Jack Holloway – pocketgazette.com

Google opens I/O with Gemini, Android updates expected

Google I/O 2026 kicks off Tuesday, May 19, with the main keynote starting at 10 a.m. Pacific Time at the company’s annual developer conference in Mountain View. The two-day event runs through May 20 and is free to watch online via Google’s YouTube channel and the I/O website, so nobody needs a badge to follow along.

Google has framed this year’s conference around what it calls the “agentic era” of development, a phrase from the company’s developer blog that signals the keynote will lean heavily on AI systems capable of handling multi-step tasks with minimal human hand-holding. That framing puts Gemini at the center of the show, and previews from several technology outlets suggest the AI model family will be the through line connecting nearly every product announcement on the agenda.

Android 17 is the other headliner everyone is watching. The next major version of Google’s mobile operating system is widely expected to get stage time, though the specifics of what changes it brings remained unconfirmed before the keynote. Google has not released a public timeline for Android 17, and I/O has historically been where the company lays out its roadmap for developers who need to prepare their apps for whatever comes next.

Then there are the glasses. Google has confirmed plans to release smart glasses built around Android XR in 2026, according to Mashable, which noted that the company has not settled on a final product name. Wired said attendees and viewers should expect another look at the Android XR smart glasses, which have been previewed before but not yet shipped. CNET framed the wearable question more broadly, asking whether smart glasses could become a genuine pillar of Google’s device lineup rather than a niche curiosity. That is a reasonable question given how the category has underdelivered for years, and Google’s answer, or at least its pitch, is apparently coming today.

PCMag said before the keynote that it expected significant changes to Gemini and Android Auto, along with possible details on a device it referred to as a Googlebook running something called Aluminum OS. That item was firmly in the realm of pre-show speculation rather than confirmed news, and it is the kind of thing that either gets announced or quietly disappears from the conversation by Tuesday afternoon.

Beyond hardware, Google’s developer blog described the event as a chance to showcase tools designed to automate complex workflows and simplify building AI-ready applications. That language is aimed squarely at the developer community, and there is typically a second keynote focused specifically on APIs, tools, and features that engineers can actually use. Mashable noted that this developer-focused session usually runs separately from the main consumer keynote, giving technical attendees something more substantive than product trailers.

Watching options are broad by design. The keynote livestreams on Google’s official YouTube channel and at the I/O site. CNET planned a YouTube viewing party with live updates from reporters on the ground in Mountain View, plus coverage before and after the main event. Canaltech reported that Google will also offer an American Sign Language version of the livestream alongside the standard broadcast, which is a straightforward accessibility measure worth mentioning for anyone who needs it.

Chrome and Cloud were also listed by Google as areas that would see updates across the two-day schedule, though details ahead of the keynote were thin. Broader Gemini integrations across Google’s product suite were another widely cited expectation, consistent with the company’s push over the past year to embed AI into Search, Maps, Workspace, and essentially everything else it ships.

The conference arrives at a moment when Google is under genuine competitive pressure in AI, with rivals moving fast and the search business facing structural questions it has not had to answer before. I/O is as much a statement of direction as it is a product showcase, and the choice to lead with an “agentic era” keynote suggests Google wants to make clear it sees the next phase of AI as less about chatbots answering questions and more about systems that take action. Whether the announcements live up to that framing is what the next several hours will determine.

iOS 27 to add new custom wallpaper feature, more: report

Apple is three weeks out from its annual developer conference, and Bloomberg is already filling in the picture. A new report from the outlet details three features coming to iOS 27: a smarter grammar checker baked into Writing Tools, a natural-language interface for building Shortcuts automations, and an AI-powered custom wallpaper generator tied to the existing Image Playground app.

The grammar checker is probably the most immediately practical of the three. According to Bloomberg, it works similarly to Grammarly, the popular third-party writing assistant, except it lives natively in the operating system. The feature surfaces in a translucent menu that slides up from the bottom of the screen, showing original text next to suggested revisions. Users can accept individual suggestions one at a time, approve everything in bulk, or dismiss the whole thing if autocorrect has already burned them one too many times. There are also controls to pause grammar checking mid-document and jump between flagged sections. For anyone who has fumbled through an important email on their phone, this is the kind of thing that quietly improves daily life without requiring much adjustment.

Alongside the grammar checker, Apple is reportedly rethinking how Writing Tools surface to users in the first place. The company has been testing a "Write With Siri" toggle that appears directly at the top of the keyboard, and a "Help Me Write" button that shows up when a user activates Siri while working inside a text field. The underlying capability is not new, but it has been tucked away enough that plenty of people have never found it. Making it a persistent, visible option at the keyboard level is a reasonable fix to an adoption problem Apple has had with its AI writing features since their introduction.

The Shortcuts update is where things get genuinely interesting for power users and casual users alike. Right now, building a useful Shortcut requires at least some tolerance for the app's logic-flow interface, which can feel like assembling furniture without the right instructions. iOS 27 apparently changes that by letting people describe what they want a shortcut to do in plain language. The app presents a prompt asking "What do you want your shortcut to do?" and a text field to type the answer. From there, the system builds and installs the shortcut automatically.

That is a meaningful shift. Shortcuts has always been one of those features iPhone users appreciate in theory but rarely dig into in practice. Reducing the barrier to entry from "learn the interface" to "describe what you want" could push the app into genuinely mainstream use, rather than remaining the territory of automation enthusiasts who post their setups on Reddit.

The wallpaper feature rounds out the trio. Bloomberg reports that iOS 27 will include an option in the standard wallpaper picker to generate custom lock screen and home screen backgrounds using AI, built on top of the Image Playground app Apple introduced with iOS 18. The framing here is straightforward: rather than browsing static options or pulling from your photo library, you could describe or prompt an image and have it generated on the spot.

Apple has been careful about how it positions image generation, keeping it largely contained to Image Playground and specific creative contexts rather than scattering it across the operating system. Folding it into the wallpaper picker is a logical expansion that keeps the feature anchored in a place where people already expect to be creative, and where the stakes of a weird or off-putting result are low. You can always just pick a different wallpaper.

Taken together, these three features point toward a consistent theme in how Apple is building out its AI layer: less about dramatic new capabilities and more about making existing things easier to find, easier to use, and more useful in the moments where people actually need them. Grammar help when you are typing, Shortcuts you can build without a manual, wallpapers you can generate without leaving the settings screen. None of it is science fiction, but that is somewhat the point.

Apple will formally unveil iOS 27 at its Worldwide Developers Conference, which begins in roughly three weeks. The full scope of what the company has planned will become clearer then, but Bloomberg's reporting suggests the software update leans heavily into making Apple Intelligence feel less like a feature to seek out and more like infrastructure that is simply there when you reach for it.

Watch Vega C rocket launch European-Chinese space weather satellite to orbit tonight

A European-Chinese spacecraft built to map the invisible battle between the sun and Earth's magnetic field made it to orbit Monday night, lifted there by a Vega C rocket that departed French Guiana right on schedule.

The SMILE satellite lifted off from Europe's Spaceport in Kourou at 11:52 p.m. EDT on May 18, and about 56 minutes later it was circling Earth in a clean circular orbit 439 miles (707 kilometers) up. No drama, no anomalies. For a mission that has been years in the making between two of the world's biggest space agencies, a smooth launch was exactly the right start.

SMILE stands for Solar wind Magnetosphere Ionosphere Link Explorer, which is a mouthful, but the science behind it is genuinely fascinating. The sun is constantly exhaling a stream of charged particles called the solar wind. When that stream hits Earth's magnetosphere, the invisible magnetic bubble surrounding our planet, the results range from spectacular auroras to crippling geomagnetic storms that can knock out power grids and fry satellites. SMILE's job is to watch that interaction in real time, giving scientists a much clearer picture of how and why it happens.

The mission is a collaboration between the European Space Agency and the Chinese Academy of Sciences, and the division of labor is worth noting. China's academy was responsible for the satellite platform, spacecraft operations, and three of the four onboard science instruments: the Ultraviolet Imager, the Light Ion Analyser, and the Magnetometer. ESA contributed the payload module, the fourth instrument (a Soft X-ray Imager), the rocket itself, and integration and testing services. ESA also had a hand in building the UV imager and will share operations duties once SMILE is up and running.

That four-instrument suite is what makes SMILE more than just another monitoring satellite. Together, the tools will allow scientists to image the magnetosphere's response to solar activity rather than just measuring it at a single point, which is how most space weather observations have worked until now. Think of the difference between a thermometer outside your window and a full weather radar sweep of the entire region.

SMILE is not quite ready to start science operations yet. Over the next 25 days, the spacecraft will conduct 11 engine burns to reshape its orbit from that initial circular path into a highly elliptical one. At its peak, SMILE will swing 75,185 miles (121,000 kilometers) above the North Pole while dipping to just 3,107 miles (5,000 kilometers) over the South Pole. That lopsided orbit is intentional: it gives the spacecraft the wide vantage point needed to watch large-scale magnetospheric dynamics play out.

Once in its working orbit, the mission team will run through a series of instrument checkouts before declaring the spacecraft fit for science. ESA says the first X-ray and ultraviolet images should arrive roughly three months after launch, after which the real work begins. The planned mission lifetime is three years.

The rocket that got SMILE there is itself something of a comeback story. The Vega C is a 115-foot-tall launcher developed by ESA and built by Italian company Avio. It made its debut in July 2022 but suffered a launch failure in December of that year, grounding the vehicle for nearly two years while engineers identified and corrected the problem. Monday's mission was Vega C's seventh flight overall and its sixth success. It was also the first Vega C launch managed by Avio directly; previous missions were handled by Arianespace, the French launch services company.

The practical stakes for missions like SMILE are higher than they might initially seem. Space weather is not an abstract concern for astronomers. A severe geomagnetic storm can induce currents in long electrical conductors on the ground, potentially damaging transformers across wide areas of the power grid. GPS and radio communications degrade during solar events. Satellites in low Earth orbit experience increased atmospheric drag when the upper atmosphere expands during geomagnetic activity, shortening their operational lives. The more precisely scientists can model these interactions, the better operators can prepare and respond.

For ESA and the Chinese Academy of Sciences, SMILE also represents something beyond the science. Large-scale space science collaborations between Europe and China are not especially common, and a successful mission here carries diplomatic weight alongside its research value. The fact that both agencies split hardware responsibilities rather than one simply buying services from the other makes the partnership more substantive than many international space arrangements tend to be.

For now, SMILE is in orbit and healthy, quietly firing its engines over the coming weeks to settle into its final post. By late summer, if everything goes to plan, scientists should be looking at the first X-ray portraits of Earth's magnetosphere taken from a genuinely novel vantage point. The sun will keep blowing. Earth will keep pushing back. And for the first time, there will be an eye in orbit watching the whole thing from above.

California jury rejects Musk lawsuit against OpenAI

A nine-person federal jury in Oakland, California, unanimously rejected Elon Musk's lawsuit against OpenAI, Sam Altman, Greg Brockman, the OpenAI Foundation, and Microsoft on Monday, ending a three-week trial with a verdict that took jurors less than two hours to reach. The outcome turned not on whether OpenAI had betrayed its founding mission, but on something far more procedural: Musk waited too long to complain about it.

U.S. District Judge Yvonne Gonzalez Rogers adopted the advisory verdict and dismissed all of Musk's claims on statute-of-limitations grounds. The applicable deadlines were three years for breach of charitable trust claims and two years for unjust enrichment claims. By the time Musk filed suit, the clock had already run out.

The core of Musk's case was that Sam Altman and Greg Brockman had steered OpenAI away from its original nonprofit mission without his knowledge, enriching themselves and their investors in the process. Musk helped found the company in 2015 and contributed roughly $38 million in its early years, and he argued that those contributions came with an implicit understanding that OpenAI would remain a public-interest organization rather than a for-profit enterprise.

OpenAI pushed back hard on that framing. The company argued that there was never any promise to remain a nonprofit in perpetuity, and that Musk had been aware of discussions about restructuring dating back to 2017. A for-profit subsidiary was created in 2019. OpenAI's lawyers also argued that Musk filed the lawsuit not out of principled concern for the company's mission but because he failed to gain unilateral control of one of the most valuable AI organizations in the world. That argument landed well enough that it apparently influenced the broader narrative of the trial, even if the jury's final decision rested on timing alone.

Microsoft, which has made substantial investments in OpenAI, was also named in the suit. Jurors rejected the claim that Microsoft aided and abetted a breach of charitable trust, again on statute-of-limitations grounds. Microsoft said after the verdict that the facts and timeline had "long been clear" and that it remained committed to its work with OpenAI.

The trial drew testimony from some of Silicon Valley's most prominent figures and unfolded over three weeks in Oakland. By the standards of tech industry litigation, which tends to be expensive, slow, and often anticlimactic, this one delivered a fairly decisive conclusion. The jury's sub-two-hour deliberation suggests the limitations question was not a close call.

Musk posted on X, the platform he owns, that he would appeal. Whether that appeal has a realistic path forward is a separate question, but it is consistent with how he has approached this dispute since filing the original complaint. The case has always been as much about public positioning as legal remedy, and an appeal keeps the narrative alive even after the verdict.

For OpenAI, the ruling clears one significant obstacle as the company pursues a conversion to a for-profit structure that has generated scrutiny from regulators and competitors alike. The Straits Times noted that the verdict removes a legal cloud that had been hanging over any potential IPO plans. None of that is settled by Monday's jury decision, but losing a $150 billion lawsuit, as the New York Times characterized the stakes, would have been a more complicated starting point for whatever comes next.

The verdict does not resolve the deeper tensions behind the case. Questions about how AI companies govern themselves, what obligations early donors and co-founders retain when an organization pivots toward commercial priorities, and whether nonprofit origins can constrain a company indefinitely are genuinely unresolved in American law. Those questions just will not be answered in this particular courtroom.

What the verdict does establish is that if you believe you have been wronged, the law expects you to say so within a reasonable window of time. OpenAI's transformation from nonprofit research lab to one of the most commercially valuable AI companies in the world happened gradually and, according to the jury, with sufficient visibility that Musk should have acted sooner. The case he eventually brought arrived after the statute of limitations had expired, and that was enough to end it.

Jury rejects Musk’s OpenAI lawsuit as too late

A nine-person federal jury in Oakland, California, took less than two hours on Monday to reject every claim Elon Musk brought against OpenAI, Sam Altman, Greg Brockman, and Microsoft. The verdict was unanimous and swift: Musk waited too long to file his 2024 lawsuit, and the applicable statutes of limitations had expired. U.S. District Judge Yvonne Gonzalez Rogers adopted the advisory verdict as the court's decision and dismissed all claims on the spot.

The suit had grown from a genuine dispute over what OpenAI was supposed to be. Musk co-founded the company alongside Altman and Brockman and contributed roughly $38 million in its early years, when OpenAI operated as a nonprofit with a stated mission to develop artificial intelligence for the broad benefit of humanity. His core allegation was that Altman and Brockman betrayed that founding mission by attaching a for-profit corporate structure to the organization and accepting large investments from Microsoft, among others. Musk's legal theories included breach of charitable trust and unjust enrichment against the OpenAI executives, plus an aiding-and-abetting claim against Microsoft.

The central dispute at trial was not really whether any of that happened. It was about when Musk found out. His legal team argued that Microsoft's $10 billion investment in 2023 was the moment he realized OpenAI had strayed from its founding terms, making that the clock's start date. OpenAI countered with earlier evidence of restructuring that, in its telling, Musk had known about for years before finally deciding to sue.

The jury sided with OpenAI's timeline. Judge Gonzalez Rogers, who presided over three weeks of testimony from Musk, Altman, and Microsoft CEO Satya Nadella, said afterward that there was "a substantial amount of evidence to support the jury's finding," adding that she had been prepared to dismiss the case on the spot once deliberations concluded. That is not a judge softening a loss for the plaintiff.

Musk took to X to offer his own framing. The judge and jury, he wrote, "never actually ruled on the merits of the case, just on a calendar technicality," and he announced plans to appeal. Calling a statute of limitations a technicality is a bit like calling the deadline for a tax return a technicality. The law treats timeliness as substantive, not administrative, and courts have long held that plaintiffs who sleep on their rights do so at their own peril. Whether an appeals court sees enough reason to revisit the question remains to be seen.

The practical stakes beyond the courtroom are significant. OpenAI has been working toward a transition to a fully for-profit structure, and a live lawsuit from one of its co-founders challenging the legitimacy of that evolution was the kind of legal cloud that complicates those conversations considerably. As Fast Company noted, the verdict clears a meaningful obstacle on the path toward a potential IPO. Investors and partners interested in OpenAI's long-term trajectory will notice that the company's founding narrative is no longer subject to active litigation.

What the trial did not settle is the underlying philosophical argument Musk has been making in public for years. OpenAI began as a nonprofit precisely because its founders, Musk among them, believed the technology was too consequential to be steered purely by profit motives. The shift toward a commercial model and deep partnership with Microsoft is exactly what Musk claimed he feared. That argument has merit as a policy debate. It just wasn't, in the end, a viable legal claim given when he chose to make it.

Three weeks of courtroom proceedings did surface some texture about how OpenAI operates and how its relationship with Microsoft evolved, though any revelations were largely absorbed into the verdict's procedural conclusion rather than producing a definitive public accounting. Altman and Nadella testified about governance and investment structure; Musk testified about his own understanding of the organization's direction. None of it overcame the timeline problem.

For now, the score is OpenAI 1, Musk 0, with an appeal pending and the broader rivalry between the two men's AI ventures very much ongoing. Musk's xAI has been building out its Grok models and competing directly for enterprise customers and talent. The legal route having closed, at least for the moment, the competition shifts back to the product and market fronts where it was always going to be decided anyway.

Apple sets June 8 keynote for WWDC26

Apple has locked in June 8 for the WWDC26 keynote, kicking off a five-day developer conference that runs through June 12 at Apple Park. The opening keynote starts at 10 a.m. PDT, with the Platforms State of the Union following at 1 p.m. PDT the same day. Apple says the event will cover AI advancements, software features, and developer tools across its platform lineup, which is roughly what you'd expect from a company that has spent the better part of two years trying to convince the world its AI ambitions are serious.

The invite Apple sent to media may be doing some of that convincing on its own. The tagline reads "Coming bright up," and the design leans heavily on a bright, Swift-themed visual. Macworld noted the Swift imagery was unusually prominent even by WWDC standards, where Swift updates are more or less a given. Mashable, TechRadar, and others have read the bright visual language as a signal that AI-related announcements will take center stage. Whether that interpretation holds up on June 8 remains to be seen, but Apple is rarely subtle when it wants to set expectations through an event invite.

More than 1,000 developers, designers, and students will be at Apple Park for the opening-day events. The broader developer program stays online and free, keeping global access intact for anyone who wants to follow sessions, labs, and one-on-one meetings with Apple engineers and designers without booking a flight to Cupertino.

There is an additional layer of significance this year that goes beyond software versions and framework updates. Both CNET and CNA have described this as potentially Tim Cook's final WWDC keynote as Apple CEO, and Macworld referred to John Ternus as incoming CEO. Cook has led Apple since 2011, and if a leadership transition is in fact underway, WWDC26 becomes a notable marker regardless of what gets announced on stage. Apple has not made any official statement on that front, so the reporting should be taken as what it is: informed speculation based on signals the company has not confirmed.

On the software side, iOS 27 is widely expected to be among the headline announcements, alongside updates to macOS, iPadOS, watchOS, and the rest of the platform stack. Apple Intelligence features, including a more capable Siri, have been previewed through leaks and earlier reporting, so WWDC will likely serve as the formal reveal for capabilities that are already reasonably well understood by anyone paying close attention to the coverage.

AppleInsider offered a useful note of caution for anyone nursing hardware hopes: do not expect new Macs at WWDC 2026. The site pointed out that predictions of Mac releases tend to surface before the developer conference with some regularity and tend not to materialize. WWDC is, at its core, a software event, and Apple has generally kept it that way outside of the occasional chip transition announcement or surprise product drop. The odds favor software dominating the agenda this year.

The timing also puts Apple squarely in the middle of the annual developer conference season. Google I/O has already come and gone, and Android Central noted that Apple is revealing WWDC details in Google I/O's shadow. That framing is not entirely unfair. Google has been aggressive about AI feature announcements this year, and Apple is walking into June with some catching up to do in public perception, even if its internal development timeline tells a different story.

For developers, the practical draw is the same as it has been for years: access to new APIs, updated frameworks, direct time with Apple's engineering and design teams, and early looks at OS betas that will ship to consumers later in the fall. Apple framed WWDC26 on its newsroom as an opportunity for developers worldwide to explore tools and technologies and connect directly with Apple engineers and designers. That description is accurate as far as it goes, even if it undersells how much the industry reads these events as a proxy for Apple's strategic direction.

Mark June 8 at 10 a.m. PDT on whatever calendar you use. Whether you're a developer waiting on new APIs, a Tim Cook watcher keeping score, or just someone curious whether Apple's Siri finally catches up to the moment, this one has reasons to tune in.

Apple announces return of popular MagSafe iPhone stand and grip

Apple is bringing back the Hikawa Grip & Stand for iPhone, a MagSafe-compatible accessory that sold out within days of its debut last November. The grip is now available worldwide through Apple's online store, and for the first time, it's being distributed globally thanks to a collaboration between its designer and PopSockets.

The accessory was originally released in November 2025 as part of Apple's celebration of 40 years of accessibility work. It was designed by Bailey Hikawa, a Los Angeles-based artist and industrial designer, and built from the ground up with input from people who have disabilities affecting muscle strength, dexterity, and hand control. The goal was straightforward: make it easier to hold a phone without the constant low-grade effort most people don't even notice they're exerting.

In practice, the Hikawa Grip & Stand does a few things at once. It attaches to the back of any MagSafe-compatible iPhone using magnets, meaning it snaps on securely and can be removed without a fight. Beyond the grip itself, it folds out into a stand that props the phone up in both portrait and landscape orientations, which covers most of the situations where you'd want your hands free. The whole thing is wrapped in a soft-touch premium silicone, which is a small but meaningful detail for anyone who needs a reliable, non-slip surface.

When the accessory launched last fall, it was only available directly from Apple and only in limited quantities. It sold out in a matter of days, which left a lot of interested buyers empty-handed. That kind of fast sellout isn't unusual for Apple limited releases, but this one stung a bit more given the specific audience the accessory was trying to serve. Accessibility tools that aren't available aren't particularly useful.

The return launch coincides with Apple's announcement of a broader slate of new accessibility features coming later this year with iOS 27, visionOS 27, and other software updates. The Hikawa Grip & Stand is a hardware companion to that broader push, though it exists independently of any specific software feature.

The partnership with PopSockets is what enables the worldwide availability this time around. Previously, the accessory was sold only through Apple's own channels in limited markets. With PopSockets involved in the collaboration, the distribution footprint has expanded considerably, which should help with both reach and stock levels, though Apple has not confirmed exactly how much inventory is available. The first run disappeared fast, so if you've been waiting, "order it and then overthink it" is probably the more practical sequence here.

The Hikawa Grip & Stand comes in three colors: blue, orange, and stone. It's priced at $54.95 and available to order now through Apple's website. That puts it in the range of a deliberate purchase rather than an impulse add-on, but for users who genuinely benefit from the ergonomic design, the price is in line with similar MagSafe accessories that lack the same accessibility-focused engineering behind them.

What makes the Hikawa accessory a bit different from the broader ecosystem of phone grips and PopSocket-style attachments is the design process behind it. According to Apple, Hikawa worked directly with people with disabilities to understand how grip mechanics and phone weight interact with limited hand strength and reduced dexterity. The result isn't just a grip that happens to be accessible, it's one where accessibility was the starting constraint rather than an afterthought. That distinction matters, even if the final product looks like a premium silicone ring on the back of your phone.

For most iPhone users, the Hikawa Grip & Stand will work just like any other MagSafe accessory: it snaps on, stays put, and pulls off cleanly. For the users it was specifically designed around, it does more than that. Either way, Apple appears to have learned from the first launch that "make it and sell out" isn't a great strategy for an accessibility product, and the worldwide availability and PopSockets partnership suggest a more serious attempt to keep it in stock this time.

The last six months in LLMs in five minutes

Simon Willison used a five-minute lightning talk at PyCon US 2026 to compress six months of LLM history into something digestible, and his annotated slide deck tells the story well: November 2025 was a turning point, coding agents quietly crossed a quality threshold that actually matters, and the open-weight models running on laptop hardware have become surprisingly hard to dismiss.

The period Willison covers begins with what he calls the November 2025 inflection point. The title of "best model" changed hands five times that month alone, trading between Anthropic, OpenAI, and Google in a sequence that started with Claude Sonnet 4.5, jumped to GPT-5.1, then Gemini 3, then GPT-5.1 Codex Max, before Anthropic reclaimed it with Claude Opus 4.5. Willison tracks this with his signature benchmark: asking each model to generate an SVG of a pelican riding a bicycle. The logic is sound. Pelicans are hard to draw, bicycles are hard to draw, the combination is physically impossible, and no AI lab would ever specifically train for it. Gemini 3 drew the best pelican of the November batch, he notes, though Opus 4.5 held the broader practitioner consensus for the next couple of months.

The pelican derby is fun, but the more consequential development from November was quieter and took longer to register. OpenAI and Anthropic had spent most of 2025 applying Reinforcement Learning from Verifiable Rewards to their models, specifically to improve code quality when paired with their respective agent frameworks. By November, the results were visible. Coding agents moved, in Willison's framing, from "often-work" to "mostly-work." That sounds like modest progress, but crossing the threshold where you can trust an agent as a daily driver without spending the majority of your time cleaning up its errors is a genuine shift in what the tools are actually useful for.

Also in November, someone named Pete made an initial commit to an obscure repository called Warelay. Willison includes this as a data point worth flagging, because Warelay went through several name changes over December and January before arriving at its final identity: OpenClaw, a personal AI assistant that, by February, had become a phenomenon well beyond what a three-month-old project typically achieves. The "Claw" as a category term took hold alongside it, driven by spinoff projects like NanoClaw and ZeroClaw. Mac Minis started selling out around Silicon Valley as people bought them specifically to run their Claws locally. Drew Breunig, as Willison recounts, joked that they had become the new digital pets, the Mac Mini functioning as the perfect aquarium. Willison's own preferred metaphor is Alfred Molina's Doc Ock from Spider-Man 2: AI-powered appendages that work fine until something damages the inhibitor chip.

The December and January holiday period brought its own pattern. A lot of people, Willison included, used the break to push the new coding agents to their limits and got somewhat carried away. His own holiday project was a vibe-coded JavaScript interpreter written in Python, a loose port of MicroQuickJS he named micro-javascript. The demo runs JavaScript in Python, inside Pyodide, inside WebAssembly, inside JavaScript, inside a browser. Technically layered and genuinely impressive. Also, by his own admission, something nobody actually needed. He has since quietly retired that one and several others from the same period.

February brought Gemini 3.1 Pro, which produced what Willison considers the best pelican-on-a-bicycle result he has seen, complete with a fish in the basket. Google's Jeff Dean amplified the moment by posting an animated version that included a frog on a penny-farthing, a giraffe driving a tiny car, an ostrich on roller skates, a turtle kickflipping a skateboard, and a dachshund driving a stretch limousine. Whether the AI labs had been watching Willison's benchmark or simply got very good at drawing improbable animals on vehicles, the output quality was hard to argue with.

April brought more developments worth noting. Google released the Gemma 4 series, which Willison describes as the most capable open-weight models he has seen from a US company. Chinese lab GLM released GLM-5.1, an open-weight model at 754 billion parameters and 1.51 terabytes in size — effective, but requiring hardware most people cannot casually afford. GLM-5.1 produced a competent pelican, though it struggled to animate it cleanly. It also, when prompted by a Bluesky commenter, generated an illustrated and animated SVG of a North Virginia Opossum on an e-scooter, complete with the tagline "Cruising the commonwealth since dusk." Willison reports that other models cannot come close to that result.

The other notable April contribution came from Qwen. The Qwen3.6-35B-A3B model, a 20.9GB file that runs on a laptop, drew a better pelican-on-a-bicycle than Claude Opus 4.7. Willison acknowledges this probably says as much about the limits of the benchmark as it does about the model, but the underlying point stands: locally runnable open-weight models have started producing results that were implausible a year ago.

The two themes Willison pulls out of the six months are straightforward. Coding agents got genuinely good enough to use. And laptop-scale open-weight models, while still well behind the frontier, have started wildly outperforming expectations. Both of those developments have practical weight, and neither looks like a temporary blip.

Here’s why I won’t be switching on auto-deleting Siri chats

Apple's next version of Siri is shaping up to be, by most measurable standards, the most privacy-conscious AI assistant on the market. Bloomberg reported this week that the forthcoming Siri app will include an option to automatically delete conversation history, offering users the same retention choices already available in the Messages app: wipe chats after 30 days, after a year, or leave them in place indefinitely.

It is a genuinely welcome feature, and it fits neatly into the broader privacy framework Apple has been building around the new Siri. The assistant will be powered by Google's Gemini model, but Apple has been explicit that this does not mean your queries will be routed through Google's infrastructure. Apple says AI tasks will run on-device wherever the hardware allows, with overflow handled through its own Private Cloud Compute servers. That arrangement also comes with a contractual restriction: Apple's agreement with Google prohibits Siri from being used to train the Gemini model. Outgoing CEO Tim Cook described the partnership this way: "We believe that we can unlock a lot of experiences and innovate in a key way due to the collaboration. We'll continue to run on the device and run in Private Cloud Compute, and maintain our industry-leading privacy standards in doing so."

So the privacy story looks reasonably solid. Auto-deletion is a sensible addition to that story, giving users meaningful control over what sticks around after a session ends. For anyone who asks Siri something sensitive and would rather not have that conversation collecting dust in a server somewhere, the option to have it disappear automatically is a practical and appropriate tool.

That said, the author of the original piece makes a compelling case for why enabling auto-delete may actually work against you, and it is worth sitting with that argument. AI assistants get more useful the more context they carry. The value is not just in answering one-off questions well; it is in building a picture of who you are, what you prefer, and how you like information presented. Conversations are how that picture gets painted. Delete them and you are back to square one every time.

The point comes through clearly in how the author describes their own setup with Claude, which they use as their primary AI. Over time, through explicit instructions and implicit feedback, they have shaped how Claude responds: succinct answers, bullet points where useful, no flowery language, no sycophancy, links to sources when making factual claims. That kind of personalization does not happen in a single session. It accumulates. And Claude has apparently internalized context the author never stated directly, drawing on past queries and search history to tailor responses in ways that were noticeably more relevant.

None of that works if you are wiping the slate after every conversation. Auto-delete is, in effect, a trade: you gain privacy assurance and lose continuity. For many use cases, that is the right trade. A one-time question about a medical symptom or a sensitive personal situation is probably not the kind of thing you want feeding a persistent AI profile. But for daily use, where you are genuinely trying to build a productive working relationship with an assistant, nuking the history is counterproductive.

The practical middle ground is manual deletion, applied with some discretion. Delete the sessions that contain genuinely sensitive material. Leave the rest. Most AI apps, including the new Siri, support this approach, and it gives you targeted control without throwing out everything that makes a chatbot actually useful.

There is a separate but related piece of advice from the article that deserves attention: start a new conversation for each distinct topic. It sounds trivial but it matters. When you keep stacking questions into a single long thread, the model processes the entire conversation history every time you add a message, which burns through tokens faster. It also makes your history harder to navigate. Keeping topics separate means you can return to a previous thread cleanly, without wading through unrelated context. A well-named sidebar of distinct conversations is a much more useful archive than one sprawling mega-thread. Most AI assistants, including Claude according to the author, do a decent job of auto-naming chats in a way that makes retrieval practical.

The timing of this Bloomberg report is a reminder that the new Siri is still not here yet. Apple has been tight-lipped about exactly when it will land, and the privacy features described this week are based on reporting rather than a product that anyone can use today. But the picture being assembled is at least coherent. On-device processing, Private Cloud Compute for heavier tasks, no training on your data by Google, and now optional auto-deletion of chat history. Whether that combination delivers a Siri that is actually worth talking to remains to be seen, but Apple's privacy architecture around it looks, on paper, more thoughtfully constructed than most of what the competition has put forward.

Auto-delete is a good option to have. It is just not obviously the right default for every user, and treating it as a privacy must-have overlooks what you give up when context disappears.

SpaceX Starlink and other satellite megaconstellations are creating an ‘unregulated geoengineering experiment’, scientists say

The satellite industry has a bold vision for the future: hundreds of thousands, maybe millions, of spacecraft circling Earth, beaming internet to remote communities, running orbital data centers, harvesting solar power. It is an ambitious picture. It is also, according to atmospheric researchers, quietly running an uncontrolled experiment on the planet's climate.

A new study published May 13 in the journal Earth's Future finds that the rapid growth of satellite megaconstellations is pumping significant amounts of pollutants into the upper atmosphere, and that without action, this pollution will eventually be enough to alter Earth's climate. The researchers describe what is already happening as "a small-scale, unregulated geoengineering experiment that could have many unintended and serious environmental consequences" — the words of Eloise Marais, a professor of atmospheric chemistry and air quality at University College London who led the research team.

The problem is not just that rockets produce exhaust. It is where that exhaust ends up. Most megaconstellation launches today depend on SpaceX's Falcon 9 rocket, which burns kerosine fuel and releases black carbon — soot — into the upper layers of the atmosphere. Unlike pollution from cars, ships, or power plants, which gets scrubbed out of the lower atmosphere relatively quickly, black carbon released at high altitude lingers for two and a half to three years. Marais notes that this makes it roughly 540 times more potent as a climate forcer than the same soot released at ground level. The upper atmosphere is, in this respect, a place where pollution goes to accumulate.

Satellite re-entries add a separate concern. When a satellite burns up on its way back into the atmosphere, it releases aluminum oxides, which have the potential to damage the ozone layer. Launches heat; re-entries corrode. Megaconstellations do both at scale and on an accelerating schedule, because the satellites themselves are designed with short operational lifespans — typically around five years — and get swapped out regularly for newer models. That churn means more rockets going up and more hardware coming down, year after year.

The numbers behind all of this are striking. Since 2020, when the megaconstellation era began in earnest, concentrations of high-altitude air pollution from the space sector have risen significantly, according to the research. On conservative estimates — and the team emphasizes conservative, since actual satellite deployment has consistently outpaced their projections — the global space industry will have released more climate-altering chemicals into the atmosphere by 2030 than the United Kingdom produces in a year. By 2029, pollution specifically from megaconstellation launches is projected to account for more than 40 percent of all space-sector air pollution.

The megaconstellation players driving this growth include SpaceX's Starlink, Amazon's LEO network, and Chinese operators Guowang and Qianfan. Starlink alone now counts more than 10,000 active satellites. The European Space Agency puts the total number of operational satellites currently in orbit above 15,000 — three times the figure from 2020. By 2030, that number could reach 100,000, with continued steep growth expected afterward.

To be clear, the researchers are not saying the sky is falling today. By 2029, the accumulated pollutants will represent roughly one hundredth of the quantity that would be needed to produce meaningful geoengineering effects. But the trajectory matters. Geoengineering — in particular, the concept of Stratospheric Aerosol Injection, which involves deliberately seeding the stratosphere with reflective particles to reduce incoming heat — is already understood to carry serious unpredictable risks: disrupted rain patterns, droughts, unforeseen weather changes. The concern Marais raises is that the space industry is inadvertently drifting toward similar territory, without the careful study, the international deliberation, or the regulatory guardrails that any intentional intervention would require.

The modeling her team uses is able to track both climate effects and ozone depletion from projected pollution loads, giving a reasonably clear picture of where the trend leads. What it cannot do is keep pace with the speed of deployment. The satellite industry is simply moving faster than the science can follow, which is itself part of the problem Marais is trying to highlight.

"We need to be taking it far more seriously in terms of regulating the pollution that's coming from launches and reentries," she told Space.com. "There also needs to be far more funding funneled into research to study this because we can't keep up with the space industry."

That gap between industrial momentum and scientific oversight is the crux of the issue. The companies building these constellations are working within existing launch regulations, which were not designed with high-volume, high-frequency megaconstellation operations in mind. Nobody sat down and decided to run an atmospheric experiment. It is an emergent consequence of thousands of individual launch decisions, each one unremarkable on its own, adding up to something researchers now feel compelled to name and study before it becomes much harder to reverse.