VentureBeat

Join the event trusted by enterprise leaders for nearly two decades. VB Transform brings together the people building real enterprise AI strategy. Learn more Google DeepMind announced Thursday what it claims is a major breakthrough in hurricane forecasting, introducing an artificial intelligence system that can predict both the path and intensity of tropical cyclones with unprecedented accuracy — a longstanding challenge that has eluded traditional weather models for decades. The company launched Weather Lab, an interactive platform showcasing its experimental cyclone prediction model, which generates 50 possible storm scenarios up to 15 days in advance. More significantly, DeepMind announced a partnership with the U.S. National Hurricane Center, marking the first time the federal agency will incorporate experimental AI predictions into its operational forecasting workflow. “We are presenting three different things,” said Ferran Alet, a DeepMind research scientist leading the project, during a press briefing Wednesday. “The first one is a new experimental model tailored specifically for cyclones. The second one is, we’re excited to announce a partnership with the National Hurricane Center that’s allowing expert human forecasters to see our predictions in real time.” The announcement marks a critical juncture in the application of artificial intelligence to weather forecasting, an area where machine learning models have rapidly gained ground against traditional physics-based systems. Tropical cyclones — which include hurricanes, typhoons, and cyclones — have caused $1.4 trillion in economic losses over the past 50 years, making accurate prediction a matter of life and death for millions in vulnerable coastal regions. Why traditional weather models struggle with both storm path and intensity The breakthrough addresses a fundamental limitation in current forecasting methods. Traditional weather models face a stark trade-off: global, low-resolution models excel at predicting where storms will go by capturing vast atmospheric patterns, while regional, high-resolution models better forecast storm intensity by focusing on turbulent processes within the storm’s core. “Making tropical cyclone predictions is hard because we’re trying to predict two different things,” Alet explained. “The first one is track prediction, so where is the cyclone going to go? The second one is intensity prediction, how strong is the cyclone going to get?” DeepMind’s experimental model claims to solve both problems simultaneously. In internal evaluations following National Hurricane Center protocols, the AI system demonstrated substantial improvements over existing methods. For track prediction, the model’s five-day forecasts were on average 140 kilometers closer to actual storm positions than ENS, the leading European physics-based ensemble model. More remarkably, the system outperformed NOAA’s Hurricane Analysis and Forecast System (HAFS) on intensity prediction — an area where AI models have historically struggled. “This is the first AI model that we are now very skillful as well on tropical cyclone intensity,” Alet noted. How AI forecasts beat traditional models on speed and efficiency Beyond accuracy improvements, the AI system demonstrates dramatic efficiency gains. While traditional physics-based models can take hours to generate forecasts, DeepMind’s model produces 15-day predictions in approximately one minute on a single specialized computer chip. “Our probabilistic model is now even faster than the previous one,” Alet said. “Our new model, we estimate, is probably around one minute” compared to the eight minutes required by DeepMind’s previous weather model. This speed advantage allows the system to meet tight operational deadlines. Tom Anderson, a research engineer on DeepMind’s AI weather team, explained that the National Hurricane Center specifically requested forecasts be available within six and a half hours of data collection — a target the AI system now meets ahead of schedule. National Hurricane Center partnership puts AI weather forecasting to the test The partnership with the National Hurricane Center validates AI weather forecasting in a major way. Keith Battaglia, senior director leading DeepMind’s weather team, described the collaboration as evolving from informal conversations to a more official partnership allowing forecasters to integrate AI predictions with traditional methods. “It wasn’t really an official partnership then, it was just sort of more casual conversation,” Battaglia said of the early discussions that began about 18 months ago. “Now we’re sort of working toward a kind of a more official partnership that allow us to hand them the models that we’re building, and then they can decide how to use them in their official guidance.” The timing proves crucial, with the 2025 Atlantic hurricane season already underway. Hurricane center forecasters will see live AI predictions alongside traditional physics-based models and observations, potentially improving forecast accuracy and enabling earlier warnings. Dr. Kate Musgrave, a research scientist at the Cooperative Institute for Research in the Atmosphere at Colorado State University, has been evaluating DeepMind’s model independently. She found it demonstrates “comparable or greater skill than the best operational models for track and intensity,” according to the company. Musgrave stated she’s “looking forward to confirming those results from real-time forecasts during the 2025 hurricane season.” The training data and technical innovations behind the breakthrough The AI model’s effectiveness stems from its training on two distinct datasets: vast reanalysis data reconstructing global weather patterns from millions of observations, and a specialized database containing detailed information about nearly 5,000 observed cyclones from the past 45 years. This dual approach is a departure from previous AI weather models that focused primarily on general atmospheric conditions. “We are training on cyclone specific data,” Alet explained. “We are training on IBTracs and other types of data. So IBTracs provides latitude and longitude and intensity and wind radii for multiple cyclones, up to 5000 cyclones over the last 30 to 40 years.” The system also incorporates recent advances in probabilistic modeling through what DeepMind calls Functional Generative Networks (FGN), detailed in a research paper released alongside the announcement. This approach generates forecast ensembles by learning to perturb the model’s parameters, creating more structured variations than previous methods. Past hurricane predictions show promise for early warning systems Weather Lab launches with over two years of historical predictions, allowing experts to evaluate the model’s performance across all ocean basins. Anderson demonstrated the system’s capabilities using Hurricane Beryl from 2024 and the notorious Hurricane Otis from 2023.

Join the event trusted by enterprise leaders for nearly two decades. VB Transform brings together the people building real enterprise AI strategy. Learn more Generating voices that are not only humanlike and nuanced but diverse continues to be a struggle in conversational AI.  At the end of the day, people want to hear voices that sound like them or are at least natural, not just the 20th-century American broadcast standard.  Startup Rime is tackling this challenge with Arcana text-to-speech (TTS), a new spoken language model that can quickly generate “infinite” new voices of varying genders, ages, demographics and languages just based on a simple text description of intended characteristics.  The model has helped boost customer sales — for the likes of Domino’s and Wingstop — by 15%.  “It’s one thing to have a really high-quality, life-like, real person-sounding model,” Lily Clifford, Rime CEO and co-founder, told VentureBeat. “It’s another to have a model that can not just create one voice, but infinite variability of voices along demographic lines.” A voice model that ‘acts human’  Rime’s multimodal and autoregressive TTS model was trained on natural conversations with real people (as opposed to voice actors). Users simply type in a text prompt description of a voice with desired demographic characteristics and language.  For instance: ‘I want a 30 year old female who lives in California and is into software,’ or ‘Give me an Australian man’s voice.’  “Every time you do that, you’re going to get a different voice,” said Clifford.  Rime’s Mist v2 TTS model was built for high-volume, business-critical applications, allowing enterprises to craft unique voices for their business needs. “The customer hears a voice that allows for a natural, dynamic conversation without needing a human agent,” said Clifford.  For those looking for out-of-the-box options, meanwhile, Rime offers eight flagship speakers with unique characteristics:  Luna (female, chill but excitable, Gen-Z optimist) Celeste (female, warm, laid-back, fun-loving) Orion (male, older, African-American, happy) Ursa (male, 20 years old, encyclopedic knowledge of 2000s emo music) Astra (female, young, wide-eyed) Esther (female, older, Chinese American, loving) Estelle (female, middle-aged, African-American, sounds so sweet) Andromeda (female, young, breathy, yoga vibes) The model has the ability to switch between languages, and can whisper, be sarcastic and even mocking. Arcana can also insert laughter into speech when given the token <laugh>. This can return varied, realistic outputs, from “a small chuckle to a big guffaw,” Rime says. The model can also interpret <chuckle>, <sigh> and even <hum> correctly, even though it wasn’t explicitly trained to do so.  “It infers emotion from context,” Rime writes in a technical paper. “It laughs, sighs, hums, audibly breathes and makes subtle mouth noises. It says ‘um’ and other disfluencies naturally. It has emergent behaviors we are still discovering. In short, it acts human.”  Capturing natural conversations Rime’s model generates audio tokens that are decoded into speech using a codec-based approach, which Rime says provides for “faster-than-real-time synthesis.” At launch, time to first audio was 250 milliseconds and public cloud latency was roughly 400 milliseconds.  Arcana was trained in three stages: Pre-training: Rime used open-source large language models (LLMs) as a backbone and pre-trained on a large group of text-audio pairs to help Arcana learn general linguistic and acoustic patterns. Supervised fine-tuning with a “massive” proprietary dataset.  Speaker-specific fine-tuning: Rime identified the speakers it found “most exemplary” among its dataset, conversations and reliability.  Rime’s data incorporates sociolinguistic conversation techniques (factoring in social context like class, gender, location), idiolect (individual speech habits) and paralinguistic nuances (non-verbal aspects of communication that go along with speech).   The model was also trained on accent subtleties, filler words (those subconscious ‘uhs’ and ‘ums’) as well as pauses, prosodic stress patterns (intonation, timing, stressing of certain syllables) and multilingual code-switching (when multilingual speakers switch back and forth between languages).  The company has taken a unique approach to collecting all this data. Clifford explained that, typically, model builders will gather snippets from voice actors, then create a model to reproduce the characteristics of that person’s voice based on text input. Or, they’ll scrape audiobook data.  “Our approach was very different,” she explained. “It was, ‘How do we create the world’s largest proprietary data set of conversational speech?’”  To do so, Rime built its own recording studio in a basement in San Francisco and spent several months recruiting people off Craigslist, through word-of-mouth, or just causally gathered themselves and friends and family. Rather than scripted conversations, they recorded natural conversations and chitchat.  They then annotated voices with detailed metadata, encoding gender, age, dialect, speech affect and language. This has allowed Rime to achieve 98 to 100% accuracy.  Clifford noted that they are constantly augmenting this dataset.  “How do we get it to sound personal? You’re never going to get there if you’re just using voice actors,” she said. “We did the insanely hard thing of collecting really naturalistic data. The huge secret sauce of Rime is that these aren’t actors. These are real people.” A ‘personalization harness’ that creates bespoke voices Rime intends to give customers the ability to find voices that will work best for their application. They built a “personalization harness” tool to allow users to do A/B testing with various voices. After a given interaction, the API reports back to Rime, which provides an analytics dashboard identifying the best-performing voices based on success metrics.  Of course, customers have different definitions of what constitutes a successful call. In food service, that might be upselling an order of fries or extra wings.  “The goal for us is how do we create an application that makes it easy for our customers to run those experiments themselves?,” said Clifford. “Because our customers aren’t voice casting directors, neither are we. The challenge becomes how to make that personalization analytics layer really intuitive.” Another KPI customers are maximizing for is the caller’s willingness to talk to the AI. They’ve found that, when switching to Rime, callers are 4X more likely to talk to the bot.  “For the first time ever, people are like, ‘No,

Presented by SAP At this year’s SAP Sapphire conference, the company set the stage for a new era in enterprise intelligence. SAP showcased a significant expansion in the capabilities for SAP Business Data Cloud, with intelligent applications that are prebuilt and composable to meet the needs of every business leader. The keynote emphasized a future where organizations use applications fueled with trusted business data to adapt in real time, anticipate customer needs, and strengthen operational efficiency.  As businesses face market volatility and evolving regulatory demands, SAP is making the case that intelligent applications will be the defining factor for competitive success. Here are the key announcements from this year’s event. Transforming outcomes with intelligent applications One of the standout introductions was People Intelligence, an offering grounded in workforce composition, skills, and compensation data products. It provides composition insights, compensation analytics, and skills optimization to help companies refine talent strategies, boost engagement, and ensure compliance. SAP is also rolling out additional intelligent applications that bring real-time insights to different aspects of enterprise operations. These include solutions focused on cloud ERP, customer intelligence, finance, and procurement management. Each application integrates structured business data with AI-driven recommendations, ensuring business leaders can move beyond analysis and into actionable execution. Revolutionizing business data with AI and semantic intelligence Beyond applications, SAP is enhancing its knowledge graph capabilities within Business Data Cloud. This technology enables companies to map business metadata across multiple sources, improving data accuracy, connectivity, and AI-powered insights. When paired with Joule, SAP’s conversational AI tool, businesses gain the ability to interact with complex datasets naturally, making decision-making faster and more intuitive. Expanding SAP’s reach with multi-cloud availability Another major announcement is SAP Business Data Cloud being available on AWS, with Google Cloud and Microsoft Azure expected to follow later this year. This step ensures businesses can integrate SAP’s intelligent applications within their preferred cloud environments. Partnering for a smarter future SAP’s intelligent applications will benefit from an expanding partner ecosystem, with companies collaborating to extend SAP’s capabilities. Accenture is developing intelligent applications focused on supply chain, finance, and procurement challenges. Adobe and SAP are working together on a new solution that synchronizes marketing and operational data for real-time demand forecasting. Palantir is supporting SAP customers in modernizing cloud migration efforts, ensuring seamless data integration. Additional partnerships include Thomson Reuters, which is launching an application for navigating global trade policies, and Moody’s, which is integrating financial risk datasets with SAP accounts receivable data for improved cash flow forecasting. Collibra is enhancing data quality within Business Data Cloud, allowing companies to execute advanced compliance and decision-making workflows more efficiently. Intelligent applications as the future of business strategy SAP Sapphire 2025 reinforced the growing shift toward AI-powered enterprise intelligence. With the expansion of Business Data Cloud, multi-cloud accessibility, and a broad network of partners, SAP is making a case for intelligent applications as the new standard for business strategy. Organizations looking to gain a competitive advantage should explore how these AI-driven tools can integrate into their workflows, assess multi-cloud deployment options, and engage with SAP’s evolving ecosystem of intelligent applications. As AI continues to redefine business, those who embrace AI-powered decision-making will be best positioned to lead in the digital economy. Sponsored articles are content produced by a company that is either paying for the post or has a business relationship with VentureBeat, and they’re always clearly marked. For more information, contact [email protected]. source

Join the event trusted by enterprise leaders for nearly two decades. VB Transform brings together the people building real enterprise AI strategy. Learn more Regular ChatGPT users (including the author of this article) may or may not have noticed that the hit chatbot from OpenAI allows users to enter into a “temporary chat” designed to wipe all the information exchanged between the user and the underlying AI model as soon as the chat session is closed. In addition, OpenAI also allows users to manually delete prior ChatGPT sessions from the left sidebar on the web and desktop/mobile apps by left-clicking or control-clicking, or holding down/long pressing on them from the selector. However, this week, OpenAI found itself facing criticism from some ChatGPT users after they discovered that the company has not actually been deleting these chat logs as previously indicated. “You’re telling me my deleted chatgpt chats are actually not deleted and [are] being saved to be investigated by a judge?” posted X user @ns123abc. The comment drew more than a million views. Another user, @kepano, added, “you can ‘delete’ a ChatGPT chat, however all chats must be retained due to legal obligations ?”. As AI influencer and software engineer Simon Willison wrote on his personal blog: “Paying customers of [OpenAI’s] APIs may well make the decision to switch to other providers who can offer retention policies that aren’t subverted by this court order!” Instead, OpenAI confirmed it has been preserving deleted and temporary user chat logs since mid-May 2025 in response to a federal court order, although it did not disclose this to users until June 5th. The order, embedded below and issued on May 13, 2025, by U.S. Magistrate Judge Ona T. Wang, requires OpenAI to “preserve and segregate all output log data that would otherwise be deleted on a going forward basis,” including chats deleted by user request or due to privacy obligations. The court’s directive stems from The New York Times (NYT) v. OpenAI and Microsoft, a now year-and-a-half old copyright case still being argued. The NYT’s lawyers allege that OpenAI’s language models regurgitate copyrighted news content verbatim. The plaintiffs argue that logs, including those that users may have deleted, could contain infringing outputs relevant to the lawsuit. While OpenAI complied with the order immediately, it did not publicly notify affected users for more than three weeks, issuing a blog post and FAQ describing the legal mandate and outlining who is impacted. However, OpenAI is placing the blame squarely on the NYT and the judge’s order, saying it believes the preservation demand to be “baseless.” OpenAI clarifies what’s going on with the court order to preserve ChatGPT user logs — including which chats are impacted In a blog post published yesterday, OpenAI COO Brad Lightcap defended the company’s position and stated that it was advocating for user privacy and security against an over-broad judicial order, writing: “The New York Times and other plaintiffs have made a sweeping and unnecessary demand in their baseless lawsuit against us: retain consumer ChatGPT and API customer data indefinitely. This fundamentally conflicts with the privacy commitments we have made to our users.” The post clarified that ChatGPT Free, Plus, Pro and Team users, along with API customers without a zero data retention (ZDR) agreement, are affected by the preservation order, meaning even if users on these plans delete their chats or use temporary chat mode, their chats will be stored for the foreseeable future. However, subscribers to the ChatGPT Enterprise and Edu users, as well as API clients using ZDR endpoints, are not impacted by the order and their chats will be deleted as directed. The retained data is held under legal hold, meaning it is stored in a secure, segregated system and only accessible to a small number of legal and security personnel. “This data is not automatically shared with The New York Times or anyone else,” Lightcap emphasized in OpenAI’s blog post. Sam Altman floats new concept of ‘AI privilege’ allowing for confidential conversations between models and users, similar to speaking to a human doctor or lawyer OpenAI CEO and co-founder Sam Altman also addressed the issue publicly in a post from his account on the social network X last night, writing: “recently the NYT asked a court to force us to not delete any user chats. we think this was an inappropriate request that sets a bad precedent. we are appealing the decision. we will fight any demand that compromises our users’ privacy; this is a core principle.” He also suggested a broader legal and ethical framework may be needed for AI privacy: “we have been thinking recently about the need for something like ‘AI privilege’; this really accelerates the need to have the conversation.” “imo talking to an AI should be like talking to a lawyer or a doctor.” “i hope society will figure this out soon.“ The notion of AI privilege — as a potential legal standard — echoes attorney-client and doctor-patient confidentiality. Whether such a framework would gain traction in courtrooms or policy circles remains to be seen, but Altman’s remarks indicate OpenAI may increasingly advocate for such a shift. What comes next for OpenAI and your temporary/deleted chats? OpenAI has filed a formal objection to the court’s order, requesting that it be vacated. In court filings, the company argues that the demand lacks a factual basis and that preserving billions of additional data points is neither necessary nor proportionate. Judge Wang, in a May 27 hearing, indicated that the order is temporary. She instructed the parties to develop a sampling plan to test whether deleted user data materially differs from retained logs. OpenAI was ordered to submit that proposal by today (June 6) but I have yet to see the filing. What it means for enterprises and decision-makers in charge of ChatGPT usage in corporate environments While the order exempts ChatGPT Enterprise and API customers using ZDR endpoints, the broader legal and reputational implications matter deeply for professionals responsible for deploying and scaling AI

Join the event trusted by enterprise leaders for nearly two decades. VB Transform brings together the people building real enterprise AI strategy. Learn more Zencoder, the artificial intelligence coding startup founded by serial entrepreneur Andrew Filev, announced today the public beta launch of Zentester, an AI-powered agent designed to automate end-to-end software testing. This critical but often sluggish step can delay product releases by days or weeks. The new tool represents Zencoder’s latest attempt to distinguish itself in the increasingly crowded AI coding assistant market, where companies are racing to automate not just code generation but entire software development workflows. Unlike existing AI coding tools that focus primarily on writing code, Zentester targets the verification phase — ensuring software works as intended before it reaches customers. “Verification is the missing link in scaling AI-driven development from experimentation to production,” said Filev in an exclusive interview with VentureBeat. The CEO, who previously founded project management company Wrike and sold it to Citrix for $2.25 billion in 2021, added: “Zentester doesn’t just generate tests—it gives developers the confidence to ship by validating that their AI-generated or human-written code does what it’s supposed to do.” The announcement comes as the AI coding market undergoes rapid consolidation. Last month, Zencoder acquired Machinet, another AI coding assistant with over 100,000 downloads. At the same time, OpenAI reached an agreement to acquire coding tool Windsurf for approximately $3 billion (the deal was completed in May). The moves underscore how companies are rushing to build comprehensive AI development platforms rather than point solutions. Why software testing has become the biggest roadblock in AI-powered development Zentester addresses a persistent challenge in software development: the lengthy feedback loops between developers and quality assurance teams. In typical enterprise environments, developers write code and send it to QA teams for testing, often waiting several days for feedback. By then, developers have moved on to other projects, creating costly context switching when issues are discovered. “In a typical engineering process, after a developer builds a feature and sends it to QA, they receive feedback several days later,” Filev told VentureBeat. “By then, they’ve already moved on to something else. This context switching and back-and-forth—especially painful during release crunches—can stretch simple fixes into week-long ordeals.” Early customer Club Solutions Group reported dramatic improvements, with CEO Mike Cervino stating, “What took our QA team a couple of days now takes developers 2 hours.” The timing is particularly relevant as AI coding tools generate increasingly large volumes of code. While tools like GitHub Copilot and Cursor have accelerated code generation, they have also created new quality assurance challenges. Filev estimates that if AI tools increase code generation by 10x, testing requirements will similarly increase by 10x — overwhelming traditional QA processes. How Zentester’s AI agents click buttons and fill forms like human testers Unlike traditional testing frameworks that require developers to write complex scripts, Zentester operates on plain English instructions. The AI agent can interact with applications like a human user—clicking buttons, filling forms, and navigating through software workflows—while validating both frontend user interfaces and backend functionality. The system integrates with existing testing frameworks, including Playwright and Selenium, rather than replacing them entirely. “We absolutely do not like people abandoning stuff that’s part of our DNA,” Filev said. “We feel that AI should leverage the processes and tools that already exist in industry.” Zentester offers five core capabilities: developer-led quality testing during feature development, QA acceleration for comprehensive test suite creation, quality improvement for AI-generated code, automated test maintenance, and autonomous verification in continuous integration pipelines. The tool represents the latest addition to Zencoder’s broader multi-agent platform, which includes coding agents for generating software and unit testing agents for basic verification. The company’s “Repo Grokking” technology analyzes entire code repositories to provide context, while an error-correction pipeline aims to reduce AI-generated bugs. The launch intensifies competition in the AI development tools market, where established players like Microsoft’s GitHub Copilot and newer entrants like Cursor are vying for developer mindshare. Zencoder’s approach of building specialized agents for different development phases contrasts with competitors focused primarily on code generation. “At this point, there are three strong coordination products in the market that are production grade: it’s us, Cursor, and Windsurf,” Filev said in a recent interview. “For smaller companies, it’s becoming harder and harder to compete.” The company claims superior performance on industry benchmarks, reporting 63% success rates on SWE-Bench Verified tests and approximately 30% on the newer SWE-Bench Multimodal benchmark — results Filev says double previous best performances. Industry analysts note that end-to-end testing automation represents a logical next step for AI coding tools, but successful implementation requires a sophisticated understanding of application logic and user workflows. What enterprise buyers need to know before adopting AI testing platforms Zencoder’s approach offers both opportunities and challenges for enterprise customers evaluating AI testing tools. The company’s SOC 2 Type II, ISO 27001 and ISO 42001 certifications address security and compliance concerns critical for large organizations. However, Filev acknowledges that enterprise caution is warranted. “For enterprises, we don’t advocate changing software development lifecycles completely, yet,” he said. “What we advocate is AI-augmented, where now they can have quick AI code review and acceptance testing that reduces the amount of work that needs to be done by the next party in the pipeline.” The company’s integration strategy — working within existing development environments like Visual Studio Code and JetBrains IDEs rather than requiring platform switches — may appeal to enterprises with established toolchains. The race to automate software development from idea to deployment Zentester’s launch positions Zencoder to compete for a larger share of the software development workflow as AI tools expand beyond simple code generation. The company’s vision extends to full automation from requirements to production deployment, though Filev acknowledges current limitations. “The next jump is going to be requirements to production — the whole thing,” Filev said. “Can you now pipe it so that you could have natural language requirements and then AI could help you break it

Join the event trusted by enterprise leaders for nearly two decades. VB Transform brings together the people building real enterprise AI strategy. Learn more Days after OpenAI and Google Cloud announced a partnership to support the growing use of generative AI platforms, much of the AI-powered web and tools went down due to an outage of the leading cloud providers. Google Cloud Service Platform (GCP) and some Cloudflare services began experiencing issues around 10:00 a.m. PT today, affecting several AI development tools and data storage services, including ChatGPT and Claude, as well as a variety of other AI platforms.  We are aware of a service disruption to some Google Cloud services and we are working hard to get you back up and running ASAP. Please view our status dashboard for the latest updates: https://t.co/sT6UxoRK4R — Google Cloud (@googlecloud) June 12, 2025 A GCP spokesperson confirmed the outage to VentureBeat, urging users to check its public status dashboard.  GCP said affected services include API Gateway, Agent Assist, Cloud Data Fusion, Contact Center AI Platform, Google App Engine, Google BigQuery, Google Cloud Storage, Identity Platform, Speech-to-Text, Text-to-Speech and Vertex AI Search, among other tools. Google’s mobile development platform, Firebase, also went down. VentureBeat staffers had trouble accessing Google Meet, but other Google services on Workspace remained online.  A Cloudflare spokesperson told VentureBeat only “a limited number of services at Cloudflare use Google Cloud and were impacted. We expect them to come back shortly. The core Cloudflare services were not impacted.” Despite media reports and user-provided feedback on Down Detector, AWS stated that its service remains up, including AI platforms such as Bedrock and Sagemaker.  OpenAI acknowledged some users had issues logging into their platforms but have since resolved the problem. Anthropic noted on its status page that Claude experienced “elevated error rates on the API, console and Claude AI.” We are aware of issues affecting multiple external internet providers, impacting the availability of our services such as single sign-on (SSO) and other log-in methods. Our engineering teams are working to mitigate these issues. Thank you for your continued patience. For the… — OpenAI (@OpenAI) June 12, 2025 Developer tools like LlamaIndex’s LlamaCloud, Weights & Biases, Windsurf, Supabase and Replit reported issues. Other platforms like Character AI also announced they were affected.  Hi folks – LlamaCloud (https://t.co/DHMd6BFO0l) is currently down due to the ongoing global AWS/GCP/Firebase outage. We are closely monitoring the solution and will keep you posted when it’s resolved! — Jerry Liu (@jerryjliu0) June 12, 2025 ? We’re aware of the Google Cloud outage affecting various web services, including Weights & Biases products like W&B Models and @weave_wb. Our team is monitoring the situation and will provide updates. Thank you for your patience. — Weights & Biases (@weights_biases) June 12, 2025 Our upstream cloud providers are currently experiencing a major outage. We are working as best we can to restore Replit services. — Replit ⠕ (@Replit) June 12, 2025 In addition to AI tools, other websites and internet services, such as Spotify and Discord, also reportedly went down.  Needing more cloud  In many ways, the outage highlights the challenges of relying on a single cloud service or database provider and the risks associated with an interconnected Internet. If one of your cloud services goes down, it could impact some users whose log-in or data stream is hosted there.  Google Cloud has been gradually wresting market leadership in enterprise AI from its competitors, thanks to the large number of developer and database tools it has begun offering organizations. On the other hand, Cloudflare has been partnering with companies like Hugging Face to deploy AI apps faster.  First reported by Reuters, Google and OpenAI have struck a deal that will allow OpenAI to utilize Google Cloud to meet the growing demand on its platform.  But that’s not to say Google or Cloudflare may lose an edge among enterprise AI users who depend on consistent uptime. While the company continues to investigate the cause of the outage, enterprises often have, and should have, redundancies in case their provider goes down. Outages happen, and they happen far too frequently. The last massive outage happened around the same time last year, in July, when CrowdStrike accidentally triggered outages that impacted Microsoft Windows users. In typical fashion, many people saw the outages as an opportunity for comedy, or at least to catch up on tasks they’d been putting off.  much of the AI internet is down now firebase is down, cursor is down, lovable is down, supabase is down, google ai is down, cursor is down, aws is down… almost everything is down. finally time to catch up on the 87 tools, 14 models, and 12 AI startup ideas we want to build. — GREG ISENBERG (@gregisenberg) June 12, 2025 So yes, it seems like the digital universe is giving everyone a forced break today! — Murugan (MGA) (@murugan_mga) June 12, 2025 source

Join the event trusted by enterprise leaders for nearly two decades. VB Transform brings together the people building real enterprise AI strategy. Learn more Zip, the $2.2 billion procurement platform startup, unveiled a suite of 50 specialized artificial intelligence agents on Tuesday designed to automate the tedious manual work that plagues enterprise purchasing departments worldwide, marking what industry analysts call the most significant advancement in procurement technology in decades. The AI agents, announced at Zip’s inaugural AI Summit in Brooklyn, can autonomously handle complex tasks ranging from contract reviews and tariff assessments to regulatory compliance checks — work that currently consumes millions of hours across corporate America. Early adopters including OpenAI, Canva, and Webflow are already testing the technology, which Zip says represents a fundamental shift from AI-assisted workflows to fully autonomous task completion. “Today Zip is cutting through the agentic AI hype with AI agents that actually work,” said Rujul Zaparde, Zip’s co-founder and CEO, in an exclusive interview with VentureBeat. “Not vague chatbots. Not generic assistants. Real, specialized AI agents that do one job and do it perfectly.” The announcement comes as enterprises increasingly struggle with procurement bottlenecks that can involve 30 or more approval steps for major purchases, particularly in heavily regulated industries like financial services. Procurement represents the second-largest corporate expense category after payroll, yet remains largely managed through manual, error-prone processes that leave trillions of dollars inefficiently managed. How AI agents tackle the 30-step procurement approval process Zip’s approach centers on what the company calls “agentic procurement orchestration” — embedding specialized AI agents directly into existing procurement workflows rather than requiring employees to adopt separate AI tools. The system addresses a critical enterprise challenge: while companies have top-down mandates to adopt AI, most employees don’t know how to effectively integrate tools like ChatGPT into their daily procurement tasks. “The unique insight we’ve had is that the technology actually is good enough to solve very specific tasks,” explained Lu Cheng, Zip’s co-founder and chief technology officer, in an interview with VentureBeat. “It’s effectively a junior level employee that’s very good at following specific instructions.” The agents tackle diverse procurement pain points with surgical precision. A tariff analysis agent dynamically assesses how global trade policies affect vendor pricing, while a GDPR compliance agent flags potential privacy risks in vendor documents. An intake validation agent can spot discrepancies in purchase requests — for instance, catching when an employee claims a software purchase won’t involve customer data sharing while the vendor’s documentation indicates otherwise. One enterprise customer processing 1,410 procurement requests in their first month with Zip would traditionally require human review of every request’s pricing, categorization, and compliance details. With Zip’s agents, that work happens automatically. From 4.6 million AI insights to $4.4 billion in enterprise savings Zip’s aggressive push into AI automation builds on substantial momentum. The company has already delivered over 4.6 million AI insights to customers and helped enterprises save $4.4 billion in trackable procurement costs since its 2020 founding. In 2024 alone, Zip processed 14 million reviews across its customer base — work that previously required human analysts to manually examine contracts, security documentation, and compliance materials. “We had a customer just went live — an 8,000 person, well-regarded tech company — in their first month they processed 1,410 requests,” Zaparde said. “The first step for all 1,410 requests was someone in procurement checking if the price was correct, if the categories aligned. With this agent, they basically don’t have to do that 1,410 times.” The company has set an ambitious goal: within five years, as Zip processes over one billion reviews annually, 90% should be handled entirely by AI agents. That scale of automation could reshape how enterprises manage supplier relationships and spending decisions. Why Zip’s data access gives it an edge over Coupa Zip’s agents gain their effectiveness through privileged access to comprehensive enterprise data that competitors cannot easily replicate. As the orchestration layer connecting finance, legal, procurement, IT, and security teams, Zip already integrates with an average of seven enterprise systems per customer — including contract management platforms, risk assessment tools, and ERP financial systems. “We have a really deep understanding of what a legal review, what a security review actually constitutes because we literally have the documents that they’re reviewing thousands or hundreds of thousands of times across our customer base,” Zaparde explained. This data advantage allows Zip agents to access contract renewal dates, payment histories, vendor relationships spanning decades, and real-time regulatory changes — context that isolated enterprise systems cannot provide. The company built its agents using a no-code platform that enterprise customers can customize for their specific needs. Configuration typically takes two to four hours per agent, though complex implementations can require up to 20 hours for customers with intricate approval processes. OpenAI and Canva lead early adoption of automated procurement OpenAI, which has partnered closely with Zip through the startup’s AI Lab initiative, exemplifies the early adoption trend. “We’ve worked closely with the Zip team to power their agentic platform and it’s been really exciting to see how quickly they’ve turned real-world procurement pain points into focused AI task agents,” said Kathryn Devlin, Head of Procure-to-Pay Operations, Travel and Expense at OpenAI. The collaboration reflects a broader enterprise imperative: as companies face mounting pressure to optimize spending and control costs, procurement automation has become strategically critical. Research firm IDC projects the global procurement software market will grow from $8.03 billion in 2024 to $18.28 billion by 2032, with AI-powered solutions driving much of that expansion. Wiz Technology Procurement leader Idan Cohen highlighted the strategic shift AI enables: “We’ll save so much time on the technical work and day-to-day tasks that we need to do as part of the procurement process, and be enabled to really focus on what we’re supposed to do — being a true partner to the business and to our vendors.” Building enterprise trust with citations and human oversight Zip has designed its agent architecture to address enterprise concerns about AI accuracy and data

Join the event trusted by enterprise leaders for nearly two decades. VB Transform brings together the people building real enterprise AI strategy. Learn more Mistral AI, the French artificial intelligence startup, announced Wednesday a sweeping expansion into AI infrastructure that positions the company as Europe’s answer to American cloud computing giants, while simultaneously unveiling new reasoning models that rival OpenAI’s most advanced systems. The Paris-based company revealed Mistral Compute, a comprehensive AI infrastructure platform built in partnership with Nvidia, designed to give European enterprises and governments an alternative to relying on U.S.-based cloud providers like Amazon Web Services, Microsoft Azure, and Google Cloud. The move represents a significant strategic shift for Mistral from purely developing AI models to controlling the entire technology stack. “This move into AI infrastructure marks a transformative step for Mistral AI, as it allows us to address a critical vertical of the AI value chain,” said Arthur Mensch, CEO and co-founder of Mistral AI. “With this shift comes the responsibility to ensure that our solutions not only drive innovation and AI adoption, but also uphold Europe’s technological autonomy and contribute to its sustainability leadership.” How Mistral built reasoning models that think in any language Alongside the infrastructure announcement, Mistral unveiled its Magistral series of reasoning models — AI systems capable of step-by-step logical thinking similar to OpenAI’s o1 model and China’s DeepSeek R1. But Guillaume Lample, Mistral’s chief scientist, says the company’s approach differs from competitors in crucial ways. “We did everything from scratch, basically because we wanted to learn the expertise we have, like, flexibility in what we do,” Lample told me in an exclusive interview. “We actually managed to be, like, a really, very efficient on the stronger online reinforcement learning pipeline.” Unlike competitors that often hide their reasoning processes, Mistral’s models display their full chain of thought to users — and crucially, in the user’s native language rather than defaulting to English. “Here we have like the full chain of thought which is given to the user, but in their own language, so they can actually read through it, see if it makes sense,” Lample explained. The company released two versions: Magistral Small, a 24-billion parameter open-source model, and Magistral Medium, a more powerful proprietary system available through Mistral’s API. Why Mistral’s AI models gained unexpected superpowers during training The models demonstrated surprising capabilities that emerged during training. Most notably, Magistral Medium retained multimodal reasoning abilities — the capacity to analyze images — even though the training process focused solely on text-based mathematical and coding problems. “Something we realized, not exactly by mistake, but something we absolutely did not expect, is that if at the end of the reinforcement learning training, you plug back the initial vision encoder, then you suddenly, kind of out of nowhere, see the model being able to do reasoning over images,” Lample said. The models also gained sophisticated function-calling abilities, automatically performing multi-step internet searches and code execution to answer complex queries. “What you will see is a model doing this, thinking, then realizing, okay, this information might be updated. Let me do like a web search,” Lample explained. “It will search on like internet, and then it will actually pass the results, and it will result over it, and it will say, maybe, maybe the answer is not in this results. Let me search again.” This behavior emerged naturally without specific training. “It’s something that whether or not on things to do next, but we found that it’s actually happening kind of naturally. So it was a very nice surprise for us,” Lample noted. The engineering breakthrough that makes Mistral’s training faster than competitors Mistral’s technical team overcame significant engineering challenges to create what Lample describes as a breakthrough in training infrastructure. The company developed a system for “online reinforcement learning” that allows AI models to continuously improve while generating responses, rather than relying on pre-existing training data. The key innovation involved synchronizing model updates across hundreds of graphics processing units (GPUs) in real-time. “What we did is that we found a way to just unscrew the model through GPUs. I mean, from GPU to GPU,” Lample explained. This allows the system to update model weights across different GPU clusters within seconds rather than the hours typically required. “There is no like open source infrastructure that will do this properly,” Lample noted. “Typically, there are a lot of like open source attempts to do this, but it’s extremely slow. Here, we focused a lot on the efficiency.” The training process proved much faster and cheaper than traditional pre-training. “It was much cheaper than regular pre training. Pre training is something that would take weeks or months on other GPUs. Here, we are nowhere close to this. It was like, I depend on how many people we put on this. But it was more like, it was like, fairly less than one week,” Lample said. Nvidia commits 18,000 chips to European AI independence The Mistral Compute platform will run on 18,000 of Nvidia’s newest Grace Blackwell chips, housed initially in a data center in Essonne, France, with plans for expansion across Europe. Nvidia CEO Jensen Huang described the partnership as crucial for European technological independence. “Every country should build AI for their own nation, in their nation,” Huang said at a joint announcement in Paris. “With Mistral AI, we are developing models and AI factories that serve as sovereign platforms for enterprises across Europe to scale intelligence across industries.” Huang projected that Europe’s AI computing capacity would increase tenfold over the next two years, with more than 20 “AI factories” planned across the continent. Several of these facilities will have more than a gigawatt of capacity, potentially ranking among the world’s largest data centers. The partnership extends beyond infrastructure to include Nvidia’s work with other European AI companies and Perplexity, the search company, to develop reasoning models in various European languages where training data is often limited. How Mistral plans to solve AI’s environmental and sovereignty problems Mistral

Join the event trusted by enterprise leaders for nearly two decades. VB Transform brings together the people building real enterprise AI strategy. Learn more Major AI providers like OpenAI, Google, xAI and others have all launched various AI agents that conduct exhaustive or “deep” research across the web on behalf of users, spending minutes at a time to compile extensively cited white papers and reports that, in their best case versions, are ready to be circulated to colleagues, customers and business partners without any human editing or reworking. But they all have a significant limitation out-of-the-box: they are only able to search the web and the many public facing websites on it — not any of the enterprise customer’s internal databases and knowledge graphs. Unless, of course, the enterprise or their consultants take the time to build a retrieval augmented generation (RAG) pipeline using something like OpenAI’s Responses API, but this would require a fair bit of time, expense, and developer expertise to set up. But now AlphaSense, an early AI platform for market intelligence, is trying to do enterprises — particularly those in financial services and large enterprises (it counts 85% of the S&P 100 as its customers) — one better. Today the company announced its own “Deep Research,” an autonomous AI agent designed to automate complex research workflows that extends across the web, AlphaSense’s catalog of continuously updated, non-public proprietary data sources such as Goldman Sachs and Morgan Stanley research reports, and the enterprise customers’ own data (whatever they hook the platform up to, it’s their choice). Now available to all AlphaSense users, the tool helps generate detailed analytical outputs in a fraction of the time traditional methods require. “Deep Research is our first autonomous agent that conducts research in the platform on behalf of the user—reducing tasks that once took days or weeks to just minutes,” said Chris Ackerson, Senior Vice President of Product at AlphaSense, in an exclusive interview with VentureBeat. Underlying model architecture and performance optimization To power its AI tools — including Deep Research — AlphaSense relies on a flexible architecture built around a dynamic suite of large language models. Rather than committing to a single provider, the company selects models based on performance benchmarks, use case fit, and ongoing developments in the LLM ecosystem. Currently, AlphaSense draws on three primary model families: Anthropic, accessed via AWS Bedrock, for advanced reasoning and agentic workflows; Google Gemini, valued for its balanced performance and ability to handle long-context prompts; and Meta’s Llama models, integrated through a partnership with AI hardware startup Cerebras. Through that collaboration, AlphaSense uses Cerebras Inference running on WSE-3 (Wafer-Scale Engine) hardware, optimizing inference speed and efficiency for high-volume tasks. This multi-model strategy enables the platform to deliver consistently high-quality outputs across a range of complex research scenarios. New AI agent aims to replicate the work of a skilled analyst team with speed and high accuracy Ackerson emphasized the tool’s unique combination of speed, depth, and transparency. “To reduce hallucinations, we ground every AI-generated insight in source content, and users can trace any output directly to the exact sentence in the original document,” he said. This granular traceability is aimed at building trust among business users, many of whom rely on AlphaSense for high-stakes decisions in volatile markets. Every report generated by Deep Research includes clickable citations to underlying content, enabling both verification and deeper follow-up. Building on a decade of AI development AlphaSense’s launch of Deep Research marks the latest step in a multi-year evolution of its AI offerings. “From the founding of the company, we’ve been leveraging AI to support financial and corporate professionals in the research process, starting with better search to eliminate blind spots and control-F nightmares,” Ackerson said. He described the company’s path as one of continuous improvement: “As AI improved, we moved from basic information discovery to true analysis—automating more of the workflow, always directed by the user.” AlphaSense has introduced several AI tools over the past few years. “We’ve launched tools like Generative Search for fast Q&A across all AlphaSense content, Generative Grid to analyze documents side by side, and now Deep Research for long-form synthesis across hundreds of documents,” he added. Use cases: from M&A analysis to executive briefings Deep Research is designed to support a range of high-value workflows. These include generating company and industry primers, screening for M&A opportunities, and preparing detailed board or client briefings. Users can issue natural language prompts, and the agent returns tailored outputs complete with supporting rationale and source links. Proprietary data and internal integration set it apart One of AlphaSense’s primary advantages lies in its proprietary content library. “AlphaSense aggregates over 500 million premium and proprietary documents, including exclusive content like sell-side research and expert call interviews—data you can’t find on the public web,” Ackerson explained. The platform also supports integration of clients’ internal documentation, creating a blended research environment. “We allow customers to integrate their own institutional knowledge into AlphaSense, making internal data more powerful when combined with our premium content,” he said. This means firms can feed internal reports, slide decks, or notes into the system and have them analyzed alongside external market data for deeper contextual understanding. Commitment to continuous information updates and a security focus All data sources in AlphaSense are continuously updated. “All of our content sets are growing—hundreds of thousands of documents added daily, thousands of expert calls every month, and continuous licensing of new high-value sources,” Ackerson said. AlphaSense also places significant emphasis on enterprise security. “We’ve built a secure, enterprise-grade system that meets the requirements of the most regulated firms. Clients retain control of their data, with full encryption and permissions management,” Ackerson noted. Deployment options are designed to be flexible. “We offer both multi-tenant and single-tenant deployments, including a private cloud option where the software runs entirely within the client’s infrastructure,” he said. Growing precision, custom enterprise AI demand The launch of Deep Research responds to a broader enterprise trend toward intelligent automation. According to a Gartner prediction

Join the event trusted by enterprise leaders for nearly two decades. VB Transform brings together the people building real enterprise AI strategy. Learn more European AI powerhouse Mistral today launched Magistral, a new family of large language models (LLMs) that marks the first from the company to enter the increasingly competitive space of “reasoning,” or models that take time to reflect on their thinking to catch errors and solve more complex tasks than basic text-based LLMs. The announcement features a strategic dual release: a powerful, proprietary Magistral Medium for enterprise clients, and, notably, a 24-billion parameter open-source version, Magistral Small. The latter release appears calculated to reinforce the company’s commitment to its foundational roots, following a period where it faced criticism for leaning into more closed, proprietary models such as its Medium 3 for enterprises, launched back in May 2025. A return to open source roots In a move that will undoubtedly be celebrated by developers and the wider AI community, Mistral is releasing Magistral Small under the permissive open source Apache 2.0 license. This is a crucial detail. Unlike more restrictive licenses, Apache 2.0 allows anyone to freely use, modify, and distribute the model’s source code, even for commercial purposes. This empowers startups and established companies alike to build and deploy their own applications on top of Mistral’s latest reasoning architecture without licensing fees or fear of vendor lock-in. This open approach is particularly significant given the context. While Mistral built its reputation on powerful open models, its recent release of Medium 3 as a purely proprietary offering drew concern from some quarters of the open-source community, who worried the company was drifting towards a more closed ecosystem, similar to competitors like OpenAI. The release of Magistral Small under such a permissive license serves as a powerful counter-narrative, reaffirming Mistral’s dedication to arming the open community with cutting-edge tools. Competitive performance against formidable foes Mistral isn’t just talking a big game; it came with receipts. The company released a suite of benchmarks pitting Magistral-Medium against its own predecessor, Mistral-Medium 3, and competitors from Deepseek. The results show a model that is fiercely competitive in the reasoning arena. On the AIME-24 mathematics benchmark, Magistral-Medium scores an impressive 73.6% on accuracy, neck-and-neck with its predecessor and significantly outperforming Deepseek’s models. When using majority voting (a technique where the model generates multiple answers and the most common one is chosen), its performance on AIME-24 jumps to a staggering 90%. The new model also holds its own across other demanding tests, including GPQA Diamond, a graduate-level question-answering benchmark, and LiveCodeBench for coding challenges. While Deepseek-V3 shows strong performance on some benchmarks, Magistral-Medium consistently proves itself to be a top-tier reasoning model, validating Mistral’s claims of its advanced capabilities. Enterprise power While Magistral Small caters to the open-source world, the benchmark-validated Magistral Medium is aimed squarely at the enterprise. It’s acessible via Mistral’s Le Chat interface and La Plateforme API, it delivers the top-tier performance needed for mission-critical tasks. Mistral is making this model available on major cloud platforms, including Amazon SageMaker, with Azure AI, IBM WatsonX, and Google Cloud Marketplace to follow. This dual-release strategy allows Mistral to have its cake and eat it too: fostering a vibrant ecosystem around its open models while monetizing its most powerful, performance-tested technology for corporate clients. Cost comparison When it comes to cost, Mistral is positioning Magistral Medium as a distinct, premium offering, even compared to its own models. At $2 per million input tokens and $5 per million output tokens, it represents a significant price increase from the older Mistral Medium 3, which costs just $0.40 for input and $2 for output. However, when placed against its external rivals, Magistral Medium’s pricing strategy appears highly aggressive. Its input cost matches that of OpenAI’s latest model and sits within the range of Gemini 2.5 Pro, yet its $5 output price substantially undercuts both, which are priced at $8 and upwards of $10, respectively. Magistral API cost compared to other leading LLM reasoners. Credit: VentureBeat made with Google Gemini 2.5 Pro (Preview) While it is considerably more expensive than specialized models like DeepSeek-Reasoner, it is an order of magnitude cheaper than Anthropic’s flagship Claude Opus 4, making it a compelling value proposition for customers seeking state-of-the-art reasoning without paying the absolute highest market prices. Reasoning you can view, understand and use Mistral is pushing three core advantages with the Magistral line: transparency, multilingualism, and speed. Breaking away from the “black box” nature of many AI models, Magistral is designed to produce a traceable “chain-of-thought.” This allows users to follow the model’s logical path, a critical feature for high-stakes professional fields like law, finance, and healthcare, where conclusions must be verifiable. Furthermore, these reasoning capabilities are global. Mistral emphasizes the model’s “multilingual dexterity,” highlighting high-fidelity performance in languages including French, Spanish, German, Italian, Arabic, Russian, and Simplified Chinese. On the performance front, the company claims a major speed boost. A new “Think mode” and “Flash Answers” feature in Le Chat reportedly enables Magistral Medium to achieve up to 10 times the token throughput of competitors, facilitating real-time reasoning at a scale previously unseen. From code gen to creative strategy and beyond The applications for Magistral are vast. Mistral is targeting any use case that demands precision and structured thought, from financial modeling and legal analysis to software architecture and data engineering. The company even showcased the model’s ability to generate a one-shot physics simulation, demonstrating its grasp of complex systems. But it’s not all business. Mistral also recommends the model as a “creative companion” for writing and storytelling, capable of producing work that is either highly coherent or, as the company puts it, “delightfully eccentric.” With Magistral, Mistral AI is making a strategic play to not just compete, but lead in the next frontier of AI. By re-engaging its open-source base with a powerful, permissively licensed model while simultaneously pushing the envelope on enterprise-grade performance, the company is signaling that the future of reasoning AI will be both