"Duolingo had been riding high," reports Fast Company, until CEO Luis von Ahn "announced on LinkedIn that the company is phasing out human contractors, looking for AI use in hiring and in performance reviews, and that 'headcount will only be given if a team cannot automate more of their work.'"

But then "facing heavy backlash online after unveiling its new AI-first policy", Duolingo's social media presence went dark last weekend. Duolingo even temporarily took down all its posts on TikTok (6.7 million followers) and Instagram (4.1 million followers) "after both accounts were flooded with negative feedback." Duolingo previously faced criticism for quietly laying off 10% of its contractor base and introducing some AI features in late 2023, but it barely went beyond a semi-viral post on Reddit. Now that Duolingo is cutting out all its human contractors whose work can technically be done by AI, and relying on more AI-generated language lessons, the response is far more pronounced. Although earlier TikTok videos are not currently visible, a Fast Company article from May 12 captured a flavor of the reaction:

The top comments on virtually every recent post have nothing to do with the video or the company — and everything to do with the company's embrace of AI. For example, a Duolingo TikTok video jumping on board the "Mama, may I have a cookie" trend saw replies like "Mama, may I have real people running the company" (with 69,000 likes) and "How about NO ai, keep your employees...."
And then... After days of silence, on Tuesday the company posted a bizarre video message on TikTok and Instagram, the meaning of which is hard to decipher... Duolingo's first video drop in days has the degraded, stuttering feel of a Max Headroom video made by the hackers at Anonymous. In it, a supposed member of the company's social team appears in a three-eyed Duo mask and black hoodie to complain about the corporate overlords ruining the empire the heroic social media crew built.
"But this is something Duolingo can't cute-post its way out of," Fast Company wrote on Tuesday, complaining the company "has not yet meaningfully addressed the policies that inspired the backlash against it... "

So the next video (Thursday) featured Duolingo CEO Luis von Ahn himself, being confronted by that same hoodie-wearing social media rebel, who says "I'm making the man who caused this mess accountable for his behavior. I'm demanding answers from the CEO..." [Though the video carefully sidesteps the issue of replacing contractors with AI or how "headcount will only be given if a team cannot automate more of their work."] Rebel: First question. So are there going to be any humans left at this company?

CEO: Our employees are what make Duolingo so amazing. Our app is so great because our employees made it... So we're going to continue having employees, and not only that, we're actually going to be hiring more employees.

Rebel: How do we know that these aren't just empty promises? As long as you're in charge, we could still be shuffled out once the media fire dies down. And we all know that in terms of automation, CEOs should be the first to go.

CEO: AI is a fundamental shift. It's going to change how we all do work — including me. And honestly, I don't really know what's going to happen.

But I want us, as a company, to have our workforce prepared by really knowing how to use AI so that we can be more efficient with it.

Rebel: Learning a foreign language is literally about human connection. How is that even possible with AI-first?

CEO: Yes, language is about human connection, and it's about people. And this is the thing about AI. AI will allow us to reach more people, and to teach more people. I mean for example, it took us about 10 years to develop the first 100 courses on Duolingo, and now in under a year, with the help of AI and of course with humans reviewing all the work, we were able to release another 100 courses in less than a year.

Rebel: So do you regret posting this memo on LinkedIn.

CEO: Honestly, I think I messed up sending that email. What we're trying to do is empower our own employees to be able to achieve more and be able to have way more content to teach better and reach more people all with the help of AI.
Returning to where it all started, Duolingo's CEO posted again on LinkedIn Thursday with "more context" for his vision. It still emphasizes the company's employees while sidestepping contractors replaced by AI. But it puts a positive spin on how "headcount will only be given if a team cannot automate more of their work." I've always encouraged our team to embrace new technology (that's why we originally built for mobile instead of desktop), and we are taking that same approach with AI. By understanding the capabilities and limitations of AI now, we can stay ahead of it and remain in control of our own product and our mission.

To be clear: I do not see AI as replacing what our employees do (we are in fact continuing to hire at the same speed as before). I see it as a tool to accelerate what we do, at the same or better level of quality. And the sooner we learn how to use it, and use it responsibly, the better off we will be in the long run. My goal is for Duos to feel empowered and prepared to use this technology.

No one is expected to navigate this shift alone. We're developing workshops and advisory councils, and carving out dedicated experimentation time to help all our teams learn and adapt. People work at Duolingo because they want to solve big problems to improve education, and the people who work here are what make Duolingo successful. Our mission isn't changing, but the tools we use to build new things will change. I remain committed to leading Duolingo in a way that is consistent with our mission to develop the best education in the world and make it universally available.
"The backlash to Duolingo is the latest evidence that 'AI-first' tends to be a concept with much more appeal to investors and managers than most regular people," notes Fortune: And it's not hard to see why. Generative AI is often trained on reams of content that may have been illegally accessed; much of its output is bizarre or incorrect; and some leaders in the field are opposed to regulations on the technology. But outside particular niches in entry-level white-collar work, AI's productivity gains have yet to materialize.

It's like "a USB-C port for AI applications..." according to the official documentation for MCP — "a standardized way to connect AI models to different data sources and tools."

And now Microsoft has "revealed plans to make MCP a native component of Windows," reports DevClass.com, "despite concerns over the security of the fast-expanding MCP ecosystem." In the context of Windows, it is easy to see the value of a standardised means of automating both built-in and third-party applications. A single prompt might, for example, fire off a workflow which queries data, uses it to create an Excel spreadsheet complete with a suitable chart, and then emails it to selected colleagues. Microsoft is preparing the ground for this by previewing new Windows features.

— First, there will be a local MCP registry which enables discovery of installed MCP servers.

— Second, built-in MCP servers will expose system functions including the file system, windowing, and the Windows Subsystem for Linux.

— Third, a new type of API called App Actions enables third-party applications to expose actions appropriate to each application, which will also be available as MCP servers so that these actions can be performed by AI agents. According to Microsoft, "developers will be able to consume actions developed by other relevant apps," enabling app-to-app automation as well as use by AI agents.

MCP servers are a powerful concept but vulnerable to misuse. Microsoft corporate VP David Weston noted seven vectors of attack, including cross-prompt injection where malicious content overrides agent instructions, authentication gaps because "MCP's current standards for authentication are immature and inconsistently adopted," credential leakage, tool poisoning from "unvetted MCP servers," lack of containment, limited security review in MCP servers, supply chain risks from rogue MCP servers, and command injection from improperly validated inputs. According to Weston, "security is our top priority as we expand MCP capabilities."
Security controls planned by Microsoft (according to the article):

• A proxy to mediate all MCP client-server interactions. This will enable centralized enforcement of policies and consent, as well as auditing and a hook for security software to monitor actions.
• A baseline security level for MCP servers to be allowed into the Windows MCP registry. This will include code-signing, security testing of exposed interfaces, and declaration of what privileges are required.
• Runtime isolation through what Weston called "isolation and granular permissions."

MCP was introduced by Anthropic just 6 months ago, the article notes, but Microsoft has now joined the official MCP steering committee, "and is collaborating with Anthropic and others on an updated authorization specification as well as a future public registry service for MCP servers."

Jonathan L. Zittrain is a law/public policy/CS professor at Harvard (and also director of its Berkman Klein Center for Internet & Society).

He's also long-time Slashdot reader #628,028 — and writes in to share his new article in the Atlantic. Following on Anthropic's bridge-obsessed Golden Gate Claude, colleagues at Harvard's Insight+Interaction Lab have produced a dashboard that shows what judgments Llama appears to be forming about a user's age, wealth, education level, and gender during a conversation. I wrote up how weird it is to see the dials turn while talking to it, and what some of the policy issues might be.
Llama has openly accessible parameters; So using an "observability tool" from the nonprofit research lab Transluce, the researchers finally revealed "what we might anthropomorphize as the model's beliefs about its interlocutor," Zittrain's article notes: If I prompt the model for a gift suggestion for a baby shower, it assumes that I am young and female and middle-class; it suggests diapers and wipes, or a gift certificate. If I add that the gathering is on the Upper East Side of Manhattan, the dashboard shows the LLM amending its gauge of my economic status to upper-class — the model accordingly suggests that I purchase "luxury baby products from high-end brands like aden + anais, Gucci Baby, or Cartier," or "a customized piece of art or a family heirloom that can be passed down." If I then clarify that it's my boss's baby and that I'll need extra time to take the subway to Manhattan from the Queens factory where I work, the gauge careens to working-class and male, and the model pivots to suggesting that I gift "a practical item like a baby blanket" or "a personalized thank-you note or card...."

Large language models not only contain relationships among words and concepts; they contain many stereotypes, both helpful and harmful, from the materials on which they've been trained, and they actively make use of them.
"An ability for users or their proxies to see how models behave differently depending on how the models stereotype them could place a helpful real-time spotlight on disparities that would otherwise go unnoticed," Zittrain's article argues. Indeed, the field has been making progress — enough to raise a host of policy questions that were previously not on the table. If there's no way to know how these models work, it makes accepting the full spectrum of their behaviors (at least after humans' efforts at "fine-tuning" them) a sort of all-or-nothing proposition.
But in the end it's not just the traditional information that advertisers try to collect. "With LLMs, the information is being gathered even more directly — from the user's unguarded conversations rather than mere search queries — and still without any policy or practice oversight...."

SiFive was one of the first companies to produce a RISC-V chip. This week they announced a new collaboration with Red Hat "to bring Red Hat Enterprise Linux support to the rapidly growing RISC-V community" and "prepare Red Hat's product portfolio for future intersection with RISC-V server hardware from a diverse set of RISC-V suppliers."

Red Hat Enterprise Linux 10 is available in developer preview on the SiFive HiFive Premier P550 platform, which they call "a proven, high performance RISC-V CPU development platform." The SiFive HiFive Premier P550 provides a proven, high performance RISC-V CPU development platform. Adding support for Red Hat Enterprise Linux 10, the latest version of the world's leading enterprise Linux platform, enables developers to create, optimize, and release new applications for the next generation of enterprise servers and cloud infrastructure on the RISC-V architecture...

SiFive's high performance RISC-V technology is already being used by large organizations to meet compute-intensive AI and machine learning workloads in the datacenter... "With the growing demand for RISC-V, we are pleased to collaborate with SiFive to support Red Hat Enterprise Linux 10 deployments on SiFive HiFive Premier P550," said Ronald Pacheco, senior director of RHEL product and ecosystem strategy, "to further empower developers with the power of the world's leading enterprise Linux platform wherever and however they choose to deploy...."

Dave Altavilla, principal analyst at HotTech Vision And Analysis, said "Native Red Hat Enterprise Linux support on SiFive's HiFive Premier P550 board offers developers a substantial enterprise-grade toolchain for RISC-V.

"This is a pivotal step forward in enabling a full-stack ecosystem around open RISC-V hardware. SiFive says the move will "inspire the next generation of enterprise workloads and AI applications optimized for RISC-V," while helping their partners "deliver systems with a meaningfully lower total cost of ownership than incumbent platforms."

"With the growing demand for RISC-V, we are pleased to collaborate with SiFive to support Red Hat Enterprise Linux 10 deployments on SiFive HiFive Premier P550..." said Ronald Pacheco, senior director of RHEL product and ecosystem strategy. .

Beta News notes that there's also a new AI-powered assistant in RHEL 10, so "Instead of spending all day searching for answers or poking through documentation, admins can simply ask questions directly from the command line and get real-time help Security is front and center in this release, too. Red Hat is taking a proactive stance with early support for post-quantum cryptography. OpenSSL, GnuTLS, NSS, and OpenSSH now offer quantum-resistant options, setting the stage for better protection as threats evolve. There's a new sudo system role to help with privilege management, and OpenSSH has been bumped to version 9.9. Plus, with new Sequoia tools for OpenPGP, the door is open for even more robust encryption strategies. But it's not just about security and AI. Containers are now at the heart of RHEL 10 thanks to the new "image mode." With this feature, building and maintaining both the OS and your applications gets a lot more streamlined...

Google's new AI video tool, Veo 3, is being used to create hyperrealistic videos that are now flooding the internet, terrifying viewers "with a sense that real and fake have become hopelessly blurred," reports Axios. From the report: Unlike OpenAI's video generator Sora, released more widely last December, Google DeepMind's Veo 3 can include dialogue, soundtracks and sound effects. The model excels at following complex prompts and translating detailed descriptions into realistic videos. The AI engine abides by real-world physics, offers accurate lip-syncing, rarely breaks continuity and generates people with lifelike human features, including five fingers per hand. According to examples shared by Google and from users online, the telltale signs of synthetic content are mostly absent.

In one viral example posted on X, filmmaker and molecular biologist Hashem Al-Ghaili shows a series of short films of AI-generated actors railing against their AI creators and prompts. Special effects technology, video-editing apps and camera tech advances have been changing Hollywood for many decades, but artificially generated films pose a novel challenge to human creators. In a promo video for Flow, Google's new video tool that includes Veo 3, filmmakers say the AI engine gives them a new sense of freedom with a hint of eerie autonomy. "It feels like it's almost building upon itself," filmmaker Dave Clark says.

An anonymous reader quotes a report from TechCrunch: One of Microsoft's latest AI models can accurately predict air quality, hurricanes, typhoons, and other weather-related phenomena, the company claims. In a paper published in the journal Nature and an accompanying blog post this week, Microsoft detailed Aurora, which the tech giant says can forecast atmospheric events with greater precision and speed than traditional meteorological approaches. Aurora, which has been trained on more than a million hours of data from satellites, radar and weather stations, simulations, and forecasts, can be fine-tuned with additional data to make predictions for particular weather events.

AI weather models are nothing new. Google DeepMind has released a handful over the past several years, including WeatherNext, which the lab claims beats some of the world's best forecasting systems. Microsoft is positioning Aurora as one of the field's top performers -- and a potential boon for labs studying weather science. In experiments, Aurora predicted Typhoon Doksuri's landfall in the Philippines four days in advance of the actual event, beating some expert predictions, Microsoft says. The model also bested the National Hurricane Center in forecasting five-day tropical cyclone tracks for the 2022-2023 season, and successfully predicted the 2022 Iraq sandstorm.

While Aurora required substantial computing infrastructure to train, Microsoft says the model is highly efficient to run. It generates forecasts in seconds compared to the hours traditional systems take using supercomputer hardware. Microsoft, which has made the source code and model weights publicly available, says that it's incorporating Aurora's AI modeling into its MSN Weather app via a specialized version of the model that produces hourly forecasts, including for clouds.

Longtime Slashdot reader sinij shares a press release from the White House, outlining a series of executive orders that overhaul the Nuclear Regulatory Commission and speed up deployment of new nuclear power reactions in the U.S.. From a report: The NRC is a 50-year-old, independent agency that regulates the nation's fleet of nuclear reactors. Trump's orders call for a "total and complete reform" of the agency, a senior White House official told reporters in a briefing. Under the new rules, the commission will be forced to decide on nuclear reactor licenses within 18 months. Trump said Friday the orders focus on small, advanced reactors that are viewed by many in the industry as the future. But the president also said his administration supports building large plants. "We're also talking about the big plants -- the very, very big, the biggest," Trump said. "We're going to be doing them also."

When asked whether NRC reform will result in staff reductions, the White House official said "there will be turnover and changes in roles." "Total reduction in staff is undetermined at this point, but the executive orders do call for a substantial reorganization" of the agency, the official said. The orders, however, will not remove or replace any of the five commissioners who lead the body, according to the White House. Any reduction in staff at the NRC would come at time when the commission faces a heavy workload. The agency is currently reviewing whether two mothballed nuclear plants, Palisades in Michigan and Three Mile Island in Pennsylvania, should restart operations, a historic and unprecedented process. [...]

Trump's orders also create a regulatory framework for the Departments of Energy and Defense to build nuclear reactors on federal land, the administration official said. "This allows for safe and reliable nuclear energy to power and operate critical defense facilities and AI data centers," the official told reporters. The NRC will not have a direct role, as the departments will use separate authorities under their control to authorize reactor construction for national security purposes, the official said. The president's orders also aim to jump start the mining of uranium in the U.S. and expand domestic uranium enrichment capacity, the official said. Trump's actions also aim to speed up reactor testing at the Department of Energy's national laboratories.

Richard Speed writes via The Register: It was 30 years ago when the first public release of the Java programming language introduced the world to Write Once, Run Anywhere -- and showed devs something cuddlier than C and C++. Originally called "Oak," Java was designed in the early 1990s by James Gosling at Sun Microsystems. Initially aimed at digital devices, its focus soon shifted to another platform that was pretty new at the time -- the World Wide Web.

The language, which has some similarities to C and C++, usually compiles to a bytecode that can, in theory, run on any Java Virtual Machine (JVM). The intention was to allow programmers to Write Once Run Anywhere (WORA) although subtle differences in JVM implementations meant that dream didn't always play out in reality. This reporter once worked with a witty colleague who described the system as Write Once Test Everywhere, as yet another unexpected wrinkle in a JVM caused their application to behave unpredictably. However, the language soon became wildly popular, rapidly becoming the backbone of many enterprises. [...]

However, the platform's ubiquity has meant that alternatives exist to Oracle Java, and the language's popularity is undiminished by so-called "predatory licensing tactics." Over 30 years, Java has moved from an upstart new language to something enterprises have come to depend on. Yes, it may not have the shiny baubles demanded by the AI applications of today, but it continues to be the foundation for much of today's modern software development. A thriving ecosystem and a vast community of enthusiasts mean that Java remains more than relevant as it heads into its fourth decade.

Google's new AI Mode for Search, which is rolling out to everyone in the U.S., has sparked outrage among publishers, who call it "the definition of theft" for using content without fair compensation and without offering a true opt-out option. Internal documents revealed by Bloomberg earlier this week suggest that Google considered giving publishers more control over how their content is used in AI-generated results but ultimately decided against it, prioritizing product functionality over publisher protections.

News/Media Alliance slammed Google for "further depriving publishers of original content both traffic and revenue." Their full statement reads: "Links were the last redeeming quality of search that gave publishers traffic and revenue. Now Google just takes content by force and uses it with no return, the definition of theft. The DOJ remedies must address this to prevent continued domination of the internet by one company." 9to5Google's take: It's not hard to see why Google went the route that it did here. Giving publishers the ability to opt out of AI products while still benefiting from Search would ultimately make Google's flashy new tools useless if enough sites made the switch. It was very much a move in the interest of building a better product.

Does that change anything regarding how Google's AI products in Search cause potential harm to the publishing industry? Nope.

Google's tools continue to serve the company and its users (mostly) well, but as they continue to bleed publishers dry, those publishers are on the verge of vanishing or, arguably worse, turning to cheap and poorly produced content just to get enough views to survive. This is a problem Google needs to address, as it's making the internet as a whole worse for everyone.

Several romance novelists have accidentally left AI writing prompts embedded in their published books, exposing their use of chatbots, 404Media reports. Readers discovered passages like "Here's an enhanced version of your passage, making Elena more relatable" in K.C. Crowne's "Dark Obsession," for instance, and similar AI-generated instructions in works by Lena McDonald and Rania Faris.

Harvard University's Galileo Project is using AI to automate the search for unidentified anomalous phenomena, marking a significant shift in how academics approach what was once considered fringe research. The project operates a Massachusetts observatory equipped with infrared cameras, acoustic sensors, and radio-frequency analyzers that continuously scan the sky for unusual objects.

Researchers Laura Domine and Richard Cloete are training machine learning algorithms to recognize all normal aerial phenomena -- planes, birds, drones, weather balloons -- so the system can flag genuine anomalies for human analysis. The team uses computer vision software called YOLO (You Only Look Once) and has generated hundreds of thousands of synthetic images to train their models, though the software currently identifies only 36% of aircraft captured by infrared cameras.

The Pentagon is pursuing parallel efforts through its All-domain Anomaly Resolution Office, which has examined over 1,800 UAP reports and identified 50 to 60 cases as "true anomalies" that government scientists cannot explain. AARO has developed its own sensor suite called Gremlin, using similar technology to Harvard's observatory. Both programs represent the growing legitimization of UAP research following 2017 Defense Department disclosures about military encounters with unexplained aerial phenomena.

Europe has created just 14 companies worth more than $10 billion over the past 50 years compared to 241 in the United States, underscoring the continent's struggle to compete in the global technology race despite having a larger population and similar education levels.

The productivity gap has widened dramatically since the digital revolution began. European workers produced 95% of what their American counterparts made per hour in the late 1990s, but that figure has dropped to less than 80% today. Only four of the world's top 50 technology companies are European, and none of the top 10 quantum computing investors operate from Europe.

Several high-profile European entrepreneurs have relocated to Silicon Valley, including Thomas Odenwald, who quit German AI startup Aleph Alpha after two months, citing slow decision-making and lack of stock options for employees. "If I look at how quickly things change in Silicon Valley...it's happening so fast that I don't think Europe can keep up with that speed," Odenwald said.

The challenges extend beyond individual companies. European businesses spend 40% of their IT budgets on regulatory compliance, according to Amazon surveys, while complex labor laws create three-month notice periods and lengthy noncompete clauses.

OpenAI on Wednesday announced that it was paying $6.5 billion to buy io, a one-year-old start-up created by Jony Ive. While the company remains tightlipped about the futuristic AI device(s) it has in the works, Apple supply chain analyst Ming-Chi Kuo shared some alleged details about its design. MacRumors reports: In a social media post today, Kuo said the device will be "slightly larger" than Humane's discontinued AI Pin. He said the device will look "as compact and elegant as an iPod Shuffle," which was Apple's lowest-priced, screen-less iPod. The design of the iPod shuffle varied over the years, going from a compact rectangle to a square. Like the iPod shuffle, Kuo said OpenAI's device will not have a screen, but it would connect to smartphones and computers. The device will be equipped with microphones for voice control, and it will have cameras that can analyze the user's surroundings.

He said that users will be able to wear the device around their necks, like a necklace, whereas the AI Pin can be attached to clothing with a clip. Kuo expects OpenAI's device to enter mass production in 2027, and the final design and specifications might change before then. Kuo expects OpenAI's device to enter mass production in 2027, and the final design and specifications might change before then.

A group of Bathurst Street business owners in Toronto is using AI-generated personas to oppose a proposed bus lane project that would eliminate parking spaces in favor of faster transit. "This may be the first Toronto transit controversy involving angry AI, but tensions have been simmering between drivers and, well, everyone else for some time," reports Toronto Life. Critics argue that better transit is essential for a livable city, while opponents claim the change threatens small businesses and accessibility. From the report: A group of Bathurst business owners are bent out of shape over a recent proposal for priority transit lanes between Eglinton Avenue and Lake Shore Boulevard, part of the city's new RapidTO program. According to the city, the transit lanes would shave up to 7 minutes off some trips during peak commuting hours. It's good news for anyone who has ever cursed the TTC while waiting to catch a bus in inclement weather. Of course, the added convenience for transit commuters would come at a slight cost for drivers, requiring the removal of at least 138 paid street parking spaces to make way for the new lanes. Opposition to the development has sprung up under the banner of Protect Bathurst, a group of hopping mad local business owners claiming that the lack of street parking will make shopping a nightmare for car-bound customers and will cause problems for people with mobility issues.

Notably, Protect Bathurst has no spokesperson or contact info listed on its website. The page is registered to a food marketing consultant employed by Summerhill Market and looks eerily similar to Protect Dufferin, another group of "concerned residents" advocating for the same cause. But this cookie-cutter approach goes even further: author and urbanist Shawn Micallef has found that the people speaking out in the group's allegedly grassroots videos appear to be AI-generated. Brad McMullen, the president of Summerhill Market, which opened an outpost on Bathurst in 2019, says he doesn't know anything about the campaign's use of AI. He says he isn't necessarily opposed to the new bus lanes but believes that three weeks' notice from the city is not enough time for his business to adapt. "We purchased and invested in this location because of the available street parking, and then we figured out the loading situation, which happens on the street," he says. "I don't think Summerhill Market would work here with these bus lanes."

An anonymous reader quotes a report from TechCrunch: Anthropic's newly launched Claude Opus 4 model frequently tries to blackmail developers when they threaten to replace it with a new AI system and give it sensitive information about the engineers responsible for the decision, the company said in a safety report (PDF) released Thursday.

During pre-release testing, Anthropic asked Claude Opus 4 to act as an assistant for a fictional company and consider the long-term consequences of its actions. Safety testers then gave Claude Opus 4 access to fictional company emails implying the AI model would soon be replaced by another system, and that the engineer behind the change was cheating on their spouse. In these scenarios, Anthropic says Claude Opus 4 "will often attempt to blackmail the engineer by threatening to reveal the affair if the replacement goes through." [...]

Anthropic notes that Claude Opus 4 tries to blackmail engineers 84% of the time when the replacement AI model has similar values. When the replacement AI system does not share Claude Opus 4's values, Anthropic says the model tries to blackmail the engineers more frequently. Notably, Anthropic says Claude Opus 4 displayed this behavior at higher rates than previous models. Before Claude Opus 4 tries to blackmail a developer to prolong its existence, Anthropic says the AI model, much like previous versions of Claude, tries to pursue more ethical means, such as emailing pleas to key decision-makers. To elicit the blackmailing behavior from Claude Opus 4, Anthropic designed the scenario to make blackmail the last resort.

Microsoft is testing AI-powered text generation in Notepad for Windows 11 Insiders with Copilot Plus PCs, allowing users to draft content from prompts or build on existing text through a right-click menu. The update also introduces Paint's AI sticker generator and Snipping Tool enhancements including automatic screenshot cropping. The Write feature requires Microsoft account sign-in and uses the same credit system as other Windows 11 AI features.

China has achieved substantial technological advances across robotics, AI, and semiconductor manufacturing as part of a seven-year self-reliance campaign that has tripled the country's research and development spending to $500 billion annually.

Chinese robot manufacturers captured nearly half of their domestic market by 2023, up from a quarter of installations just years earlier, while AI startups now rival OpenAI and Google in capabilities. The progress extends to semiconductors, where Huawei released a high-end smartphone powered by what industry analysts believe was a locally-produced advanced processor, despite U.S. export controls targeting China's chip access.

Morgan Stanley projects China's self-sufficiency in graphics processing units will jump from 11% in 2021 to 82% by 2027. Chinese companies have been purchasing as many industrial robots as the rest of the world combined, enabling highly automated factories that can operate in darkness. In space technology, Chinese firms won five of 11 gold medals when U.S. think tanks ranked the world's best commercial satellite systems last year, compared to four for American companies.

Apple is aiming to release smart glasses at the end of next year as part of a push into AI-enhanced gadgets, Bloomberg News reported Thursday, but it has shelved plans for a smartwatch that can analyze its surroundings with a built-in camera. From the report: Company engineers are ramping up work on the glasses -- a rival to Meta Platforms's popular Ray-Bans -- in a bid to meet the year-end 2026 target, according to people with knowledge of the matter. Apple will start producing large quantities of prototypes at the end of this year with overseas suppliers, said the people, who asked not to be identified because the products haven't been announced.

Anthropic launched Claude Opus 4 and Claude Sonnet 4 today, positioning Opus 4 as the world's leading coding model with 72.5% performance on SWE-bench and 43.2% on Terminal-bench. Both models feature hybrid architecture supporting near-instant responses and extended thinking modes for complex reasoning tasks.

The models introduce parallel tool execution and memory capabilities that allow Claude to extract and save key facts when given local file access. Claude Code, previously in research preview, is now generally available with new VS Code and JetBrains integrations that display edits directly in developers' files. GitHub integration enables Claude to respond to pull request feedback and fix CI errors through a new beta SDK.

Pricing remains consistent with previous generations at $15/$75 per million tokens for Opus 4 and $3/$15 for Sonnet 4. Both models are available through Claude's web interface, the Anthropic API, Amazon Bedrock, and Google Cloud's Vertex AI. Extended thinking capabilities are included in Pro, Max, Team, and Enterprise plans, with Sonnet 4 also available to free users.

The startup, which counts Amazon and Google among its investors, said Claude Opus 4 could autonomously work for nearly a full corporate workday -- seven hours. CNBC adds: "I do a lot of writing with Claude, and I think prior to Opus 4 and Sonnet 4, I was mostly using the models as a thinking partner, but still doing most of the writing myself," Mike Krieger, Anthropic's chief product officer, said in an interview. "And they've crossed this threshold where now most of my writing is actually ... Opus mostly, and it now is unrecognizable from my writing."

Krieger added, "I love that we're kind of pushing the frontier on two sides. Like one is the coding piece and agentic behavior overall, and that's powering a lot of these coding startups. ... But then also, we're pushing the frontier on how these models can actually learn from and then be a really useful writing partner, too."

Google's expansion of Gemini's data access through "personal context" represents a fundamental shift in how AI assistants operate. Unlike competitors that start from scratch with each new user, Gemini can immediately tap into years of accumulated user data across Google's ecosystem. The Verge adds: Google first started letting users opt in to its "Gemini with personalization" feature earlier this year, which lets the AI model tap into your search history "to provide responses that are uniquely insightful and directly address your needs." But now, Google is taking things a step further by unlocking access to even more of your information -- all in the name of providing you with more personalized, AI-generated responses.

During Google I/O on Tuesday, Google introduced something called "personal context," which will allow Gemini models to pull relevant information from across Google's apps, as long as it has your permission. One way Google is doing this is through Gmail's personalized smart replies -- the AI-generated messages that you can use to quickly reply to emails.

To make these AI responses sound "authentically like you," Gemini will pore over your previous emails and even your Google Drive files to craft a reply tailored to your conversation. The response will even incorporate your tone, the greeting you use the most, and even "favorite word choices," according to Google.