Tag: VentureBeat

Andrej Karpathy, known for his work at Tesla and OpenAI, has unveiled an open-source project called autoresearch. This project aims to automate the scientific method using AI agents, allowing them to run experiments autonomously while humans sleep.

The core concept involves an AI agent with a fixed compute budget making autonomous decisions to optimize a training script, evaluating results based on validation loss. In a single overnight run, Karpathy’s agent completed 126 experiments, showcasing a significant efficiency gain in model tuning.

The impact of autoresearch extends beyond machine learning, as demonstrated by its application in marketing. Eric Siu, founder of Single Grain, highlighted the potential for marketing teams to run 36,500+ experiments per year, transforming the industry’s approach to data-driven decision-making.

The rapid adoption and adaptation of autoresearch across various fields underscores its potential. By automating and accelerating experimentation processes, this innovative approach enhances efficiency and challenges traditional research practices.

Source: VentureBeat

Recent research by Morgan Stanley indicates a potential shift in U.S. commerce spending towards agentic AI by 2030, reaching up to $385 billion. In response to this trend, Azoma, an e-commerce startup, has introduced the Agentic Merchant Protocol (AMP). This framework aims to streamline product visibility for high-volume retailers in an increasingly autonomous shopping ecosystem.

AMP addresses the challenge of maintaining brand integrity and control in a landscape where AI agents often rely on unverified sources, leading to a ‘black box’ effect. By centralizing product intelligence and legal guidelines, AMP provides a standardized system to ensure consistent brand representation.

The platform caters specifically to consumer goods companies, emphasizing the importance of maintaining a unified brand identity across various AI platforms.

Key features of AMP include Canonical Machine-Native Catalogues, Programmatic Open Web Distribution, Agent-Agnostic Infrastructure, and Performance Visibility, offering brands a comprehensive toolkit for managing their presence in the agentic commerce landscape.

Moreover, Azoma’s proprietary RegGuard™ Compliance engine enables automated content audits against regulatory standards, driving performance improvements for early partners. By optimizing content visibility and addressing technical barriers, Azoma empowers brands to enhance their AI engagement and drive revenue growth.

Azoma’s pricing strategy aims to transition towards outcome-based pricing, aligning costs with measurable performance outcomes. This shift reflects the company’s vision of tying revenue directly to successful agentic interactions, creating a more transparent and value-driven e-commerce model.

Source: VentureBeat

Microsoft has introduced Copilot Cowork, a cloud-based AI automation tool that extends across various Microsoft applications, revolutionizing how users interact with AI technology. This new feature, developed in collaboration with Anthropic, enhances Microsoft’s existing AI tool 365 Copilot, enabling users to delegate complex, multi-step tasks to an AI agent that seamlessly navigates and utilizes the functionalities of Microsoft’s suite of apps including Outlook, Teams, Excel, and PowerPoint.

Copilot Cowork is a key part of Microsoft’s ‘Wave 3’ update for Microsoft 365 Copilot, offering agentic capabilities within individual Office apps, integrating Anthropic’s Claude models into Copilot Chat, and introducing new enterprise pricing tiers that bundle AI productivity with security and governance features.

While resembling Anthropic’s ‘Claude Cowork’ applications, Microsoft’s Copilot Cowork operates uniquely in the cloud within Microsoft 365’s infrastructure. This distinction allows Copilot Cowork to access a user’s enterprise data graph, combining signals from various Microsoft applications for seamless task execution.

This move signifies Microsoft’s shift towards transforming Copilot into an ‘execution layer’ AI, capable of proactively completing tasks on behalf of users rather than merely providing responses.

With Copilot Cowork currently in Research Preview, Microsoft aims to offer wider access through its Frontier program by late March 2026. The company’s strategic approach emphasizes deep integration with the existing M365 ecosystem, catering to enterprise users who prioritize seamless AI task automation within a secure and governed environment.

Source: VentureBeat

OpenAI has introduced a significant upgrade to its AI model, GPT-5.4, which promises to revolutionize computer use and financial analysis. The new model, available in two versions – GPT-5.4 Thinking and GPT-5.4 Pro, introduces groundbreaking features that enhance productivity and efficiency across various industries.

One of the key capabilities of GPT-5.4 is its ‘native’ Computer Use mode, allowing the AI to navigate a user’s computer seamlessly and work across applications. This marks a step towards autonomous workflows, enabling agents to carry out multi-step tasks efficiently.

Additionally, OpenAI has integrated financial plugins that directly integrate GPT-5.4 into Microsoft Excel and Google Sheets, empowering users with advanced financial reasoning and modeling capabilities. These integrations aim to provide granular analysis and automated task completion, enhancing productivity in the financial sector.

The model also boasts impressive performance improvements, using fewer tokens and supporting up to 1 million tokens of context. Its enhanced web browsing capabilities and document handling further solidify its position as a versatile and reliable AI solution.

Developers and coders will also benefit from GPT-5.4’s enhanced tool search functionality, which reduces token usage by 47% while maintaining accuracy. The model’s coding prowess, combined with its state-of-the-art computer-use capabilities, make it a valuable asset for complex, multi-step tasks.

OpenAI’s pricing strategy for GPT-5.4 reflects its advanced capabilities, with the Pro version catering to more complex tasks at a higher cost. Despite the premium pricing, the model remains competitive in the AI landscape, offering value for its cutting-edge features and performance.

Overall, GPT-5.4 represents a significant advancement in AI technology, empowering users with unprecedented computer-use capabilities and enhanced financial analysis tools. The model’s focus on efficiency, reliability, and reduced errors underscores its potential to transform professional workflows across various industries.

Source: VentureBeat

San Francisco-based Anthropic has announced the launch of Claude Marketplace, a platform aimed at simplifying AI procurement for enterprises. Despite its dispute with the U.S. Department of War, the company’s new offering allows businesses with existing Anthropic commitments to allocate a portion of their spending towards tools powered by Anthropic’s Claude models but developed by external partners like GitLab, Harvey, and Replit. The initiative is designed to streamline procurement processes and consolidate AI spending for enterprises, as highlighted in Anthropic’s Claude Marketplace FAQ.

Anthropic’s move with Claude Marketplace raises questions about how enterprises will choose to leverage Claude – either directly through Anthropic’s products and APIs or through third-party applications embedding Claude for specialized workflows. The platform’s integration capabilities and focus on customizability align with current trends in enterprise AI adoption, providing users with access to a range of tools for tailored workflows. This development comes amidst a growing landscape of AI marketplaces, with efforts from companies like OpenAI, Lightning AI, and Salesforce to surface AI agents catering to diverse customer needs.

The introduction of Claude Marketplace signifies Anthropic’s commitment to enhancing AI tool accessibility for enterprises, enabling them to leverage the best Claude-powered solutions seamlessly. With the potential for Claude to act as an orchestrator, managing multiple tools within enterprise workflows, the platform offers a centralized approach to AI integration and procurement.

While adoption remains a key challenge, Anthropic’s strategic move with Claude Marketplace reflects a broader industry shift towards more efficient AI tool procurement and utilization within enterprise settings.

Source: VentureBeat

Researchers at MIT have introduced a new technique called Attention Matching that enables the compression of the Key-Value (KV) cache by up to 50 times with minimal loss in quality, significantly enhancing memory efficiency for large language models without compromising accuracy. The KV cache, crucial for processing sequential responses efficiently, grows in size as the conversation lengthens, posing a significant hurdle for serving models with ultra-long contexts.

Attention Matching focuses on preserving specific mathematical properties during compression, such as attention output and attention mass, ensuring that the compressed memory behaves identically to the original, even with unpredictable user prompts. This method bypasses the computationally intensive gradient-based optimization of previous techniques, making it orders of magnitude faster while maintaining high compaction ratios and quality.

Experiments by the researchers demonstrate that Attention Matching can compress the KV cache by 50 times, offering substantial memory savings and processing speed advantages over existing methods. Enterprises exploring AI applications that demand efficient memory utilization can leverage the benefits of this innovative technique to optimize performance without sacrificing accuracy.

Source: VentureBeat

Google’s product manager, Shubham Saboo, has introduced a new open-source project that aims to redefine persistent memory technology. Saboo presented the ‘Always On Memory Agent’ on Google Cloud Platform’s Github, marking a shift from traditional vector databases to a novel LLM-driven approach.

The project, developed using Google’s Agent Development Kit and Gemini 3.1 Flash-Lite model, aims to address the challenge of continuously ingesting and storing information without relying on vector databases, offering a fresh perspective on agent infrastructure.

This agent system, designed for autonomy and consolidation of memories, emphasizes simplicity by utilizing SQLite for structured memory storage and eschewing traditional retrieval stacks in favor of a specialized memory layer.

Google’s Flash-Lite model complements the Always On Memory Agent by providing high-speed, cost-efficient processing tailored for tasks like translation and UI generation.

While the release showcases the potential of continuous memory for enterprise AI applications, it also raises governance concerns around data retention, compliance, and scalability, prompting a deeper exploration of the trade-offs in memory design.

Source: VentureBeat

Alibaba’s renowned Qwen AI team, known for its impactful open-source generative models, is experiencing significant upheaval following the departure of key members after the release of the Qwen3.5 small model series. The exit of technical lead Junyang ‘Justin’ Lin, along with other team members, has raised concerns about the team’s future and commitment to open-source efforts.

The Qwen3.5 models, recognized for their efficient reasoning capabilities, represent a milestone in algorithm-hardware co-design. However, the departures have cast uncertainty over the team’s trajectory, especially with the appointment of a new leader, potentially indicating a shift towards metric-driven strategies.

Amidst speculations of a ‘Gemini-fication’ trend, reminiscent of industry shifts seen at other tech giants, concerns loom over the fate of Qwen’s open-source ethos. The enterprise community faces uncertainties about the future accessibility of Qwen models, hinting at a possible transition towards proprietary offerings to meet business objectives.

As internal tensions and structural changes unfold at Alibaba, the AI community closely monitors how Qwen’s legacy of openness and innovation will evolve in the face of leadership transitions and strategic realignments.

Source: VentureBeat

Google has introduced a new command-line interface (CLI) for Google Workspace, providing a unified interface for accessing applications like Gmail, Docs, Sheets, and more. This move aims to streamline interactions for both human users and AI agents, enabling developers to automate tasks more efficiently.

The CLI, named googleworkspace/cli, offers structured JSON output and agent-centric workflows, making it easier for users to execute tasks directly within the terminal. Features like per-resource help, dry-run previews, and schema inspection enhance the ability of developers and AI systems to interact with Workspace data effectively.

One of the key advantages of the CLI is the reduction in maintenance overhead and the simplification of Workspace as a programmable runtime environment. By providing a common interface for accessing Workspace APIs, the CLI aims to enhance the development of internal automation and agent-driven workflows.

While the CLI is not officially supported by Google, it presents a valuable tool for enterprise teams looking to optimize their workflow automation processes. The release emphasizes the importance of a cleaner, more efficient interface for interacting with Workspace applications, improving developer productivity and operational simplicity.

Source: VentureBeat

Databricks, a leading technology company, has introduced KARL, a reinforcement learning agent designed to enhance enterprise search capabilities. The traditional enterprise search pipelines often struggle with various search behaviors, leading to inefficiencies and breakdowns. In response to this challenge, Databricks developed KARL to address these shortcomings through simultaneous training across six distinct enterprise search behaviors using a new reinforcement learning algorithm.

KARL’s capabilities are impressive, claiming to outperform the renowned Claude Opus 4.6 model at a significantly lower cost per query and latency. The agent’s strength lies in its ability to handle complex tasks, such as synthesizing intelligence across meeting notes, reconstructing deal outcomes, and generating insights from unstructured data, which often lack clear right or wrong answers.

The key innovation of KARL is its ability to generalize across diverse tasks, demonstrating superior performance even on tasks it was not explicitly trained on. Powered by the Optimal Advantage-based Policy Optimization with Lagged Inference policy, KARL’s training efficiency and sample reuse make it a practical solution for enterprise teams.

Furthermore, KARL’s approach to contextual memory, grounded reasoning, and compression layers showcases its adaptability and problem-solving capabilities in handling ambiguous and challenging queries. While KARL excels in many areas, it still faces challenges in addressing queries with significant ambiguity and expanding its capabilities to include SQL queries and file search functionalities.

For enterprise data teams, KARL’s introduction prompts a reevaluation of pipeline architecture, the significance of reinforcement learning in search behavior development, and the practical implications of training specialized search agents. By embracing multi-task training and purpose-built search agents, enterprises can enhance their retrieval infrastructure and improve search efficiency.

Source: VentureBeat

EY, a global professional services firm, has significantly increased coding productivity by four to five times through the strategic integration of AI agents with its engineering standards. This integration has revolutionized the efficiency of teams working on EY’s suite of audit, tax, and financial platforms.

The key challenge faced by coding agents was the inability to deploy generated code effectively. Stephen Newman, EY Global CTO Engineering Leader, emphasized the importance of generating code that is integratable, compliant, and does not create additional cleanup work in the long run.

The solution involved connecting coding agents to EY’s engineering standards, code repositories, and compliance frameworks. This integration led to substantial productivity gains, but it was not a simple plug-and-play process. EY invested 18 to 24 months in building the necessary cultural foundation and technical integrations to enable semi-autonomous coding at scale.

EY adopted a cultural shift approach by introducing GitHub Copilot-style tools to familiarize engineers with prompt engineering and AI assistance organically. This approach ensured that AI adoption was user-driven rather than imposed from the top.

The success of this initiative hinged on granting agents access to EY’s code repositories and engineering standards, allowing them to generate deployable code efficiently. By evaluating multiple agent platforms, EY selected Factory, which demonstrated tangible value to developers and significantly improved productivity.

Under Newman’s guidance, EY established a workload classification framework to delineate tasks suitable for autonomous agent handling and those requiring human oversight. This restructuring of developer roles to orchestrators directing agents resulted in efficiency gains ranging from 15% to 60% in the early adoption phase.

EY’s journey showcases the potential of integrating AI technologies with established engineering practices, paving the way for enhanced productivity and efficiency in software development.

Source: VentureBeat

Microsoft has unveiled Phi-4-reasoning-vision-15B, a compact yet powerful multimodal AI model that challenges the industry’s reliance on massive AI systems. The 15-billion-parameter model, available through Microsoft Foundry, HuggingFace, and GitHub, excels at tasks like reasoning through math problems, interpreting charts, and handling visual tasks. What sets this model apart is its efficiency, requiring far less training data than its competitors, potentially reshaping how AI deployment is viewed economically.

The model’s innovative approach to reasoning, balancing structured reasoning for tasks like math and science with direct responses for tasks like image captioning, showcases Microsoft’s pragmatic view on AI model design. By leveraging a mid-fusion architecture and careful data curation, Microsoft has created a model that excels in efficiency, speed, and accuracy, making it a compelling option for edge devices, interactive applications, and on-premise servers.

This release marks a shift in the AI industry’s paradigm, emphasizing the importance of meticulous engineering over sheer scale. Microsoft’s open-weight release strategy positions Phi-4-reasoning-vision-15B as a foundational model for various applications, offering developers a high-performing yet resource-efficient solution. While the model faces challenges on certain benchmarks, its real-world impact and deployment scenarios remain to be seen.

Source: VentureBeat

German AI startup Black Forest Labs has unveiled a novel technique, Self-Flow, that aims to revolutionize the training of multimodal AI models. Traditionally, generative AI models have relied on external ‘teachers’ for semantic understanding, leading to limitations in scalability. Self-Flow addresses this challenge by enabling models to learn representation and generation simultaneously, achieving state-of-the-art results across images, video, and audio without external supervision.

Self-Flow tackles the ‘semantic gap’ in generative training by introducing an ‘information asymmetry’ approach. Through Dual-Timestep Scheduling, the model learns to generate outputs while predicting what a ‘cleaner’ version of itself would see, fostering deep internal semantic understanding.

The practical implications of Self-Flow are significant, with the technique converging 2.8x faster than current industry standards and continuously improving with increased compute and parameters. The framework excels in typography, temporal consistency in videos, and joint video-audio synthesis, surpassing competitive baselines in quantitative metrics.

Looking ahead, Self-Flow paves the way for world models capable of understanding the underlying physics and logic of scenes for planning and robotics. Black Forest Labs has made the research paper and official inference code available on GitHub, hinting at future commercial applications.

Source: VentureBeat

OpenAI, a leader in AI technology, has developed an AI data agent that is transforming enterprise data analysis. Built by two engineers at OpenAI, this tool has become an integral part of the company’s operations, serving thousands of employees daily. The agent, powered by GPT-5.2, offers a user-friendly interface that allows employees to access and analyze vast amounts of corporate data with simple, plain-English queries.

This innovative tool streamlines data analysis processes and enables employees across various departments to gain valuable insights autonomously. From revenue breakdowns to latency debugging, the agent handles a wide range of analytical tasks efficiently, saving significant time and effort for users.

The agent’s ability to operate seamlessly across organizational boundaries provides a comprehensive view of data insights to users company-wide. The system’s use of Codex, OpenAI’s AI coding agent, further enhances its capabilities by automating code generation and data mapping processes.

While the commercial potential of this internal data agent is evident, OpenAI has chosen not to sell the tool but instead encourages enterprises to build their own versions using available APIs and technologies. This approach aligns with OpenAI’s strategy of empowering businesses to harness AI for their specific needs and underscores the company’s commitment to advancing AI technology for broader adoption.

This development signifies a shift in how enterprises can leverage AI to enhance data analysis and decision-making processes. By focusing on data governance and accessibility, companies can unlock the full potential of AI-driven tools like OpenAI’s data agent, paving the way for accelerated innovation and competitiveness in the digital era.

Source: VentureBeat

Google has unveiled its latest AI model, Gemini 3.1 Flash-Lite, offering enhanced cost-efficiency and speed for enterprises and developers seeking advanced reasoning and multimodal capabilities. Positioned as the most budget-friendly and responsive option in the Gemini 3 series, this model is tailored for large-scale intelligence applications.

Designed to optimize the “time to first token,” Flash-Lite focuses on reducing latency for real-time applications like customer support and content moderation. It outperforms its predecessor, Gemini 2.5 Flash, with a 2.5X faster response time and a 45% increase in overall output speed.

A notable feature is the introduction of thinking levels, allowing developers to dynamically adjust the model’s reasoning intensity based on task complexity. Flash-Lite’s performance metrics, including an Elo score of 1432 and specialized strengths in various cognitive domains, demonstrate its competitive edge in the AI landscape.

Compared to its flagship counterpart, Gemini 3.1 Pro, Flash-Lite stands out as a cost-effective solution, priced at $0.25 per 1 million input tokens and $1.50 per 1 million output tokens. This pricing strategy positions it as a more affordable option than many competitors, offering substantial cost savings without compromising performance.

By leveraging a dual-model approach with Flash-Lite for high-volume tasks and Pro for complex reasoning, enterprises can achieve a balance between cost efficiency and cognitive processing power. Feedback from the community and developers has highlighted Flash-Lite’s speed, intelligence-to-speed ratio, and reliability in data tagging, making it a preferred choice for diverse applications.

Released through Google AI Studio and Vertex AI, Flash-Lite and Pro cater to enterprise requirements, ensuring secure and efficient AI operations. The models represent a shift towards utility-grade AI, enabling reliable autonomy and high-precision task execution at scale.

Source: VentureBeat

Alibaba’s latest release, the Qwen3.5 Small Model Series, has made a significant impact in the AI sector. This series, which includes models like Qwen3.5-9B, has outperformed OpenAI’s gpt-oss-120B while being significantly smaller in size. The key to this success lies in a hybrid architecture that combines Gated Delta Networks and sparse Mixture-of-Experts, enabling higher throughput and lower latency.

These models are natively multimodal, showcasing a level of visual understanding previously unseen in models of their size. Benchmark data reveals exceptional performance across various tasks, from visual reasoning to mathematical prowess, positioning the Qwen3.5 series as a notable development in the AI landscape.

Moreover, the release of these models under the Apache 2.0 license is a positive step for the open ecosystem, allowing for commercial use, modification, and distribution without royalty payments. This move enhances accessibility and fosters innovation in the AI community.

Enterprise applications of the Qwen3.5 series span a wide range of functions, from visual workflow automation to real-time edge analysis. However, teams must be mindful of operational challenges that come with deploying small-scale models, such as the risk of a ‘Hallucination Cascade’ in multi-step workflows.

The Qwen3.5 series represents a shift towards localized deployment of powerful AI models, enabling organizations to streamline tasks that previously relied on cloud-based solutions.

Source: VentureBeat

OpenAI’s recent $110 billion funding injection from SoftBank, Nvidia, and Amazon marks a significant development in enterprise artificial intelligence. While the influx of capital is noteworthy, the real game-changer is OpenAI’s collaboration with Amazon, introducing a ‘Stateful Runtime Environment’ on Amazon Web Services (AWS), the leading cloud platform globally.

This move signals a shift towards autonomous ‘AI coworkers’ and a need for a new architectural foundation different from GPT-4. For businesses on AWS, this means upcoming access to a stateful runtime environment, promising a significant evolution in agentic intelligence capabilities.

The core innovation lies in the distinction between ‘stateless’ and ‘stateful’ environments. The Stateful Runtime Environment on Amazon Bedrock will enable AI models to maintain persistent context, memory, and identity, revolutionizing developer workflows and reducing the complexity of maintaining context.

OpenAI’s platform, Frontier, designed to streamline AI agent development and deployment, empowers enterprises to bridge the ‘AI opportunity gap’ by offering shared business context, a robust agent execution environment, and built-in governance. While Frontier resides on Microsoft Azure, AWS will serve as the exclusive cloud distribution provider, allowing AWS customers to leverage agentic workloads seamlessly.

Enterprises interested in adopting the new Stateful Runtime Environment can register their interest via OpenAI’s dedicated Enterprise Interest Portal, signaling a shift towards production-grade agentic workflows.

The partnership dynamics between OpenAI, Amazon, and Microsoft present strategic choices for CTOs and decision-makers. While Azure remains the go-to for standard tasks, AWS’s Stateful Runtime Environment excels in complex, long-running agent scenarios, offering a cost-efficient solution for enterprises looking to scale OpenAI models.

Despite the Amazon investment, Microsoft’s commercial and revenue share relationship with OpenAI remains intact, underscoring the intricate ties between the two tech giants. As OpenAI positions itself as a key infrastructure player straddling Azure and AWS, the enterprise AI landscape is evolving towards tailored solutions based on specific technical requirements.

Source: VentureBeat