Anthropic researchers have developed a breakthrough "cross-layer transcoder" (CLT) that functions like an fMRI for large language models, [1]mapping how they process information internally . Testing on Claude 3.5 Haiku, researchers discovered the model performs longer-range planning for specific tasks -- such as selecting rhyming words before constructing poem sentences -- and processes multilingual concepts in a shared neural space before converting outputs to specific languages.

The team also confirmed that LLMs can fabricate reasoning chains, either to please users with incorrect hints or to justify answers they derived instantly. The CLT identifies interpretable feature sets rather than individual neurons, allowing researchers to trace entire reasoning processes through network layers.



[1] https://fortune.com/2025/03/27/anthropic-ai-breakthrough-claude-llm-black-box/

An anonymous reader [1]shares a report :

> For decades, India's economic promise has rested on its demographic dividend -- the competitive edge of a massive, young, and increasingly educated workforce. Economists and policymakers have routinely cited the country's population profile as its ticket to economic superpower status, with projections of reaching $10 trillion in GDP and achieving high-income status by 2047. These forecasts depend heavily on a critical assumption: that roughly 500 million Indians currently aged 5-24 will find productive employment as they enter the workforce over the next two decades. But a sobering new analysis from Bernstein suggests this fundamental premise [2]may be crumbling under the weight of rapid advances in AI .

>

> "The advent of AI threatens to erode all the advantages of India's rich demographic dividend," write Bernstein analysts Venugopal Garre and Nikhil Arela, who characterize their assessment as a potential "doomsday scenario" for a nation that has hitched its economic wagon to services-led growth. At stake is India's $350 billion services export sector -- a sprawling ecosystem of IT outsourcing, business process management, and offshore knowledge centers that employs over 10 million workers, mostly in jobs that place them in the top 25% of the country's income distribution.

>

> While India's IT giants have successfully navigated previous technological shifts -- from basic call centers in the late 1980s to cloud computing and data analytics more recently -- AI poses a fundamentally different challenge. Unlike earlier transitions that required human adaptation, today's AI systems threaten to replace rather than complement the workforce. "AI subscriptions that come at a fraction of the costs of India's entry level engineers can be deployed to perform tasks at higher precision and speed," the report note.



[1] https://indiadispatch.com/p/the-500-million-worker-problem

[2] https://indiadispatch.com/p/the-500-million-worker-problem

An anonymous reader quotes a report from the Washington Post:

> Virginia is set to become the first state in the country to require some reckless drivers to put devices on their cars that [1]make it impossible to drive too fast . D.C. passed similar legislation last year. Several other states, including Maryland, are considering joining them. It's an embrace of a technological solution to a human problem: Speeding contributes to more than 10,000 deaths a year. Under the Virginia legislation, a judge can decide to order drivers to install the speed limiters in their vehicles in lieu of taking away their driving privileges or sending them to jail. It takes effect in July 2026.

>

> Del. Patrick A. Hope (D-Arlington) said various advocacy groups, including Mothers Against Drunk Driving and the National Safety Council, gave him the idea. He drove a car outfitted with the technology and was impressed. "It was easy to use, and once you're engaged it's impossible to go over the speed limit," he said. "It will make our streets safer." He thinks the device is preferable to suspending drivers' licenses, a punishment that people frequently ignore because they have no other way of getting to work or the store or taking their children to school. It's an approach similar to using an interlock device that requires a person convicted of drunken driving to pass a Breathalyzer test to start their car.

>

> Hope wanted anyone convicted of reckless driving after going 100 mph or more to be required to use a limiter for two to six months, but Gov. Glenn Youngkin (R) struck that part of the bill, leaving all use of the limiting technology up to the state courts. Hope expressed concern about the governor's amendment but will urge the General Assembly to accept it, as the legislature typically does when the bill's sponsor signals support. Drivers must pay for the speed limiters themselves. (As in D.C., indigent defendants are exempt from paying.) The limiters won't be used in Virginia on commercial vehicles. Attempting to evade the speed limiter by tampering with it or driving a different car is a misdemeanor punishable by up to a year in jail.



[1] https://www.washingtonpost.com/dc-md-va/2025/03/27/virginia-speed-limit-device/

Following OpenAI's [1]viral Studio Ghibli moment , CEO Sam Altman says it has [2]temporarily limited image generation in ChatGPT due to the overwhelming demand on its GPU infrastructure. "It's super fun seeing people love images in ChatGPT, but our GPUs are melting," he [3]posted on X today. The Verge reports:

> The demand crunch already caused the artificial intelligence company to push back availability of the built-in image generator for users on ChatGPT's free tier. But apparently that measure alone wasn't enough to ease the stress on OpenAI's infrastructure. (Altman said free users will "soon" be able to generate up to three images per day.)



[1] https://slashdot.org/story/25/03/27/0023207/openais-viral-studio-ghibli-moment-highlights-ai-copyright-concerns

[2] https://www.theverge.com/news/637542/chatgpt-says-our-gpus-are-melting-as-it-puts-limit-on-image-generation-requests

[3] https://x.com/sama/status/1905296867145154688?s=46

U.S. robotics companies, including Tesla and Boston Dynamics, are [1]urging lawmakers to establish a national robotics strategy to keep pace with China's aggressive investment in AI-driven robotics. The Associated Press reports:

> Jeff Cardenas, co-founder and CEO of humanoid startup Apptronik, of Austin, Texas, pointed out to lawmakers that it was American carmaker General Motors that deployed the first industrial robot at a New Jersey assembly plant in 1961. But the U.S. then ceded its early lead to Japan, which remains a powerhouse of industrial robotics, along with Europe. The next robotics race will be powered by artificial intelligence and will be "anybody's to win," Cardenas said in an interview after the closed-door meeting. "I think the U.S. has a great chance of winning. We're leading in AI, and I think we're building some of the best robots in the world. But we need a national strategy if we're going to continue to build and stay ahead."

>

> The Association for Advancing Automation said a national strategy would help U.S. companies scale production and drive the adoption of robots as the "physical manifestation" of AI. The group made it clear that China and several other countries already have a plan in place. Without that leadership, "the U.S. will not only lose the robotics race but also the AI race," the association said in a statement. The group also suggested tax incentives to help drive adoption, along with federally-funded training programs and funding for both academic research and commercial innovation. A new federal robotics office, the association argued, is necessary partly because of "the increasing global competition in the space" as well as the "growing sophistication" of the technology.



[1] https://apnews.com/article/united-states-robotics-competition-china-tech-702796f1584fe1920e5fd86f15a99b4f

Ubisoft is [1]launching a new subsidiary focused on Assassin's Creed, Far Cry, and Rainbow Six , backed by a 1.16 billion-euro investment from Tencent. "The as-yet-unnamed subsidiary will fold in the teams working on those three series, including Ubisoft studios in Montreal, Quebec, Sherbrooke, Saguenay, Barcelona and Sofia," reports Engadget. From the report:

> This new business will receive an investment of 1.16 billion-euro (roughly $1.25 billion) from its longstanding partner Tencent, granting the conglomerate a minority ownership stake. Following the transaction, Ubisoft will narrow focus to its other franchises, such as The Division and Tom Clancy's Ghost Recon. [...] There is some extra good news in the announcement. The description of the new subsidiary does specify that "it will drive further increases in quality of narrative solo experiences." So while we can expect to also see multiplayer and free-to-play offerings from the Ubisoft umbrella, they aren't giving up on single-player games.

"Today Ubisoft is opening a new chapter in its history," CEO and Co-Founder Yves Guillemot said. "As we accelerate the company's transformation, this is a foundational step in changing Ubisoft's operating model that will enable us to be both agile and ambitious."



[1] https://www.engadget.com/gaming/ubisoft-spins-out-subsidiary-with-a-billion-dollar-investment-from-tencent-183516992.html

An anonymous reader quotes a report from the BBC:

> Fashion retailer H&M is to use artificial intelligence (AI) to [1]create digital "twins" of 30 models . It says it will use the AI doppelgangers in some social media posts and marketing in the place of humans, if given permission by models. "We are curious to explore how to showcase our fashion in new creative ways -- and embrace the benefits of new technology -- while staying true to our commitment to personal style," said its chief creative officer Jorgen Andersson in a statement.

>

> The initiative was first reported by industry publication [2]Business of Fashion . H&M told the outlet that models would retain rights over their digital replicas and their use by the company and other brands for purposes such as marketing. Its images are likely to be initially used in social media posts, with watermarks that make their AI use clear, it added. H&M also said models would be compensated for use of their digital twins in a similar way to current arrangements -- which sees them paid for use of their images based on rates agreed by their agent.



[1] https://www.bbc.com/news/articles/c3vwg73xndeo

[2] https://www.businessoffashion.com/articles/technology/hm-plans-to-use-ai-models/

Google is rolling out new AI-powered features across Maps, Search, and Hotels to simplify travel planning, including a screenshot-detection tool in Maps that [1]identifies and saves locations mentioned in image text . The Verge reports:

> Once the new screenshot list is enabled in Maps, the Gemini-powered feature will detect places that are mentioned in text within screenshots on the device, show users the locations on the map, and allow them to review and save locations to a sharable list. The screenshot list feature will start rolling out in English this week to iOS users in the US, with Android support "coming soon."

>

> AI Overviews for Google Search are also being updated to expand travel planning tools, with itinerary-building features rolling out in English to mobile and desktop devices in the US this week that can create trip ideas for "distinct regions or entire countries." Users can use terms like "create a vacation itinerary for Greece that focuses on history" to explore reviews and photos from other users alongside a map of location recommendations, which can be saved to Google Maps or exported to Docs or Gmail.



[1] https://www.theverge.com/news/637137/google-maps-screenshot-searchhotels-travel-features

Columbia University has [1]suspended the student who created an AI tool designed to help job candidates cheat on technical coding interviews, according to disciplinary documents seen by Business Insider. Chungin "Roy" Lee received a yearlong suspension for "publishing unauthorized documents" from a disciplinary hearing about his product, Interview Coder, not for creating the tool itself. Lee had signed a form agreeing not to disclose his disciplinary record or post hearing materials online.

Interview Coder, which sells for $60 monthly, is on track to generate $2 million in annual revenue, Lee said. The university initially placed him on probation after finding him responsible for "facilitation of academic dishonesty." Lee had already submitted paperwork for a leave of absence before his suspension. He told BI he plans to move to San Francisco, which "was my plan all along."



[1] https://www.businessinsider.com/columbia-suspends-student-ai-interview-coder-cheat-tool-chungin-lee-2025-3

[1]MojoKid writes:

> Similar to Nvidia's recent desktop graphics launches, there are four initial GeForce RTX 50 series laptop GPUs coming to market, starting this month. At the top of the stack is the [2]GeForce RTX 5090 laptop GPU , which is equipped with 10,496 CUDA cores and is paired to 24GB of memory. Boost clocks top out around 2,160MHz and GPU power can range from 95-150 watts, depending on the particular laptop model. GeForce RTX 50 series GPUs for both laptops and desktops feature updated shader cores with support for neural shading, in addition to 4th gen ray tracing cores and 5th gen Tensor cores with support for DLSS 4. The GeForce RTX 50 series features a native PCIe gen 5 interface, in addition to support for DisplayPort 2.1b (up to UHBR20). These GPUs are also fed by the latest high speed GDDR7 memory, which offers efficiency benefits that are pertinent to laptop designs as well. Performance-wise, NVIDIA's mobile GeForce RTX 5090 is the new king of the hill in gaming laptops, and it [3]easily bests all other discrete mobile graphics options on the market currently .



[1] https://slashdot.org/~MojoKid

[2] https://www.nvidia.com/en-au/geforce/laptops/50-series/

[3] https://hothardware.com/reviews/nvidia-geforce-rtx-5090-mobile-gpu-preview

An anonymous reader quotes a report from BleepingComputer:

> Despite Oracle denying a breach of its Oracle Cloud federated SSO login servers and the [1]theft of account data for 6 million people , BleepingComputer has confirmed with multiple companies that associated data samples shared by the threat actor are valid. Last week, a person named 'rose87168' claimed to have breached Oracle Cloud servers and began selling the alleged authentication data and encrypted passwords of 6 million users. The threat actor also said that stolen SSO and LDAP passwords could be decrypted using the info in the stolen files and offered to share some of the data with anyone who could help recover them.

>

> The threat actor released multiple text files consisting of a database, LDAP data, and a list of 140,621 domains for companies and government agencies that were allegedly impacted by the breach. It should be noted that some of the company domains look like tests, and there are multiple domains per company. In addition to the data, rose87168 shared an [2]Archive.org URL with BleepingComputer for a text file hosted on the "login.us2.oraclecloud.com" server that contained their email address. This file indicates that the threat actor could create files on Oracle's server, indicating an actual breach. However, Oracle has denied that it suffered a breach of Oracle Cloud and has refused to respond to any further questions about the incident.

>

> "There has been no breach of Oracle Cloud. The published credentials are not for the Oracle Cloud. No Oracle Cloud customers experienced a breach or lost any data," the company told BleepingComputer last Friday. This denial, however, contradicts findings from BleepingComputer, which received additional samples of the leaked data from the threat actor and contacted the associated companies. Representatives from these companies, all who agreed to confirm the data under the promise of anonymity, confirmed the authenticity of the information. The companies stated that the associated LDAP display names, email addresses, given names, and other identifying information were all correct and belonged to them. The threat actor also shared emails with BleepingComputer, claiming to be part of an exchange between them and Oracle.



[1] https://www.bleepingcomputer.com/news/security/oracle-customers-confirm-data-stolen-in-alleged-cloud-breach-is-valid/

[2] https://web.archive.org/web/20250301161517/http:/login.us2.oraclecloud.com/oamfed/x.txt?x

Microsoft plans to roll out a new Windows scheduled task in May that launches automatically to [1]help Microsoft Office apps load faster . From a report:

> The company says the "Startup Boost" task will launch in the background on logon, with the roll-out to start in mid-May and worldwide general availability to be reached by late May 2025. On systems where it's toggled on, users will see new Office Startup Boost and Office Startup Boost Logon tasks in the Windows Task Scheduler, which will ensure that Office apps can preload "performance enhancements."

>

> "We are introducing a new Startup Boost task from the Microsoft Office installer to optimize performance and load-time of experiences within Office applications," Microsoft says on the Microsoft 365 message center. "After the system performs the task, the app remains in a paused state until the app launches and the sequence resumes, or the system removes the app from memory to reclaim resources. The system can perform this task for an app after a device reboot and periodically as system conditions allow."



[1] https://www.bleepingcomputer.com/news/microsoft/microsoft-new-windows-scheduled-task-will-launch-office-apps-faster/

Microsoft CEO Satya Nadella has told employees that DeepSeek's R1 AI model has [1]set "the new bar" for his company's AI ambitions , citing the startup's ability to reach the top of app store rankings. "What's most impressive about DeepSeek is that it's a great reminder of what 200 people can do when they come together with one thought and one play," The Verge cited Nadella as saying.

"Most importantly, not just leaving it there as a research project or an open source project, but to turn it into a product that was number one in the App Store. That's the new bar to me," he added. Microsoft quickly deployed DeepSeek's R1 on its Azure platform in January. The AI model gained recognition for its optimization below Nvidia's CUDA layer, enabling greater efficiency.



[1] https://www.theverge.com/notepad-microsoft-newsletter/637496/microsoft-satya-nadella-deepseek-chatgpt-ai-investments-notepad

An anonymous reader shares a report:

> You can now choose WhatsApp as your iPhone's [1]default app for calls and text messages , as noted by WABetaInfo. After updating WhatsApp to version 25.8.74, you'll see the app appear as an option in your Messaging and Calling default app settings.

>

> Apple first announced that it would let iPhone users in the European Union change their default phone and messaging apps, but it later said that everyone would be able to do the same in iOS 18.2.



[1] https://www.theverge.com/news/637489/iphone-whatsapp-default-calling-texting-app

The Metropolitan Police has confirmed its first permanent installation of live facial recognition (LFR) cameras is [1]coming this summer and the location will be the South London suburb of Croydon. From a report:

> The two cameras will be installed in the city center in an effort to combat crime and will be attached to buildings and lamp posts on North End and London Road. According to the police they will only be turned on when officers are in the area and in a position to make an arrest if a criminal is spotted. The installation follows a two-year trial in the area where police vans fitted with the camera have been patrolling the streets matching passersby to its database of suspects or criminals, leading to hundreds of arrests. The Met claims the system can alert them in seconds if a wanted wrong'un is spotted, and if the person gets the all-clear, the image of their face will be deleted.



[1] https://www.theregister.com/2025/03/27/uk_facial_recognition/

Nintendo has announced plans to introduce Virtual Game Cards for its Switch console in late April, allowing users to [1]share digital games across multiple systems , the Japanese gaming company said during its Nintendo Direct event. The new feature will enable players to virtually load and eject digital games between Nintendo Switch consoles, mimicking the flexibility of physical game cartridges.

Users can play a single digital title on up to two systems, requiring only a one-time local connection between devices. The company has also confirmed that Virtual Game Cards will be compatible with both current and next-generation hardware. The system will also feature a family sharing option, allowing users to lend digital games to family members for two-week periods before automatically returning to the owner's account.



[1] https://venturebeat.com/games/nintendo-unveils-virtual-game-cards-for-the-switch-for-family-sharing/

Paul Ginsparg, a physics professor at Cornell University, created arXiv nearly 35 years ago as a digital repository where researchers could share their findings before peer review. Today, the platform hosts more than 2.6 million papers, receives 20,000 new submissions monthly, and serves 5 million active users, Wired [1]writes in a profile of the platform .

"Just when I thought I was out, they pull me back in!" Ginsparg quotes from The Godfather, reflecting his inability to fully hand over the platform despite numerous attempts. If arXiv stopped functioning, scientists worldwide would face immediate disruption. "Everybody in math and physics uses it," says Scott Aaronson, a computer scientist at the University of Texas at Austin. "I scan it every night."

ArXiv revolutionized academic publishing, previously dominated by for-profit giants like Elsevier and Springer, by allowing instant and free access to research. Many significant discoveries, including the "transformers" paper that launched the modern AI boom, first appeared on the platform. Initially a collection of shell scripts on Ginsparg's NeXT machine in 1991, arXiv followed him from Los Alamos National Laboratory to Cornell, where it found an institutional home despite administrative challenges. Recent funding from the Simons Foundation has enabled a hiring spree and long-needed technical updates.



[1] https://www.wired.com/story/inside-arxiv-most-transformative-code-science/

China's ambitious AI infrastructure push has [1]resulted in hundreds of idle data centers with local media reporting up to 80% of newly built computing resources remaining unused. The country announced over 500 data center projects during 2023-2024, with at least 150 completed facilities now struggling to secure customers in a rapidly changing market.

The rise of DeepSeek's open-source reasoning model R1, which matches ChatGPT o1's performance at a fraction of the cost, has fundamentally altered hardware demand. Computing needs now prioritize low-latency infrastructure for real-time reasoning rather than facilities optimized for large-scale training workloads.

Technical misalignment compounds the problem, as many centers were constructed by companies with little AI expertise, MIT Technology Review reports. The facilities, often built in remote regions to capitalize on cheaper electricity and land, now face obsolescence as AI companies require proximity to tech hubs to minimize transmission delays. GPU rental prices have collapsed, with eight-GPU Nvidia H100 server clusters now leasing for 75,000 yuan ($10,333) monthly, down from peaks of 180,000 yuan, making operations financially unsustainable for many data center operators.



[1] https://www.technologyreview.com/2025/03/26/1113802/china-ai-data-centers-unused/

An anonymous reader quotes a report from Tom's Hardware:

> Qualcomm has reportedly filed secret complaints against Arm with the European Commission, the U.S. Federal Trade Commission (FTC), and the Korea Fair Trade Commission. Qualcomm argues that Arm's open licensing approach helped build a robust hardware and software ecosystem. However, this ecosystem is under threat now as Arm [1]moves to restrict that access to benefit its chip design business , namely compute subsystems (CSS) reference designs for client and datacenter processors and custom silicon based on CSS for large-scale clients.

>

> Qualcomm has presented its case to the EC, U.S. FTC, and Korea FTC behind closed doors and through formal filings, so it does not comment on the matter now. Arm rejected the accusations, stating that it is committed to innovation, competition, and upholding contract terms. The company called Qualcomm's move an attempt to shift attention from a wider commercial dispute between the two companies and use regulatory pressure for its benefit.

>

> Indeed, the antitrust complaints align with Qualcomm's arguments in a recent legal clash with Arm in Delaware. Qualcomm [2]won that trial , as the court ruled that the company did not break the terms of its architecture license agreement (ALA) and technology license agreement (TLA) by acquiring Nuvia and using its IP in its Snapdragon X processors for client PCs. Arm said it would seek a retrial. However, Qualcomm seems to want to ensure that it will have access to Arm's instruction set architecture and technologies by filing complaints with antitrust regulators.



[1] https://www.tomshardware.com/tech-industry/qualcomm-launches-global-antitrust-campaign-against-arm-accuses-arm-of-restricting-access-to-technology

[2] https://yro.slashdot.org/story/24/12/20/2216253/qualcomm-processors-properly-licensed-from-arm-us-jury-finds

The U.S. Space Force has [1]certified United Launch Alliance's Vulcan Centaur rocket for national security missions after successful test flights and resolution of a booster nozzle issue. This certification allows ULA to join SpaceX in conducting launches under the National Security Space Launch program, with Vulcan missions expected to begin this summer. SpaceNews reports:

> "Thank you to all our customer partners who have worked hand-in-hand with us throughout this comprehensive certification process. We are grateful for the collaboration and excited to reach this critical milestone in Vulcan development," said Tory Bruno, president and chief executive of United Launch Alliance in a ULA statement about the vehicle's certification.

>

> Bruno said at the roundtable that the next launch by ULA will be of its Atlas 5, carrying a set of Project Kuiper broadband satellites for Amazon. That launch is expected as soon as next month. He said then that would be followed by the first two Vulcan national security launches, missions designated USSF-106 and USSF-87. ULA did not give a schedule for those upcoming Vulcan launches but Space Systems Command, in a summary accompanying its press release, said the first NSSL mission on Vulcan is planned for the summer. Bruno said at the roundtable that the payloads for those missions have "complex processing" requirements beyond a typical mission, and did not state how long it would take them to be ready for a launch.

>

> Bruno said ULA is projecting a dozen launches this year, split roughly evenly between Atlas and Vulcan and between national security and commercial missions. ULA has been stockpiling components, such as BE-4 engines and solid rocket boosters, needed for those missions. "We're all staged up and ready, and as spacecraft show up, we'll be able to fly them," he said. He noted the company wants to get to a "baseline tempo" of two launches a month by the end of this year and perform 20 launches next year.



[1] https://spacenews.com/space-force-certifies-vulcan-for-national-security-launches/