VentureBeat

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More A San Francisco startup wants to make artificial intelligence development as easy as writing in a word processor. Wordware announced today a $30 million seed round led by Spark Capital, marking one of Y Combinator’s largest initial investments to date. The company has built what it calls a full-stack operating system for AI development, enabling users to create sophisticated AI agents using natural language instead of traditional programming code. With hundreds of thousands of users already on its platform, including enterprise customers like Instacart and Runway, Wordware is betting that the future of AI development belongs to domain experts rather than traditional software engineers. How natural language could replace traditional programming for AI “We are not a code-gen application,” Filip Kozera, co-founder and CEO of Wordware, told VentureBeat, distinguishing his company’s approach from other no-code tools. “We believe we’re witnessing a paradigm shift, and AI agents represent a new kind of software. Rather than focusing on code-gen, we’ve chosen to prioritize AI agents because we believe they will play a central role in driving the economy and automation in the future.” The company’s emergence comes at a critical moment in enterprise technology. Current workplace statistics suggest that 81% of workers spend less than 3 hours daily on creative work, with inefficiencies in meaningful work costing the global economy $8.9 trillion annually. Traditional AI development requires scarce and expensive engineering talent, creating a bottleneck for companies trying to implement AI solutions. Kozera draws an ambitious parallel to Microsoft Excel’s impact on data analytics: “Excel has 750 million monthly active users. What they’ve done to data analytics back in the 80s, we are trying to do to AI.” Why enterprise leaders are building AI without engineering teams The platform is already seeing adoption from major companies. “The C-suite executive comes in, spends couple days iterating on their AI agent, and then outputs an API and puts it into production,” Kozera explained. He cites an example where an Instacart founder “locked himself in his office and produced a new feature for their app” in just four days without hiring AI engineers. Another customer, Metadata, uses Wordware to build AI systems that optimize advertising spend. Kozera described how their AI agent works: “The agent takes a query from the customer, such as, ‘If I wanted to sell XYZ product in Brazil with this budget, how should I [allocate] my resources?’ It then writes code, queries multiple databases in real time, and generates a detailed report—all in under a minute.” The battle to become the operating system for AI development Despite competition from tech giants like Microsoft, Wordware is betting on its ability to move faster. “When I think about competition, I don’t necessarily worry about other startups in the space, but Microsoft is one of the players that has secured access to multiple model providers,” Kozera said. “The answer here is, as always when a startup is competing against a larger incumbent: delivery, the fact that we can take risks where they cannot.” “You have to be a little delusional in order to think that you can rebuild the whole development ecosystem that has been in the works for last 30 years for software,” he added. “This is what we’re trying to do.” Unlike typical no-code platforms, Wordware maintains a balance between accessibility and power. “Because we approached it in a way that we don’t want to have a graduation problem, it is not as simple as most no-code tools,” Kozera explained. “It does employ some programming concepts, and this is the price we pay in order to have the ability to build actually serious infrastructure.” The platform includes features like reflection loops for self-checking AI agents, comprehensive evaluation frameworks, and a GitHub-like repository system for sharing and customizing solutions. These capabilities have attracted significant attention from enterprise customers looking to accelerate their AI initiatives without building large specialized teams. Looking ahead, Wordware plans to expand its reach in early 2025 by enabling individual users to automate personal workflows using its engine. The company is actively hiring and building what Kozera describes as a unique company culture focused on transforming the AI development landscape. The $30 million investment, which includes participation from Felicis, Y-Combinator, Day One Ventures, and notable angels such as Paul Graham and Webflow’s Vlad Magdalin, suggests growing confidence in tools that bridge the gap between technical and non-technical users in AI development. As organizations increasingly seek to implement AI solutions, Wordware’s approach could reshape how enterprises approach AI implementation in the coming years. “In the space of next year, we want to build the best factory for building the AI engine,” Kozera said. “There is a potential to build a multi-trillion dollar company in the space of AI development — It’s going to be a battle, but it’s a battle I want to fight.” source

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More DeepSeek, an AI offshoot of Chinese quantitative hedge fund High-Flyer Capital Management focused on releasing high-performance open-source tech, has unveiled the R1-Lite-Preview, its latest reasoning-focused large language model (LLM), available for now exclusively through DeepSeek Chat, its web-based AI chatbot. Known for its innovative contributions to the open-source AI ecosystem, DeepSeek’s new release aims to bring high-level reasoning capabilities to the public while maintaining its commitment to accessible and transparent AI. And the R1-Lite-Preview, despite only being available through the chat application for now, is already turning heads by offering performance nearing and in some cases exceeding OpenAI’s vaunted o1-preview model. Like that model released in Sept. 2024, DeepSeek-R1-Lite-Preview exhibits “chain-of-thought” reasoning, showing the user the different chains or trains of “thought” it goes down to respond to their queries and inputs, documenting the process by explaining what it is doing and why. While some of the chains/trains of thoughts may appear nonsensical or even erroneous to humans, DeepSeek-R1-Lite-Preview appears on the whole to be strikingly accurate, even answering “trick” questions that have tripped up other, older, yet powerful AI models such as GPT-4o and Claude’s Anthropic family, including “how many letter Rs are in the word Strawberry?” and “which is larger, 9.11 or 9.9?” See screenshots below of my tests of these prompts on DeepSeek Chat: A new approach to AI reasoning DeepSeek-R1-Lite-Preview is designed to excel in tasks requiring logical inference, mathematical reasoning, and real-time problem-solving. According to DeepSeek, the model exceeds OpenAI o1-preview-level performance on established benchmarks such as AIME (American Invitational Mathematics Examination) and MATH. DeepSeek-R1-Lite-Preview benchmark results posted on X. Its reasoning capabilities are enhanced by its transparent thought process, allowing users to follow along as the model tackles complex challenges step by step. DeepSeek has also published scaling data, showcasing steady accuracy improvements when the model is given more time or “thought tokens” to solve problems. Performance graphs highlight its proficiency in achieving higher scores on benchmarks such as AIME as thought depth increases. Benchmarks and Real-World Applications DeepSeek-R1-Lite-Preview has performed competitively on key benchmarks. The company’s published results highlight its ability to handle a wide range of tasks, from complex mathematics to logic-based scenarios, earning performance scores that rival top-tier models in reasoning benchmarks like GPQA and Codeforces. The transparency of its reasoning process further sets it apart. Users can observe the model’s logical steps in real time, adding an element of accountability and trust that many proprietary AI systems lack. However, DeepSeek has not yet released the full code for independent third-party analysis or benchmarking, nor has it yet made DeepSeek-R1-Lite-Preview available through an API that would allow the same kind of independent tests. In addition, the company has not yet published a blog post nor a technical paper explaining how DeepSeek-R1-Lite-Preview was trained or architected, leaving many question marks about its underlying origins. Accessibility and Open-Source Plans The R1-Lite-Preview is now accessible through DeepSeek Chat at chat.deepseek.com. While free for public use, the model’s advanced “Deep Think” mode has a daily limit of 50 messages, offering ample opportunity for users to experience its capabilities. Looking ahead, DeepSeek plans to release open-source versions of its R1 series models and related APIs, according to the company’s posts on X. This move aligns with the company’s history of supporting the open-source AI community. Its previous release, DeepSeek-V2.5, earned praise for combining general language processing and advanced coding capabilities, making it one of the most powerful open-source AI models at the time. Building on a Legacy DeepSeek is continuing its tradition of pushing boundaries in open-source AI. Earlier models like DeepSeek-V2.5 and DeepSeek Coder demonstrated impressive capabilities across language and coding tasks, with benchmarks placing it as a leader in the field. The release of R1-Lite-Preview adds a new dimension, focusing on transparent reasoning and scalability. As businesses and researchers explore applications for reasoning-intensive AI, DeepSeek’s commitment to openness ensures that its models remain a vital resource for development and innovation. By combining high performance, transparent operations, and open-source accessibility, DeepSeek is not just advancing AI but also reshaping how it is shared and used. The R1-Lite-Preview is available now for public testing. Open-source models and APIs are expected to follow, further solidifying DeepSeek’s position as a leader in accessible, advanced AI technologies. source

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Gary Shapiro, the CEO of the CTA, has seen tech change over decades. I talked to him about his latest views of politics, tech and economic growth. He was blunt in his responses, and that reminded me of the title of his new book on innovation, dubbed Pivot or Die: How Leaders Thrive When Everything Changes. I spoke with Shapiro a day before the U.S. presidential election. I asked him about politics and tech in one of my early questions. But he first went on to tell me about CES 2025, the big tech trade show in Las Vegas that will happen in early January. After all, that’s job one for Shapiro, who, as CEO of the Consumer Technology Association (CTA), has to make sure the massive trade show for technologists goes off without a hitch. Pivot or Die is Gary Shapiro’s new book. We eventually circled back and talked about some political issues. He was particularly concerned about the U.S. Federal Trade Commission’s “excessive” antitrust enforcement against the tech giants under the leadership of Democratic appointee Lina Khan. We also talked about the turbulent tech economy and how its been affected by the shadow of two major wars in the world. And we addressed the impact of AI on the tech industry. That led us to a discussion of the role of government in the tech industry, when it comes to both support and oversight. And I asked him about the chance we have for balanced growth — where revenues grow, AI gets accepted, and jobs grow too. Here’s an edited transcript of our interview. VentureBeat: It’s good to see you getting another book out. Gary Shapiro is the face of CES. Gary Shapiro: It’s still relevant. We just had a meeting today, the day before the election and 60 days before CES. We talked about the pivots we had made previously in the last few years as an organization. Given the uncertainty of what’s going to happen tomorrow, given the economy and everything else, the only thing that’s certain–well, something will happen. VentureBeat: This is going to run after the election, but what are some of your thoughts that link what you have in the book to your views of politics and tech? Shapiro: The CES is an amazing, powerful tech event. I was looking back at what you had written last year about it, before and after. A lot of people go with a very full agenda, but we always say you have to have time for serendipity and discovery. We have a new look, a new feel. We focused the campaign on “Dive in.” We’re inviting attendees to do three things: connect, solve, and discover. “Connect,” in the technology world–we want people to get together, through B2B and B2C. The statistic we’ve used before is that the average attendee has about 29 meetings during the show. It’s an important business event. You get that face to face. About 75% of attendees say their business is primarily B2B, or both B2B and B2C. We struggle with the name. Some people call it the Consumer Electronics Show, but it’s just CES. There’s so much there that’s B2B. That’s one of the changes that occurs for people when they discover. Another theme of the show is not only to connect, but to solve. We talked about human security for all, our work with the United Nations focusing on the sustainable development goals, focusing on fundamental human security in areas like health care, finances, personal safety. We see that in all sorts of ways. Of course accessibility is another big thing. Even for me, attending the show in 2024, that was one of the biggest surprises and themes. I don’t think I even talked about that before the show – how many people were there looking out on behalf of the disability community, and how many companies were responding to that with all sorts of technology. We have a meeting of the disability community before the show. We had to cut off attendance because there wasn’t enough room. Getting ready to ride in the Goodyear Blimp at CES 2024. Discovery or serendipity is the third theme, the trends we see. Some of it is a continuation, but some of it is new. Obviously AI is still a big thing. It pervades almost every category. Digital health is also very big. Mobility is huge with electric vehicles and connected cars, autonomy, sustainability. We’ve done our own pivots as well. The concept of electricity being available is not something we’ve talked about at CES before. Now we have a whole conference track on it. With electric cars, generative AI, and quantum, they all use tremendous amounts of electricity that we’re not prepared for. Not surprising to me, a number of companies in the last month have announced deals with nuclear power plants and things like that, which is totally new, but it’s a way of dealing with it. We’ll have that. We’ll have exhibitors focusing on energy savings on the supply side, heat reduction, local production. We’ll also have panels talking about how we can look at the electric grid. Similarly to that vein, we’ll have a shift that we haven’t had before to quantum computing. We have a half day of programming on that. While AI and generative AI is currently the thing, now generative AI doesn’t get you to the finish line with a lot of things, for example in health care. It’s a whole shift upwards in computing that we haven’t had in a long time. The other great thing about CES, we have the most powerful group of keynotes we’ve ever had. You heard about Delta earlier this week. We’ll continue to announce more. This is our first keynote in the Sphere, Ed Bastian. Delta is celebrating its 100th anniversary. Their shift to becoming a technology company is

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More A modern-day AI agent consists of, at least, a large language model (LLM) that has been enabled to call some tools. Given the right set of tools for coding, it would start by generating the code, be able to run it in a container, observe the results, modify the code and therefore have a better chance of producing useful code. By contrast, a generative AI model takes some input and, through the process of predicting expectations, produces an output. For example, we give it a coding task, it produces some code, and, depending on the complexity of the task, the code may be usable as is. As they take on different tasks, agents should be allowed to talk to each other. For example, imagine your company intranet with its useful search box directing you to the apps and resources you need. If you are a large enough company, these apps owned by different departments each have their own search boxes. It makes a lot of sense to create agents, maybe by using techniques like retrieval augmented generation (RAG), to augment the search boxes. What does not make sense is to force the user to repeat their query once the search box has identified it as useful given the initial query. Rather, we would prefer the top agent to coordinate with other agents representing various apps and present a consolidated and unified chat interface to you, the user. A multi-agent system representing software or an organization’s various workflows can have several interesting advantages, including improved productivity and robustness, operational resilience and the ability ability to perform faster upgrades of different modules. Hopefully, this article will help you see how this is achieved. But first, how should we go about building these multi-agent systems? Capturing the organization and roles First we should capture the processes, roles, responsible nodes and connections of various actors in the organization. By actors, I mean individuals and/or software apps that act as knowledge workers within the organization. An organizational chart might be a good place to start, but I would suggest starting with workflows, as the same people within an organization tend to act with different processes and people depending on workflows. There are available tools that use AI to help identify workflows, or you can build your own gen AI model. I’ve built one as a GPT which takes the description of a domain or a company name and produces an agent network definition. Because I’m utilizing a multi-agent framework built in-house at my company, the GPT produces the network as a Hocon file, but it should be clear from the generated files what the roles and responsibilities of each agent are and what other agents it is connected to. Note that we want to make sure that the agent network is a directed acyclic graph (DAG). This means that no agent can simultaneously become down-chain and up-chain to any other agent, whether directly or indirectly. This greatly reduces the chances that queries in the agent network fall into a tailspin. In the examples outlined here, all agents are LLM-based. If a node in the multi-agent organization can have zero autonomy, then that agent paired with its human counterpart, should run everything by the human. We will need all processing nodes, be they apps, humans or existing agents, to be represented as agents. Lately there have been many announcements by companies offering specialized agents. We would, of course, want to make use of such agents, if available. We can pull in a preexisting agent and wrap its API into one of our agents so we can make use of our inter-agent communication protocols. This means that such third-party agents will need to have their API available for us to use. How to define agents Various agent architectures have been proposed in the past. For instance, a blackboard architecture requires a centralized point of communication where various agents declare their roles and capabilities, and the blackboard calls them depending on how it plans to fulfill a request (see OAA). I prefer a more distributed architecture that respects the encapsulation of responsibilities. Each agent, having received a request, decides whether it can process it or not, and what it requires to do to process the request, then returns its list of requirements to its requesting up-chain agent. If the agent has down-chains, it asks them if they can help fulfill all or part of the request. If it receives any requirements from the contacted down-chains, it checks with other agents to see if they can fulfill them; if not, it sends them up-chain so that they can ask the human user. This architecture is called the AAOSA architecture and — fun fact — was the architecture used in early versions of Siri. Here is a sample system prompt that can be used to turn an agent into an AAOSA agent. When you receive an inquiry, you will: Call your tools to determine which down-chain agents in your tools are responsible for all or part of it Ask down-chain agents what they need to handle their part of the inquiry. Once requirements are gathered, you will delegate the inquiry and the fulfilled requirements to the appropriate down-chain agents. Once all down-chain agents respond, you will compile their responses and return the final response. You may, in turn, be called by other agents in the system and have to act as a down-chain to them. In addition to the set of roles and responsibilities defined in natural language in each agent’s system prompt, agents may or may not include tools that they can call, with various arguments being passed to the tools. For instance, a product manager agent may need to be able to process various tickets on a virtual Kanban board, or an alerts agent may need to call a tool to issue alerts in an alerting system. Current

Presented by Celonis Is the smart use of process intelligence the best route to digital transformation? BMW Group thinks so. Over the last eight years, the €155-billion ($US167-billion) automaker has increasingly focused its efforts to create a more agile, efficient and innovative global company by giving employees advanced tools and AI to model, analyze and optimize manufacturing, sales, service and other end-to-end processes, including its supply chains. BMW’s commitment is not merely a cost-cutting exercise — it’s a forward-thinking strategy designed to maintain competitive edge in a world of “VUCA” (volatility, uncertainty, complexity and ambiguity) says Dr. Patrick Lechner, BMW Head of Process Intelligence, Robotics Process Automation, & Low-Code/No-Code.  By embracing process intelligence (PI) and artificial intelligence (AI) across its operations, BMW aims to streamline processes, boost efficiency and enable data-driven decisions at every level of the company. “The automotive world is changing so fast, and the nature of change has changed,” Lechner says. “Electric vehicle adoption is increasingly fast. There are new players and new sales models, including online.  So, we really have to adapt constantly to challenges by new competitors, to new demands in the outside world. By optimizing our processes, we create additional benefit for our customers.” BMW’s benefits from optimized processes BMW credits process improvements for impressive and growing benefits across the company. Among them: Enhanced production efficiency. By analyzing the minutiae of its production line processes, BMW has optimized resource allocation, minimizing delays and waste. This process overhaul has led to improved manufacturing timelines, shorter production cycles and cost efficiencies.                     Supply chain optimization. BMW’s global supply chain is complex and involves hundreds of suppliers and distributors. By utilizing process intelligence to track each step and transaction, BMW has streamlined its inventory management, reducing excess stock and ensuring parts are available when needed.       Increased Automation. Through its citizen developer program, BMW employees have implemented over 1,100 business process automations, reducing manual work and fostering a culture of continuous improvement. Customer service and experience improvements. By examining the processes behind customer support and warranty claims, BMW has been able to speed up service response times and resolve issues faster. This not only reduces costs but also boosts customer satisfaction. BMW leads a quiet global revolution In an era when automotive innovation extends far beyond engines and chassis, BMW’s commitment, scale of adoption and value realized has made it a leader in a quiet global revolution that uses advanced tools, process intelligence and AI to find and unlock hidden value, notes Lars Reinkemeyer, editor of Process Intelligence in Action, and chief evangelist at Celonis, the pioneer and leading global market vendor in the space.  Every one of the 2.5 million cars sold in 140 countries by BMW Group has been touched by at least one optimized Celonis process. BMW officials say that’s just the start. Eight years after two initial proof-of-concepts in a Munich plant, business and technology leaders at the automaker continue to put the “pedal to the metal” to transformation driven by applying process intelligence to high-value business cases. Recent implementation of the Celonis Process Intelligence Platform provides a 360-degree digital overview of the entire supply chain — from vehicle development to customer delivery and subsequent service. Having a holistic view lets BMW visualize, analyze and refine its entire operational landscape, enabling faster, data-driven decision-making that keeps the company agile. In March, BMW and Celonis announced the deepening of a strategic alliance to develop new process innovations. And at the annual Celosphere event in Munich in October, the automaker discussed several new initiatives to expand ecosystem-wide process optimization. BMW’s process intelligence initiatives roll on Long known for its engineering prowess, BMW continues to leverage advanced process intelligence tools to ensure that every facet of its business, from parts procurement to customer support, is running as smoothly and efficiently as its legendary engines. Among key initiatives for 2025: Expanded use of predictive analytics. The company plans to use predictive insights provided by process intelligence to anticipate and mitigate supply chain disruptions. The goal is to forecast demand and supply with greater accuracy, enabling the company to adjust its logistics strategies and avoid bottlenecks before they arise. Real-time process monitoring for customer service. With customer expectations continually rising, BMW is extending its process intelligence capabilities into customer service workflows, tracking the progress of each case in real-time to minimize delays and deliver a superior service experience. Sustainability monitoring across processes. A major initiative aims to make all processes more sustainable. By using process intelligence to identify high-energy consumption points within its operations, BMW plans to take targeted actions to cut down on emissions and waste. Continued expansion to non-expert users. Democratizing the use of process tools has been a key goal of BMW efforts from the beginning. To this end, the company is rolling out Copilots and conversational AI agents to help non-expert users understand how to optimize their daily work and improve training, self-service and user experience. Extending beyond the organization. BMW is working to extend process intelligence capabilities across its entire network of suppliers and dealers to ensure their operations align with BMW’s high standards. By extending process intelligence across this ecosystem, BMW can ensure consistent quality, minimize delivery delays, better manage inventory and reduce costs throughout the production lifecycle. Innovating beyond core products BMW’s ongoing commitment to process intelligence and optimization is a world-class example of how a century-old legacy brand is innovating beyond its core products. As a highly complex organization, the BMW Group faces processes that span multiple divisions, from vehicle manufacturing to supply chain management. As global markets change, customer demands increase and new technologies emerge, Lechner says, BMW recognizes the critical importance of making these processes as agile and efficient as possible. Process intelligence takes data from systems like ERPs, CRMS, and Excel and uses process intelligence and AI technology, augmented with business context to create a living, moving digital twin of a business’s end-to-end processes. It then lets BMW extract valuable insights to pinpoint inefficiencies that would be hard to see with traditional analysis methods.

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Adobe researchers have created a breakthrough AI system that processes documents directly on smartphones without internet connectivity, potentially transforming how businesses handle sensitive information and how consumers interact with their devices. The system, called SlimLM, represents a major shift in artificial intelligence deployment — away from massive cloud computing centers and onto the phones in users’ pockets. In tests on Samsung’s latest Galaxy S24, SlimLM demonstrated it could analyze documents, generate summaries, and answer complex questions while running entirely on the device’s hardware. “While large language models have attracted significant attention, the practical implementation and performance of small language models on real mobile devices remain understudied, despite their growing importance in consumer technology,” explained the research team, led by scientists from Adobe Research, Auburn University, and Georgia Tech. How small language models are disrupting the cloud computing status quo SlimLM enters the scene at a pivotal moment in the tech industry’s shift toward edge computing — a model in which data is processed where it’s created, rather than in distant data centers. Major players like Google, Apple, and Meta have been racing to push AI onto mobile devices, with Google unveiling Gemini Nano for Android and Meta working on LLaMA-3.2, both aimed at bringing advanced language capabilities to smartphones. What sets SlimLM apart is its precise optimization for real-world use. The research team tested various configurations, finding that their smallest model — at just 125 million parameters, compared to models like GPT-4o, which contain hundreds of billions — could efficiently process documents up to 800 words long on a smartphone. Larger SlimLM variants, scaling up to 1 billion parameters, were also able to approach the performance of more resource-intensive models, while still maintaining smooth operation on mobile hardware. This ability to run sophisticated AI models on-device without sacrificing too much performance could be a game-changer. “Our smallest model demonstrates efficient performance on [the Samsung Galaxy S24], while larger variants offer enhanced capabilities within mobile constraints,” the researchers wrote. Why on-device AI could reshape enterprise computing and data privacy The business implications of SlimLM extend far beyond technical achievement. Enterprises currently spend millions on cloud-based AI solutions, paying for API calls to services like OpenAI or Anthropic to process documents, answer questions, and generate reports. SlimLM suggests a future where much of this work could be done locally on smartphones, significantly reducing costs while improving data privacy. Industries that handle sensitive information — such as healthcare providers, law firms, and financial institutions — stand to benefit the most. By processing data directly on the device, companies can avoid the risks associated with sending confidential information to cloud servers. This on-device processing also helps ensure compliance with strict data protection regulations like GDPR and HIPAA. “Our findings provide valuable insights and illuminate the capabilities of running advanced language models on high-end smartphones, potentially reducing server costs and enhancing privacy through on-device processing,” the team noted in their paper. Inside the technology: How researchers made AI work without the cloud The technical breakthrough behind SlimLM lies in how the researchers rethought language models to meet the hardware limitations of mobile devices. Instead of merely shrinking existing large models, they conducted a series of experiments to find the “sweet spot” between model size, context length, and inference time, ensuring that the models could deliver real-world performance without overloading mobile processors. Another key innovation was the creation of DocAssist, a specialized dataset designed to train SlimLM for document-related tasks like summarization and question answering. Instead of relying on generic internet data, the team tailored their training to focus on practical business applications, making SlimLM highly efficient for tasks that matter most in professional settings. The future of AI: Why your next digital assistant might not need the internet SlimLM’s development points to a future where sophisticated AI doesn’t require constant cloud connectivity, a shift that could democratize access to AI tools while addressing growing concerns about data privacy and the high costs of cloud computing. Consider the potential applications: smartphones that can intelligently process emails, analyze documents, and assist with writing — all without sending sensitive data to external servers. This could transform how professionals in industries like law, healthcare, and finance interact with their mobile devices. It’s not just about privacy; it’s about creating more resilient and accessible AI systems that work anywhere, regardless of internet connectivity. For the broader tech industry, SlimLM represents a compelling alternative to the “bigger is better” mentality that has dominated AI development. While companies like OpenAI are pushing toward trillion-parameter models, Adobe’s research demonstrates that smaller, more efficient models can still deliver impressive results when optimized for specific tasks. The end of cloud dependence? The (soon-to-be) public release of SlimLM’s code and training dataset could accelerate this shift, empowering developers to build privacy-preserving AI applications for mobile devices. As smartphone processors continue to evolve, the balance between cloud-based and on-device AI processing could tip dramatically toward local computing. What SlimLM offers is more than just another step forward in AI technology; it’s a new paradigm for how we think about artificial intelligence. Instead of relying on vast server farms and constant internet connections, the future of AI could be personalized, running directly on the device in your pocket, maintaining privacy, and reducing dependence on cloud computing infrastructure. This development marks the beginning of a new chapter in AI’s evolution. As the technology matures, we may soon look back on cloud-based AI as a transitional phase, with the true revolution being the moment AI became small enough to fit in our pockets. source

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Mistral, the French startup that made waves last year with a record-setting seed funding amount for Europe, has launched a slew of updates today including a new, large foundational model named Pixtral Large. The company is further upgrading its free web-chased chatbot, Le Chat, adding image generation, web search, and an interactive “canvas,” matching the features of and turning it into a more serious and direct competitor to OpenAI’s ChatGPT. As Mistral AI CEO and co-founder Arthur Mensch wrote on his account on the social network X, “At Mistral, we’ve grown aware that to create the best AI experience, one needs to co-design models and product interfaces. Pixtral was trained with high-impact front-end applications in mind and is a good example of that.” Users who want to try out the new Le Chat features will need to enable them as beta features on the web interface. Note that Le Chat access does require a free Mistral, Google, or Microsoft account to use. Pixtral Large — open source multimodal AI Pixtral Large, Mistral’s new 124-billion-parameter model, builds upon its predecessor, Mistral Large 2, unveiled over the summer 2024, as well as its first multimodal model, Pixtral 12-B, released in September. It includes a 123-billion-parameter decoder and a 1-billion-parameter vision encoder, enabling it to excel in both text and visual data processing. Parameters, as you’ll recall, refer to the number of settings that govern a model’s inputs and outputs, with more parameters generally connoting a more capable, knowledgable and performant model. According to a post by Mistral Head of Developer Relations Sophia Yang to her X account, Pixtral Large excels at “multilingual OCR [optical character recognition], reasoning, chart understanding, and more.” Yang included a screenshot of Pixtral Large in Le Chat analyzing a receipt uploaded by a user using OCR, showing its capabilities for ingesting and documenting expenses, as well as in this case, splitting a bill with a tip included. With a context window of 128,000 tokens, Pixtral Large is able to handle up to 30 high-resolution images per input or around a 300-page book, again equivalent to leading OpenAI GPT series models. The model demonstrates state-of-the-art performance across diverse benchmarks, including MathVista, DocVQA, and VQAv2, making it ideal for tasks like chart interpretation, document analysis, and image understanding. While the model and weights are available for download freely on Hugging Face, they are released under a custom Mistral AI Research License, which specifies only non-commercial, research-focused applications. Those looking to use it commercially will need to do so through Mistral’s API on its Le Platforme managed web service, or obtain a separate license from the company directly through a contact form, meaning it is not actually fully open source. Still, by offering Pixtral Large, Mistral AI empowers researchers and developers to harness advanced multimodal AI while ensuring responsible and ethical use. Le Chat comes for ChatGPT with rival matching features At the center of Mistral’s AI tools is Le Chat, a free platform now enhanced with new features powered by Pixtral Large. Designed for diverse use cases like research, ideation, and automation, Le Chat integrates text, vision, and interactive functionalities into a seamless productivity experience. New Features of Le Chat: 1. Web Search with Citations: Users can supplement the AI’s knowledge with real-time web searches, complete with source citations for transparency. 2. Canvas for Ideation: This innovative interface allows users to create, modify, and collaborate on documents, presentations, and designs in an interactive new space that appears to the left of the chatbot interface. As Yang wrote about it on X: Le Chat Canvas is “great for creative ideation. You can use Canvas to create documents, presentations, code, mockups… the list goes on.” It comes just six weeks after OpenAI released its own Canvas sidebar interactive element for ChatGPT, which many viewed as a feature designed to rival Anthropic’s earlier Artifacts release for its Claude chatbot. 3. Advanced Document and Image Analysis: With Pixtral Large, Le Chat can now process and summarize complex PDFs, extracting insights from graphs, tables, equations, and more. 4. Image Generation: Through a partnership with separate image model startup Black Forest Labs, Le Chat now includes image generation capabilities powered by the Flux Pro model, enabling users to produce high-quality visuals directly in the chat interface. This is a clear answer to OpenAI’s DALL-E 3 integration in ChatGPT (both models from OpenAI, however) as well as the second big integration of Black Forest Labs’ new models into a leading AI foundation model provider’s offerings, following its earlier team-up with Elon Musk’s xAI to power image generation in that company’s Grok-2 chatbot available through X, the social network Musk also owns. 5. Task Agents for Automation: Customizable agents automate repetitive tasks like summarizing meeting minutes, processing invoices, or scanning receipts, saving users time and effort. These features position Le Chat as a versatile AI assistant, capable of handling tasks traditionally requiring multiple tools. Mistral AI highlights Le Chat’s comprehensive feature set and its accessibility compared to platforms like ChatGPT, Perplexity, and Claude. While competitors may require premium subscriptions for similar functionalities, Le Chat provides an integrated, multimodal experience entirely for free during its beta phase. Mistral is coming to play hard With Pixtral Large and the enhanced Le Chat, Mistral is flexing its research and development muscles. Even as some in the tech industry believe that the cost of intelligence is being driven down and making life more difficult for model providers to find revenue streams, Mistral isn’t giving up on advancing its offerings to compete with the other leaders in the field, and doing so on fewer parameters — 124 billion compared to say, 405 billion from Meta’s latest Llama 3.1 release. However, Mistral is still missing some of the advanced voice and audio features found on rivals such as OpenAI’s ChatGPT Advanced Voice Mode or Google’s Gemini Live. A recent survey by Kong showed despite

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More There is no doubt AI agents will continue to be a fast-growing trend in enterprise AI. But as more companies look to deploy agents, they’re also looking for a way to help them make sense of the many actions these autonomous or semi-autonomous, AI guided bots will take, and avoid conflicts. To combat the potential sprawl of different AI agents deployed by users, service providers and enterprises alike have been building another type of AI agent: the orchestrator agent. Enter the orchestrator: these type of agents function as managers of other, more specialized agents, understanding each one’s role and activating each based on the next steps needed to finish a task. Most orchestrator agents, sometimes called meta agents, monitor if an agent succeeded or failed and choose the following agent to trigger to get the desired outcome. Good orchestrator agents exhibit certain features that make these work different from other agents, and for enterprises, elements make them work much better.  Integration Agentic ecosystems would eventually bring workflows together, even if the task involves talking to an agent outside the current platform. Orchestrator agents need to have robust integrations with other systems. Otherwise, agents remain an island able to communicate only with itself.  ServiceNow vice president of AI and Innovation Dorit Zilbershot said enterprises need to investigate if the orchestration agents they’re building or buying offer integration points to other systems.  “Effective orchestration agents support integrations with multiple enterprise systems, enabling them to pull data and execute actions across the organizations,” Zllbershot said. “This holistic approach provides the orchestration agent with a deep understanding of the business context, allowing for intelligent, contextual task management and prioritization.” For now, AI agents exist in islands within themselves. However, service providers like ServiceNow and Slack have begun integrating with other agents. Slack announced it offers integration for agents from Salesforce, Workday, Asana and Cohere. Full stack AI company Writer connects its agents to Amazon and Macy’s APIs so customers can directly sell products.  Don Schuerman, CTO at Pega, echoed the sentiment, saying an ideal orchestration agent is “API-centric so it can work both across agents but also across human-centric channels so that humans can be pulled in when needed.”  Knowledge of enterprise processes Like all agents, orchestrator agents need to know how the business works.  Orchestrator agents need a more holistic view of the best next step while moving the process forward. Zilbershot said a good orchestration agent “should be able to quickly analyze the context to determine both the best-suited AI agent and the optimal sequence of AI agent assignments to optimize workflows and minimize delays.” It’s not just about having insight into company data — though that is another essential component for agentic ecosystems — it’s also about understanding the processes enterprises do to run their business.  Writer CEO May Habib told VentureBeat in an earlier interview that enterprises that want an effective agentic system provide the workflow for an orchestrator agent to follow, not the other way around.  “If you don’t get the nodes in a workflow right, then the automated workflow is just moving crap from one system to another,” Habib said. “Over time, we built an application that, automatically with AI, knows based on the workflow suggests which tools to access.”  Reasoning capabilities Due to its nature, orchestrator agents make reasoning decisions more than other AI agents. As AI agents are tasked with more complex tasks, so will the orchestrator agents that help manage them.  Large language models underpin agent creation, and models with greater reasoning capabilities can run different scenarios before triggering the next agent. Orchestrator agents must have strong reasoning skills to ensure the workflow doesn’t break down.  Smooth communication between agents and human employees ServiceNow’s Zilbershot pointed out that orchestration agents are primarily responsible for the interaction between humans and agents. She said enterprises deploying AI agents would benefit from orchestrator agents with user-friendly interfaces and feedback networks. Hence, the agents continue to improve based on how employees interact and use them.  “By serving as the connective tissue between specialized AI agents and human operators, orchestration agents make it exponentially easier to not only streamline operations but also enhance the overall effectiveness of an organization’s agentic AI system,” she said.  Although AI agents are designed to go through workflows automatically, experts said it’s still important that the handoff between human employees and AI agents goes smoothly. The orchestration agent allows humans to see where the agents are in the workflow and lets the agent figure out its path to complete the task.  “An ideal orchestration agent allows for visual definition of the process, has rich auditing capability, and can leverage its AI to make recommendations and guidance on the best actions. At the same time, it needs a data virtualization layer to ensure orchestration logic is separated from the complexity of back-end data stores,” said Pega’s Schuerman.  Orchestrator agents already ship out in many agent frameworks. It can even be a differentiator for many agent libraries in the future. As enterprises continue experimenting more with agents, orchestrator agents may improve.  source

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More The artificial intelligence arms race heated up Friday as Amazon announced an additional $4 billion investment in Anthropic, doubling its stake to $8 billion in a move that signals the cloud computing giant’s ambitious bid to compete with Microsoft and Google in the fast-evolving AI landscape. The deal, which maintains Amazon as a minority investor, establishes AWS as Anthropic’s primary cloud and training partner. Most significantly, it commits Anthropic to using Amazon’s custom-designed Trainium and Inferentia chips for training and deploying its advanced AI models — a major win for Amazon’s semiconductor strategy. Amazon’s calculated investment positions the company at the center of AI infrastructure development. While Microsoft has captured headlines and market momentum through its OpenAI partnership, Amazon is taking a different approach by building a comprehensive AI stack from silicon to software. The commitment to use AWS Trainium chips is particularly notable, as it gives Amazon’s custom silicon program the kind of high-profile validation it needs to compete with Nvidia’s dominance in AI acceleration. How Amazon plans to challenge Microsoft’s AI dominance The expanded partnership has already shown promising results. According to Anthropic, tens of thousands of customers are using its Claude models through Amazon Bedrock, including major enterprises like Pfizer, which reported tens of millions in operational cost savings. The European Parliament has also adopted Claude to power a document analysis system that processes 2.1 million official documents. The timing of this deal aligns with a crucial shift in enterprise AI adoption. As companies move from AI experimentation to production deployment, they’re increasingly focused on security, scalability, and cost-effectiveness. By integrating Anthropic’s technology directly into the AWS ecosystem, Amazon is positioning itself to capture this next wave of enterprise AI adoption. Inside the battle for AI cloud supremacy This move significantly reshapes the competitive dynamics in cloud AI services. While Microsoft’s OpenAI partnership gave it an early lead in the generative AI race, Amazon’s deeper integration with Anthropic could prove more sustainable in the long run. The focus on custom silicon and tight hardware-software integration mirrors the successful playbook Apple used in personal computing — but at cloud scale. The deal also creates an interesting dynamic with Google, which invested $2 billion in Anthropic last year. With both tech giants now holding significant stakes, Anthropic has effectively positioned itself as a Switzerland of sorts in the AI wars, maintaining independence while leveraging the resources of multiple tech giants. What Amazon’s AI investment means for enterprise technology For enterprise customers, this partnership addresses several critical concerns. First, it promises more cost-effective AI deployment through optimization for AWS’s custom chips. Second, it provides a clear path to scale AI applications through Amazon’s global infrastructure. Perhaps most importantly, it offers a more secure and compliant way to adopt advanced AI capabilities. Anthropic’s latest Computer Use feature, which allows AI to operate computers like humans, will be available first to AWS customers. This exclusivity period could give Amazon’s enterprise customers a significant head start in automating complex workflows. The future of cloud computing: AI takes center stage The real significance of this deal lies in its long-term implications for the cloud computing industry. As AI becomes increasingly central to enterprise operations, the ability to offer optimized, integrated AI services could become the key differentiator in the cloud market. Amazon’s investment suggests a belief that the future of cloud computing will be built on AI infrastructure. The deal also reflects a broader industry trend toward vertical integration in AI, with major players seeking to control every layer of the stack from chips to applications. This could lead to a more concentrated market structure, with a few large players dominating the AI infrastructure landscape. As enterprise AI adoption accelerates, this partnership could prove pivotal in determining which technology giants emerge as the dominant forces in the AI era. With the generative AI market projected to exceed $1 trillion within the decade, Amazon’s expanded investment in Anthropic represents a strategic bet on shaping the future of artificial intelligence. source

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More As we’ve covered here before at VentureBeat, the cloud computing wars have swiftly morphed into the AI wars, with leading cloud computing divisions Google Cloud, Microsoft Azure, and Amazon Web Services (AWS) all rolling out new tools for customers to access, use, deploy, and build atop a range of AI models. Therefore, it was not too surprising to learn his week that Google Cloud was offering a new AI agent ecosystem program called AI Agent Space. This initiative empowers businesses to discover, deploy, and co-create AI agents designed to automate tasks, enhance customer experiences, and optimize operations. With a growing focus on the enterprise, Google’s announcement positions it as a major player alongside competitors like Microsoft, SAP, and Salesforce. Google’s ecosystem is built around enabling partners to develop highly customizable AI agents by providing them with robust tools and resources, including early previews of Google’s AI technologies, direct support from engineering teams, and best practices to streamline development. In addition, Google says it will promote new agents through its Google Cloud Marketplace to allow partners to scale the agents they make to new, interested audiences. It makes sense and is a win-win for Google, showing that it has a robust catalog of agents for companies to select from and ideally, more users coming to it and Google Cloud as a result. Current agents built by enterprises atop Google AI models and Google Cloud Interestingly, Google chose the announcement of AI Agent Space to highlight other agents built internally by its Google Cloud customers and partners, though these solutions aren’t yet, for the most part, available on the AI Agent Space. Presumably, Google wanted to show what is possible with its tools and how businesses could port their internal agents over to the AI Agent Space and sell them as SaaS solutions there. Among them are: “Accenture is transforming customer support at a major retailer by offering convenient self-service options through virtual assistants, enhancing the overall customer experience. Bain supports SEB’s wealth management division with an AI agent that enhances end-customer conversations with suggested responses and generates call summaries that help increase efficiency by 15%.   BCG provides a sales optimization tool to improve the effectiveness and impact of insurance advisors.  Capgemini optimizes the ecommerce experience by helping retailers accept customer orders through new revenue channels and to accelerate the order-to-cash process for digital stores. Cognizant helps legal teams draft contracts, assigning risk scores and recommendations for how to optimize operational impact.   Deloitte offers a “Care Finder” agent as part of its Agent Fleet, helping care seekers find in-network providers often in less than a minute — significantly faster than the average call time of 5-8 minutes. HCLTech helps predict and eliminate different types of defects on manufacturing products with its manufacturing quality agent, Insight. Infosys optimizes digital marketplaces for a leading consumer brand manufacturer, providing actionable insights on inventory planning, promotions, and product descriptions.  PwC uses AI agent technology to help oncology clinics streamline administrative work so that doctors can optimize their time with patients. TCS helps build persona-based AI agents contextualized with enterprise knowledge to accelerate software development. Wipro supports a national healthcare provider in using agent technology to develop and adjust contracts, streamlining a complex and time-consuming task while improving accuracy.“ Competing Solutions from Microsoft, SAP, and Salesforce Google’s announcement came on the heels of similarly positioned AI agent initiatives from rivals including Microsoft with its Copilot Studio, which has emerged as a leader in the AI agent space. More than 100,000 organizations creating or editing AI agents since its launch. At its Ignite conference in November 2024, Microsoft announced major updates, including integration with 1,800 large language models (LLMs) in Azure, offering enterprises unparalleled flexibility, autonomous agents, and multi-agent collaboration through its “agent mesh” architecture. This makes it a strong option for large organizations with complex IT landscapes Meanwhile, SAP recently updated its Joule AI assistant to offer collaborative AI agents designed to break down silos and unify workflows across business functions. Joule’s agents work collectively to address challenges such as payment disputes or supply chain disruptions, leveraging SAP’s deep integration with ERP, CRM, and HR systems. Furthermor, Salesforce’s Agentforce, launched in September 2024, integrates AI agents into its vast ecosystem, leveraging its Data Cloud to enhance service, sales, and marketing functions. While Microsoft and SAP focus on enterprise integration and cross-functional workflows, and Salesforce leans on low-code accessibility, Google stands out for its partner-driven flexibility and open ecosystem. By empowering partners to co-develop agents tailored to specific industries, Google fosters innovation while offering customers a diverse range of solutions. Google’s marketplace model ensures that businesses can choose from a variety of pre-built agents or work with partners to create custom solutions. This contrasts with Microsoft’s large-scale, infrastructure-driven approach and SAP’s process-unifying agents, making Google’s ecosystem particularly appealing to organizations with diverse and evolving needs. Yet, in terms of raw numbers, Google’s AI Studio remains behind the competition. The Google blog post announcement noted that its AI Agent Space is already live with “solutions from select partners” and the company “plan[s] to add hundreds of additional AI agents over the coming months.” But at present, VentureBeat’s viewing of the AI Agent Space revealed only 19 distinct agent models available for use, a fry cry from the hundreds or thousands available on cloud rivals’ solutions. source