VentureBeat

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More The Trump administration walked back an Executive Order from former President Joe Biden that created rules around the development and deployment of AI. Since then, the government has stepped back from regulating the technology.  In a more than three-hour hearing at the Senate Committee on Commerce, Science and Transportation, executives like OpenAI CEO Sam Altman, AMD CEO Lisa Su, Coreweave co-founder and CEO Michael Intrator and Microsoft Vice Chair and President Brad Smith urged policymakers to ease the process of building infrastructure around AI development. The executives told policymakers that speeding up permitting could make building new data centers, power plants to energize data centers and even chip fabricators crucial in shoring up the AI Tech Stack and keeping the country competitive against China. They also spoke about the need for more skilled workers like electricians, easing software talent immigration and encouraging “AI diffusion” or the adoption of generative AI models in the U.S. and worldwide.  Altman, fresh from visiting the company’s $500 billion Stargate project in Texas, told senators that the U.S. is leading the charge in AI, but it needs more infrastructure like power plants to fuel its next phase. “I believe the next decade will be about abundant intelligence and abundant energy. Making sure that America leads in both of those, that we are able to usher in these dual revolutions that will change the world we live in incredibly positive ways is critical,” Altman said.  The hearing came as the Trump administration is determining how much influence the government will have in the AI space. Sen. Ted Cruz of Texas, chair of the committee, said he proposed creating an AI regulatory sandbox.  Microsoft’s Smith said in his written testimony that American AI companies need to continue innovating because ” it is a race that no company or country can win by itself.” Supporting the AI tech stack Microsoft’s Smith laid out the AI Tech Stack, which he said shows how important each segment of the sector is to innovation. “We’re all in this together. If the United States is gonna succeed in leading the world in AI, it requires infrastructure, it requires success at the platform level, it requires people who create applications,” Smith said.  He added, “Innovation will go faster with more infrastructure, faster permitting and more electricians.”  AMD’s Su reiterated that “maintaining our lead actually requires excellence at every layer of the stack.”  “I think open ecosystems are really a cornerstone of U.S. leadership, and that allows ideas to come from everywhere and every part of the innovation sector,” Su said. “It’s reducing barriers to entry and strengthening security as well as creating a competitive marketplace for ideas.” With AI models needing more and more GPUs for training, the need to improve the production of chips, build more data centers, and find ways to power them has become even more critical. The Chips and Science Act, a Biden-era law, was meant to jumpstart semiconductor production in the U.S., but making the needed chips to power the world’s most powerful models locally is proving to be slow and expensive.  In recent months, companies like Cerebras have announced plans to build more data centers to help process model training and inference.  A break from current policies The Senate majority of Republican policymakers made it clear during the hearing that the Trump administration would prefer not to regulate AI development, preferring a more market-driven, hands-off approach. This administration has also pushed for more U.S.-focused growth, demanding businesses use American products and create more American jobs.  However, the executives noted that for American AI to remain competitive, companies need access to international talent and, more importantly, clear export policies so models made in the U.S. can be attractive to other countries.  “We need faster adoption, what people refer to as AI diffusion. The ability to put AI to work across every part of the American economy to boost productivity, to boost economic growth, to enable people to innovate in their work,” Smith said. “If America is gonna lead the world, we need to connect with the world. Our global leadership relies on our ability to serve the world with the right approach and on our ability to sustain the trust of the rest of the world.” He added that removing quantitative caps for tier two countries is essential because these policies “sent a message to 120 nations that couldn’t count on us to provide the AI they want and need.”  Altman noted, “There will be great chips and models trained around the world,” reiterating American companies’ leading position in the space.  There’s some good news in the area of AI diffusion because while the hearing was ongoing, the Commerce Department announced it was modifying rules from the Biden administration that limited which countries could receive chips made by American companies. The rule was set to take effect on May 15. While the executives said government standards would be helpful, they decried any move to “pre-approve” model releases, similar to the EU. Open ecosystem Generative AI occupies a liminal space in tech regulation. On the one hand, the comparative lack of rules has allowed companies like OpenAI to develop technology without much fear of repercussions. On the other hand, AI, like the internet and social media before it, touches people’s lives professionally and personally.  In some ways, the executives veered away from how the Trump administration has positioned U.S. growth. The hearing showed that while AI companies want support from the government to speed up the process of expanding the AI infrastructure, they also need to be more open to the rest of the world. It requires talent from abroad. It needs to sell products and platforms to other countries.  Social media commentary varied, with some pointing out that executives, in particular Altman, had different opinions on regulation before. 2023 Sam Altman: Tells Congress a new agency should be

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More It’s a conundrum: Customer teams have more data than they can ever begin to use—from Salesforce notes, Jira tickets, project dashboards and Google Docs—but they struggle to combine it all when crafting customer messaging that really resonates.  Existing tools often rely on generic templates or slides and fail to provide a complete picture of customer journeys, roadmaps, project goals and business objectives.  Korl, a startup launched today, hopes to overcome these challenges with a new platform that works across multiple systems to help create highly customized communications. The multi-agent, multimodal tool uses a mix of models from OpenAI, Gemini, and Anthropic to source and contextualize data.  “Engineers have powerful AI tools, but customer-facing teams are stuck with shallow, disconnected solutions,” Berit Hoffmann, CEO and co-founder of Korl, told VentureBeat in an exclusive interview. “Korl’s core innovation is rooted in our advanced multi-agent pipelines designed to build the customer and product context that generic presentation tools lack.”  Creating tailored customer materials through a multi-source view Korl’s AI agents aggregate information from across different systems — such as engineering documentation from Jira, outlines from Google Docs, designs from Figma, and project data from Salesforce — to build a multi-source view.  For instance, once a customer connects Korl to Jira, its agent studies existing and planned product capabilities to figure out how to map data and import new product capabilities, Hoffmann explained. The platform matches product data with customer information—such as usage history, business priorities and lifecycle stage—filling in gaps with AI.  “Korl’s data agents automatically gather, enrich, and structure diverse datasets from internal sources and external public data,” said Hoffmann.  The platform then automatically generates personalized quarterly business reviews (QBRs), renewal pitches, tailored presentations and other materials for use in important customer milestones.  Hoffmann said the company’s core differentiator is its ability to deliver “polished, customer-ready materials” such as slides, narratives and emails, “rather than merely analytics or raw insights.” “We think this delivers a level of operational value that customer-facing teams need today given the pressures to do more with less,” she said.  Switching between OpenAI, Gemini, Anthropic, based on performance Korl orchestrates an “ensemble of models” across OpenAI, Gemini and Anthropic, selecting the best model for the job at the time based on speed, accuracy and cost, Hoffmann explained. Korl needs to perform complex, diverse tasks — nuanced narratives, data computation, visuals — so each use case is matched with the most performant model. The company has implemented “sophisticated fallback mechanisms” to mitigate failures; early on, they observed high failure rates when relying on a single provider, Hoffman reported. The startup developed a proprietary auto-mapper fine-tuned to handle diverse enterprise data schemas across Jira, Salesforce and other systems. The platform automatically maps to relevant fields in Korl.  “Rather than just semantic or field-name matching, our approach evaluates additional factors like data sparsity to score and predict field matches,” said Hoffmann.  To speed the process, Korl combines low-latency, high-throughput models (such as GPT-4o for rapid, context-building responses) with deeper analytical models (Claude 3.7 for more complex, customer-facing communications).  “This ensures that we optimize for the best end user experience, making context-driven tradeoffs between immediacy and accuracy,” Hoffmann explained.  Because “security is paramount,” Korl seeks enterprise-grade privacy guarantees from vendors to ensure customer data is excluded from training datasets. Hoffmann pointed out that its multi-vendor orchestration and contextual prompting further limit inadvertent exposure and data leaks. Grappling with data that is ‘too messy’ or ‘incomplete’ Hoffman noted that, early on, Korl heard from customers that they worried their data would be “too messy” or “incomplete” to be put to good use. In response, the company built pipelines to understand business object relationships and fill in gaps — such as how to position features externally, or how to align values around desired outcomes.  “Our presentation agent is what leverages that data to generate customer slides and talk track [guide conversations with potential customers or leads] dynamically when needed,” said Hoffmann.  She also said Korl features “true multimodality.” The platform isn’t just pulling data from various sources; it’s interpreting different types of information such as text, structured data, charts or diagrams.  “The critical step is moving beyond the raw data to answer: What story does this graph tell? What are the deeper implications here, and will they actually resonate with this specific customer?,” she said. “We’ve built our process to perform that crucial due diligence, ensuring the output isn’t just aggregated data, but genuinely rich content delivered with meaningful context.” Two of Korl’s close competitors include Gainsight and Clari; however, Hoffmann said Korl differentiates itself by incorporating deep product and roadmap context. Effective customer renewal and expansion strategies require a deep understanding of what a product does, and this should be coupled with an analysis of customer data and behavior. Further, Hoffmann said Korl addresses two “foundational shortcomings” of existing platforms: deep business context and brand accuracy. Korl’s agents gather business context from multiple systems. “Without this comprehensive data intelligence, automated decks lack strategic business value,” she said.  When it comes to branding, Korl’s proprietary technology extracts and replicates guidelines from existing materials. Reducing deck prep time from ‘multiple hours to minutes’  Early indications suggest Korl can unlock at least a 1-point improvement in net revenue retention (NRR) for mid-market software companies, said Hoffmann. This is because it uncovers previously unrealized product value and makes it easy to communicate that to customers before they churn or make renewal or expansion decisions.  The platform also improves efficiency, reducing deck preparation time for each customer call from “multiple hours to minutes,” according to Hoffman.  Early customers include skills-building platform Datacamp and gifting and direct mail company Sendoso.  “They tackle a critical and overlooked challenge: Too often, product features are released while go-to-market (GTM) teams are not prepared to sell, support or communicate them effectively,” said Amir Younes, Sendoso’s chief customer officer. “With Korl’s AI, [go-to-market] GTM enablement and asset

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Researchers at Mem0 have introduced two new memory architectures designed to enable Large Language Models (LLMs) to maintain coherent and consistent conversations over extended periods.  Their architectures, called Mem0 and Mem0g, dynamically extract, consolidate and retrieve key information from conversations. They are designed to give AI agents a more human-like memory, especially in tasks requiring recall from long interactions.  This development is particularly significant for enterprises looking to deploy more reliable AI agents for applications that span very long data streams. The importance of memory in AI agents LLMs have shown incredible abilities in generating human-like text. However, their fixed context windows pose a fundamental limitation on their ability to maintain coherence over lengthy or multi-session dialogues.  Even context windows that reach millions of tokens aren’t a complete solution for two reasons, the researchers behind Mem0 argue. As meaningful human-AI relationships develop over weeks or months, the conversation history will inevitably grow beyond even the most generous context limits. Second, Real-world conversations rarely stick to a single topic. An LLM relying solely on a massive context window would have to sift through mountains of irrelevant data for each response. Furthermore, simply feeding an LLM a longer context doesn’t guarantee it will effectively retrieve or use past information. The attention mechanisms that LLMs use to weigh the importance of different parts of the input can degrade over distant tokens, meaning information buried deep in a long conversation might be overlooked. “In many production AI systems, traditional memory approaches quickly hit their limits,” Taranjeet Singh, CEO of Mem0 and co-author of the paper, told VentureBeat.  For example, customer-support bots can forget earlier refund requests and require you to re-enter order details each time you return. Planning assistants may remember your trip itinerary but promptly lose track of your seat or dietary preferences in the next session. Healthcare assistants can fail to recall previously reported allergies or chronic conditions and give unsafe guidance.  “These failures stem from rigid, fixed-window contexts or simplistic retrieval methods that either re-process entire histories (driving up latency and cost) or overlook key facts buried in long transcripts,” Singh said. In their paper, the researchers argue that a robust AI memory should “selectively store important information, consolidate related concepts, and retrieve relevant details when needed—mirroring human cognitive processes.” Mem0 Mem0 architecture Credit: arXiv Mem0 is designed to dynamically capture, organize and retrieve relevant information from ongoing conversations. Its pipeline architecture consists of two main phases: extraction and update. The extraction phase begins when a new message pair is processed (typically a user’s message and the AI assistant’s response). The system adds context from two sources of information: a sequence of recent messages and a summary of the entire conversation up to that point. Mem0 uses an asynchronous summary generation module that periodically refreshes the conversation summary in the background.  With this context, the system then extracts a set of important memories specifically from the new message exchange. The update phase then evaluates these newly extracted “candidate facts” against existing memories. Mem0 leverages the LLM’s own reasoning capabilities to determine whether to add the new fact if no semantically similar memory exists; update an existing memory if the new fact provides complementary information; delete a memory if the new fact contradicts it; or do nothing if the fact is already well-represented or irrelevant.  “By mirroring human selective recall, Mem0 transforms AI agents from forgetful responders into reliable partners capable of maintaining coherence across days, weeks, or even months,” Singh said. Mem0g Mem0g architecture Credit: arXiv Building on the foundation of Mem0, the researchers developed Mem0g (Mem0-graph), which enhances the base architecture with graph-based memory representations. This allows for a more sophisticated modeling of complex relationships between different pieces of conversational information. In a graph-based memory, entities (like people, places, or concepts) are represented as nodes, and the relationships between them (like “lives in” or “prefers”) are represented as edges. As the paper explains, “By explicitly modeling both entities and their relationships, Mem0g supports more advanced reasoning across interconnected facts, especially for queries that require navigating complex relational paths across multiple memories.” For example, understanding a user’s travel history and preferences might involve linking multiple entities (cities, dates activities) through various relationships. Mem0g uses a two-stage pipeline to transform unstructured conversation text into graph representations. First, an entity extractor module identifies key information elements (people, locations, objects, events, etc.) and their types. Then, a relationship generator component derives meaningful connections between these entities to create relationship triplets that form the edges of the memory graph. Mem0g includes a conflict detection mechanism to spot and resolve conflicts between new information and existing relationships in the graph. Impressive results in performance and efficiency The researchers conducted comprehensive evaluations on the LOCOMO benchmark, a dataset designed for testing long-term conversational memory. In addition to accuracy metrics, they used an “LLM-as-a-Judge” approach for performance metrics, where a separate LLM assesses the quality of the main model’s response. They also tracked token consumption and response latency to evaluate the techniques’ practical implications. Mem0 and Mem0g were compared against six categories of baselines, including established memory-augmented systems, various Retrieval-Augmented Generation (RAG) setups, a full-context approach (feeding the entire conversation to the LLM), an open-source memory solution, a proprietary model system (OpenAI’s ChatGPT memory feature) and a dedicated memory management platform. The results show that both Mem0 and Mem0g consistently outperform or match existing memory systems across various question types (single-hop, multi-hop, temporal and open-domain) while significantly reducing latency and computational costs. For instance, Mem0 achieves a 91% lower latency and saves more than 90% in token costs compared to the full-context approach, while maintaining competitive response quality. Mem0g also demonstrates strong performance, particularly in tasks requiring temporal reasoning. “These advances underscore the advantage of capturing only the most salient facts in memory, rather than retrieving large chunk of original text,” the researchers write. “By converting the conversation

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Zencoder announced today the launch of Zen Agents, a platform that enables organization-wide creation and sharing of specialized AI tools for software development. The release includes an open-source marketplace where developers can contribute and discover custom agents, marking a significant shift in how development teams leverage artificial intelligence. While existing AI coding assistants have primarily focused on boosting individual developer productivity, Zencoder’s approach addresses the collaborative reality of modern software engineering, where delays often occur between coding and feedback loops. “If you look at the tools that are used today for real AI in engineering, it’s basically coding agents with an IDE,” said Andrew Filev, CEO and founder of Zencoder, in an exclusive interview with VentureBeat. “And if you dig even one layer deeper, you’ll find they’re usually focused on the individual developer. It all makes perfect sense, because it all starts with the developer, right?” But Filev points to a critical gap in current solutions: “There’s this whole layer of things that you can do beyond individual engineers, because engineers don’t work alone. In any successful software business, development happens in teams.” How Zen Agents shrinks development cycles by automating the in-between steps The new platform addresses this gap by enabling teams to create and deploy custom agents tailored to specific frameworks, workflows, or codebases. These agents can be shared across organizations, ensuring consistent practices while eliminating repetitive tasks. What distinguishes Zen Agents technically is its implementation of the Model Context Protocol (MCP), a standard originated by Anthropic and supported by OpenAI that allows large language models to interact with external tools. “As part of this launch, we’re introducing our own registry that has over 100 MCP servers,” Filev explained. “We created this because there’s no standard registry available yet. If a standard registry existed, we would simply connect to it, since our real value comes from our agents and specialized tools.” Industry analysts see this as a natural progression in development tools. The initial wave of AI coding assistants provided immediate productivity boosts for individual tasks, but fell short in addressing the collaborative nature of enterprise software development, where time is often lost in handoffs between team members. Zen Agents aims to address these handoffs by allowing specialized agents to automate parts of the development lifecycle, from code review to testing. “For example, let’s say you have an agent that does code review,” Filev said. “Imagine there is an agent that you trust. The agent doesn’t even necessarily have to be as good as a human, because if it finds issues and provides feedback immediately, you can address those problems right away.” The platform is designed to be enterprise-ready, with Zencoder touting ISO 27001, SOC 2 Type II certification, and ISO 42001 for responsible AI management systems — necessary credentials for adoption in security-conscious organizations. Perhaps the most distinctive aspect of the launch is the open-source marketplace, which allows the broader developer community to contribute specialized agents. This approach mirrors successful open-source ecosystems like Visual Studio Code extensions or npm packages, where community contributions vastly expand the capabilities beyond what any single vendor could develop. “I’m a big believer in collective intelligence,” noted Filev. “There are so many use cases that we haven’t even thought of yet, and even if we did imagine them all, we would never have the resources to cover them ourselves.” Early adopters have already found value in creating specialized agents. “I’ve been impressed by the examples that integrate several steps in their workflow,” Filev shared. “For instance, you can pull a wireframe from Figma, automatically generate code based on it, and then submit a pull request — all as a seamless process.” Another notable example addresses accessibility requirements—an area often acknowledged as important but frequently deprioritized under tight deadlines. “Our Developer Advocate created an agent that improves the accessibility of code,” Filev said. “Everyone in software agrees that accessibility is extremely important, but in reality, teams don’t always have the time to properly address these needs.” According to Matt Walker, Co-founder and CTO of Simon Data, who was quoted in the press release, the impact has been measurable: “Zen Agents marks an important evolution in AI-assisted development. Team-shareable agents along with MCP integrations lets us build specialized AI tools that genuinely understand our unique development workflows and infrastructure. We’ve already noticed a significant reduction in context-switching across our engineering teams.” Beyond coding: The race toward AI-enhanced developer flow state Pricing for Zen Agents currently follows a simple tiered structure. “Our pricing plans are straightforward: we offer a free tier, along with $20 and $40 monthly options,” said Filev, though he noted that as usage grows, the company is considering expanded options. “The way I think about it is simple—the more you use it, the more money you save.” Looking ahead, Filev sees Zen Agents evolving toward greater autonomy, not to replace engineers but to make them dramatically more productive. “We’re racing towards autonomy—not with the goal of replacing engineers, but with the vision of making engineers 10 times more productive,” he said. This vision extends beyond just writing code to maintaining what developers call “flow state”—periods of uninterrupted, highly productive work. “Our company has Zen in its name, and it’s not productive to start working on something, then jump to something else, only to later return to the original task,” Filev explained. “If we can keep you in that flow state, then mission accomplished, right?” While Zencoder is initially focused on software engineering applications, Filev hinted at broader potential. “Many of my tech friends are already using this technology for non-engineering purposes,” he said, mentioning personal assistants and marketing automation as examples. “I’m curious to see what the community creates with it—there’s a real possibility it could gain traction in a much broader context.” As AI tools mature in the software development space, Zen Agents points to a future where the technology becomes

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Last night, OpenAI published a blog post on its official website authored by CEO and co-founder Sam Altman announcing a major new hire: Fidji Simo, currently CEO and Chair at grocery delivery company Instacart, will join OpenAI as CEO of Applications, a newly created executive position. Simo will step into her new role at OpenAI later in 2025 following a transition period. She will report directly to Altman and lead the Applications division, which includes teams responsible for translating OpenAI’s research into products used by consumers and businesses worldwide. Simo also shared a message with her team at Instacart, which she later posted publicly on her LinkedIn profile, stating: “This was an incredibly hard decision because I love this company… At the same time, you all know my passion for AI generally and in particular for the potential it has to cure diseases — the ability to lead such an important part of our collective future was a hard opportunity to pass up.” She further wrote: “Joining OpenAI at this critical moment is an incredible privilege and responsibility. This organization has the potential of accelerating human potential at a pace never seen before and I am deeply committed to shaping these applications toward the public good.” Altman was clear to state he will remain the CEO of the entire ChatGPT company, but that the organization’s growing scope across AI research, product delivery, and infrastructure meant it was time for the creation of a new division with its own executive leader. Altman said he plans to devote increased focus to Research, Compute, and Safety Systems, which will continue reporting directly to him. He reiterated that all functions will remain integrated under a single OpenAI structure, including its nonprofit arm. Simo will remain Instacart’s CEO for the coming months and help onboard her successor, who is expected to come from the company’s existing management team. She will continue serving as Chair of the Instacart Board after stepping down as CEO. OpenAI’s evolving structure and AI stack Altman emphasized the strategic importance of the new role, noting that OpenAI has evolved beyond its original identity as a research lab. Over the past two and a half years, the company has grown into a global product provider and, more recently, an infrastructure builder delivering AI tools at large scale. “We’re in a privileged position to be scaling at a pace that lets us do them all simultaneously, and bringing on exceptional leaders is a key part of doing that well,” Altman wrote in his blog post. He described Simo as uniquely qualified to lead the Applications group, citing her leadership experience, operational expertise, and alignment with OpenAI’s mission. Separately, in prepared testimony before the U.S. Congress today, Altman also included a chart showing how OpenAI views the “AI stack” it offers customers, and how the entire industry is divided into three core buckets or “layers”: Also in that testimony, Altman revealed more examples of what OpenAI considers to be part of the AI applications layer, among them, chatbots such as its hit ChatGPT and Microsoft’s Copilot (also powered in part by AI models): “Ultimately, both the infrastructure and platform layers support the applications layer. These are devices and software applications that use AI to deliver better services to people. ChatGPT and 4 Microsoft’s Copilot are both examples of AI applications. One of the amazing things about the applications layer is it’s not just companies – large or small or established or startup – that are creating AI applications. It’s everybody. It’s researchers using new AI-infused applications to change drug discovery. It’s non-profits changing the way they deliver services. It’s teachers using AI as a tool to improve the way they prepare material for a classroom. It’s governments making everything from the filing of a tax return to the renewal of a driver’s license easier and more efficient. To build a new AI economy, it’s critical to get all three of these layers working and to get a flywheel turning across the ecosystem. It’s essential to build the infrastructure layer so people can develop and deploy the models at the platform layer. It’s essential to use the AI models so that people will build the applications on top of them. And it’s essential for customers to adopt the applications, so the market can grow, and drive increased investment to expand the infrastructure further. The process repeats itself. This is how a new economy is born.” More OpenAI team members shared praise for Simo on social media, mainly X, where Chief Marketing Officer (CMO) Kate Rouch wrote of the new hire: “She’s all signal. No noise. Highest integrity leader I know. Very good news: for OpenAI, for all of us.” Caitlin Kalinowski, who is a member of the technical staff at OpenAI and leads its hardware and robotics vision, also took to X to post, in part: “We couldn’t be luckier to have her deep experience and warm leadership style. Welcome, Fidji!!!” Simo boasts an impressive track record focused on turning tech into viable businesses Simo joined OpenAI’s board in March 2024 and brings extensive leadership experience in technology, consumer platforms, and healthcare innovation. At Instacart, she oversaw the company’s public debut and a series of business milestones, including scaling its advertising business and launching new offerings. Before that, she spent a decade at Facebook, where she led the core Facebook app and was responsible for key product areas including Video, Marketplace, Groups, and Ads. She is also the co-founder of the Metrodora Institute, a clinic and research center focused on neuroimmune disorders, where she serves as President of the Metrodora Foundation. In addition to her board seat at OpenAI, Simo sits on the boards of Shopify and previously Cirque du Soleil. With this leadership appointment, OpenAI is signaling a deeper focus on scaling its consumer and enterprise-facing AI offerings. By establishing a dedicated CEO role

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More There’s a new king on the throne of AI coding models: Today, Google’s DeepMind AI research unit unveiled Gemini 2.5 Pro “I/O” edition, a new version of its hit Gemini 2.5 Pro multimodal large language model (LLM) released back in March that DeepMind CEO Demis Hassabis said on X is “the best coding model we’ve ever built!” Indeed, the initial benchmarks released by the company indicate Google has taken the lead — for the first time since the generative AI race began in earnest with the late 2022 launch of ChatGPT — above all other models on at least one important coding benchmark. The new version, labeled “gemini-2.5-pro-preview-05-06,” replaces the previous 03-25 release and is now available for indie developers in Google AI Studio and for enterprises in the Vertex AI cloud platform, as well as to individual users in the Gemini app. Google’s blog post said it also powers the Gemini mobile app’s Canvas and other features. The new version powers feature development in apps like Gemini 95, where the model helps match visual styles across components automatically. It also enables workflows like converting YouTube videos into full-featured learning applications and crafting highly styled components—such as responsive video players or animated dictation UIs—with little to no manual CSS editing. It’s a proprietary model, meaning enterprises will have to pay Google to use it and access it only through Google’s web services. However, it doesn’t alter pricing or rate limits; current users of Gemini 2.5 Pro will be automatically routed to the updated model which costs $1.25/$10 per million tokens in/out (for context lengths of 200,000 tokens) compared to Claude 3.7 Sonnet’s $3/$15. The company frames this move — ahead of Google’s annual I/O (input/output) developer conference later this month in Mountain View and online, May 20-21 — as a response to strong community feedback around Gemini’s practical utility in real-world code generation and interface design. Logan Kilpatrick, Senior Product Manager for Gemini API and Google AI Studio, confirmed in a developer blog post that the update also addresses key developer feedback around function calling, with improvements in error reduction and trigger reliability. Top scores from human raters at generating web apps On WebDev Arena Leaderboard, a third-party metric that ranks models by human preference based on their ability to generate visually appealing and functional web apps, Gemini 2.5 Pro Preview (05-06) has now overtaken Anthropic’s Claude 3.7 Sonnet at the number one spot. The new version scored 1499.95 on the leaderboard, placing it well ahead of Sonnet 3.7’s 1377.10. The previous Gemini 2.5 Pro (03-25) model held third place with a score of 1278.96, meaning the I/O edition represents a 221-point jump. As noted by the AI power user “Lisan al Gaib” on X, not even OpenAI’s GPT-4o (“o3”) was able to displace Sonnet 3.7, highlighting the significance of Gemini’s advancement. Gemini’s performance boost reflects improved reliability, aesthetics, and usability in its outputs. Already winning rave reviews Several developers and platform leaders have highlighted the model’s improved reliability and application in production scenarios. Cognition’s Silas Alberti noted that Gemini 2.5 Pro was the first model to successfully complete a complex refactoring of a backend routing system, demonstrating the kind of decision-making one would expect from a senior developer. Michael Truell, CEO of the AI coding tool Cursor, said internal testing shows a marked decrease in tool call failures, a previously noted issue. He expects users to find the latest version significantly more effective in hands-on environments. Cursor has already integrated Gemini 2.5 Pro into its own code agent, reflecting how developers are using the model as a key component in more intelligent developer workflows. Michele Catasta, President of Replit, described Gemini 2.5 Pro as the best frontier model for balancing capability with latency. His comments suggest that Replit is considering integration of the model into its own tools, especially for tasks where high responsiveness and reliability are crucial. Similarly, AI educator and BlueShell private AI chatbot founder Paul Couvert noted on X that “Its code and UI generation capabilities are impressive.’” And as Pietro Schirano, CEO of the AI art tool EverArt, noted on X, the new Gemini 2.5 Pro I/O edition was able to generate an interactive simulation of the “1 gorilla vs. 100 men” meme that’s been circulating on social media lately from a single prompt. Showing off another interactive Tetris-style puzzle game with working sound effects reportedly created in less than a minute, X user “RameshR” (@rezmeram) wrote that “the casual game industry is dead!!” These endorsements add weight to DeepMind’s claims of practical improvements and may encourage broader adoption across developer platforms. Full apps and programs from one text prompt One of the standout features of the update is its ability to build full, interactive web apps or simulations from a single prompt. This aligns with DeepMind’s vision of simplifying the prototyping and development process. Demonstrations within the Gemini app showcase how users can transform visual patterns or thematic prompts into usable code, lowering the barrier to entry for design-oriented developers and teams experimenting with new ideas. Although the architecture and under-the-hood changes of Gemini 2.5 Pro have not been detailed publicly, the emphasis remains on enabling faster, more intuitive development experiences. By leaning into its strengths in code generation and multimodal inputs, Gemini 2.5 Pro is positioned less as a research novelty and more as a practical tool for real-world coding challenges. The early release reflects a clear intention from Google DeepMind to meet developer demand and maintain momentum ahead of its major conference announcements. source

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Lightricks, the company behind popular creative apps like Facetune and VideoLeap, announced today the release of its most powerful AI video generation model to date. The LTX Video 13-billion-parameter model (LTXV-13B) generates high-quality AI video up to 30 times faster than comparable models while running on consumer-grade hardware rather than expensive enterprise GPUs. The model introduces “multiscale rendering,” a novel technical approach that dramatically increases efficiency by generating video in progressive layers of detail. This enables creators to produce professional-quality AI videos on standard desktop computers and high-end laptops instead of requiring specialized enterprise equipment. “The introduction of our 13B parameter LTX Video model marks a pivotal moment in AI video generation with the ability to generate fast, high-quality videos on consumer GPUs,” said Zeev Farbman, co-founder and CEO of Lightricks, in an exclusive interview with VentureBeat. “Our users can now create content with more consistency, better quality, and tighter control.” How Lightricks democratizes AI video by solving the GPU memory problem A major challenge for AI video generation has been the enormous computational requirements. Leading models from companies like Runway, Pika, and Luma typically run in the cloud on multiple enterprise-grade GPUs with 80GB or more of VRAM (video memory), making local deployment impractical for most users. Farbman explained how LTXV-13B addresses this limitation: “The major dividing line between consumer and enterprise GPUs is the amount of VRAM. Nvidia positions their gaming hardware with strict memory limits — the previous generation 3090 and 4090 GPUs maxed out at 24 gigabytes of VRAM, while the newest 5090 reaches 32 gigabytes. Enterprise hardware, by comparison, offers significantly more.” The new model is designed to operate effectively within these consumer hardware constraints. “The full model, without any quantization, without any approximation, you will be able to run on top consumer GPUs — 3090, 4090, 5090, including their laptop versions,” Farbman noted. Two AI-generated rabbits, rendered on a single consumer GPU, stride off after a brief glance at the camera — an unedited four-second sample from Lightricks’ new LTXV-13B model. (Credit: Lightricks) Inside ‘multiscale rendering’: The artist-inspired technique that makes AI video generation 30X faster The core innovation behind LTXV-13B‘s efficiency is its multiscale rendering approach, which Farbman described as “the biggest technical breakthrough of this release.” “It allows the model to generate details gradually,” he explained. “You’re starting on the coarse grid, getting a rough approximation of the scene, of the motion of the objects moving, etc. And then the scene is kind of divided into tiles. And every tile is filled with progressively more details.” This process mirrors how artists approach complex scenes — starting with rough sketches before adding progressively finer details. The advantage for AI is that “your peak amount of VRAM is limited by a tile size, not the final resolution,” Farbman said. The model also features a more compressed latent space, which requires less memory while maintaining quality. “With videos, you have a higher compression ratio that allows you, while you’re in the latent space, to just take less VRAM,” Farbman added. Performance metrics showing Lightricks’ LTXV-13B model generating video in just 37.59 seconds, compared to over 1,491 seconds for a competing model on equivalent hardware — a nearly 40× speed improvement. (Credit: Lightricks) Why Lightricks is betting on open source when AI markets are increasingly closed While many leading AI models remain behind closed APIs, Lightricks has made LTXV-13B fully open source, available on both Hugging Face and GitHub. This decision comes during a period when open-source AI development has faced challenges from commercial competition. “A year ago, things were closed, but things are kind of opening up. We’re seeing really a lot of cool LLMs and diffusion models opening up,” Farbman reflected. “I’m more optimistic now than I was half a year ago.” The open-source strategy also helps accelerate research and improvement. “The main rationality for open-sourcing it is to reduce the cost of your R&D,” Farbman explained. “There are a ton of people in academia that use the model, write papers, and you’re starting to become this curator that understands where the real gold is.” How Getty and Shutterstock partnerships help solve AI’s copyright challenges As legal challenges mount against AI companies using scraped training data, Lightricks has secured partnerships with Getty Images and Shutterstock to access licensed content for model training. “Collecting data for training AI models is still a legal gray area,” Farbman acknowledged. “We have big customers in our enterprise segment that care about this kind of stuff, so we need to make sure we can provide clean models for them.” These partnerships allow Lightricks to offer a model with reduced legal risk for commercial applications, potentially giving it an advantage in enterprise markets concerned about copyright issues. The strategic gamble: Why Lightricks offers its advanced AI model free to startups In an unusual move for the AI industry, Lightricks is offering LTXV-13B free to license for enterprises with under $10 million in annual revenue. This approach aims to build a community of developers and companies who can demonstrate the model’s value before monetization. “The thinking was that academia is off the hook. These guys can do whatever they want with the model,” Farbman said. “With startups and industry, you want to create win-win situations. I don’t think you can make a ton of money from a community of artists playing with AI stuff.” For larger companies that find success with the model, Lightricks plans to negotiate licensing agreements similar to how game engines charge successful developers. “Once they hit ten million in revenue, we’re going to come to talk with them about licensing,” Farbman explained. Despite the advances represented by LTXV-13B, Farbman acknowledges that AI video generation still has limitations. “If we’re honest with ourselves and look at the top models, we’re still far away from Hollywood movies. They’re not there yet,” he said. However, he sees immediate practical applications

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Enterprise AI in 2025 is moving from experimentation to implementation and deployments are evolving from AI assistants to AI agents. That’s the primary theme of the IBM Think 2025 conference, which gets underway today. At the event, IBM is announcing an extensive list of new enterprise AI services as well as enhancements to existing technologies to help move more enterprise AI efforts into real-world deployment. The core of IBM’s updates are a series of updates for its watsonx platform that was first announced at Think 2023. At the Think 2024 event, the big theme was the introduction of orchestration and the ability to help enterprise build their own AI assistants. In 2025, AI assistants are table stakes and the conversation across the industry and in every enterprise is how to build, use and benefit from agentic AI. IBM is announcing a series of agentic AI capabilities, including: AI Agent Catalog: A centralized discovery hub for pre-built agents. Agent Connect: A partner program for third-party developers to integrate their agents with watsonx Orchestrate. Domain-specific agent templates for sales, procurement and HR. No-code agent builder for business users without technical expertise. Agent development toolkit for developers. Multi-agent orchestrator with agent-to-agent collaboration capabilities. Agent Ops (in private preview) providing telemetry and observability. IBM’s fundamental goal is to help enterprises bridge the gap between experimentation, real-world deployments, and business benefits. “Over the next few years, we expect there will be over a billion new applications constructed using generative AI,” IBM CEO Arvind Krishna said in a briefing with press and analysts. “AI is one of the unique technologies that can hit at the intersection of productivity, cost savings and revenue scaling.” The enterprise AI challenge: How to get real ROI While there is no shortage of hype and interest in AI, that’s not what actually makes a real difference for an enterprise concerned with the bottom line.  Research sponsored by IBM shows that enterprises only get the return on investment (ROI) they expect approximately 25% of the time. Krishna noted that several factors impact ROI. They include access to enterprise data, the siloed nature of different applications, and the challenges of hybrid infrastructure. “Everybody is doubling down on AI investments,” Krishna said. “The only change over the last 12 months is that people are stopping experimentation and focusing very much on where is the value to the business.” From AI experimentation to enterprise production At the heart of IBM’s announcements is a recognition that organizations are shifting from isolated AI experiments to coordinated deployment strategies that require enterprise-grade capabilities. “We’re trying to bridge the gap from where we are today, which is thousands of experiments into enterprise grade deployments which require the same kind of security governance and standards that we did demand on mission critical applications,” Ritika Gunnar, general manager data and AI at IBM, told VentureBeat in an interview. The evolution of IBM’s watsonx Orchestrate platform reflects the broader maturity of AI technology. The platform was first announced by IBM in 2023, largely as a way to help build and work with AI assistants and automations. In 2024, as agentic AI first began to become mainstream, IBM started to add agentic capabilities and partnered with multiple vendors, including Crew AI. With IBM’s new agentic AI components, the direction is now to help enable multi-agent collaboration and workflows. It’s about going beyond just the ability to build and deploy agents to actually figuring out how an enterprise can generate an ROI from agents. “We really believe that we’re entering into an era of systems of true intelligence,” Gunnar said. “Because now we’re integrating AI that can do things for you and this is a big differentiation.” The technology and protocols that enable enterprise agentic AI The industry has no shortage of attempts to help enable agentic AI. Langchain is a widely used platform for building and running agents and is also part of a wider effort alongside Cisco and Galileo for the AGNTCY open framework for agentic AI. When it comes to agent-to-agent communications, Google announced Agent2Agent in April. Then, of course, there is Model Context Protocol (MCP), which has emerged to become a de facto standard for connecting agentic AI tools to services. Gunnar explained that IBM uses its own technology for the multi-agent orchestration piece. She noted that how agents work together is critical and is a point of differentiation for IBM. That said, she also emphasized that IBM is trying to take an open approach. That means enterprises can build agents with IBM’s tools, such as BeeAI, or those from other vendors, including Crew AI or Langchain, and they’ll all still work with watsonx Orchestrate. IBM is also enabling and supporting MCP. According to Gunnar, IBM is supporting MCP by making it easy for tools with an MCP interface to automatically show up and be usable in watsonx Orchestrate. Specifically, if a tool exists with an MCP interface, it will automatically be available to use in watsonx Orchestrate. “Our goal is to be open,” she said. “We want you to integrate your agents, regardless of whatever framework that you’ve built it in.” Addressing enterprise concerns: Security, governance and compliance As part of making sure agentic AI is ready for enterprise usage, there is a need to ensure trust and compliance. That’s also a critical part of IBM’s push. Gunnar explained that IBM has built guardrails and governance directly into the watsonx portfolio. “We’re expanding the capabilities that we have for governance of LLMs into agentic technology, ”  she said. “Just as we have evaluation of LLMs, you need to be able to have an evaluation of what it means for agent responses.” IBM is also extending its traditional machine learning evaluation metrics to agent technologies. Gunnar said that IBM tracks over 100 different metrics for large language models, which it is now extrapolating and extending to agentic technologies as well. Real-world impact Agentic AI is already having

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Microsoft CEO Satya Nadella’s endorsement of Google DeepMind‘s Agent2Agent (A2A) open protocol and Anthropic’s Model Context Protocol (MCP) will immediately accelerate agentic AI-based collaboration and interdependence, leading to rapid gains in agentic-based apps and platforms. Nadella’s endorsement delivers the catalyst the agentic AI development community needed to fast-track their collaborations, leading to entirely new apps, platforms and networks. Having historically been open about the potential for agentic AI to integrate across platforms, yesterday’s announcement, which also unveiled upcoming support for CoPilot Studio and Foundry, set a new precedent in how committed Microsoft is to open agentic standards.   On Wednesday, Nadella wrote on X (formerly Twitter) that “Open protocols like A2A and MCP are key to enabling the agentic web,” before announcing upcoming support in Copilot Studio and Foundry. While often agreeing with the concept of open standards for agentic AI integration, this is the first time he’s endorsed a standard publicly. Nadella’s influence on the industry will lead to a shift from proprietary ecosystems toward cross-platform, agentic AI collaboration. Open protocols like A2A and MCP are key to enabling the agentic web. With A2A support coming to Copilot Studio and Foundry, customers can build agentic systems that interoperate by design. https://t.co/8wZGYU8kkj — Satya Nadella (@satyanadella) May 7, 2025 Committing to open architectures Nadella’s history of championing open, interoperable AI architectures spans over a decade. Throughout his many keynotes and interviews, Nadella has made it clear that open standard, not proprietary silos, are among the most reliable catalysts there are to drive adoption of new AI technologies. As early as 2018, Nadella highlighted Microsoft’s collaboration with Facebook on ONNX (Open Neural Network Exchange), an open model format. He said, ONNX “is supported broadly now by all of the frameworks… as well as the hardware acceleration from Intel, Qualcomm, and Nvidia.” Late last year, he reaffirmed an open platform approach to AI development, saying, “We’re building on the open platform ethos of GitHub, bringing multi-model choice to Copilot.” One of the primary goals of that effort is to enable developers to leverage AI models from any provider instead of being locked into a single vendor.   Earlier this year, Nadella doubled down on openness in agentic AI architectures, apps and platforms, emphasizing that open source “absolutely has a massive, massive role to play” in advanced AI systems. He added that “having a posture that allows interoperability is incredibly important” for enterprise AI. Nadella has also predicted that the more AI models become capable of taking on more complex tasks, the more the “models themselves become more of a commodity, and all value gets created by how you steer, ground, and fine-tune these models with your business data and workflow.” Through dozens of interviews and keynotes, Nadella continues to underscore that the real competitive edge comes from an open, flexible ecosystem where organizations can mix, match and customize AI components to suit their needs. A2A and MCP are growing in importance Nadella’s endorsement of A2A and MCP protocols shows how far the Microsoft senior management team has agreed that an open protocol approach is the best direction for the company. Endorsing both on the same day shows how far Microsoft is in its strategies related to agentic AI collaboration, integration and how diverse agentic AI architectures can be combined. Below is a table comparing each protocol along with a short explanation of its relevance in enterprises. Protocol Origin/Definition Enterprise Relevance A2A Google DeepMind introduced Agent2Agent (A2A) in 2025. It is an open protocol that standardizes inter-agent communication. It uses a shared schema for AI agents to exchange tasks, requests and results, enabling agents from any vendor or platform to collaborate seamlessly. A2A allows multi-vendor AI workflows to integrate smoothly, reducing vendor lock-in and supporting composability across diverse platforms. This unlocks dynamic, interoperable agent ecosystems for complex enterprise automation and innovation. MCP Model Context Protocol (MCP), open-sourced by Anthropic in late 2024, specifies how AI models request context/data from external sources in a standardized, secure manner. It acts as a universal interface for connecting AI models to tools and data, similar to a “USB-C port” for AI. MCP offers a uniform method to pull in enterprise data and app functionalities, streamlining AI integration. This fuels consistent security, governance and scalability across platforms, making it easier to deploy AI assistants enterprise-wide. Addressing the past of enterprise vendor lock-in early With enterprise deployments of CoPilot being a strategic priority at Microsoft, it’s understandable that Nadella took a preemptive step of reducing enterprise buyers’ concerns over vendor lock-in. Previous generations of the company’s products sold into enterprises were known for enforcing vendor lock-in either through complex and costly pricing strategies or challenging integration techniques, especially if a comparable Microsoft-native product was available. Nadella has long signaled to enterprises that he sees their infrastructure as being heterogeneous. Stepping into the role of cross-platform enabler, Nadella positions Microsoft’s Azure AI strategy as a viable alternative for enterprise DevOps workflows and ongoing development, for example. Copilot can quickly become part of multi-agent workflows now. Interested in enabling agentic AI providers to collaborate in creating new agentic apps, systems and platforms, Microsoft’s enforcement of A2A and MCP will prove to be a noteworthy catalyst for agentic AI’s growth. A quick win for security and compliance From a compliance standpoint, standard protocols simplify auditing and streamline compliance. When agent-to-agent interaction is structured through A2A, organizations can track exactly which entities exchanged information and when. This end-to-end visibility helps regulators see how customer data flows through various agents. For instance, if a user’s personal medical records are passed between an insurance agent and a hospital’s scheduling agent, A2A logs can confirm data privacy compliance. Conversely, MCP standardizes how AI systems request and consume data from enterprise data sources. It also resembles zero trust because role-based permissions govern each request, which helps analyze data leakage. Microsoft’s reinforcement of these protocols ensures that even if