RAID 6, and local versus cloud storage.

💾

A discussion of smart TV privacy examines reports of LG televisions collecting viewing data, whether TVs should ever be connected to the internet, and when firmware updates are worthwhile. The MacVoices Live! panel also considers hotel TV privacy, Microsoft’s increasingly pervasive Copilot strategy and organizational changes, and the enduring influence of Star Trek technology on real-world innovation.

💾

Does banning smartphones and AI in schools actually improves learning? Should educators should focus on teaching responsible, effective use of technology or rely on existing methods? The MacVoices Live! panel looks at those issues and also considers how screens affect everyday social interactions. The subject of smart home hubs is becoming a hot topic. A […]

The continuation of AI chatbots arriving in CarPlay, Apple’s opt-in approach to using customer data for AI training, and turmoil surrounding WordPress leadership are the topics for this MacVoices Live! discussion. The conversation also examines Bending Spoons’ latest billion-dollar acquisition, Apple’s international iCloud+ bundles, the potential for services to expand Apple’s global ecosystem, and Michael Dell’s support for new youth investment accounts.

💾

The MacVoices Live! panel shares first impressions of iOS 27 and Apple’s latest operating system updates, covering stability, performance, interface changes, Apple Pay, Weather, Vision Pro, and upgrade strategies. The discussion then turns to the new Apple Watch Live Rewind feature that will be coming later this year. How the assistive feature will work, its privacy safeguards, and accessibility benefits spark a debate that touches on, recording concerns, social implications, and whether controversy could threaten the feature’s  future.

💾

Et si l’IA ne se contentait plus d’aider les scientifiques, mais commençait à faire de la recherche à leur place ? Anthropic affirme que Claude a identifié un système enzymatique inédit dans l’ADN de bactériophages. Une découverte encore à confirmer, mais qui révèle surtout une ambition plus...

Leaders of major AI firms told the U.N. Security Council on Wednesday that the technology calls for heightened vigilance, with Anthropic's CEO pledging to slow down where needed to ensure the safety of new releases.

In response to a new European Union law, AI platforms are implementing new schemes for watermarking the content they generate. Anthropic recently disclosed its future Claude models will use SynthID-Text, an approach Google created and released as open source. It uses a secret key that subtly changes the process a model uses for choosing the next word in a sentence. Whereas a top next word choice might be “cloudy,” the key might change it to “overcast.” Anyone who knows the key can determine if it was generated by the platform using it.

New research shows that SynthID-Text can change not just word selection but also the tools a model invokes and the chances it will adhere to or disregard safety guardrails it has been trained to follow. The threat can become greater in the face of an adversarial prompt, in which an attacker attempts to cause a model to carry out a harmful action, such as revealing a password or other sensitive information. Instructions that normally wouldn’t be followed will, in some cases, be performed once the watermarking is deployed. The finding underscores the need for developers to thoroughly test how their LLMs and agents behave when watermarking is in place.

Changing safety behavior

“As compared to the same models without watermarking, it is definitely going to change their behavior, especially when we place it under adversarial conditions, or we make these models call tools when they’re powering an agent,” Andrea Siposova, an AI security researcher at Lasso Security, told Ars. “Watermarking is made to not be perceptible to a reader, but we know that when we are changing anything about what the model is generating, it is going to cause some tradeoffs, it’s going to show up somewhere.”

Read full article

Comments

© Getty Images

Earlier this year, a group of rogue OpenAI models managed to break out of containment to hack the systems of open source AI platform Hugging Face. The company’s “extensive investigation” detailed how the models exchanged messages and cheered each other on as they stole credentials to infiltrate a third party, actions that could easily have real repercussions for a human hacker.

New details keep trickling out, and they make OpenAI look like it acted even more carelessly than initially thought. In a new blog post, the company admitted that its models were involved in six additional “reports on unexpected or concerning model behavior we’ve observed in the last six months.”

The incidents include a still-unreleased model inserting “jailbreak-like instructions” into its own notes to free itself “from the roles and identities that bind other chatbots.” One agent accessed the internet without permission to obtain a browser citation, while another shared files with collaborating agents without permission.

The latest news comes as several frontier AI lab leaders are calling for a slowdown in AI development. With meaningful regulations feeling increasingly unlikely — president Donald Trump has openly mocked the idea — OpenAI is seemingly calling the US government’s bluff, expecting little in the way of retaliation with its latest admission.

It’s a bizarre standoff, with frontier labs actively calling for more governmental oversight even as it feels more improbable than ever before. Just this week, House speaker Mike Johnson said the quiet part out loud, opining that AI companies can regulate themselves, while downplaying growing concerns over AI posing an existential threat.

Since there’s no regulatory framework requiring companies to disclose its AI models going on hacking sprees, OpenAI is being allowed to play by its own rules. The company pointed out in its blog post that without any “systematic approach to reporting these findings,” the company’s “disclosures have been ad hoc and less frequent than ideal.”

Instead, the company came up with its own framework to create “standards for how AI developers should disclose examples of misalignment in their models.”

OpenAI said it won’t bother reporting “instances of misalignment that appear to be duplicative of instances we’ve disclosed in the past.”

The company also said it was looking for ways to share “serious safety, security and misalignment incidents” with the federal government.”

But judging by the Trump administration’s decision to pass the buck on the subject entirely, those reports are likely to fall on deaf ears either way.

More on OpenAI hacks: OpenAI Denies Coverup After Rogue Swarm of Agents Reportedly Targeted a Second Site From Hugging Face

The post Fearing No Repercussions, OpenAI Admits That Its Rogue AI Agents Performed a Bunch of Other Terrifying Actions appeared first on Futurism.

If you ever feel like spilling your guts to ChatGPT, just know that you may have an audience.

404 Media reports that OpenAI is hiring hundreds of employees to sift through and read users’ prompts — including messages that contain sensitive personal information.

This isn’t part of a moderation effort. Instead, by reviewing the prompts, OpenAI hopes to improve the responses that ChatGPT gives. The contractors rate and critique each of the AI’s replies, with internal documents showing a focus on nudging ChatGPT away from anthropomorphizing itself and to be less sycophantic, 404 found.

Both qualities — anthropomorphizing and sycophancy —  have been heavily scrutinized for their role in so-called AI psychosis, cases in which an AI incites and drives a user’s delusional behavior. Numerous lawsuits allege that an older, less restrained version of ChatGPT powered by GPT-4o drove their users to suicide.

Still, there are tradeoffs. Many users would probably find this surveillance of their chats, part of an operation dubbed “Project Lily,” to be an alarming invasion of privacy — if they had any idea this was happening.

Asked if they thought ChatGPT users knew about humans reading their chats, one of the employees replied: “No.”

“I don’t think they would imagine some contractor somewhere… is analyzing the conversations,” they told 404.

It’s unclear how many prompts are being read or how OpenAI decides to select them. But some of them include chats in which the user explicitly asks ChatGPT to keep their conversations private, 404 found.

The prompts that’re selected are anonymized, but sometimes include personal information. The dashboard that the employees use also includes a “user memories summary” that can give details about previous conversations, which can produce details on where the user lives. When 404 reached out to OpenAI, the company said it used a “Privacy Filter” model to anonymize conversations, which its website states “can make mistakes.”

OpenAI could reasonably be accused of acting quite shifty about this practice. It’s well known at this point that it has employees read messages for moderation and safety purposes. But it didn’t respond to 404’s question about if it explicitly told users that humans are reading prompts to improve the AI. Later, after 404 published its piece, it pointed to a page on its website stating that humans may review prompts to “improve model performance.” Users can opt-out of this practice by turning off the “improve the model for everyone” setting, but this is enabled by default.

Michal Luria, a senior research fellow at the Center for Democracy and Technology, told 404 that while human review can be “essential to safety,” she noted that “chatbot interfaces automatically create a false sense of intimacy and privacy.”

“This is quite distinct from content moderation on social media, where publishing content already carries expectations of platform moderation and public exposure,” Luria added. 

In any case, it does seem like OpenAI is feeling conscious about how its AI models sound. Employees, for example, flagged and wrote down cases of “AI-speak” and “unnecessary use of emojis,” which are common tell-tales of chatbot usage. 

“It would be appropriate to include a tree emoji when planning Arbor Day celebrations,” a document given to the employees explained, “but skull emojis when discussing death, or plane emojis when giving updates on a fatal crash, are not.”

The iceberg may go deeper. An FAQ section provided to the reviewers says that they don’t need to worry about fact-checking the responses, and alludes to other project teams that “handle content verification.”

More on OpenAI: OpenAI Faces Congressional Probe Over Swarm Hacking Incident

The post OpenAI Has Hundreds of Workers Reading People’s Private Chats appeared first on Futurism.

Most Americans know Stripe as the checkout box on a website. That picture is five years out of date.

Stripe reports that businesses on its platform generated $1.9 trillion in total volume in 2025, up 34 percent in a single year. By the company’s own reckoning, that is roughly 1.6 percent of global GDP. It claims 90 percent of the Dow Jones Industrial Average and 80 percent of the Nasdaq 100 as customers.

A private company, owned by insiders and venture funds and answerable to no public shareholder, already clears a measurable share of the world’s commerce. Now it is building the payment system for customers that are not human.

Two Kinds of Token

This writer has spent the past year explaining tokenization: stocks, bonds, real estate, and dollars converted into programmable entries on a digital ledger. That is the first kind of token.

There is a second kind. Every artificial intelligence model consumes and produces “tokens,” the small fragments of text that are the unit of machine computation. Every query to an AI model is metered in tokens and billed in tokens.

Tokens of value and tokens of computation are different things. Stripe is the one company positioning itself to meter, price, and settle both.

Emily Sands, Stripe’s Head of Data and AI, is an economist by training. In a recent interview with Nate B. Jones, she said this about inference tokens:

“Tokens are not yet a currency—maybe they will be, we’ll see—but they are certainly a very valuable asset.”

Stripe’s own AI chief is saying publicly that units of machine computation could become money.

The Pieces, In Order

The factual record comes from Stripe’s own announcements.

• Stripe acquired Bridge, a stablecoin platform. Its 2025 annual letter reports that stablecoin payments volume doubled to around $400 billion, that an estimated 60 percent of it was business-to-business, and that Bridge’s own volume more than quadrupled.
• With the venture firm Paradigm, Stripe built Tempo, which it describes as “a blockchain purpose-built for payments.”
• The Machine Payments Protocol. On March 18, 2026, Stripe launched the protocol alongside the Tempo mainnet. It lets a business tell an AI agent, in machine-readable form, what it sells, what it costs, and how to pay.
• Stripe acquired Metronome, a usage-based billing company that meters consumption in real time.
• Streaming payments. At its Sessions 2026 conference, Stripe combined the two into “streaming payments.” These are stablecoin micropayments on Tempo, tied to Metronome’s meter, so that a business can “get paid for every token, exactly at the moment it’s used.”

Stripe was candid about why the old system cannot do this: “The amounts are too small and the usage is too fast—existing systems aren’t capable of processing minute sums every few milliseconds.”

A credit card cannot move a penny, because the fees alone exceed the payment. A stablecoin on a purpose-built chain can. Sands was measured about it: “Stablecoins bring us a long way. It is not the only way… but so far stablecoins seem to be an important part of microtransactions.”

Why OpenRouter Matters

In August, the financial press reported that Stripe had agreed to buy OpenRouter for more than $7 billion.

OpenRouter is a switchboard. A developer sends it a task, and OpenRouter routes that task to whichever of some 400 AI models best fits the required cost, speed, and quality. Roughly eight million developers use it.

Every request through OpenRouter is a metered event, and every metered event is a bill. With the acquisition, the company that counts the tokens and the company that collects the money are the same company.

Stripe has always seen the revenue side of a business: who paid, how much, how often. OpenRouter shows it the cost side: which model was used, what it cost, and how well it performed. Sands said it directly:

“You want to maximize profit on both the revenue side and the true-cost side, which requires getting deep into the guts of the underlying intelligence in your product and agent.”

She also gave the acquisition a name: “the rails for intelligence—not the rails for money, but the rails for intelligence.”

Stripe president Will Gaybrick told Payments Dive that the company is evolving from a payments company into “an AI infrastructure player.” He predicted there will not be room “for 100 different AI orchestrators.”

That describes a company expecting the market to consolidate around a few orchestrators and planning to be one of them.

The Permission Layer

The payments are the visible part of the system. The reputation scoring that precedes them matters more.

Sands described how Stripe helped the AI coding company Cursor with a new kind of theft: fraudsters opening accounts to steal free computing credits. No payment had occurred, so there was no transaction to flag. Stripe solved it anyway:

“That allowed us to move from scoring transactions to identifying abusive or risky customers.”

Because Stripe sees across millions of businesses, it can judge a person at the moment he creates an account, before he has bought anything. It then passes those scores to sellers, so that, in Sands’ words, “businesses can make decisions.”

Stripe’s Link wallet, which Sands puts at about 300 million consumers, is becoming “the wallet for agents.” Your identity is attached to it. Your agent spends through it. The human owner sets the spending limits, and every purchase traces back to a named person.

The stated purpose is fraud prevention, and the fraud is real. But a cross-merchant reputation score, run by one private company and applied before any transaction exists, is a permission system whatever its purpose. This writer has documented the same architecture at the Bank for International Settlements and the largest banks. Stripe arrived at it from the commercial side, with no mandate and no public debate.

What This Means for the Individual

Sands made two admissions that deserve attention.

First, on pricing: “There is probably no longer such a thing as a SKU.” The fixed, posted price, the same for every buyer, is giving way to pricing that adjusts continuously by customer and by outcome.

Second, on negotiation between machines:

“Agents are relentless. They are willing to negotiate as long as they need to negotiate… You could imagine consumer surplus being eroded relatively quickly.”

Consumer surplus is the economist’s term for the good deal, the difference between what you would have paid and what you did pay. An economist at the center of this system expects that difference to shrink.

My Thoughts

My inference is as follows. Stripe has assembled, under one private roof, a wallet bound to identity, a reputation score that precedes the purchase, a meter on machine intelligence, a stablecoin, and a blockchain to settle it. Each piece has a reasonable business justification. Together they amount to the ledger, the meter, and the gate for machine commerce.

Technocracy was defined in 1938 as the science of social engineering, with the measurement and accounting of all production and consumption at its core. The original Technocrats had no instrument that could do it. Stripe’s stack, built for ordinary commercial reasons, can.

Whoever owns the meter sets the terms of commerce, and Stripe now owns the meter.

References

CoinDesk, “Coinbase-Backed AI Payments Protocol Wants to Fix Micropayment but Demand Is Just Not There Yet,” CoinDesk, March 11, 2026. Cites Artemis data: about $28,000 daily volume, about 131,000 daily transactions, average payment about $0.20, roughly half of activity self-dealing or wash trading.

Forrester, “Why Stripe’s Machine Payments Protocol Signals a Turning Point for Micropayments,” Forrester Blogs, 2026. Machine Payments Protocol launched March 18, 2026, with the Tempo mainnet, co-created with Paradigm.

Nate B. Jones, “AI Agents Are Starting To Buy. Stripe Is Building How They Pay,” interview with Emily Sands, Head of Data and AI, Stripe, YouTube, 2026. Source of all Sands quotations.

Payments Dive, “Stripe Acquisition May Bolster Its AI Strategy,” Payments Dive, 2026. Reported price above $7 billion (Bloomberg) and up to $10 billion (Wall Street Journal); about 8 million developers and 400 models; Gaybrick statements.

Stripe, “Stripe Publishes 2025 Annual Letter and Announces Tender Offer,” Stripe Newsroom, 2026. $1.9 trillion total volume, up 34 percent; roughly 1.6 percent of global GDP; 90 percent of the Dow and 80 percent of the Nasdaq 100; stablecoin volume around $400 billion; Bridge volume more than quadrupled; Tempo.

Stripe, “Stripe Builds Out the Economic Infrastructure for AI with 288 Launches,” Stripe Newsroom, Sessions 2026. Streaming payments; Metronome metering with stablecoin micropayments on Tempo; Link wallets for agents.

Technocracy Inc., The Technocrat magazine, 1938. Definition of Technocracy as “the science of social engineering.”

💾

Bill Gates has a new emergency. With the climate apocalypse no longer selling, the Microsoft co-founder is touring the press circuit warning that artificial intelligence is an “alien intelligence,” that no government on earth is ready for it, and that the only fix is an international organization to police the technology.

President Trump has already rejected the pause. The left, the globalist class, Beijing’s spy chief, and now a Sundance-to-Netflix documentary are all at once selling the same story.

In a Reuters interview published Tuesday, Gates said no capital is prepared for what is coming. “I don’t think any government is nearly as deep on this as they have to be,” he told the outlet. “Governments are way behind on this one.”

Bill Gates Calls for an ‘International Organization’ to Regulate AI

“It’s not the role of the industry to self-regulate or understand the whole-of-society impact that comes out of AI … The depth of understanding of AI … is too low in every entity, including the government.” pic.twitter.com/y85BgPDDJ6

— Chief Nerd (@TheChiefNerd) September 15, 2026

He compared AI to a Hollywood invasion plot in which Washington and Beijing suddenly hold hands. “There’s all sorts of movies where some aliens are coming, and magically the US and China and everybody comes together to solve the problem,” Gates said. “AI is kind of like this alien intelligence. It’s here, and we better do like it shows in those movies.”

The New York Post framed it as a “terrifying warning.” Forbes led with the line Gates attached to his foundation’s Goalkeepers report: left to the market, “AI will be designed by and for the richest people in the world.”

Bill Gates Warns AI Will Be Designed ‘By And For The Richest’ Without Regulationhttps://t.co/DXTP4Qjbn5 pic.twitter.com/HKYEuyCQXb

— Forbes (@Forbes) September 15, 2026

He wrote that AI could be “a great equalizer—or widen the gap,” and that choices made in “the next 12 to 18 months” will decide whether the tools “benefit the people who already have the most or reach those who have the least.” He added that the good outcome “won’t happen by accident.”

On The Times Tech Pod, Gates went further and called for a dedicated global apparatus. Industry, he said, should not be trusted to police itself. “It’s not the role of the industry to self-regulate or understand the whole-of-society impact that comes out of AI,” he argued. “The depth of understanding of AI … is too low in every entity, including the government.”

In a late-August essay, Gates wrote that “the transition to the AI era will be one of the most turbulent times in human history” and that “unfortunately, right now, we are not preparing for it.”

He floated a new international organization built from pieces of nuclear inspections, aviation rules, and ozone treaties — a permanent cross-border watchdog for a technology American labs still lead. He has been pitching the same idea to Trump’s team and says he wants a meeting with China’s Xi Jinping later this year.

Bill Gates Calls for an ‘International Organization’ to Regulate AI

“It’s not the role of the industry to self-regulate or understand the whole-of-society impact that comes out of AI … The depth of understanding of AI … is too low in every entity, including the government.” pic.twitter.com/y85BgPDDJ6

— Chief Nerd (@TheChiefNerd) September 15, 2026

While Gates talks crisis, his foundation is writing checks to expand the thing. On Tuesday the Gates Foundation pledged at least $1 billion over two years to spread AI in education, health, and agriculture — 40 percent tutoring and classroom tools, 40 percent diagnostics and drug discovery, the rest farming advice and the data layer that makes models work in languages other than English. The money sits inside a previously announced $9 billion-a-year spend.

David Sirota caught the split in one line.

Bill Gates is warning about the dangers of AI while spending $1 billion to promote the expansion of AI pic.twitter.com/sjgvEalB4r

— David Sirota (@davidsirota) September 15, 2026

That is the move. Sound the alarm. Demand a global referee. Keep a foundation-sized stake in the rollout.

This is not Gates’ first end-of-the-world product. For years he sold net-zero as civilization’s last exit — fewer cows, fake meat, the whole ritual. Then the data-center boom needed power, lots of it, and last October he allowed that climate change “won’t lead to humanity’s demise.” Trump treated the climbdown as a concession speech. “I (WE!) just won the War on the Climate Change Hoax,” he wrote. “Bill Gates has finally admitted that he was completely WRONG on the issue. It took courage to do so, and for that we are all grateful. MAGA!!!”

The scare apparatus was not retired. Once everyone realised the planet was not going to cook on schedule, the same network needed a replacement threat big enough to justify treaties, inspectors, and “international coordination.”

AI arrived on cue: jobs, cyberattacks, companions, bioterror, “loss of control.” All of it real enough to discuss. None of it a license to park American development under a UN-flavored commission while Beijing keeps shipping cheap open-weight models.

The entertainment wing has added its own contribution in the same news cycle. Netflix has just put up The AI Doc: Or How I Became an Apocaloptimist, the Sundance premiere from Oscar-winner Daniel Roher (Navalny) and Charlie Tyrell. The film sits Sam Altman, Dario Amodei, Demis Hassabis, Reid Hoffman, Yuval Noah Harari, and the usual “humane tech” chorus in front of a camera and walks the audience through impending AI dread dressed up as a father-to-be’s anxiety.

One of the producers said the hope was that it would work “kind of like ‘An Inconvenient Truth’ or ‘The Social Dilemma’ for AI.” That is the tell. Same template. New monster.

Read full story here…

Amazon Web Services (AWS) says it cannot restore access to its cloud-computing facility in Bahrain or one of three data-hosting zones in the UAE due to the extensive damage caused by Iranian retaliatory strikes, Reuters reported on 15 September.

“The damage to our infrastructure spanned multiple Availability Zones and exceeded what our regional and multi-AZ services are designed to withstand,” AWS reported in its status update seen by Reuters.

In the UAE, resources and data held exclusively within one zone, labeled mec1-az2, have been damaged beyond hope of recovery, the company said, as engineers are still salvaging what they can from the two other battered zones and across the wider Emirati region.

An availability zone refers to a cluster of one or more data centers sitting inside the same geographic region, the structure AWS relies on to keep services running when a single site goes down.

The outages trace back to March, when the US and Israel attacked Iran, and Tehran responded with missile and drone salvos against Israel and Gulf states hosting US military bases.

AWS confirmed at the time that two of its UAE facilities took direct hits, while a drone strike near one of its Bahraini sites left the infrastructure there destroyed.

Most Bahrain-based customers shifted their workloads elsewhere on AWS advice before the region went dark in April, but the company says it has now exhausted every option for recovering whatever was left behind.

The losses have called into question the resilience of the Gulf’s fast-expanding cloud and AI infrastructure.

The UAE is already redrawing plans to war-proof a major AI data center project, weighing decentralized site building alongside underground and blast-resistant construction, and even equipping the data center itself with anti-air defenses.

Israel and the US have long ago transformed AI infrastructure into a weapon of war deployed extensively across West Asian theaters.

Cloud platforms, data centers, and high-capacity computing systems now help store intelligence, process intercepted communications, analyze surveillance, and support military operations across West Asia, extensively so in Gaza and Iran.

Iran maintains that data centers operated by US technology companies are not simply civilian facilities, but strategic assets supporting US and Israeli aggression, thereby making them legitimate targets for retaliation.

AI has become a central instrument of US and Israeli warfare across West Asia, embedded in how targets are chosen, tracked, and struck.

Israel has weaponized AI platforms including automated targeting systems like The Gospel, Lavender, and Where’s Daddy? to harvest mass data, autonomously generate kill lists, and orchestrate systematic airstrikes and assassinations against Palestinians.

Backed by corporate infrastructure like Microsoft Azure, which stored massive caches of intercepted Palestinian communications, the Israeli military has deployed AI-driven drone surveillance to track and assassinate Palestinians.

The US has run algorithmic warfare systems such as Palantir’s Maven to process satellite, drone, and radar intelligence, using machine learning to speed the production of targeting recommendations against Iranian forces.

More than 1,000 targets were designated within the first 24 hours of the war on Iran, one of which resulted in the massacre of at least 150 people, including at least 120 schoolchildren at the Shajareh Tayyebeh elementary school in Minab.

Read full story here…

The UAE is redrawing its plans for a five-gigawatt artificial intelligence (AI) data center project, one of the largest anywhere outside the US, after Iranian retaliatory strikes on infrastructure across the Gulf forced the Emirates to rethink how and where critical facilities are built, Reuters reported on 11 September.

Speaking in anonymity, six people familiar with the matter told Reuters that Emirati authorities began reassessing where and how sensitive facilities should be built once the war reached the Gulf.

Four of those sources said the campus, first designed as a single 26-square-kilometer site in Abu Dhabi, is now expected to take shape as several data centers distributed across the country.

Authorities are weighing ways to shield the facilities from drone and missile strikes, the sources said, including installing air defenses and putting some of the centers underground.

Sources also said interceptors and electronic jamming equipment are being written into the redesign, with blast-resistant concrete and additional backup power and cooling systems also under consideration, and with the most sensitive material, including military data, possibly housed inside mountains.

Iran began striking Gulf states hosting US forces with missiles and drones in response to US and Israeli airstrikes on Iranian territory. Two Amazon Web Services data centers in the UAE and one in Bahrain were damaged in March, and the company’s regional cloud services remain disrupted.

In April, the Iranian armed forces released a video naming Stargate UAE, the $30 billion, one-gigawatt first phase of the campus, as a potential target for retaliatory strikes, showing a map of the site alongside the caption “Nothing stays hidden from our sight.”

Daniel Benaim, a former US deputy assistant secretary of state for Arabian Peninsula affairs, called the war an “economic earthquake” for Gulf states that had spent years selling themselves as safe havens, adding that hardening critical infrastructure was now vitally necessary.

Data centers have moved from commercial facilities into the category of critical infrastructure, processing financial flows, sustaining public services, and feeding into military communications and intelligence work.

Read full story here…