Beth Washington

Good Luck, Have Fun, Don’t Die trailer ushers in AI apocalypse

Briarcliff Entertainment, culture, Entertainment, film trailers, Good Luck Have Fun Don't Die, Gore Verbinski, science fiction film, Trailers / Beth Washington / November 12, 2025

Director Gore Verbinski has racked up an impressive filmography over the years, from The Ring and the first three installments of the Pirates of the Caribbean franchise to the 2011 Oscar-nominated animated western Rango. Granted, he’s had his share of failures (*coughThe Lone Ranger *cough*), but if this trailer is any indication, Verbinski has another winner on his hands with the absurdist sci-fi dark comedy Good Luck, Have Fun, Don’t Die.

Sam Rockwell stars as the otherwise unnamed “Man from the Future,” who shows up at a Los Angeles diner looking like a homeless person but claiming to be a time traveler from an apocalyptic future. He’s there to recruit the locals into his war against a rogue AI, although the diner patrons are understandably dubious about his sanity. (“I come from a nightmare apocalypse,” he assures the crowd about his grubby appearance. “This is the height of f*@ing fashion!”) Somehow, he convinces a handful of Angelenos to join his crusade, and judging by the remaining footage, all kinds of chaos breaks out.

In addition to the eminently watchable Rockwell, the cast includes Haley Lu Richardson as Ingrid, Michael Pena as Mark, Zazie Beetz as Janet, and Juno Temple as Susan. Dino Fetscher, Anna Acton, Asim Chaudhury, Daniel Barnett, and Domonique Maher also appear in as-yet-undisclosed roles. Matthew Robinson (The Invention of Lying, Love and Monsters) penned the script. This is Verbinski’s first indie film, and Tom Ortenberg, CEO of distributor Briarcliff Entertainment, praised it as “wildly original, endlessly entertaining, and unlike anything audiences have seen before.” Color us intrigued.

Good Luck, Don’t Die, Have Fun hits theaters on February 13, 2026.

Good Luck, Have Fun, Don’t Die trailer ushers in AI apocalypse Read More »

US states could lose $21 billion of broadband grants after Trump overhaul

BEAD program, Broadband Equity Access and Deployment program, Policy / Beth Washington / November 11, 2025

The BEAD law is clear that the money can be used for more than sending subsidies to Internet service providers. The law says BEAD money can be allocated for connecting eligible community anchor institutions; data collection, broadband mapping, and planning; installing Internet and Wi-Fi infrastructure or providing reduced-cost broadband to multi-family buildings; and providing affordable Internet-capable devices.

The current law also says that if a state fails to use its full allocation, the National Telecommunications and Information Administration (NTIA) “shall reallocate the unused amounts to other eligible entities with approved final proposals.” The law gives the NTIA chief latitude to spend the money for “any use determined necessary… to facilitate the goals of the Program.”

Arielle Roth, who has overseen the BEAD overhaul in her role as head of the NTIA, has said she’s open to sending the remaining funds to states. Roth said in an October 28 speech that the NTIA is “considering how states can use some of the BEAD savings—what has commonly been referred to as nondeployment money—on key outcomes like permitting reform” but added that “no final decisions have been made.” The Ernst bill would take that decision out of the NTIA’s hands.

States still waiting after Biden plans thrown out

After Congress created BEAD, the Biden administration spent about three years developing rules and procedures for the program and then evaluating plans submitted by each US state and territory. The process included developing new maps that, while error-prone due to false submissions by ISPs, provided a more accurate view of broadband coverage gaps than was previously available.

By November 2024, the Biden administration had approved initial funding plans submitted by every state and territory. But the Trump administration rewrote the program rules, eliminating a preference for fiber and demanding lower-cost deployments.

States that could have started construction in summer 2025 had to draft new plans and keep waiting for the grant money. The Trump administration is also telling states that they must exempt ISPs from net neutrality and price laws in order to obtain grant funding.

As for when the long-delayed grants will be distributed, Roth said the NTIA is “on track to approve the majority of state plans and get money out the door this year.”

US states could lose $21 billion of broadband grants after Trump overhaul Read More »

Google announces even more AI in Photos app, powered by Nano Banana

AI, AI images, Artificial Intelligence, Google, google photos, Tech / Beth Washington / November 11, 2025

We’re running out of ways to tell you that Google is releasing more generative AI features, but that’s what’s happening in Google Photos today. The Big G is finally making good on its promise to add its market-leading Nano Banana image-editing model to the app. The model powers a couple of features, and it’s not just for Google’s Android platform. Nano Banana edits are also coming to the iOS version of the app.

Nano Banana started making waves when it appeared earlier this year as an unbranded demo. You simply feed the model an image and tell it what edits you want to see. Google said Nano Banana was destined for the Photos app back in October, but it’s only now beginning the rollout. The Photos app already had conversational editing in the “Help Me Edit” feature, but it was running an older non-fruit model that produced inferior results. Nano Banana editing will produce AI slop, yes, but it’s better slop.

Nano Banana in Help me edit

Google says the updated Help Me Edit feature has access to your private face groups, so you can use names in your instructions. For example, you could type “Remove Riley’s sunglasses,” and Nano Banana will identify Riley in the photo (assuming you have a person of that name saved) and make the edit without further instructions. You can also ask for more fantastical edits in Help Me Edit, changing the style of the image from top to bottom.

Google announces even more AI in Photos app, powered by Nano Banana Read More »

Pirelli’s Cyber Tire might become highway agencies’ newest assistant

Cars, Pirelli Cyber Tire, Traffic / Beth Washington / November 11, 2025

“Two weeks ago, a European manufacturer tested… the traction control and stability with a dramatic improvement in stability and the traction,” he said. “The nice part of the story is that there is not only an objective improvement—2 or 3 meters in braking distance—but there is also from these customers always a better feeling… which is something that is very important to us because numbers are for technicians, but from our customer’s perspective, the pleasure to drive also very important.”

The headline said something about traffic?

While the application described above mostly serves the Cyber Tire-equipped car, the smart tires can also serve the greater good. Earlier this year, we learned of a trial in the Italian region of Apulia that fitted Cyber Tires to a fleet of vehicles and then inferred the health of the road surface from data collected by the tires.

Working with a Swedish startup called Univrses, Pirelli has been fusing sensor data from the Cyber Tire with cameras. Misani offered an example.

“You have a hole [in the road]. If you have a hole, maybe the visual [system] recognizes and the tire does not because you automatically try to avoid the hole. So if the tire does not pass over the hole, you don’t measure anything,” he said. “But your visual system will detect it. On the opposite side, there are some cracks on the road that are detected from the visual system as something that is not even on the road, but they cannot say how deep, how is the step, how is it affecting the stability of the car and things like this. Matching the two things, you have the possibility to monitor in the best possible way the condition of the road.”

“Plus thanks to the vision, you have also the possibility to exploit what we call the vertical status—traffic signs, the compatibility between the condition of the road and the traffic signs,” he said.

The next step is a national program in Italy. “We are investigating and making a project to actively control not the control unit of the car but the traffic information,” Misani said. “On some roads, you can vary the speed limit according to the status; if we can detect aquaplaning, we can warn [that] at kilometer whatever, there is aquaplaning and [the speed limit will be automatically reduced]. We are going in the direction of integrating into the smart roads.”

Pirelli’s Cyber Tire might become highway agencies’ newest assistant Read More »

Researchers isolate memorization from problem-solving in AI neural networks

AI, AI alignment, AI architecture, AI behavior, AI memorization, AI research, AI safety, Allen Institute for AI, Biz & IT, computational neuroscience, copyright, generalization, Goodfire, gradient descent, K-FAC, loss curvature, machine learning, mechanistic interpretability, model editing, neural networks, OLMo, overfitting, transformer models, weight matrices / Beth Washington / November 11, 2025

The hills and valleys of knowledge

Basic arithmetic ability lives in the memorization pathways, not logic circuits.

When engineers build AI language models like GPT-5 from training data, at least two major processing features emerge: memorization (reciting exact text they’ve seen before, like famous quotes or passages from books) and what you might call “reasoning” (solving new problems using general principles). New research from AI startup Goodfire.ai provides the first potentially clear evidence that these different functions actually work through completely separate neural pathways in the model’s architecture.

The researchers discovered that this separation proves remarkably clean. In a preprint paper released in late October, they described that when they removed the memorization pathways, models lost 97 percent of their ability to recite training data verbatim but kept nearly all their “logical reasoning” ability intact.

For example, at layer 22 in Allen Institute for AI’s OLMo-7B language model, the researchers ranked all the weight components (the mathematical values that process information) from high to low based on a measure called “curvature” (which we’ll explain more below). When they examined these ranked components, the bottom 50 percent of weight components showed 23 percent higher activation on memorized data, while the top 10 percent showed 26 percent higher activation on general, non-memorized text.

In other words, the components that specialize in memorization clustered at the bottom of their ranking, while problem-solving components clustered at the top. This mechanistic split enabled the researchers to surgically remove memorization while preserving other capabilities. They found they could delete the bottom-ranked components to eliminate memorization while keeping the top-ranked ones that handle problem-solving.

Perhaps most surprisingly, the researchers found that arithmetic operations seem to share the same neural pathways as memorization rather than logical reasoning. When they removed memorization circuits, mathematical performance plummeted to 66 percent while logical tasks remained nearly untouched. This discovery may explain why AI language models notoriously struggle with math without the use of external tools. They’re attempting to recall arithmetic from a limited memorization table rather than computing it, like a student who memorized times tables but never learned how multiplication works. The finding suggests that at current scales, language models treat “2+2=4” more like a memorized fact than a logical operation.

It’s worth noting that “reasoning” in AI research covers a spectrum of abilities that don’t necessarily match what we might call reasoning in humans. The logical reasoning that survived memory removal in this latest research includes tasks like evaluating true/false statements and following if-then rules, which are essentially applying learned patterns to new inputs. This also differs from the deeper “mathematical reasoning” required for proofs or novel problem-solving, which current AI models struggle with even when their pattern-matching abilities remain intact.

Looking ahead, if the information removal techniques receive further development in the future, AI companies could potentially one day remove, say, copyrighted content, private information, or harmful memorized text from a neural network without destroying the model’s ability to perform transformative tasks. However, since neural networks store information in distributed ways that are still not completely understood, for the time being, the researchers say their method “cannot guarantee complete elimination of sensitive information.” These are early steps in a new research direction for AI.

Traveling the neural landscape

To understand how researchers from Goodfire distinguished memorization from reasoning in these neural networks, it helps to know about a concept in AI called the “loss landscape.” The “loss landscape” is a way of visualizing how wrong or right an AI model’s predictions are as you adjust its internal settings (which are called “weights”).

Imagine you’re tuning a complex machine with millions of dials. The “loss” measures the number of mistakes the machine makes. High loss means many errors, low loss means few errors. The “landscape” is what you’d see if you could map out the error rate for every possible combination of dial settings.

During training, AI models essentially “roll downhill” in this landscape (gradient descent), adjusting their weights to find the valleys where they make the fewest mistakes. This process provides AI model outputs, like answers to questions.

Figure 1: Overview of our approach. We collect activations and gradients from a sample of training data (a), which allows us to approximate loss curvature w.r.t. a weight matrix using K-FAC (b). We decompose these weight matrices into components (each the same size as the matrix), ordered from high to low curvature. In language models, we show that data from different tasks interacts with parts of the spectrum of components differently (c). — Figure 1 from the paper “From Memorization to Reasoning in the Spectrum of Loss Curvature.” Credit: Merullo et al.

The researchers analyzed the “curvature” of the loss landscapes of particular AI language models, measuring how sensitive the model’s performance is to small changes in different neural network weights. Sharp peaks and valleys represent high curvature (where tiny changes cause big effects), while flat plains represent low curvature (where changes have minimal impact). They used these curvature values to rank the weight components from high to low, as mentioned earlier.

Using a technique called K-FAC (Kronecker-Factored Approximate Curvature), they found that individual memorized facts create sharp spikes in this landscape, but because each memorized item spikes in a different direction, when averaged together they create a flat profile. Meanwhile, reasoning abilities that many different inputs rely on maintain consistent moderate curves across the landscape, like rolling hills that remain roughly the same shape regardless of the direction from which you approach them.

“Directions that implement shared mechanisms used by many inputs add coherently and remain high-curvature on average,” the researchers write, describing reasoning pathways. In contrast, memorization uses “idiosyncratic sharp directions associated with specific examples” that appear flat when averaged across data.

Different tasks reveal a spectrum of mechanisms

The researchers tested their technique on multiple AI systems to verify the findings held across different architectures. They primarily used Allen Institute’s OLMo-2 family of open language models, specifically the 7 billion- and 1 billion-parameter versions, chosen because their training data is openly accessible. For vision models, they trained custom 86 million-parameter Vision Transformers (ViT-Base models) on ImageNet with intentionally mislabeled data to create controlled memorization. They also validated their findings against existing memorization removal methods like BalancedSubnet to establish performance benchmarks.

The team tested their discovery by selectively removing low-curvature weight components from these trained models. Memorized content dropped to 3.4 percent recall from nearly 100 percent. Meanwhile, logical reasoning tasks maintained 95 to 106 percent of baseline performance.

These logical tasks included Boolean expression evaluation, logical deduction puzzles where solvers must track relationships like “if A is taller than B,” object tracking through multiple swaps, and benchmarks like BoolQ for yes/no reasoning, Winogrande for common sense inference, and OpenBookQA for science questions requiring reasoning from provided facts. Some tasks fell between these extremes, revealing a spectrum of mechanisms.

Mathematical operations and closed-book fact retrieval shared pathways with memorization, dropping to 66 to 86 percent performance after editing. The researchers found arithmetic particularly brittle. Even when models generated identical reasoning chains, they failed at the calculation step after low-curvature components were removed.

Figure 3: Sensitivity of different kinds of tasks to ablation of flatter eigenvectors. Parametric knowledge retrieval, arithmetic, and memorization are brittle, but openbook fact retrieval and logical reasoning is robust and maintain around 100% of original performance. — Figure 3 from the paper “From Memorization to Reasoning in the Spectrum of Loss Curvature.” Credit: Merullo et al.

“Arithmetic problems themselves are memorized at the 7B scale, or because they require narrowly used directions to do precise calculations,” the team explains. Open-book question answering, which relies on provided context rather than internal knowledge, proved most robust to the editing procedure, maintaining nearly full performance.

Curiously, the mechanism separation varied by information type. Common facts like country capitals barely changed after editing, while rare facts like company CEOs dropped 78 percent. This suggests models allocate distinct neural resources based on how frequently information appears in training.

The K-FAC technique outperformed existing memorization removal methods without needing training examples of memorized content. On unseen historical quotes, K-FAC achieved 16.1 percent memorization versus 60 percent for the previous best method, BalancedSubnet.

Vision transformers showed similar patterns. When trained with intentionally mislabeled images, the models developed distinct pathways for memorizing wrong labels versus learning correct patterns. Removing memorization pathways restored 66.5 percent accuracy on previously mislabeled images.

Limits of memory removal

However, the researchers acknowledged that their technique isn’t perfect. Once-removed memories might return if the model receives more training, as other research has shown that current unlearning methods only suppress information rather than completely erasing it from the neural network’s weights. That means the “forgotten” content can be reactivated with just a few training steps targeting those suppressed areas.

The researchers also can’t fully explain why some abilities, like math, break so easily when memorization is removed. It’s unclear whether the model actually memorized all its arithmetic or whether math just happens to use similar neural circuits as memorization. Additionally, some sophisticated capabilities might look like memorization to their detection method, even when they’re actually complex reasoning patterns. Finally, the mathematical tools they use to measure the model’s “landscape” can become unreliable at the extremes, though this doesn’t affect the actual editing process.

This article was updated on November 11, 2025 at 9: 16 am to clarify an explanation about sorting weights by curvature.

Benj Edwards is Ars Technica’s Senior AI Reporter and founder of the site’s dedicated AI beat in 2022. He’s also a tech historian with almost two decades of experience. In his free time, he writes and records music, collects vintage computers, and enjoys nature. He lives in Raleigh, NC.

Researchers isolate memorization from problem-solving in AI neural networks Read More »

Next-generation black hole imaging may help us understand gravity better

astrophysics, black hole, event horizon, Event Horizon Telescope, gravity, Physics, relativity, Science / Beth Washington / November 8, 2025

Right now, we probably don’t have the ability to detect these small changes in phenomena. However, that may change, as a next-generation version of the Event Horizon Telescope is being considered, along with a space-based telescope that would operate on similar principles. So the team (four researchers based in Shanghai and CERN) decided to repeat an analysis they did shortly before the Event Horizon Telescope went operational, and consider whether the next-gen hardware might be able to pick up features of the environment around the black hole that might discriminate among different theorized versions of gravity.

Theorists have been busy, and there are a lot of potential replacements for general relativity out there. So, rather than working their way through the list, they used a model of gravity (the parametric Konoplya–Rezzolla–Zhidenko metric) that isn’t specific to any given hypothesis. Instead, it allows some of its parameters to be changed, thus allowing the team to vary the behavior of gravity within some limits. To get a sense of the sort of differences that might be present, the researchers swapped two different parameters between zero and one, giving them four different options. Those results were compared to the Kerr metric, which is the standard general relativity version of the event horizon.

Small but clear differences

Using those five versions of gravity, they model the three-dimensional environment near the event horizon using hydrodynamic simulations, including infalling matter, the magnetic fields it produces, and the jets of matter that those magnetic fields power.

The results resemble the sorts of images that the Event Horizon Telescope produced. These include a bright ring with substantial asymmetry, where one side is significantly brighter due to the rotation of the black hole. And, while the differences are subtle between all the variations of gravity, they’re there. One extreme version produced the smallest but brightest ring; another had a reduced contrast between the bright and dim side of the ring. There were also differences between the width of the jets produced in these models.

Next-generation black hole imaging may help us understand gravity better Read More »

Questions swirl after Trump’s GLP-1 pricing deal announcement

drug, eli lilly, GLP-1, health, mounjaro, Novo Nordisk, Ozempic, pricing, Trump, wegovy, zepbound / Beth Washington / November 7, 2025

While some may stand to gain access to the drugs under these categories, another factor in assessing the deal’s impact is that millions are expected to lose federal health coverage under the Trump administration’s “One Big Beautiful Bill Act.”

Unmatched prices

In addition to the deals for federal programs, the administration also announced new direct-to-consumer prices. Currently, people with a prescription can buy the most popular drugs, Wegovy and Zepbound, directly from Novo Nordisk and Eli Lilly, respectively, for $499 each. Under the new deal, Wegovy will be available for $350, as will Ozempic. And Zepbound will be available at “an average” of $346. While the prices are lower, the out-of-pocket costs are still likely to be more than most people would pay if they went through an insurance plan, and paying outside their insurance policies means that the payments won’t be counted toward out-of-pocket maximums and other tallies. Generally, experts expect that direct-to-consumer sales won’t play a significant role in lowering overall drug costs.

It remains unclear if Trump’s deal will have any effect on GLP-1 prices for those on commercial insurance plans.

Trump hailed the deals, calling them “most favored-nation pricing.” But even with the lower prices for some, Americans are still paying more than foreign counterparts. As Sen. Bernie Sanders (I-Vt.) noted last year, while Novo Nordisk set Ozempic’s list price at nearly $1,000 in the US and the new deal is as low as $245, the drug costs just $155 in Canada, $122 in Italy, $71 in France, and $59 in Germany. Wegovy, similarly, is $186 in Denmark, $137 in Germany, and $92 in the United Kingdom. Eli Lilly’s Mounjaro is $94 in Japan.

A study published last year in JAMA Network Open led by researchers at Yale University estimated that the manufacturing cost for this class of drugs is under $5 for a month’s supply.

The announcement also said that future GLP-1 drugs in pill form (rather than injections) from the two companies will be priced at $150. That price will be for federal programs and direct-to-consumer sales. While such pills are nearing the market, none are currently available or approved by the Food and Drug Administration. Given that they are not yet for sale, the cost savings from this deal are unknown.

Questions swirl after Trump’s GLP-1 pricing deal announcement Read More »

With Skigill, the classic RPG skill tree becomes a crowded battlefield

achromi, gaming, roguelike, skill tree, survivors / Beth Washington / November 7, 2025

The stark presentation can threaten to be overwhelming at times, especially when nearly the entire screen descends into a mass of yellow pixels. But the simplicity of the colors also makes it easy to unfocus your eyes and instinctually plan a path away from the red and yellow threats and through to the safety of the inky black void.

Skigill also benefits from allowing players to speed up or slow down time with a quick tap of the shoulder buttons. This is a lifesaver for the tedium of the early parts of runs, when the enemies are coming in at a relative trickle and you don’t want to stand around waiting for the actual challenging part of the run.

Thankfully, you can fast-forward through the early-run doldrums… Credit: Achromi

Where Skigill suffers most, for the moment, is in an overall lack of variety. The recent Early Access release only has three playable characters and four skill-tree maps. While they differ in some important ways, the options fall far short of the dozens of arcane unlockable options that can keep players coming back to a game like Vampire Survivors for dozens of hours.

For replayability, Skigill currently relies on a difficulty system that simply scales up the quantity and health of enemies alongside a secondary skill tree that lets you purchase more permanent upgrades from the main menu. Unfortunately, even at the higher difficulty levels, you’ll see the same four yellow-on-black enemies filling your immediate vicinity, with little in the way of interesting movement or attack patterns to mix things up.

The developers promise on Steam that Skigill will see at least three new characters “along with more weapons and build possibilities” as it goes from Early Access to a Version 1.0 launch. We’re happy to hear that, because even this limited, somewhat repetitive taste of the game has had us coming back for the fun take on an old RPG staple.

With Skigill, the classic RPG skill tree becomes a crowded battlefield Read More »

Mark Zuckerberg’s illegal school drove his neighbors crazy

Features, mark zuckerberg, Meta, Policy, syndication, ZOning / Beth Washington / November 7, 2025

Neighbors complained about noise, security guards, and hordes of traffic.

An entrance to Mark Zuckerberg’s compound in Palo Alto, California. Credit: Loren Elliott/Redux

The Crescent Park neighborhood of Palo Alto, California, has some of the best real estate in the country, with a charming hodgepodge of homes ranging in style from Tudor revival to modern farmhouse and contemporary Mediterranean. It also has a gigantic compound that is home to Mark Zuckerberg, his wife Priscilla Chan, and their daughters Maxima, August, and Aurelia. Their land has expanded to include 11 previously separate properties, five of which are connected by at least one property line.

The Zuckerberg compound’s expansion first became a concern for Crescent Park neighbors as early as 2016, due to fears that his purchases were driving up the market. Then, about five years later, neighbors noticed that a school appeared to be operating out of the Zuckerberg compound. This would be illegal under the area’s residential zoning code without a permit. They began a crusade to shut it down that did not end until summer 2025.

WIRED obtained 1,665 pages of documents about the neighborhood dispute—including 311 records, legal filings, construction plans, and emails—through a public record request filed to the Palo Alto Department of Planning and Development Services. (Mentions of “Zuckerberg” or “the Zuckerbergs” appear to have been redacted. However, neighbors and separate public records confirm that the property in question belongs to the family. The names of the neighbors who were in touch with the city were also redacted.)

The documents reveal that the school may have been operating as early as 2021 without a permit to operate in the city of Palo Alto. As many as 30 students might have enrolled, according to observations from neighbors. These documents also reveal a wider problem: For almost a decade, the Zuckerbergs’ neighbors have been complaining to the city about noisy construction work, the intrusive presence of private security, and the hordes of staffers and business associates causing traffic and taking up street parking.

Over time, neighbors became fed up with what they argued was the city’s lack of action, particularly with respect to the school. Some believed that the delay was because of preferential treatment to the Zuckerbergs. “We find it quite remarkable that you are working so hard to meet the needs of a single billionaire family while keeping the rest of the neighborhood in the dark,” reads one email sent to the city’s Planning and Development Services Department in February. “Just as you have not earned our trust, this property owner has broken many promises over the years, and any solution which depends on good faith behavioral changes from them is a failure from the beginning.”

Palo Alto spokesperson Meghan Horrigan-Taylor told WIRED that the city “enforces zoning, building, and life safety rules consistently, without regard to who owns a property.” She also refuted the claim that neighbors were kept in the dark, claiming that the city’s approval of construction projects at the Zuckerberg properties “were processed the same way they are for any property owner.” She added that, though some neighbors told the city they believe the Zuckerbergs received “special treatment,” that is not accurate.

“Staff met with residents, conducted site visits, and provided updates by phone and email while engaging the owner’s representative to address concerns,” Horrigan-Taylor said. “These actions were measured and appropriate to abate the unpermitted use and responsive to neighborhood issues within the limits of local and state law.”

According to The New York Times, which first reported on the school’s existence, it was called “Bicken Ben School” and shared a name with one of the Zuckerbergs’ chickens. The listing for Bicken Ben School, or BBS for short, in a California Department of Education directory claims the school opened on October 5, 2022. This, however, is the year after neighbors claim to have first seen it operating. It’s also two and a half years after Sara Berge—the school’s point of contact, per documents WIRED obtained from the state via public record request—claims to have started her role as “head of school” for a “Montessori pod” at a “private family office” according to her LinkedIn profile, which WIRED viewed in September and October. Berge did not respond to a request to comment.

Between 2022 and 2025, according to the documents Bicken Ben filed to the state, the school grew from nine to 14 students ranging from 5 to 10 years old. Neighbors, however, estimated that they observed 15 to 30 students. Berge similarly claimed on her LinkedIn profile to have overseen “25 children” in her job. In a June 2025 job listing for “BBS,” the school had a “current enrollment of 35–40 students and plans for continued growth,” which the listing says includes a middle school.

In order for the Zuckerbergs to run a private school on their land, which is in a residential zone, they need a “conditional use” permit from the city. However, based on the documents WIRED obtained, and Palo Alto’s public database of planning applications, the Zuckerbergs do not appear to have ever applied for or received this permit.

Per emails obtained by WIRED, Palo Alto authorities told a lawyer working with the Zuckerbergs in March 2025 that the family had to shut down the school on its compound by June 30. A state directory lists BBS, the abbreviation for Bicken Ben School, as having operated until August 18, and three of Zuckerberg’s neighbors—who all requested anonymity due to the high-profile nature of the family—confirmed to WIRED in late September that they had not seen or heard students being dropped off and picked up on weekdays in recent weeks.

However, Zuckerberg family spokesperson Brian Baker tells WIRED that the school didn’t close, per se. It simply moved. It’s not clear where it is now located, or whether the school is operating under a different name.

In response to a detailed request for comment, Baker provided WIRED with an emailed statement on behalf of the Zuckerbergs. “Mark, Priscilla and their children have made Palo Alto their home for more than a decade,” he said. “They value being members of the community and have taken a number of steps above and beyond any local requirements to avoid disruption in the neighborhood.”

“Serious and untenable”

By the fall of 2024, Zuckerberg’s neighbors were at their breaking point. At some point in mid-2024, according to an email from then mayor Greer Stone, a group of neighbors had met with Stone to air their grievances about the Zuckerberg compound and the illegal school they claimed it was operating. They didn’t arrive at an immediate resolution.

In the years prior, the city had received several rounds of complaints about the Zuckerberg compound. Complaints for the address of the school were filed to 311, the nationwide number for reporting local non-emergency issues, in February 2019, September 2021, January 2022, and April 2023. They all alleged that the property was operating illegally under city code. Both were closed by the planning department, which found no rule violations. An unknown number of additional complaints, mentioned in emails among city workers, were also made between 2020 and 2024—presumably delivered via phone calls, in person, or to city departments not included in WIRED’s public record request.

In December 2020, building inspection manager Korwyn Peck wrote to code enforcement officer Brian Reynolds about an inspection he attempted to conduct around the Zuckerberg compound, in response to several noise and traffic complaints from neighbors. He described that several men in SUVs had gathered to watch him, and a tense conversation with one of them had ensued. “This appears to be a site that we will need to pay attention to,” Peck wrote to Reynolds.

“We have all been accused of ‘not caring,’ which of course is not true,” Peck added. “It does appear, however, with the activity I observed tonight, that we are dealing with more than four simple dwellings. This appears to be more than a homeowner with a security fetish.”

In a September 11, 2024, email to Jonathan Lait, Palo Alto’s director of planning and development services and Palo Alto city attorney Molly Stump, one of Zuckerberg’s neighbors alleged that since 2021, “despite numerous neighborhood complaints” to the city of Palo Alto, including “multiple code violation reports,” the school had continued to grow. They claimed that a garage at the property had been converted into another classroom, and that an increasing number of children were arriving each day. Lait and Stump did not respond to a request to comment.

“The addition of daily traffic from the teachers and parents at the school has only exacerbated an already difficult situation,” they said in the email, noting that the neighborhood has been dealing with an “untenable traffic” situation for more than eight years.

They asked the city to conduct a formal investigation into the school on Zuckerberg’s property, adding that their neighbors are also “extremely concerned” about the school, and “are willing to provide eyewitness accounts in support of this complaint.”

Over the next week, another neighbor forwarded this note to all six Palo Alto city council members, as well as then mayor Stone. One of these emails described the situation as “serious” and “untenable.”

“We believe the investigation should be swift and should yield a cease and desist order,” the neighbor wrote.

Lait responded to the neighbor who sent the original complaint on October 15, claiming that he’d had an “initial call” with a “representative” of the property owners and that he was directing the city’s code enforcement staff to reexamine the property.

On December 11, 2024, the neighbor claimed that since one of their fellow neighbors had spoken to a Zuckerberg representative, and the representative had allegedly admitted that there was a school on the property, “it seems like an open and shut case.”

“Our hope is that there is an equal process in place for all residents of Palo Alto regardless of wealth or stature,” the neighbor wrote. “It is hard to imagine that this kind of behavior would be ignored in any other circumstance.”

That same day, Lait told Christine Wade, a partner at SSL Law Firm—who, in an August 2024 email thread, said she was “still working with” the Zuckerberg family—that the Zuckerbergs lacked the required permit to run a school in a residential zone.

“Based on our review of local and state law, we believe this use constitutes a private school use in a residential zone requiring a conditional use permit,” Lait wrote in an email to Wade. “We also have not found any state preemptions that would exclude a use like this from local zoning requirements.” Lait added that a “next step,” if a permit was not obtained, would be sending a cease and desist to the property owner.

According to several emails, Wade, Lait, and Mark Legaspi, CEO of the Zuckerberg family office called West 10, went on to arrange an in-person meeting at City Hall on January 9. (This is the first time that the current name of the Zuckerberg family office, West 10, has been publicly disclosed. The office was previously called West Street.) Although WIRED did not obtain notes from the meeting, Lait informed the neighbor on January 10 that he had told the Zuckerbergs’ “representative” that the school would need to shut down if it didn’t get a conditional use permit or apply for that specific permit.

Lait added that the representative would clarify what the family planned to do in about a week; however, he noted that if the school were to close, the city may give the school a “transition period” to wind things down. Wade did not respond to a request for comment.

“At a minimum, give us extended breaks”

There was another increasingly heated conversation happening behind the scenes. On February 3 of this year, at least one neighbor met with Jordan Fox, an employee of West 10.

It’s unclear exactly what happened at this meeting, or if the neighbor who sent the September 11 complaint was in attendance. But a day after the meeting with Fox, two additional neighbors added their names to the September 11 complaint, per an email to Lait.

On February 12, a neighbor began an email chain with Fox. This email was forwarded to Planning Department officials two months later. The neighbor, who seemingly attended the meeting, said they had “connected” with fellow neighbors “to review and revise” an earlier list of 14 requests that had been reportedly submitted to the Zuckerbergs at some previous point. The note does not specify the contents of this original list of requests, but of the 19 neighbors who originally contributed to it, they claimed that 15 had contributed to the revised list.

The email notes that the Zuckerbergs had been “a part of our neighborhood for many years,” and that they “hope that this message will start an open and respectful dialogue,” built upon the “premise of how we all wish to be treated as neighbors.”

“Our top requests are to minimize future disruption to the neighborhood and proactively manage the impact of the many people who are affiliated with you,” the email says. This includes restricting parking by “security guards, contractors, staff, teachers, landscapers, visitors, etc.” In the event of major demolitions, concrete pours, or large parties, the email asks for advance notice, and for dedicated efforts to “monitor and mitigate noise.”

The email also asks the Zuckerbergs to, “ideally stop—but at a minimum give us extended breaks from—the acquisition, demolition and construction cycle to let the neighborhood recover from the last eight years of disruption.”

At this point, the email requests that the family “abide by both the letter and the spirit of Palo Alto” by complying with city code about residential buildings.

Specifically, it asks the Zuckerbergs to get a use permit for the compound’s school and to hold “a public hearing for transparency.” It also asks the family to not expand its compound any further. “We hope this will help us get back the quiet, attractive residential neighborhood that we all loved so much when we chose to move here.”

In a follow-up on March 4, Fox acknowledged the “unusual” effects that come with being neighbors with Mark Zuckerberg and his family.

“I recognize and understand that the nature of our residence is unique given the profile and visibility of the family,” she wrote. “I hope that as we continue to grow our relationship with you over time, you will increasingly enjoy the benefits of our proximity—e.g., enhanced safety and security, shared improvements, and increased property values.”

Fox said that the Zuckerbergs instituted “a revised parking policy late last year” that should address their concerns, and promised to double down on efforts to give advanced notice about construction, parties, and other potential disruptions.

However, Fox did not directly address the unpermitted school and other nonresidential activities happening at the compound. She acknowledged that the compound has “residential support staff” including “childcare, culinary, personal assistants, property management, and security,” but said that they have “policies in place to minimize their impact on the neighborhood.”

It’s unclear if the neighbor responded to Fox.

“You have not earned our trust”

While these conversations were happening between Fox and Zuckerberg’s neighbors, Lait and others at the city Planning Department were scrambling to find a solution for the neighbor who complained on September 11, and a few other neighbors who endorsed the complaint in September and February.

Starting in February, one of these neighbors took the lead on following up with Lait. They asked him for an update on February 11, and heard back a few days later. He didn’t have any major updates, “but after conversations with the family’s representatives, he said he was exploring whether a “subset of children” could continue to come to the school sometimes for “ancillary” uses.

“I also believe a more nuanced solution is warranted in this case,” Lait added. Ideally, such a solution would respond to the neighbors’ complaints, but allow the Zuckerbergs to “reasonably be authorized by the zoning code.”

The neighbor wasn’t thrilled. The next day, they replied and called the city’s plan “unsatisfactory.”

“The city’s ‘nuanced solution’ in dealing with this serial violator has led to the current predicament,” they said (referring to the nuanced solution Lait mentioned in his last email.)

Horrigan-Taylor, the Palo Alto spokesperson, told WIRED that Lait’s mention of a “nuanced” solution referred to “resolving, to the extent permissible by law, neighborhood impacts and otherwise permitted use established by state law and local zoning.”

“Would I, or any other homeowner, be given the courtesy of a ‘nuanced solution’ if we were in violation of city code for over four years?” they added.

“Please know that you have not earned our trust and that we will take every opportunity to hold the city accountable if your solution satisfies a single [redacted] property owner over the interests of an entire neighborhood,” they continued.

“If you somehow craft a ‘nuanced solution’ based on promises,” the neighbor said, “the city will no doubt once again simply disappear and the damage to the neighborhood will continue.”

Lait did not respond right away. The neighbor followed up on March 13, asking if he had “reconsidered” his plan to offer a “‘nuanced solution’ for resolution of these ongoing issues by a serial code violator.” They asked when the neighborhood could “expect relief from the almost decade long disruptions.”

Behind the scenes, Zuckerberg’s lawyers were fighting to make sure the school could continue to operate. In a document dated March 14, Wade argues that she believed the activities at “the Property” “represent an appropriate residential use based on established state law as well as constitutional principles.”

Wade said that “the Family” was in the process of obtaining a “Large Family Daycare” license for the property, which is legal for a cohort of 14 or fewer children all under the age of 10.

“We consistently remind our vendors, guests, etc. to minimize noise, not loiter anywhere other than within the Family properties, and to keep areas clean,” Wade added in the letter. Wade also attached an adjusted lease corresponding with the address of the illicit school, which promises that the property will be used for only one purpose. The exact purpose is redacted.

On March 25, Lait told the neighbor that the city’s June 30 deadline for the Zuckerbergs to shut down the school had not changed. However, the family’s representative said that they were pursuing a daycare license. These licenses are granted by the state, not the city of Palo Alto.

The subtext of this email was that if the state gave them a daycare licence, there wasn’t much the city could do. Horrigan-Taylor confirmed with WIRED that “state licensed large family day care homes” do not require city approval, adding that the city also “does not regulate homeschooling.”

“Thanks for this rather surprising information,” the neighbor replied about a week later. “We have repeatedly presented ideas to the family over the past 8 years with very little to show for it, so from our perspective, we need to understand the city’s willingness to act or not to act.”

Baker told WIRED that the Zuckerbergs never ended up applying for a daycare license, a claim that corresponds with California’s public registry of daycare centers. (There are only two registered daycare centers in Palo Alto, and neither belongs to the Zuckerbergs. The Zuckerbergs’ oldest child, Maxima, will also turn 10 in December and consequently age out of any daycare legally operating in California.)

Horrigan-Taylor said that a representative for the Zuckerbergs told the city that the family wanted to move the school to “another location where private schools are permitted by right.”

In a school administrator job listing posted to the Association Montessori International website in July 2022 for “BBS,” Bicken Ben head of school Berge claims that the school had four distinct locations, and that applicants must be prepared to travel six to eight weeks per year. The June 2025 job listing also says that the “year-round” school spans “across multiple campuses,” but the main location of the job is listed as Palo Alto. It’s unclear where the other sites are located.

Most of the Zuckerbergs’ neighbors did not respond to WIRED’s request for comment. However, the ones that did clearly indicated that they would not be forgetting the Bicken Ben saga, or the past decade of disruption, anytime soon.

“Frankly I’m not sure what’s going on,” one neighbor said, when reached by WIRED via landline. “Except for noise and construction debris.”

This story originally appeared on wired.com.

Wired.com is your essential daily guide to what’s next, delivering the most original and complete take you’ll find anywhere on innovation’s impact on technology, science, business and culture.

Mark Zuckerberg’s illegal school drove his neighbors crazy Read More »

Google plans secret AI military outpost on tiny island overrun by crabs

AI, AI and work, AI infrastructure, australia, Biz & IT, Christmas Island, Cloud Computing, crabs, datacenters, defense, Google, machine learning, Military, red crabs, SubCom, subsea cables / Beth Washington / November 6, 2025

Christmas Island Shire President Steve Pereira told Reuters that the council is examining community impacts before approving construction. “There is support for it, providing this data center actually does put back into the community with infrastructure, employment, and adding economic value to the island,” Pereira said.

That’s great, but what about the crabs?

Christmas Island’s annual crab migration is a natural phenomenon that Sir David Attenborough reportedly once described as one of his greatest TV moments when he visited the site in 1990.

Every year, millions of crabs emerge from the forest and swarm across roads, streams, rocks, and beaches to reach the ocean, where each female can produce up to 100,000 eggs. The tiny baby crabs that survive take about nine days to march back inland to the safety of the plateau.

While Google is seeking environmental approvals for its subsea cables, the timing could prove delicate for Christmas Island’s most famous residents. According to Parks Australia, the island’s annual red crab migration has already begun for 2025, with a major spawning event expected in just a few weeks, around November 15–16.

During peak migration times, sections of roads close at short notice as crabs move between forest and sea, and the island has built special crab bridges over roads to protect the migrating masses.

Parks Australia notes that while the migration happens annually, few baby crabs survive the journey from sea to forest most years, as they’re often eaten by fish, manta rays, and whale sharks. The successful migrations that occur only once or twice per decade (when large numbers of babies actually survive) are critical for maintaining the island’s red crab population.

How Google’s facility might coexist with 100 million marching crustaceans remains to be seen. But judging by the size of the event, it seems clear that it’s the crab’s world, and we’re just living in it.

Google plans secret AI military outpost on tiny island overrun by crabs Read More »

Space junk may have struck a Chinese crew ship in low-Earth orbit

china, chinese space program, human spaceflight, Science, shenzhou, shenzhou 20, shenzhou 21, Space, tiangong / Beth Washington / November 5, 2025

Three Chinese astronauts were due to depart the Tiangong space station, reenter the atmosphere, and land in the remote desert of Inner Mongolia on Wednesday. Instead, officials ordered the crew to remain at the station while engineers investigate a potential problem with their landing craft.

The China Manned Space Agency, run by the country’s military, announced the change late Tuesday in a brief statement posted to Weibo, the Chinese social media platform.

“The Shenzhou 20 manned spacecraft is suspected of being impacted by small space debris,” the statement said. “Impact analysis and risk assessment are underway. To ensure the safety and health of the astronauts and the complete success of the mission, it has been decided that the Shenzhou 20 return mission, originally scheduled for November 5, will be postponed.”

What we know

The Shenzhou 20 astronauts arrived at the Tiangong station in April. Their replacements on the Shenzhou 21 mission docked with Tiangong on Friday, temporarily raising the station’s crew size to six people. After several days of joint operations, the six astronauts held a handover ceremony early Tuesday to formally transfer command of the outpost to the new crew.

Less than 24 hours later, Chinese officials decided to call off Shenzhou 20’s departure from Tiangong. The statement from the China Manned Space Agency did not say what part of the Shenzhou 20 spacecraft may have been damaged, what evidence led engineers to suspect space debris was the culprit, or how long Shenzhou 20’s departure might be postponed.

This view shows a Shenzhou spacecraft departing the Tiangong space station in 2023. Credit: China Manned Space Agency

The ship has three sections, with a landing capsule positioned between the crew living quarters and a power and propulsion module. The modules separate from one another before reentry, and the return craft heads for a parachute-assisted landing while the other elements burn up during atmospheric reentry.

Space junk may have struck a Chinese crew ship in low-Earth orbit Read More »

Some stinkbugs’ legs carry a mobile fungal garden

agriculture, Biology, bugs, entomology, fungus, insects, Science / Beth Washington / November 4, 2025

Many insect species hear using tympanal organs, membranes roughly resembling our eardrums but located on their legs. Grasshoppers, mantises, and moths all have them, and for decades, we thought that female stinkbugs of the Dinidoridae family have them, too, although located a bit unusually on their hind rather than front legs.

Suspecting that they use their hind leg tympanal organs to listen to male courtship songs, a team of Japanese researchers took a closer look at the organs in Megymenum gracilicorne, a Dinidoridae stinkbug species native to Japan. They discovered that these “tympanal organs” were not what they seemed. They’re actually mobile fungal nurseries of a kind we’ve never seen before.

Portable gardens

Dinidoridae is a small stinkbug family that lives exclusively in Asia. The bug did attract some scientific attention, but not nearly as much as its larger relatives like Pentatomidae. Prior work looking specifically into organs growing on the hind legs of Dinidoridae females was thus somewhat limited. “Most research relied on taxonomic and morphological approaches. Some taxonomists did describe that female Dinidoridae stinkbugs have an enlarged part on the hind legs that looks like the tympanal organ you can find, for example, in crickets,” said Takema Fukatsu, an evolutionary biologist at the National Institute of Advanced Industrial Science and Technology in Tokyo.

Based on that appearance, these parts were classified as tympanal organs—the case was closed, and it stayed closed until Fukatsu’s team started examining them more closely. Most insects have tympanal organs on their front legs, not hind legs, or on abdominal segments. The initial goal of Fukatsu’s study was to figure out what impact this unusual position has on Dinidoridae females’ ability to hear sounds.

Early on in the study, it turned out that whatever Dinidoridae females have on their hind legs, they are not tympanal organs. “We found no tympanal membrane and no sensory neurons, so the enlarged parts on the hind legs had nothing to do with hearing,” Fukatsu explained. Instead, the organ had thousands of small pores filled with benign filamentous fungi. The pores were connected to secretory cells that released substances that Fukatsu’s team hypothesized were nutrients enabling the fungi to grow.

Some stinkbugs’ legs carry a mobile fungal garden Read More »

Author name: Beth Washington