us-technology

Key Takeaways Two major breaches exposed sensitive data — Tea app leaked thousands of selfies, government IDs, and over 1.1 million private messages, including names, contact details, and deeply personal conversations. Poor security practices enabled attacks — An unsecured Firebase database and an exploitable API left user data wide open, showing a lack of encryption and proper access controls. AI-driven “vibe coding” played a role — Heavy reliance on AI-generated code without audits likely introduced vulnerabilities, highlighting the security risks of unreviewed AI-assisted development. Severe privacy and safety fallout — Leaked data is being exploited online, putting users at risk of doxxing, harassment, and legal consequences in sensitive cases like abortion discussions. What if the “anonymous” dating app you used to warn women about toxic men accidentally leaked your selfie, ID, messages, and contact details?  That’s exactly what happened to users of Tea (officially Tea Dating Advice): a women-only “dating safety” app designed to help women share information about men in their area.  To join, women must upload a selfie and a government-issued ID to verify their identity. Once they’re in, they’re encouraged to share experiences, raise red flags, and connect with others for mutual support and protection. In a surprising turn for its mission, Tea has now experienced two major data breaches: one exposed photos and IDs, and the second leaked over a million messages sent in (admittedly misplaced) confidence.  Some of these messages discuss abortions, cheating partners, and personal details like car models and social handles. Both types of breaches are now being exploited online, with some images turned into public rankings and private data used to dox or mock users.  So how did an app whose mission was to make women safer end up doing the exact opposite in one of the worst privacy disasters of the year? Let’s take a closer look. What Got Leaked, and Why It’s So Serious The first Tea breach drew a lot of negative attention. An exposed Firebase database left tens of thousands of selfies and government IDs accessible to anyone. 4chan users quickly scraped the images and made mirror downloads.  They even set up a Facemash-style site where people ranked leaked selfies by attractiveness, including leaderboards. Tea’s initial response was disappointing. The company minimized the breach, claiming it only involved “legacy” data from over two years ago. Sadly for them, that defense quickly fell apart. A second, much larger breach has now exposed over 1.1 million private messages, with many of these sent as recently as last week. These weren’t just casual DMs. They included: Women discussing abortions Users realized they were dating the same men Real phone numbers, names, and social media handles Accusations of cheating, abuse, and more, often naming people directly To make matters worse, a researcher found out it was possible to use the app’s API to send a notification to every single user. Tea marketed itself as a place to stay anonymous.  The nature of these leaks showed that it was anything but: with full identities linked to deeply personal conversations, users could now face blackmail, harassment, or worse.  A Case Study in Negligence: How It Happened Tea’s backend was shockingly insecure for an app that promised safety, not once, but twice.  The initial breach involved a completely unsecured Firebase storage instance. That alone exposed over 72,000 images, including 13,000 selfies, government-issued IDs, and 59,000 images from posts, messages, and comments. In a statement, Tea claimed the breach only affected data stored on its “legacy data system.”  That claim didn’t last long, though. Just days later, security researcher Kasra Rahjerdi uncovered a second, more serious vulnerability: Tea’s API allowed any logged-in user to access a recent, unsecured database using their API key, which included private messages from as recently as last week. Rahjerdi discovered something even more alarming in his research: push notifications could be sent to all users using the same attack vector. Tea claims it has since fixed the vulnerability and contacted law enforcement. But it’s too little too late: the damage has been done. The data has already been scraped, archived, and widely shared online. The app was marketed as discreet and anonymous, but the reality was closer to leaving the door wide open and hoping no one walked in. Vibe Coding, AI Tools, and Faking Competence Tea Dating Advice didn’t just have bad luck. It also suffered from poor development practices and likely relied too much on AI-generated code. According to the original hacker who revealed the first breach on 4chan, Tea was a prime example of “vibe coding”: a rising trend where developers rely heavily on AI tools to build products without proper security checks, version control, or code reviews. Guillermo Rauch, founder and CEO of AI cloud app company Vercel, offered a sardonic take on this trend: “On Tea Dating, AI and Vibe Coding security TL;DR: the antidote for mistakes AIs make is… more AI.” Unfortunately for Tea, and even more so for the women who used it, that approach appears to have backfired. A Georgetown University study found that 48% of AI-generated code had security flaws.  Tech consultant Santiago Valdarrama gets it right: “Vibe coding is awesome, but the code these models generate is full of security holes and can be easily hacked.” This kind of AI-assisted (or, in all honesty, AI-led) development might help quickly ship features. But without oversight, it also ships vulnerabilities. The Ongoing Repercussions of Tea’s Breach Tea promised its users a private space to share sensitive stories, from relationship red flags to personal trauma. Sadly, it ended up turning those confessions into liabilities. After the initial breach, photos of women who used the app were scraped and reposted on 4chan. Soon after, they were transformed into a Facemash-style site that ranked their appearances. Many of these pictures were voted on tens of thousands of times, erasing any anonymity and dignity from these women instantly. The second breach surpasses the first by a large margin. It includes genuine conversations between real women, discussing

A jury in Florida has found Tesla partially liable for a 2019 crash involving the company’s Autopilot self-driving feature, The Washington Post reports. As a result, the company will have to pay $43 million in compensatory damages and even more in punitive damages. Autopilot comes pre-installed on Tesla’s cars and handles things like collision detection and emergency braking. Tesla has mostly avoided taking responsibility for crashes involving cars with the Autopilot enabled, but the Florida case played out differently. The jury ultimately decided that the self-driving tech enabled driver George McGee to take his eyes off the road and hit a couple, Naibel Benavides Leon and Dillon Angulo, ultimately killing one and severely injuring the other. During the case, Tesla’s lawyers argued that McGee’s decision to take his eyes off the road to reach for his phone was the cause of the crash, and that Autopilot shouldn’t be considered. The plaintiffs, Angulo and Benevides Leon’s family, argued that the way Tesla and Elon Musk talked about the feature ultimately created the illusion that Autopilot was safer than it really was. “My concept was that it would assist me should I have a failure … or should I make a mistake,” McGee said on the stand. “And in that case I feel like it failed me.” The jury ultimately assigned two-thirds of the responsibility to McGee and a third to Tesla, according to NBC News. When reached for comment, Tesla said it would appeal the decision and gave the following statement: Today’s verdict is wrong and only works to set back automotive safety and jeopardize Tesla’s and the entire industry’s efforts to develop and implement life-saving technology. We plan to appeal given the substantial errors of law and irregularities at trial. Even though this jury found that the driver was overwhelmingly responsible for this tragic accident in 2019, the evidence has always shown that this driver was solely at fault because he was speeding, with his foot on the accelerator – which overrode Autopilot – as he rummaged for his dropped phone without his eyes on the road. To be clear, no car in 2019, and none today, would have prevented this crash. This was never about Autopilot; it was a fiction concocted by plaintiffs’ lawyers blaming the car when the driver – from day one – admitted and accepted responsibility. In a National Highway Traffic Safety Administration investigation of Autopilot from 2024, crashes were blamed on driver misuse of Tesla’s system and not the system itself. The NHTSA also found that Autopilot was overly permissive and “did not adequately ensure that drivers maintained their attention on the driving task,” which lines up with the 2019 Florida crash. While Autopilot is only one component of Tesla’s larger collection of self-driving driving features, selling the idea that the company’s cars could safely driving on their own is a key part of its future. Elon Musk has claimed that Full Self-Driving (FSD), the paid upgrade to Autopilot, is “safer than human driving.” Tesla’s Robotaxi service relies on FSD being able to function with no or minimal supervision, something that produced mixed results in the first few days the service was available. Update, August 1, 6:05PM ET: This story was updated after publication to include Tesla’s statement. Read More

In partnership with developers, game marketplace GOG (Good Old Games) has launched a new website called FreedomtoBuy.games that’ll let you download select “adult” games for free. GOG believes the website takes a stand “against the quiet erasure of creative works from digital shelves,” a response of sorts to recent decisions from Steam and Itch to delist certain violent and sexuality-explicit games from their respective platforms. GOG is currently offering 13 games for free for the next 48 hours, some with well-known scandals and others that seem to fall into the “NSFW visual novel” bucket that makes up the majority of sexually-explicit games on digital storefronts. The titles available to download are: Leap of Love Being a DIK — Season 1 Leap of Faith POSTAL 2 House Party HuniePop Lust Theory Agony + Agony Unrated Treasure of Nadia Summer’s Gone — Season 1 Fetish Locator Week One Helping the Hotties Sapphire Safari POSTAL 2, a graphically violent open-world game, is a notable inclusion because it was banned in New Zealand in 2004 and delisted from the German version of Steam in 2016. HuniePop, one of several “adult-only” games Twitch streamers are explicitly forbidden to stream, makes sense on the list, too. GOG has made a concerted effort to preserve games of all types, including maintaining them so that they run on current hardware. The point of making these games available to download is as much about preservation as it is about highlighting how apparently easy it is to pressure digital storefronts to remove content, though. Valve’s decision to delist titles from Steam was chalked up to a new rule that requires games to abide by the standards set by the payment processors that work with Steam. Itch offered a similar explanation for the delistings on its storefront, pointing to pressure payment processors were receiving from a nonprofit called Collective Shout. In the process, indie games like VILE: Exhumed have been delisted, primarily for depicting things that might make a certain group of people uncomfortable. Itch, for its part, seems to be trying to bring back as many games to its storefront as it can. The platform is restoring free NSFW games, and says it’s still in talks with its payment partners about restoring paid games to its storefront. Read More

T-Mobile has sealed the deal on its UScellular acquisition. In exchange for $4.3 billion, T-Mobile gets UScellular’s customers, stores and 30 percent of its spectrum. If you’re a UScellular customer, you don’t have to do anything. “UScellular customers stay on their existing plans with no changes for now,” the carrier said. You can continue to manage your account through UScellular’s website. You can also still turn to the T-Mobile-owned carrier for customer support. The $4.3 billion wasn’t the only price T-Mobile had to pay. To gain the approval of Trump’s FCC, the carrier agreed to gut its DEI programs. That followed Verizon doing the same for its Frontier acquisition. The president has used merger approvals as a cudgel to push his agenda (including getting lawsuits settled) in the private sector. UScellular will now exist only as an infrastructure company. It can now generate revenue from licensing its remaining spectrum and towers. Read More

Hey iPhone users! If you haven’t heard yet, a slew of new exciting updates are coming to Apple’s operating system this fall. No, it won’t be called iOS 19 — it’ll be named iOS 26. The change we’re most excited for is the new Liquid Glass design (think Windows Vista, but arguably more thoughtful), which looks to be Apple’s largest visual update in years. We spent two weeks test-driving it — you can check out our detailed hands-on iOS 26 preview, or you can try it out yourself by downloading and installing the public beta, available now. (While the beta is open to the public and stable, always remember there’s a degree of risk involved with beta software.) Not ready to upgrade your smartphone this year? No worries, we’ll help you find out if your phone will be able to run iOS 26. Last year, Apple didn’t nix any iPhones from its eligibility list, but that’s not the case for 2025 — a few models are getting cut this time. All iPhone 8 models and the iPhone X were the last to get the boot in 2023, and this year the 2018 models are getting left behind. If you have an ineligible device, you won’t be able to download iOS 26 when it becomes available this fall. We’ll get to the bottom of which iPhones will support iOS 26 this year. To see what’s coming with the latest OS and more, check out everything announced at Apple’s WWDC June 9 event. These three iPhones won’t be compatible with iOS 26 Unlike last year, several iPhone models won’t be eligible to download the newest iOS when it makes its debut this fall. This trio of models first released in 2018 won’t be coming to the iOS 26 party: iPhone XR iPhone XS iPhone XS Max iPhones compatible with iOS 26 While we don’t yet know the new iPhones Apple will be dropping this fall — though there are iPhone 17 rumors — we do know, per Apple’s site, that the phones listed below will be compatible with iOS 26. Basically, if you have an iPhone that was announced in 2019 or later, you’re in the clear: iPhone SE (second generation or later) iPhone 11 iPhone 11 Pro iPhone 11 Pro Max iPhone 12 iPhone 12 mini iPhone 12 Pro iPhone 12 Pro Max iPhone 13 iPhone 13 mini iPhone 13 Pro iPhone 13 Pro Max iPhone 14 iPhone 14 Plus iPhone 14 Pro iPhone 14 Pro Max iPhone 15 iPhone 15 Plus iPhone 15 Pro iPhone 15 Pro Max iPhone 16e iPhone 16 iPhone 16 Plus iPhone 16 Pro iPhone 16 Pro Max What if I don’t want to buy a new iPhone? If you want to continue using your older iPhone that isn’t supported by iOS 26, that’s fine. However, you’ll miss out on security updates which could potentially put your phone at risk for malware and other threats. Additionally, some apps may stop working if they require a certain version of iOS or later. And of course, you won’t be able to access the latest features iOS 26 offers. When will iOS 26 become available? Apple usually rolls out its latest iOS in mid-September, just a few days before the new iPhones hit store shelves. Last year, it released iOS 18 on Monday, Sept. 16. Expect a confirmation of the release date at the iPhone 17 event, expected in early September. iOS 26 features we’re excited about Liquid Glass design: Your home screen is getting revamped with new app icons, including dark mode and all-clear options. You’ll also notice buttons with a new floating design. Liquid Glass was designed to make all of Apple’s OSes more cohesive. Phone app redesign: You’ll finally be able to scroll through contacts, recent calls and voicemail messages all on one screen. It also comes with a new feature called Hold Assist that’ll notify you when an agent comes to the phone so you can avoid the elevator music. Live Translate: iOS 26 is bringing the ability to have a conversation via phone call or text message with someone who speaks another language. Live Translate will translate your conversation in real time. Polls feature: Coming to group messages in the Messages app, chat members will be able to create polls. This can help prevent the unwanted 30+ messages when it comes to deciding which restaurant you’re meeting at this weekend. Read More

Liquid Glass is a huge new change coming to iOS 26. (Apple) Waiting until the fall can feel like ages when you’re ready to upgrade your iPhone to iOS 26. But there’s good news: you can test out all the features now by downloading and installing Apple’s public beta, which CEO Tim Cook says is (with the other current beta operating systems) “by far the most popular developer betas we’ve had,” 9to5Mac reports. We also previewed the iOS 26 public beta release, which shows off the fresh home and lock screen redesign we’ve been asking to see for years. Called Liquid Glass, the new translucent look will extend across all of Apple’s upcoming operating systems. The overhaul is one of several big changes coming to iOS, macOS, iPadOS and the rest of Apple’s software suite, all of which were showcased during the company’s WWDC keynote on June 9. After overpromising on AI plans last year, Apple kept its iOS roadmap focused more on basic quality of life improvements this year. There are multiple useful additions coming to the Phone and Messages apps on your iPhone, for instance: Apple execs outlined the ability to weed out spam texts or other unknown senders and an option to hold your spot on a phone call when you’ve been waiting for a representative to pick up. Plus, a treasured feature that we took for granted is coming back (hint: it’s in the Photos app). Siri, meanwhile, is in a holding pattern. Apple has previously specified that its smarter voice assistant — first promised at WWDC 2024 — is delayed until some point “in the coming year,” so you shouldn’t expect any major changes in the current betas. But there are reports that Apple is aiming to give Siri a bigger brain transplant by basing it on third-party artificial intelligence models like OpenAI’s ChatGPT or Anthropic’s Claude, which could make 2026 a pivotal year. With each beta, it seems like additional new improvements are popping up, like a newly discovered FaceTime feature that’ll freeze your video if it detects nudity. Most newer iPhone models are eligible to download iOS 26 (both the betas and final version). Want to see the full list of new features coming this fall? Read on. What is iOS 26? The current iPhone operating system is iOS 18, and Apple is still actively updating it — version 18.6 was just recently released. But don’t expect to see iOS 19. Instead, Apple is skipping the numbering ahead to iOS 26 later this year. The company has decided to line up its iOS version numbers with a year-based system, similar to car model years. So while iOS and its sibling operating systems will be released in late 2025, they’re all designated “26” to reflect the year ahead. (Meanwhile, iOS 18 is still getting new versions this summer, too.) It’s official, we’re moving to iOS 26. (Apple) What is Liquid Glass design? Let’s be honest. Out of everything announced at WWDC this year, the new Liquid Glass design was the star of the show. The iPhone’s home and lock screens have looked pretty much the same year after year — the last exciting thing (in my opinion) was the option to add your own aesthetic to your home screen by customizing your apps and widgets. So seeing the home and lock screens’ new facelift is refreshing. So what exactly is Liquid Glass? Apple calls it a “new translucent material” since, well, the apps and widgets are clear. However, the screen can still adapt to dark and light modes, depending on surroundings. You’ll also notice buttons with a new floating design in several apps, like Phone and Maps. They’re designed to be less distracting than the current buttons, but are still easy to see. While the design overhaul has proven to be controversial since its announcement, some — including Engadget’s own Devindra Hardawar — like the new direction, even if it’s somewhat reminiscent of Microsoft’s translucent Windows Vista Aero designs from nearly twenty years ago. That said, as of the release of the iOS 26 beta 2, Apple has already incorporated some user feedback into the design, dialing back the transparency in at least some places. And while it will continue to evolve, Apple users won’t be able to escape it: Liquid Glass was designed to make all of Apple’s OSes more cohesive. Here’s a look at how the translucent aesthetic will look with the new macOS Tahoe 26 on your desktop. What are the new and notable features of iOS 26? iOS 26 has a laundry list of new features. Among the most worthwhile: Phone app redesign: You’ll finally be able to scroll through contacts, recent calls and voicemail messages all on one screen. It also comes with a new feature called Hold Assist that’ll notify you when an agent comes to the phone so you can avoid the elevator music and continue on with other tasks. Live Translation in Phone, FaceTime and Messages: iOS 26 is bringing the ability to have a conversation via phone call or text message with someone who speaks another language. Live Translation will translate your conversation in real time, which results in some stop-and-go interactions in the examples Apple shared during its presentation. Polls in group chats: Tired of sorting through what seems like hundreds of messages in your group chat? You and your friends will soon be able to create polls in group messages for deciding things like which brunch spot you’re eating at or whose car you’re taking on a road trip. Filtering unknown senders in Messages: If you haven’t received spam texts about unpaid tolls or other citations, you’re lucky. For those of us who have, those annoying messages will soon be filtered away in a separate folder. Visual Intelligence: Similar to a reverse Google image search, this new feature will allow you to search for anything that’s on your iPhone screen. For instance, if you spot a pair of shoes someone is wearing

August 1, 2025 3:05 PM Image credit: VentureBeat with DALL-E 3 Want smarter insights in your inbox? Sign up for our weekly newsletters to get only what matters to enterprise AI, data, and security leaders. Subscribe Now The rise in Deep Research features and other AI-powered analysis has given rise to more models and services looking to simplify that process and read more of the documents businesses actually use.  Canadian AI company Cohere is banking on its models, including a newly released visual model, to make the case that Deep Research features should also be optimized for enterprise use cases.  The company has released Command A Vision, a visual model specifically targeting enterprise use cases, built on the back of its Command A model. The 112 billion parameter model can “unlock valuable insights from visual data, and make highly accurate, data-driven decisions through document optical character recognition (OCR) and image analysis,” the company says. “Whether it’s interpreting product manuals with complex diagrams or analyzing photographs of real-world scenes for risk detection, Command A Vision excels at tackling the most demanding enterprise vision challenges,” the company said in a blog post.  The AI Impact Series Returns to San Francisco – August 5 The next phase of AI is here – are you ready? Join leaders from Block, GSK, and SAP for an exclusive look at how autonomous agents are reshaping enterprise workflows – from real-time decision-making to end-to-end automation. Secure your spot now – space is limited: https://bit.ly/3GuuPLF This means Command A Vision can read and analyze the most common types of images enterprises need: graphs, charts, diagrams, scanned documents and PDFs.  ? @cohere just dropped Command A Vision on @huggingface ? Designed for enterprise multimodal use cases: interpreting product manuals, analyzing photos, asking about charts… ❓?? A 112B dense vision-language model with SOTA performance – check out the benchmark metrics in… pic.twitter.com/ORMfM5f8cF — Jeff Boudier ? (@jeffboudier) July 31, 2025 Since it’s built on Command A’s architecture, Command A Vision requires two or fewer GPUs, just like the text model. The vision model also retains the text capabilities of Command A to read words on images and understands at least 23 languages. Cohere said that, unlike other models, Command A Vision reduces the total cost of ownership for enterprises and is fully optimized for retrieval use cases for businesses.  How Cohere is architecting Command A Cohere said it followed a Llava architecture to build its Command A models, including the visual model. This architecture turns visual features into soft vision tokens, which can be divided into different tiles.  These tiles are passed into the Command A text tower, “a dense, 111B parameters textual LLM,” the company said. “In this manner, a single image consumes up to 3,328 tokens.” Cohere said it trained the visual model in three stages: vision-language alignment, supervised fine-tuning (SFT) and post-training reinforcement learning with human feedback (RLHF). “This approach enables the mapping of image encoder features to the language model embedding space,” the company said. “In contrast, during the SFT stage, we simultaneously trained the vision encoder, the vision adapter and the language model on a diverse set of instruction-following multimodal tasks.” Visualizing enterprise AI  Benchmark tests showed Command A Vision outperforming other models with similar visual capabilities.  Cohere pitted Command A Vision against OpenAI’s GPT 4.1, Meta’s Llama 4 Maverick, Mistral’s Pixtral Large and Mistral Medium 3 in nine benchmark tests. The company did not mention if it tested the model against Mistral’s OCR-focused API, Mistral OCR.  It enables agents to securely see inside your organization’s visual data, unlocking the automation of tedious tasks involving slides, diagrams, PDFs, and photos. pic.twitter.com/iHZnUWekrk — cohere (@cohere) July 31, 2025 Command A Vision outscored the other models in tests such as ChartQA, OCRBench, AI2D and TextVQA. Overall, Command A Vision had an average score of 83.1% compared to GPT 4.1’s 78.6%, Llama 4 Maverick’s 80.5% and the 78.3% from Mistral Medium 3.  Most large language models (LLMs) these days are multimodal, meaning they can generate or understand visual media like photos or videos. However, enterprises generally use more graphical documents such as charts and PDFs, so extracting information from these unstructured data sources often proves difficult.  With Deep Research on the rise, the importance of bringing in models capable of reading, analyzing and even downloading unstructured data has grown. Cohere also said it’s offering Command A Vision in an open weights system, in hopes that enterprises looking to move away from closed or proprietary models will start using its products. So far, there is some interest from developers. Very impressed at its accuracy extracting hand handwritten notes from an image! — Adam Sardo (@sardo_adam) July 31, 2025 Finally, an AI that won’t judge my terrible doodles. — Martha Wisener ? (@martwisener) August 1, 2025 Daily insights on business use cases with VB Daily If you want to impress your boss, VB Daily has you covered. We give you the inside scoop on what companies are doing with generative AI, from regulatory shifts to practical deployments, so you can share insights for maximum ROI. Read our Privacy Policy Thanks for subscribing. Check out more VB newsletters here. An error occured. Read More

August 1, 2025 12:07 PM VentureBeat/Midjourney Want smarter insights in your inbox? Sign up for our weekly newsletters to get only what matters to enterprise AI, data, and security leaders. Subscribe Now When President Trump released the U.S. AI Action Plan last week, many were surprised to see “encourage open-source and open-weight AI,” as one of the administration’s top priorities. The White House has elevated what was once a highly technical topic into an urgent national concern — and a key strategy to winning the AI race against China. China’s emphasis on open source, also highlighted in its own Action Plan released shortly after the U.S., makes the open-source race imperative. And the global soft power that comes with more open models from China makes their recent leadership even more notable.  When DeepSeek-R1, a powerful open-source large language model (LLM) out of China, was released earlier this year, it didn’t come with a press tour. No flashy demos. No keynote speeches. But it was open weights and open science. Open weight means anyone with the right skills and computing resources can run, replicate, or make a model their own; open science shares some of the tricks behind the model development. Within hours, researchers and developers seized on it. Within days, it became the most-liked model of all time on Hugging Face — with thousands of variants created and used across major tech companies, research labs and startups. Most strikingly, this explosion of adoption happened not just abroad, but in the U.S. For the first time, American AI was being built on Chinese foundations. The AI Impact Series Returns to San Francisco – August 5 The next phase of AI is here – are you ready? Join leaders from Block, GSK, and SAP for an exclusive look at how autonomous agents are reshaping enterprise workflows – from real-time decision-making to end-to-end automation. Secure your spot now – space is limited: https://bit.ly/3GuuPLF DeepSeek wasn’t the only one Within a week, the U.S. stock market — sensing the tremor — took a tumble. It turns out Deepseek was just the opening act. Dozens of Chinese research groups are now pushing the frontiers of open-source AI, sharing not only powerful models, but the data, code and scientific methods behind them. They’re moving quickly — and they’re doing it in the open. Meanwhile, U.S.-based companies — many of which pioneered the modern AI revolution — are increasingly closing up. Flagship models like GPT-4, Claude and Gemini are no longer released in ways that allow builders more control. They’re accessible only through chatbots or APIs: Gated interfaces that let you interact with a model but not see how it works, retrain it or use it freely. The model’s weights, training data and behavior remain proprietary, tightly controlled by a few tech giants. This is a dramatic reversal. Between 2016 and 2020, the U.S. was the global leader in open-source AI. Research labs from Google, OpenAI, Stanford and elsewhere released breakthrough models and methods that laid the foundation for everything we now call “AI.” The transformer — the “T” in ChatGPT — was born out of this open culture. Hugging Face was created during this era to democratize access to these technologies. Now, the U.S. is slipping, and the implications are profound. American scientists, startups and institutions are increasingly driven to build on Chinese open models because the best U.S. models are locked behind APIs. As each new open model emerges from abroad, Chinese companies like DeepSeek and Alibaba strengthen their positions as foundational layers in the global AI ecosystem. The tools that power America’s next generation of AI products, research and infrastructure are increasingly coming from overseas. And at a deeper level, there’s a more fundamental risk: Every advancement in AI — including the most closed systems — is built on open foundations. Proprietary models depend on open research, from transformer architecture to training libraries and evaluation frameworks. But more importantly, open-source increases a country’s velocity in building AI. It fuels rapid experimentation, lowers barriers to entry and creates compounding innovation. When openness slows down, the entire ecosystem follows. If the U.S. falls behind in open-source today, it may find itself falling behind in AI altogether. Moving away from black box AI This matters not just for innovation, but for security, science and democratic governance. Open models are transparent and auditable. They allow governments, educators, healthcare institutions and small businesses to adapt AI to their needs, without vendor lock-in or black-box dependencies. We need more and better U.S.-developed open source models and artifacts. U.S. institutions already pushing for openness must build on their success. Meta’s open-weight Llama family has led to tens of thousands of variations on Hugging Face. The Allen Institute for AI continues to publish excellent fully open models. Promising startups like Black Forest are building open multimodal systems. Even OpenAI has suggested it may release open weights soon. With more public and policy support for open-source AI, as demonstrated by the U.S. AI Action Plan, we can restart a decentralized movement that will ensure America’s leadership. It’s time for the American AI community to wake up, drop the “open is not safe” narrative, and return to its roots: Open science and open-source AI, powered by an unmatched community of frontier labs, big tech, startups, universities and non‑profits. We can restart a decentralized movement that will ensure U.S. leadership, built on openness, competition and scientific inquiry, and empower the next generation of builders. If we want AI to reflect democratic principles, we have to build it in the open. And if the U.S. wants to lead the AI race, it must lead the open-source AI race. Clément Delangue is the co-founder and CEO of Hugging Face. Daily insights on business use cases with VB Daily If you want to impress your boss, VB Daily has you covered. We give you the inside scoop on what companies are doing with generative AI, from regulatory shifts to practical deployments, so you can share insights for

Data platform vendor Informatica is expanding its AI capabilities as the needs of gen AI continue to increase enterprise requirements. Informatica is no stranger to the world of AI; in fact, the company debuted its first Claire AI tool for data in 2018. In the modern generative AI era, the company has expanded its technology with improved natural language capabilities in Claire GPT, as part of Informatica’s Intelligent Data Management Cloud (IDMC), which debuted in 2023. The fundamental premise is all about making it easier, faster and more intelligent to access and use data. It’s a value proposition that has made the company an attractive acquisition target, with Salesforce announcing in May that it intends to acquire the company for $8 billion. While that acquisition proceeds through approvals and regulatory processes, enterprises still face data challenges that need to be addressed. Today, Informatica announced its Summer 2025 release, showcasing how the company’s AI journey over the past seven years has evolved to address enterprise data needs. The update introduces natural language interfaces that can build complex data pipelines from simple English commands, AI-powered governance that automatically tracks data lineage to machine learning models and auto-mapping capabilities that compress week-long schema mapping projects into minutes.  The AI Impact Series Returns to San Francisco – August 5 The next phase of AI is here – are you ready? Join leaders from Block, GSK, and SAP for an exclusive look at how autonomous agents are reshaping enterprise workflows – from real-time decision-making to end-to-end automation. Secure your spot now – space is limited: https://bit.ly/3GuuPLF The release addresses a persistent enterprise data challenge that generative AI has made more urgent.  “The thing that has not changed is the data continues to be fragmented in the enterprise and that fragmentation is still at a rapid scale, it’s not converging whatsoever,” Pratik Parekh, SVP and GM of Cloud Integration at Informatica told VentureBeat. “So that means that you have to bring all of this data together.” From machine learning to gen AI for enterprise data To better understand what Informatica is doing now, it’s critical to understand how it has gotten to this point. Informatica’s initial Claire implementation in 2018 focused on practical machine learning (ML) problems that plagued enterprise data teams. The platform used accumulated metadata from thousands of customer implementations to provide design-time recommendations, runtime optimizations and operational insights. The foundation was built on what Parekh calls a “metadata system of intelligence” containing 40 petabytes of enterprise data patterns. This wasn’t abstract research, but instead applied machine learning that addressed specific bottlenecks in data integration workflows. That metadata system of intelligence has continued to improve over the years, and in the summer 2025 release, the platform includes auto-mapping capabilities that solve a persistent data problem. This feature automatically maps fields between different enterprise systems using machine learning algorithms trained on millions of existing data integration patterns. “If you have worked with data management, you know mapping is a pretty time-consuming work,” Parekh said. Auto mapping is all about taking data from a source system, such as SAP, and then using that data with other enterprise data to create a Master Data Management (MDM) record. MDM for enterprise data professionals is the so-called ‘golden record’ as it is intended to be the source of truth about a certain entity. The auto mapping feature can understand the schemas of the different systems and create the correct data field in the MDM. The results demonstrate the value of Informatica’s long-term investment in AI. Tasks that previously required deep technical expertise and significant time investment now happen automatically with high accuracy rates. “Our professional services have done some work mapping that typically takes seven days to build,” Parekh said. “This is now being done in less than five minutes,” Parekh said. A core element of any modern AI system is a natural language interface, typically accompanied by some form of copilot to assist users in executing tasks. In that regard, Informatica is no different than any other enterprise software vendor. Where it differs, though, is still on the metadata and machine learning technology. The summer 2025 release enhances Claire Copilot for Data Integration, which became generally available in May 2025 after nine months in early access and preview. The copilot enables users to type requests, such as “bring all Salesforce data into Snowflake,” and have the system orchestrate the necessary pipeline components.  The summer 2025 release adds new interactive capabilities to the copilot, including enhanced question-and-answer features that help users understand how to use the product, with answers sourced directly from documentation and help articles. The technical implementation required developing specialized language models fine-tuned for data management tasks using what Parekh calls – Informatica grammar. “The natural language translated into Informatica grammar is where our secret sauce comes in,” Parekh explained. “Our whole platform is a metadata driven platform. So underneath we have our own grammar as to how this describes the mapping, what describes the data quality rule, what describes an MDM asset.” Market timing: Enterprise AI demands explode The timing of Informatica’s AI evolution aligns with fundamental changes in how enterprises consume data.  Brett Roscoe, SVP & GM, Cloud Data Governance and Cloud Ops at Informatica, noted that a big difference in the enterprise data landscape over the last several years has been the scale, with more people than ever needing more access to data. Previously, data requests came primarily from centralized analytics teams with technical expertise; in the gen AI era, those requests come from everywhere. “All of a sudden, with the world of gen AI, you’ve got your marketing team and your finance team all asking for data to go drive their generative AI projects,” Roscoe explained. The summer release’s AI Governance Inventory and Workflows capabilities tackle this challenge directly. The platform now automatically catalogs AI models, tracks their data sources and maintains lineage from source systems through to AI applications. This addresses enterprise concerns about maintaining visibility and control as AI projects proliferate beyond

Google DeepMind announced today a breakthrough artificial intelligence system that transforms how organizations analyze Earth’s surface, potentially revolutionizing environmental monitoring and resource management for governments, conservation groups, and businesses worldwide. The system, called AlphaEarth Foundations, addresses a critical challenge that has plagued Earth observation for decades: making sense of the overwhelming flood of satellite data streaming down from space. Every day, satellites capture terabytes of images and measurements, but connecting these disparate datasets into actionable intelligence has remained frustratingly difficult. “AlphaEarth Foundations functions like a virtual satellite,” the research team writes in their paper. “It accurately and efficiently characterizes the planet’s entire terrestrial land and coastal waters by integrating huge amounts of Earth observation data into a unified digital representation.” The AI system reduces error rates by approximately 23.9% compared to existing approaches while requiring 16 times less storage space than other AI systems. This combination of accuracy and efficiency could dramatically lower the cost of planetary-scale environmental analysis. The AI Impact Series Returns to San Francisco – August 5 The next phase of AI is here – are you ready? Join leaders from Block, GSK, and SAP for an exclusive look at how autonomous agents are reshaping enterprise workflows – from real-time decision-making to end-to-end automation. Secure your spot now – space is limited: https://bit.ly/3GuuPLF How the AI compresses petabytes of satellite data into manageable intelligence The core innovation lies in how AlphaEarth Foundations processes information. Rather than treating each satellite image as a separate piece of data, the system creates what researchers call “embedding fields” — highly compressed digital summaries that capture the essential characteristics of Earth’s surface in 10-meter squares. “The system’s key innovation is its ability to create a highly compact summary for each square,” the research team explains. “These summaries require 16 times less storage space than those produced by other AI systems that we tested and dramatically reduces the cost of planetary-scale analysis.” This compression doesn’t sacrifice detail. The system maintains what the researchers describe as “sharp, 10×10 meter” precision while tracking changes over time. For context, that resolution allows organizations to monitor individual city blocks, small agricultural fields, or patches of forest — critical for applications ranging from urban planning to conservation. Brazilian researchers use the system to track Amazon deforestation in near real-time More than 50 organizations have been testing the system over the past year, with early results suggesting transformative potential across multiple sectors. In Brazil, MapBiomas uses the technology to understand agricultural and environmental changes across the country, including within the Amazon rainforest. “The Satellite Embedding dataset can transform the way our team works,” Tasso Azevedo, founder of MapBiomas, said in a statement. “We now have new options to make maps that are more accurate, precise and fast to produce — something we would have never been able to do before.” The Global Ecosystems Atlas initiative employs the system to create what it calls the first comprehensive resource for mapping the world’s ecosystems. The project helps countries classify unmapped regions into categories like coastal shrublands and hyper-arid deserts — crucial information for conservation planning. “The Satellite Embedding dataset is revolutionizing our work by helping countries map uncharted ecosystems — this is crucial for pinpointing where to focus their conservation efforts,” said Nick Murray, Director of the James Cook University Global Ecology Lab and Global Science Lead of Global Ecosystems Atlas. The system solves satellite imagery’s biggest problem: clouds and missing data The research paper reveals sophisticated engineering behind these capabilities. AlphaEarth Foundations processes data from multiple sources — optical satellite images, radar, 3D laser mapping, climate simulations, and more — weaving them together into a coherent picture of Earth’s surface. What sets the system apart technically is its handling of time. “To the best of our knowledge, AEF is the first EO featurization approach to support continuous time,” the researchers note. This means the system can create accurate maps for any specific date range, even interpolating between observations or extrapolating into periods with no direct satellite coverage. The model architecture, dubbed “Space Time Precision” or STP, simultaneously maintains highly localized representations while modeling long-distance relationships through time and space. This allows it to overcome common challenges like cloud cover that often obscures satellite imagery in tropical regions. Why enterprises can now map vast areas without expensive ground surveys For technical decision-makers in enterprise and government, AlphaEarth Foundations could fundamentally change how organizations approach geospatial intelligence. The system excels particularly in “sparse data regimes” — situations where ground-truth information is limited. This addresses a fundamental challenge in Earth observation: while satellites provide global coverage, on-the-ground verification remains expensive and logistically challenging. “High-quality maps depend on high-quality labeled data, yet when working at global scales, a balance must be struck between measurement precision and spatial coverage,” the research paper notes. AlphaEarth Foundations’ ability to extrapolate accurately from limited ground observations could dramatically reduce the cost of creating detailed maps for large areas. The research demonstrates strong performance across diverse applications, from crop type classification to estimating evapotranspiration rates. In one particularly challenging test involving evapotranspiration — the process by which water transfers from land to atmosphere — AlphaEarth Foundations achieved an R² value of 0.58, while all other methods tested produced negative values, indicating they performed worse than simply guessing the average. Google positions Earth monitoring AI alongside its weather and wildfire systems The announcement places Google at the forefront of what the company calls “Google Earth AI” — a collection of geospatial models designed to tackle planetary challenges. This includes weather predictions, flood forecasting, and wildfire detection systems that already power features used by millions in Google Search and Maps. “We’ve spent years building powerful AI models to solve real-world problems,” write Yossi Matias, VP & GM of Google Research, and Chris Phillips, VP & GM of Geo, in an accompanying blog post published this morning. “These models already power features used by millions, like flood and wildfire alerts in Search and Maps; they also provide actionable insights through Google