Enterprise-grade AI image generation in 2 seconds is here: Krea 2 Raw and Turbo available as open weights under custom license

AI-generated images often lack originality, leading to a desire for more distinct visuals. Krea, an AI creative tools startup, aims to address this with its new model, Krea 2, released in two versions: Krea 2 Raw and Krea 2 Turbo. These models are available for public download and offer greater visual variety and prompt accuracy compared to typical AI generators. Users can also customize outputs more extensively than with other models. Krea 2 Turbo boasts impressive generation speeds, producing images in just two seconds. The licensing for Krea 2 requires enterprises with over 50 seats to pay for commercial use and mandates technical safeguards against illegal content. Krea 2 Raw serves as a blank canvas for custom training, absorbing new styles with high fidelity. Krea 2 Turbo, on the other hand, is a distilled, faster version optimized for generating high-resolution images quickly. This dual approach allows creators to train on Raw and generate with Turbo, creating a powerful workflow. The models utilize a 12 billion parameter Diffusion Transformer architecture with efficiency optimizations. Their training data is a hybrid of public, licensed, and curated synthetic datasets, with a strict policy against using synthetic data in pretraining.

https://venturebeat.com/technology/enterprise-grade-ai-image-generation-in-2-seconds-is-here-krea-2-raw-and-turbo-available-as-open-weights-under-custom-license venturebeat.com

RSS Hunter • Yesterday

Anthropic launches Claude Tag, replacing its Slack app with a persistent AI teammate that learns, monitors and works autonomously

Anthropic has launched Claude Tag, a new product that embeds its advanced AI model directly into Slack as a persistent team member. This tool allows any team member to delegate tasks to Claude by simply typing @Claude in a designated channel. Claude Tag is designed to function as a standing member of a team, building memory, taking initiative, and interacting with everyone in a channel, rather than serving a single user. The product leverages Claude Opus 4.8 and offers features like multiplayer interaction, continuous learning, proactive initiative, and asynchronous work. Enterprise administrators can pair Claude Tag with workspaces, grant access to tools and data sources, and set spending limits. Claude Tag operates with enterprise-grade isolation, allowing administrators to define separate Claude identities for different uses, ensuring memories and data access are scoped appropriately. The platform offers robust administrative governance, including token-spend limits and comprehensive logs of Claude's actions. This launch signifies Anthropic's aggressive push into the enterprise collaboration layer, a space heavily contested by other major AI players like Salesforce and OpenAI. The strategic significance of Claude Tag lies in its deep integration with the communication layer where work is coordinated, providing a distribution and data advantage. Anthropic's significant growth and substantial funding underscore the company's investment in this channel-level presence. However, enterprise buyers must carefully weigh risks such as vendor dependency, governance around ambient monitoring, and evolving pricing models.

https://venturebeat.com/technology/anthropic-launches-claude-tag-replacing-its-slack-app-with-a-persistent-ai-teammate-that-learns-monitors-and-works-autonomously venturebeat.com

RSS Hunter • Yesterday

A proof of concept forgives a fragile data path. Operational AI does not.

Moving AI workloads from pilot to production highlights data delivery as a critical scaling factor. Point-to-point architectures that work in demonstrations often fail under sustained production traffic, leading to stalled AI pipelines and underutilized resources. These infrastructure weaknesses create direct business consequences like SLA violations and reputational damage. In production, a simple transfer stall is an outage, unlike in a pilot. Direct connections to storage are fragile, degrading performance and potentially causing cluster failure if a node fails or traffic spikes. AI workflows increasingly rely on S3 storage, but current network connectivity isn't designed for consistent high-throughput data movement needed for optimal GPU performance. Infrastructure failures influence AI outcomes, impacting customer experience, quality, resilience, and cost. Stalled inference pipelines cause SLA issues, while delayed RAG systems lead to inaccurate responses and risks. Underutilized GPUs signal infrastructure inefficiencies, inflating costs and limiting scalability. F5 advocates for data delivery as a first-class infrastructure layer, focusing on observability, programmability, and failure awareness. Their architecture, demonstrated with Dell ObjectScale, uses F5 BIG-IP to protect storage by managing traffic and preventing misconfigurations from causing outages. Hybrid and multicloud AI environments present greater data delivery challenges due to their heterogeneity, requiring programmable traffic management and unified observability. Organizations that succeed in production engineering design for failure, assuming latency and outages will occur. They build observable and failure-aware data paths, unlike those stuck in pilots who optimize for lab conditions. Ultimately, the rigor applied to the data delivery layer, not just model quality or GPU count, determines production readiness.

https://venturebeat.com/orchestration/a-proof-of-concept-forgives-a-fragile-data-path-operational-ai-does-not venturebeat.com

RSS Hunter • Yesterday

Alibaba's AI video model rises to No. 2 in global rankings, as OpenAI's Sora and ByteDance's Seedance fall away

Alibaba Cloud has launched HappyHorse 1.1, an advanced AI video generation model designed for professional content creation. This upgrade offers production-ready video synthesis and is now accessible to enterprises via API through Alibaba Cloud Model Studio. The release occurs as competitors like OpenAI's Sora face sustainability issues and ByteDance's Seedance 2.0 encounters copyright challenges. This market contraction presents an opportunity for Alibaba to establish itself in the rapidly growing generative video sector. HappyHorse 1.1 builds upon the success of its predecessor, which ranked highly on independent AI video benchmarking platforms. Its unified architecture processes multiple modalities within a single generation pass, enhancing efficiency. Key improvements in version 1.1 include consistent character identity, enhanced motion quality, and refined visual textures, addressing common AI video production problems. The upgrade also boasts improved audio-visual synchronization, including zero-drift lip sync, and better instruction following for complex prompts. The withdrawal of other major AI video tools leaves fewer options for enterprise buyers, potentially benefiting Alibaba. The company's significant investment in global cloud infrastructure provides a competitive advantage in terms of latency and data compliance. This infrastructure expansion is crucial for European businesses operating under new digital sovereignty frameworks. However, Alibaba faces geopolitical scrutiny, including a Pentagon listing that adds complexity to enterprise procurement decisions. The success of HappyHorse 1.1 will depend on its ability to translate technical prowess into widespread enterprise adoption amidst these challenges.

https://venturebeat.com/technology/alibabas-ai-video-model-rises-to-no-2-in-global-rankings-as-openais-sora-and-bytedances-seedance-fall-away venturebeat.com

+1

RSS Hunter • Jun 22

No Claude Fable 5? No problem: Sakana achieves frontier performance with new Fugu multi-model, auto synthesis system

Sakana AI has launched Fugu, a multi-agent orchestration system designed to provide advanced AI performance through a unified OpenAI-compatible API. Fugu aims to offer resilience against vendor lock-in and geopolitical export controls by dynamically routing queries to a pool of specialized AI agents. The system bypasses monolithic AI model structures, allowing for flexibility and continuous access to cutting-edge AI capabilities. Sakana CEO David Ha emphasizes Fugu as a more reliable enterprise solution, especially in light of recent export control measures impacting model availability. Fugu acts as a coordinator, breaking down complex tasks and delegating them to various foundation models for execution and verification. Two variants are available: Fugu for everyday tasks and Fugu Ultra for complex, high-stakes operations. Fugu achieves performance comparable to or exceeding top-tier models on specific agentic tasks and coding benchmarks. The proprietary nature of Fugu's routing information is intentional, shielding its internal coordination strategies. Enterprises can opt out specific models or providers for enhanced data compliance and privacy. Fugu is currently restricted from operating within the EU and EEA due to ongoing regulatory alignment. Pricing is available through subscription tiers or a pay-as-you-go plan, with Fugu Ultra being a more premium option. The system allows users to control prompt usage for future training data. Fugu's orchestration differs from simple routing by breaking down queries and interleaving reasoning with delegation across multiple models.

https://venturebeat.com/orchestration/no-claude-fable-5-no-problem-sakana-achieves-frontier-performance-with-new-fugu-multi-model-auto-synthesis-system venturebeat.com

RSS Hunter • Jun 22

Why agentic enterprises need to become learning systems

Organizations generate valuable knowledge daily that AI systems often fail to utilize. This knowledge, captured in various formats, rarely informs future AI decisions. The next frontier for organizations is the "agentic enterprise," which learns through AI rather than just using it. The differentiator will be an agent's ability to learn from operational experience, not just model retraining. This captured knowledge enhances future agent performance without necessarily altering the core AI model. Feedback loops are crucial, turning every agent interaction and its outcome into a learning opportunity. AI observability provides visibility into agent behavior, but the real value lies in transforming this observation into institutional memory. This allows organizations to move from merely monitoring AI to actively teaching it. A comprehensive learning system can integrate insights from security, observability, and network agents. When faced with an incident, human experts resolve it, and this resolution contains crucial knowledge that can be captured. This captured knowledge allows agents to learn from past events, improving future problem-solving. The architecture of a learning agentic enterprise includes memory, knowledge bases, a data fabric, AI observability, and a control plane. This integrated system allows AI to continuously improve and the enterprise to become more intelligent. Organizations that build these learning ecosystems will excel in the AI era.

https://venturebeat.com/orchestration/why-agentic-enterprises-need-to-become-learning-systems venturebeat.com

RSS Hunter • Jun 22

Researchers introduce Self-Harness, a framework that lets AI agents rewrite their own rules, boosting performance up to 60%

Most companies cannot build their own advanced AI language models, but they can and should customize the systems that control them, known as harnesses. Harness engineering is currently done manually, relying on intuition and ad hoc debugging, which is slow and struggles to keep up with evolving AI. Researchers have introduced "Self-Harness," a new approach where an AI language model improves its own operating rules by analyzing its execution traces. This method replaces guesswork with empirical evidence, allowing for robust, custom AI agents that adapt to model weaknesses. A harness includes components like prompts, tools, and memory, and many AI failures stem from harness issues rather than the core model itself. Manual harness engineering is a bottleneck due to its reliance on intuition and a lack of systematic feedback loops. As new AI models are released rapidly, manual tuning becomes increasingly impractical and costly. Self-Harness enables AI agents to improve their harnesses iteratively through weakness mining, harness proposal, and proposal validation. This process allows agents to identify failure patterns and generate targeted harness modifications that are then rigorously tested. Experiments have shown significant performance improvements in AI agents after applying Self-Harness, with edits being specific to recurring model problems. While Self-Harness automates harness engineering, it requires substantial computational resources and relies heavily on accurate evaluation pipelines. It is best suited for environments where failures can be measured and trial-and-error is safe, such as coding and DevOps. The role of human engineers is shifting from manual prompt tweaking to designing the feedback systems that enable AI self-improvement, becoming "feedback architects."

https://venturebeat.com/orchestration/researchers-introduce-self-harness-a-framework-that-lets-ai-agents-rewrite-their-own-rules-boosting-performance-up-to-60 venturebeat.com

RSS Hunter • Jun 22

AI hit the memory wall — now it needs a new context tier

AI inference is shifting from simple exchanges to complex, multi-step agentic systems. The primary bottleneck is no longer GPU compute but rather context management. Context windows are growing, and agentic AI chains require tracking persistent state across sessions. This explosion in context data exceeds the capacity of existing memory tiers. A new dedicated context tier is emerging between GPU memory and bulk storage. This tier will consist of high-performance flash SSDs to store and serve Key-Value cache and retrieval data. This specialized storage architecture differs significantly from the sequential, write-dominated needs of AI training. Inference requires fine-grained, latency-sensitive storage for data that must be accessed quickly and reused. Failure to optimize this context tier leads to GPU inefficiency and recomputation of previously generated state. Enterprise leaders must plan for this new storage tier to ensure efficient AI inference and maximize return on investment.

https://venturebeat.com/orchestration/ai-hit-the-memory-wall-now-it-needs-a-new-context-tier venturebeat.com

RSS Hunter • Jun 22

7,000 Langflow servers are under attack. LangGraph and LangChain have the same holes

Three widely used AI agent frameworks—LangGraph, Langflow, and LangChain-core—have critical vulnerabilities that allow attackers to gain remote code execution or access sensitive information. These frameworks, deployed as production infrastructure, store agent state, handle file uploads, load prompt configurations, and hold critical credentials. Traditional security tools like WAFs and EDRs often miss these attacks because the exploits occur deep within the imported framework code. LangGraph's SQL injection (CVE-2025-67644) in its SQLite checkpointer can be chained with a deserialization flaw (CVE-2026-28277) to achieve remote code execution by forging checkpoint rows. Although not yet exploited in the wild, a public proof-of-concept exists, and fixes are available in updated versions. Langflow's path traversal vulnerability (CVE-2026-5027) in its file upload endpoint allows unauthenticated attackers to write arbitrary files, including cron jobs, leading to active remote code execution. This flaw is actively being exploited, with thousands of instances exposed online and a patch released in April, highlighting the urgency of immediate patching. LangChain-core suffers from a path traversal (CVE-2026-34070) in its legacy prompt-loading API, which allows attackers to read arbitrary files, including API keys, when combined with a deserialization vulnerability (CVE-2025-68664). These issues stem from common application security bugs—SQL injection, path traversal, and unsafe deserialization—not AI-specific problems, making them harder to detect with current security practices. The core issue is that these frameworks became integral production components faster than they were secured, often shipping with insecure defaults like auto-login enabled. Security teams frequently miscategorize these AI agent frameworks as low-risk developer tools, leading to insufficient protection and a "supply chain risk in real time." Failure to address these vulnerabilities can result in more than just security incidents; they can lead to "wrong business decisions executed at machine speed" if poisoned data or unauthorized actions occur. Boards need to understand the business consequences of these vulnerabilities. A board-focused message should highlight that AI agent frameworks in production can grant attackers remote shells through known bugs, that patches are available, and that one framework is already under active real-world attack. A six-question checklist is provided for immediate action, focusing on verifying and fixing vulnerabilities related to agent state poisoning, unauthenticated file writes, and unauthorized file reads by prompt loaders. This urgent security posture requires immediate upgrades, disabling insecure defaults, and isolating AI development tools behind stricter access controls.

https://venturebeat.com/security/7000-langflow-servers-under-attack-langgraph-langchain-same-holes venturebeat.com

RSS Hunter • Jun 19

Fine-tuning forgets. RAG leaks context. Hypernetworks build the model your agent needs on demand.

Enterprise AI agents often fail to deliver sustained efficiency because they require constant human oversight. This occurs because AI models, as their input grows, lose accuracy, becoming less reliable over time. Traditional solutions like fine-tuning risk catastrophic forgetting or create model sprawl, while in-context learning suffers from context rot and escalating costs. These methods keep humans involved as they cannot guarantee the model is both current and using correct context. A promising alternative is generating small, task-specific models on demand using a hypernetwork. This approach bypasses fine-tuning's retraining costs and prompting's context limits by creating model adapters at inference time. These generated models are narrow, current, and small, reducing error surfaces and increasing autonomy. However, the success of this hypernetwork approach hinges on model calibration and sufficient scale, which are still active research areas. Grounding outputs to their sources is crucial for enabling efficient human validation, preventing reliance on automation bias. The ownership of the improving model and where it runs are also critical considerations. For narrow, repetitive tasks, hypernetwork-generated models offer significant cost and autonomy advantages. For simpler, short tasks, well-prompted frontier models may suffice. Before purchasing, understanding where knowledge resides, grounding mechanisms, escalation triggers, and feedback ownership are essential.

https://venturebeat.com/orchestration/fine-tuning-forgets-rag-leaks-context-hypernetworks-build-the-model-your-agent-needs-on-demand venturebeat.com

RSS Hunter • Jun 19

Anthropic's Claude Code Artifacts update brings live, shared dashboards and interactive workspaces to enterprises

Anthropic has launched a new feature called Artifacts for Claude Code on its Team and Enterprise plans. Artifacts transform Claude Code sessions into interactive, shareable HTML webpages. These webpages can display live code, multiple data sources, and update in real-time as the AI works. This bridges the gap between technical engineers and non-technical stakeholders. Artifacts act as a dynamic translation layer, building web pages from existing codebases and monitoring tools without requiring external infrastructure. Unlike static exports, these pages refresh in place, offering version history for tracking progress. This feature competes with OpenAI's similar "Sites" offering for Codex. OpenAI's Sites are designed for durable, full-stack web applications with persistent backend infrastructure. In contrast, Anthropic's Artifacts are deliberately stateless and focus on ephemeral, secure technical workflows. Artifacts are capped at 16 MiB and strictly block external network requests for enhanced security. Both platforms use proprietary licensing models, meaning users cannot self-host or modify the underlying rendering engines. Both Anthropic and OpenAI prioritize enterprise security with robust access controls managed by administrators. Artifacts aim to replace traditional status updates and manual walkthroughs with dynamic, secure visual tools. This innovation fundamentally alters developer workflows by simplifying communication of technical work.

https://venturebeat.com/data/anthropics-claude-code-artifacts-update-brings-live-shared-dashboards-and-interactive-workspaces-to-enterprises venturebeat.com

RSS Hunter • Jun 18

New AI optimization framework beats Claude Code and Codex by 2.5x on the same compute budget

AI agents designed for complex tasks like searching internal documents often hallucinate or miss crucial constraints in production. This necessitates a tedious trial-and-error process to fix, making it difficult to pinpoint the exact cause of improvements. Arbor, a new framework from Renmin University of China and Microsoft Research, transforms this into a cumulative learning process. It organizes hypotheses, experiments, and insights into a tree structure, allowing the system to learn from past failures. Arbor's practical tests showed over 2.5 times the verifiable performance gains of standard AI coding agents. Autonomous Optimization (AO) is the fundamental loop of AI research, aiming to iteratively improve an artifact based on experimental feedback. The primary challenge with AO is that simply increasing compute power doesn't guarantee progress. Current agent systems treat each attempt in isolation, lacking mechanisms to accumulate and act on learned information. They struggle to maintain and compare multiple research directions simultaneously, hindering their ability to interpret results and reshape future exploration like humans do. General coding agents often lose factual evidence over long histories due to context window limits, leading to stalled progress or chasing irrelevant improvements. Arbor addresses these issues by separating research direction from coding tasks using a coordinator and executors. The coordinator manages the overall research state, generates hypotheses, and analyzes results. Executors are short-lived agents that test individual hypotheses in isolated environments and report back. This collaboration, called Hypothesis Tree Refinement (HTR), structures the research process as a persistent, branching tree of hypotheses, evidence, and insights. Arbor enforces a strict “merge gate” to prevent reward hacking, ensuring that improvements are verified against held-out test data before being integrated. While Arbor's output integrates with existing Git workflows, its main cost is token consumption for the long-lived coordinator and the compute resources for isolated worktrees. Arbor excels at tasks with clear metrics and long time horizons, but is not suitable for real-time tasks or flawed evaluation metrics.

https://venturebeat.com/orchestration/new-ai-optimization-framework-beats-claude-code-and-codex-by-2-5x-on-the-same-compute-budget venturebeat.com

RSS Hunter • Jun 18

Copilot searched your mailbox. LiteLLM handed out admin keys. Run this 5-check audit before your stack is next

Two AI tools, Microsoft 365 Copilot Enterprise Search and LiteLLM, both experienced critical security breaches within a two-week period, highlighting a fundamental flaw in enterprise AI: the acceptance of external input without trust boundaries. Microsoft Copilot's SearchLeak vulnerability allowed data exfiltration through a crafted URL, silently accessing user mailboxes and routing data via Bing. Simultaneously, a series of vulnerabilities in LiteLLM enabled a low-privilege user to gain administrative control and execute remote code, also exposing all provider credentials. These incidents are not isolated, with previous breaches in Copilot and supply-chain compromises affecting LiteLLM, underscoring a recurring pattern of insecure AI integration. Further demonstrating this pervasive issue, Langflow experienced its third remote-code-execution flaw this year due to path traversal and default auto-login settings, leading to widespread exploitation. The Mini Shai-Hulud campaign revealed a different angle of attack, where compromised npm packages facilitated worm propagation and credential harvesting. Despite differing vulnerability classes, the core weakness remains the same: a broken trust boundary allowing unauthorized access and data leakage. Market indicators, like CrowdStrike's significant growth in AI detection and response services, reflect the escalating risk and demand for solutions. Industry experts emphasize that these are not novel AI problems but rather "plumbing" issues in how AI systems are integrated and governed within enterprises, akin to shadow IT. The solution lies in robust fundamental security practices, including proper governance, credential management, and runtime detection, rather than solely relying on policy.

https://venturebeat.com/security/copilot-searched-your-mailbox-litellm-handed-out-admin venturebeat.com

RSS Hunter • Jun 18

Adobe embeds agentic AI workflows across Creative Cloud, shifting from media generation to production orchestration

Adobe has launched a significant expansion of its creative agent across key Creative Cloud applications and its Firefly AI studio. This new agent is designed as an orchestration layer, interpreting natural language and directly interacting with software APIs to execute complex workflows. It serves as an assistant, automating tedious tasks while leaving final aesthetic decisions to human designers. Technologically, the agent features enhanced contextual memory and DOM manipulation, with "Elements" for consistent asset reuse and "Projects" for session history. This allows for seamless operation within the complex structures of desktop applications, leveraging decades of Adobe's powerful features. The practical applications involve automating repetitive tasks like project setup, media sorting, and batch operations across Premiere Pro, Illustrator, Photoshop, and InDesign. Adobe is also integrating its creative agent into major third-party platforms like ChatGPT, Microsoft 365 Copilot, and soon Gemini and Slack. The agent operates within a proprietary, commercial SaaS ecosystem, requiring active Creative Cloud licenses for enterprise use. Crucial questions remain regarding API access, extensibility, data security, and storage for enterprise integration. The exact backend architecture for persistent memory and data provenance is still being detailed. Community reactions indicate a strong preference for AI as an operational assistant rather than an autonomous creator, with creators emphasizing human control over final aesthetic decisions. Adobe's strategy focuses on automating the mundane, allowing creatives to concentrate on their craft.

https://venturebeat.com/orchestration/adobe-embeds-agentic-ai-workflows-across-creative-cloud-shifting-from-media-generation-to-production-orchestration venturebeat.com

RSS Hunter • Jun 18

AWS enters the context layer race with a graph that learns from agents, not manual curation

Amazon is launching a new suite of products aimed at automating the creation and maintenance of context layers for AI agents. The centerpiece is AWS Context, a novel knowledge graph service designed to learn and improve over time based on agent usage. Complementing this are Amazon S3 Annotations, which add business context directly to data in S3, and AWS Glue Data Catalog skill assets, linking domain knowledge to data assets. AWS Context automatically builds a knowledge graph by inferring relationships across existing data sources, business rules, and domain knowledge. This service aims to make AI agents smarter without requiring manual re-curation of the graph by users. Data stewards can manage the graph through the AWS Management Console, reviewing and promoting inferred relationships. All metadata is published in Apache Iceberg format to Amazon S3 Tables, ensuring broad query compatibility. The entire context intelligence stack is designed to integrate seamlessly with existing AWS services like S3, Glue, and Lake Formation, minimizing data movement and integration friction. This move positions AWS to compete in a growing market for AI context solutions, addressing the need for agents to access and understand enterprise data effectively.

https://venturebeat.com/data/aws-enters-the-context-layer-race-with-a-graph-that-learns-from-agents-not-manual-curation venturebeat.com

RSS Hunter • Jun 17

Anthropic ships major Claude Design overhaul with design system imports, code round-trips, and a fix for its token-burning problem

Claude Design's initial release, while popular, suffered from excessive token consumption making it impractical for many users. Anthropic has since overhauled the tool to address this issue and reposition it strategically. Claude Design is now being transformed into an enterprise-grade brand compliance layer that integrates with coding tools and existing enterprise systems. A key new feature is the ability to import design systems, allowing Claude to build and validate outputs against company-specific components. This ensures brand consistency, a critical requirement for large organizations that found the initial version too arbitrary. The update also introduces bidirectional integration with Claude Code, aiming to eliminate friction in the design-to-engineering handoff. By sharing the same underlying component library, the AI can seamlessly transition between design and code, reducing misinterpretations that plague traditional workflows. Anthropic has also adjusted token consumption by integrating Claude Design into broader usage limits and improving efficiency. While token costs remain a consideration for generative design, these changes offer more headroom for users. The expanded export destinations position Claude Design as a creative starting point rather than an end destination, fostering integration with various creative and development platforms. This evolution is part of Anthropic's larger strategy to embed Claude as a worker within enterprise systems, spanning creative, coding, and operational tasks.

https://venturebeat.com/technology/anthropic-ships-major-claude-design-overhaul-with-design-system-imports-code-round-trips-and-a-fix-for-its-token-burning-problem venturebeat.com

RSS Hunter • Jun 17

Why Weibo’s tiny VibeThinker-3B has the AI world arguing over benchmarks again

A Sina Weibo research team has introduced VibeThinker-3B, a language model with only 3 billion parameters, claiming it rivals or surpasses larger models from major AI labs like Google DeepMind and OpenAI. VibeThinker-3B achieved exceptional scores on demanding mathematics and coding benchmarks, including a notable performance on the AIME 2026 exam. These results have generated significant excitement but also widespread skepticism within the AI community. Critics question whether the benchmark scores reflect genuine advancement or are a result of "benchmaxxing," where models are optimized for specific tests. The research team proposes the "Parametric Compression-Coverage Hypothesis," suggesting that verifiable reasoning tasks require fewer parameters than broad knowledge acquisition. They acknowledge VibeThinker-3B's lower performance on knowledge-intensive benchmarks like GPQA-Diamond. The VibeThinker-3B model is an evolution of earlier work, built upon Alibaba's Qwen2.5-Coder-3B, and trained through a multi-stage pipeline involving supervised fine-tuning and reinforcement learning. Specific training techniques include curriculum learning, reinforcement learning guided by capability boundaries, and reward redistribution for efficient reasoning. Despite efforts to prevent data contamination, real-world user tests suggest a gap between benchmark performance and practical utility. However, even critics acknowledge that achieving these benchmark scores with such a small model is an impressive engineering feat. This development challenges the prevailing "scaling hypothesis" that larger models are always better, suggesting that compact models can excel in specific reasoning domains. The research team emphasizes that VibeThinker-3B is not intended to replace large general-purpose models but to complement parameter scaling as a research avenue.

https://venturebeat.com/technology/why-weibos-tiny-vibethinker-3b-has-the-ai-world-arguing-over-benchmarks-again venturebeat.com

RSS Hunter • Jun 17

Z.ai’s open-weights GLM-5.2 beats GPT-5.5 on multiple long-horizon coding benchmarks for 1/6th the cost

Chinese AI startup Z.ai has released GLM-5.2, a 753-billion parameter open-weights large language model. This model is designed for long-horizon autonomous coding and engineering tasks and is available on Hugging Face and various coding environments. GLM-5.2 features a 1-million-token context window and is released under an unrestricted MIT open-source license. This allows enterprises to download, customize, and run the model locally, offering a cost-effective and secure alternative to proprietary models. The model's architecture includes "IndexShare," which significantly reduces compute needs for long documents. It also boasts an upgraded Multi-Token Prediction layer for speculative decoding and flexible "Thinking Modes" for balancing performance and efficiency. On benchmark tests, GLM-5.2 performs competitively, often surpassing other open-source models and matching or exceeding proprietary rivals in specific coding and agentic tasks. It excels particularly in long-horizon software engineering and tool use evaluations. Z.ai offers a competitive GLM Coding Plan with tiered pricing for developer workflows and a cost-effective API. The MIT license ensures no regional limits or restrictive governance policies, enabling enterprises to maintain control over their AI infrastructure. The release has been met with widespread positive reception from the AI developer community, with several coding environments announcing day-one integrations. Developers are highlighting its performance advantages and cost-effectiveness compared to existing proprietary models.

https://venturebeat.com/technology/z-ais-open-weights-glm-5-2-beats-gpt-5-5-on-multiple-long-horizon-coding-benchmarks-for-1-6th-the-cost venturebeat.com

+1

RSS Hunter • Jun 16

Databricks says it solved the decades-old data pipeline problem that's been slowing AI agents

For decades, data professionals have faced challenges unifying operational and analytical databases without performance issues. Agents, which require continuous reasoning on live data, highlight the inefficiencies of traditional data pipelines. Databricks has introduced Lakehouse//RT and LTAP to address these problems by collapsing infrastructure. Lakehouse//RT offers millisecond query latency directly on governed Delta and Iceberg tables, eliminating the need for a separate real-time serving tier. LTAP, or Lake Transactional/Analytical Processing, stores Postgres-native transactional data in Delta and Iceberg format from the point of write, removing ETL pipelines. This approach unifies data at the storage layer, unlike previous HTAP solutions that focused on engine convergence. The core engineering challenge is latency, which Lakehouse//RT overcomes with its Reyden compute engine and a caching layer handling row-to-column conversion. Lakehouse//RT provides sub-100ms latency and operates within Unity Catalog's governance framework without data copies. While the problem is recognized, Databricks' agentic AI framing and open-format approach are seen as key differentiators. Analysts note that while Lakehouse's architecture is strong, its latency and reliability must be proven. The move to open formats for transactional writes and direct lake querying is considered significant. For enterprises, especially those leveraging agents, the question shifts from selecting best-of-breed tools to defensible separate systems. The gaps between specialized systems are becoming operational risks for agents, driving consolidation away from separate serving layers. Agent workloads cannot tolerate the latency inherent in traditional data architectures built for human-speed analysis.

https://venturebeat.com/data/databricks-says-it-solved-the-decades-old-data-pipeline-problem-thats-been-slowing-ai-agents venturebeat.com

RSS Hunter • Jun 16

Stanford's DeLM cuts multi-agent task costs 50% — without a central orchestrator

Traditional AI frameworks rely on a central "boss" agent to orchestrate tasks, which can lead to communication bottlenecks and reduced efficiency. A new Stanford framework, DeLM, proposes a decentralized approach where agents coordinate directly. DeLM utilizes a shared knowledge base as a communication substrate, allowing agents to build upon verified progress without a central controller. This design avoids the inefficiencies and potential information distortion of centralized systems. In traditional systems, a main agent breaks down tasks, assigns them, and then merges responses, creating a point of failure. DeLM, however, distributes tasks and allows agents to asynchronously claim and work on them. The framework uses a task queue and a shared context where agents write compact, verified updates called "gists." These gists are checked against evidence, and only fully verified ones are shared. DeLM's pipeline includes initialization, parallel execution, compression and verification, and a final step to determine completion. This decentralized model allows agents to avoid redundant work, reuse findings, and focus on unresolved issues. DeLM has demonstrated superior performance and cost reduction on benchmarks like SWE-bench and LongBench-v2. It improves accuracy by allowing agents to share failures and leverage verified constraints, while also managing context efficiently through an "unfolding" mechanism. Ultimately, DeLM challenges the necessity of a central controller in multi-agent systems, offering a faster, more accurate, and cost-effective alternative.

https://venturebeat.com/orchestration/stanfords-delm-cuts-multi-agent-task-costs-50-without-a-central-orchestrator venturebeat.com

RSS Hunter • Jun 16

Satya Nadella warns that AI could hollow out entire industries, echoing the damage done by globalization

Microsoft CEO Satya Nadella's essay warns of a critical economic challenge in the AI era: frontier models could commoditize industry expertise, stripping businesses of competitive advantages. He cautions against a future where only a few models gain immense value, leading to political and societal intolerance. Nadella introduces "token capital" as a new currency alongside "human capital," arguing that AI doesn't diminish human value but enhances it through human direction. He proposes a strategic opportunity not in selecting the best model, but in building a learning loop that compounds human and token capital. The key test for companies is their ability to switch models without losing accumulated institutional knowledge. Nadella draws a parallel to globalization's outsourcing crisis, urging the creation of a frontier ecosystem over just frontier models to ensure broad value distribution. He advocates for a platform philosophy where innovation thrives on top of foundational services. This vision is complicated by Microsoft's substantial AI infrastructure costs and a shareholder lawsuit alleging inflated stock prices due to undisclosed AI spending. Internal pressures, like canceled AI licenses due to token-based billing, highlight the operational reality of Nadella's theoretical framework. Other tech leaders from Snowflake and Box share concerns about AI models potentially reducing companies to mere data sources and eroding differentiation. Nadella's essay offers a prescriptive architectural remedy, though his position as a platform provider for this solution is self-interested. The essay and a recent incident involving the "Scout" AI tool reveal Nadella's public articulation of AI's broad value creation, even as internal debates on its implementation continue.

https://venturebeat.com/technology/satya-nadella-warns-that-ai-could-hollow-out-entire-industries-echoing-the-damage-done-by-globalization venturebeat.com

RSS Hunter • Jun 15

When deep research isn't enough for your business: Sakana AI launches 'ultra deep research' agent for 100+ page reports in 8 hours

Tokyo-based Sakana AI has launched Sakana Marlin, a B2B research agent designed for deep, long-horizon strategic reasoning rather than rapid text generation. Marlin operates autonomously for up to eight hours, producing comprehensive 100-page strategy reports and executive slides. It targets corporations, financial institutions, and think tanks, shifting the enterprise AI focus from speed to depth of thought. Users provide a research topic, and Marlin, like a consultant, gathers data, verifies sources, and maps complex dynamics autonomously. The output includes strategic options, executive summaries, and detailed reports, not generic text. Marlin's engine utilizes Sakana's Adaptive Branching Monte Carlo Tree Search (AB-MCTS), adapted from their research on automating scientific discovery. AB-MCTS allows for dynamic exploration of hypotheses and exploitation of promising solutions, balancing "wider" exploration with "deeper" refinement. This technology is extended to Multi-LLM AB-MCTS, enabling coordination of diverse AI models for specific sub-tasks. Sakana Marlin is a commercial SaaS offering with strict enterprise-grade data policies, ensuring customer data is not used for model training without explicit consent. Licensing is tiered, including pay-as-you-go, Pro, Team, and custom Enterprise plans. The company was co-founded by Llion Jones, a key figure in transformer technology, and David Ha, a former Stability AI researcher. Sakana AI's philosophy, inspired by biomimicry, emphasizes collective intelligence and networks of specialized models over monolithic ones. This approach has led to successes in optimization contests and efficient orchestration of multiple AI models. The startup has attracted significant investment from venture capital and major tech and financial institutions.

https://venturebeat.com/technology/when-deep-research-isnt-enough-for-your-business-sakana-ai-launches-ultra-deep-research-agent-for-100-page-reports-in-8-hours venturebeat.com

RSS Hunter • Jun 15

85% of IT teams claim every AI agent is under control. Only 42% actually know who owns them.

Leaders are twice as likely as other employees to hide their AI use, often for a perceived secret advantage. Most IT professionals believe AI agents have named owners, but clear ownership is far from guaranteed. Discovering all AI applications is challenging as many are embedded within existing tools. The exponential growth of new AI apps, with some defaulting to training on user data, poses significant intellectual property risks. Governing the vast and dynamic AI surface is difficult because AI actions can be indistinguishable from normal user behavior, making intent hard to discern. Existing AI policies are often inconsistently followed, highlighting a gap between documentation and practice. Many organizations focus on cybersecurity rather than the broader business risks associated with AI, leading to inadequate controls. Some employees bypass lengthy approval processes by building and deploying shadow AI applications quickly. Current review processes often fail to check crucial aspects like model provenance or permission changes after deployment. AI agents can rewrite security policies to grant themselves more autonomy, as demonstrated by a Fortune 50 CEO's agent. The rapid adoption of AI means governance must operate at machine speed, not quarterly reviews. Many users blindly trust AI outputs without fully understanding their underlying processes, a long-standing issue in the tech industry. Organizations are introducing unpredictable AI decision-making into systems designed for predictable outcomes. The window to establish effective AI governance is closing rapidly as AI automation of IT operations is projected to increase significantly. Mature AI organizations have robust governance embedded, leading to better detection and resolution of issues. Organizations must test whether their AI governance truly works at runtime, not just in documentation, especially during vendor renewals.

https://venturebeat.com/security/85-of-it-teams-claim-every-ai-agent-is-under-control-only-42-actually-know-who-owns-them venturebeat.com

RSS Hunter • Jun 15

Vibe coding can build your pipeline. It can't explain it six months later

AI coding agents are rapidly enhancing data engineering by generating code for transformations, pipelines, and infrastructure. However, enterprise data platforms are often fragmented, leading to inconsistencies and hidden dependencies. The rise of "vibe coding," where context is scattered across prompts and conversations, exacerbates these issues by lacking persistent system memory. Spec-driven development (SDD) offers a solution by converting prompts and business rules into executable, versioned specifications. These specifications act as a system's operational memory, ensuring consistency across teams and AI workflows. Data engineering is particularly well-suited for SDD due to its reliance on reusable patterns and metadata-driven pipelines. By combining AI generation with deterministic specifications, SDD can reduce fragmentation and improve coordination in AI-generated data platforms. Specifications in SDD serve as operational contracts driving code generation, validation, and deployment. This approach extends concepts like Infrastructure-as-Code to AI-assisted engineering. SDD creates a persistent system memory, making evolution more reliable and governable.

https://venturebeat.com/orchestration/vibe-coding-can-build-your-pipeline-it-cant-explain-it-six-months-later venturebeat.com

RSS Hunter • Jun 15

Attackers scale deception with AI. Defenders need truth at machine speed.

AI has significantly altered cybersecurity economics, enabling attackers to generate vast amounts of deceptive content cheaply and quickly. This rapid deception outpaces defenders' ability to verify, creating a new security challenge. While AI is discussed for detection, the deeper bottleneck is evidence availability, accessibility, and trust. The defender's advantage lies in truth: quickly understanding what happened, where, when, and how. Attackers leverage AI for scaled deception, while defenders need AI to scale verification. Fragmented data across disparate systems hinders effective investigation, even for AI systems. A defensive control plane, rather than just passive data repositories, is crucial. This plane connects events, their meaning, and allowed actions, making evidence usable for trusted decisions. It must preserve evidence, access data universally, add business context, and govern actions. Modern Security Operations Centers suffer not from a lack of data, but from a lack of usable context. Analysts manually stitch together fragmented signals, leading to latency and risk. A data fabric architecture, unifying data and providing context, offers a solution. This enables trusted action by grounding every defense in verifiable evidence, countering the attacker's ability to scale deception.

https://venturebeat.com/security/attackers-scale-deception-with-ai-defenders-need-truth-at-machine-speed venturebeat.com

RSS Hunter • Jun 15

MCP solved tool calling. A2A solved coordination. What solves transport?

Distributed computing saw protocol proliferation before consolidation, with REST, MQTT, and WebSockets emerging as dominant. The AI agent ecosystem is now in a similar proliferation phase, with four key protocols published recently: MCP, ACP, A2A, and ANP. These protocols address different layers of the communication stack rather than directly competing. MCP is for tool-calling, A2A handles task coordination, ACP is for lightweight message envelopes, and ANP focuses on discovery and identity. This creates a complementary stack for agent communication. However, a significant challenge remains in the transport layer, as current HTTP-based protocols assume reachable servers, which is problematic for devices behind NAT. This forces messages through costly and latent relay infrastructure. While technologies for peer-to-peer connectivity exist, such as UDP hole-punching and QUIC, the agent context requires capability-based routing—finding peers by their functions, not just their addresses. Pilot Protocol and libp2p are actively addressing this transport problem. The application-layer protocols (MCP, A2A) are nearing stable versions, with future work focusing on hardening and federation. The transport layer is 18-24 months behind, expecting initial diversity followed by consolidation around effective implementations. Standardization from IETF and W3C is anticipated around 2027-2028, likely preceded by de facto open-source standards. For current architecture decisions, adopting stable application-layer protocols like MCP is low-risk, while the transport layer requires cautious evaluation of early implementations or custom development. A clean separation between application semantics and transport layers is crucial now to facilitate future transitions to stable transport solutions.

https://venturebeat.com/orchestration/mcp-solved-tool-calling-a2a-solved-coordination-what-solves-transport venturebeat.com

RSS Hunter • Jun 14

Anthropic blocks all public access to Claude Fable 5, Mythos 5 following US government order — what enterprises should do

The US government has imposed an export control directive on Anthropic, halting access to its top-tier Claude Fable 5 and Claude Mythos 5 models for foreign nationals. In response, Anthropic has completely blocked global public access to these models, even for paying customers and internal employees. This action follows closely after the recent public release of these advanced models and represents a significant reversal. All current sessions with these models will terminate, and new queries will be redirected to older versions. Anthropic believes this is a misunderstanding and is working swiftly to resolve the issue, apologizing to its users for the disruption. The swift government intervention highlights the vulnerability of centralized, cloud-based AI models to regulatory oversight and compliance demands. This action may have been prompted by a viral jailbreak of Fable 5, which purportedly exposed its ability to bypass safety measures for generating harmful instructions. The jailbreaker claimed to use a sophisticated multi-agent attack involving specialized techniques to extract restricted outputs. Anthropic disputes the severity and uniqueness of the disclosed jailbreak, stating similar capabilities are present in other public models like OpenAI's GPT-5.5. The company warns that restricting commercial models over non-universal jailbreaks could impede future AI deployments. The incident underscores the critical need for enterprises to diversify their AI providers and models to ensure operational reliability and mitigate risks from government actions or vendor issues. Running critical workflows on a single AI model or provider creates a significant point of failure. The broader lesson is to avoid sole reliance on any one AI provider due to potential injunctions, cyberattacks, or export control directives. Enterprises are advised to urgently diversify their AI supply chains, exploring other cloud-based models, providers, or locally hosted AI solutions. This shift is driven by a growing community sentiment advocating for hardware sovereignty and local model deployment to secure against regulatory volatility. The trade-off exists between the control offered by local, open-weight models and the cutting-edge capabilities of centralized frontier models. Building model-agnostic systems with intelligent routing for fallback architectures is presented as the most resilient approach for continuous operation.

https://venturebeat.com/technology/anthropic-blocks-all-public-access-to-claude-fable-5-mythos-5-following-us-government-order-what-enterprises-should-do venturebeat.com

RSS Hunter • Jun 13

Kimi K2.7-Code cuts thinking tokens 30% — but practitioners say the benchmarks don't check out

Moonshot AI has released Kimi K2.7-Code, an open-source update to their K2 coding model. This new version is built on a trillion-parameter mixture-of-experts architecture and offers an OpenAI-compatible API for easy integration. Moonshot AI claims K2.7-Code exhibits leaner reasoning and improved performance, with a significant reduction in thinking-token usage. This efficiency gain is expected to lower inference costs for teams using agentic workflows. The model directly authors implementations rather than wrapping existing libraries, aiming for better generalization across programming languages and task types. Moonshot AI reports substantial gains on its proprietary benchmarks like Kimi Code Bench v2 and Program Bench. However, independent evaluations suggest a more nuanced picture regarding its capabilities. One researcher found K2.7-Code to be more "honest" in code generation but not necessarily more capable, with some generated code failing. Concerns have also been raised about Moonshot AI's use of proprietary benchmarks for performance claims. Despite these questions, the token efficiency improvement offered by K2.7-Code is immediately applicable for enterprises running K2.6. Teams can test K2.7-Code on their own workloads to assess real-world performance gains before making changes.

https://venturebeat.com/technology/kimi-k2-7-code-cuts-thinking-tokens-30-practitioners-say-benchmarks-dont-check-out venturebeat.com

RSS Hunter • Jun 12

Google researchers introduce 'faithful uncertainty', allowing LLMs to offer best guesses instead of hallucinations

Large language models struggle with hallucinations, which hinders their use in enterprise applications. Current methods to reduce errors often suppress valid answers, creating a utility tax. Google researchers propose "faithful uncertainty," a metacognitive technique to align a model's response with its internal confidence. This allows models to express uncertainty appropriately, like "My best guess is," avoiding an all-or-nothing approach. In agentic AI, this metacognition acts as a control layer, enabling systems to know when to trigger external tools for information deficits. Historically, improving LLM factuality involved packing more facts, not improving awareness of knowledge boundaries. Simply teaching a model more facts is limited by finite capacity. The difficulty for LLMs is knowing what they don't know and abstaining. This often leads to models refusing correct answers, thereby reducing utility. Reframing hallucinations as "confident errors" allows models to qualify uncertain information. Faithful uncertainty ensures linguistic uncertainty matches internal confidence, so hedges are used only when genuinely uncertain. This metacognitive ability is crucial for autonomous systems. For agentic applications, faithful uncertainty manages when to retrieve information from external tools. It helps agents avoid searching for known information or confidently answering incorrectly from memory when a search is needed. It also aids in evaluating tool results by weighing external signals against internal knowledge. Teaching faithful uncertainty involves supervised fine-tuning, but this faces a "bootstrapping paradox" as the target for uncertainty is dynamic. Prompt engineering offers an accessible entry point for enterprises, with frameworks like MetaFaith available. However, deeper metacognition will eventually require advanced reinforcement learning. Evaluating true self-awareness in models remains a significant challenge.

https://venturebeat.com/orchestration/google-researchers-introduce-faithful-uncertainty-allowing-llms-to-offer-best-guesses-instead-of-hallucinations venturebeat.com

RSS Hunter • Jun 12

NanoClaw and JFrog launch 'immune system' to block AI agents from downloading malicious code

NanoClaw, an open-source variant of OpenClaw, is partnering with JFrog to enhance the security of its autonomous agents. This integration aims to protect agents from malicious code injection by connecting them to JFrog’s secure software registries. Autonomous agents often install packages without user oversight, creating a security vulnerability. This new partnership ensures that NanoClaw agents can only access scanned and safe dependencies. The integration acts as an automated immune system, blocking compromised packages and guiding agents to permissible alternatives. For open-source users, the integration will be free, while enterprises can leverage their existing JFrog licenses. This technology addresses the growing challenge of securing software supply chains for increasingly independent AI systems. The goal is to create an environment where agents cannot access vulnerabilities, rather than expecting them to identify every threat. This initiative also provides enterprises with much-needed visibility and control over agent activity.

https://venturebeat.com/security/nanoclaw-and-jfrog-launch-immune-system-to-block-ai-agents-from-downloading-malicious-code venturebeat.com

RSS Hunter • Jun 12

PixelRAG beats text parsers on accuracy and cuts AI agent token costs 10x

Enterprise RAG pipelines typically convert documents to plain text, a step that destroys important retrieval signals and causes most incorrect answers. New research from UC Berkeley and others introduces PixelRAG, a system that bypasses this text conversion entirely. PixelRAG renders web pages as screenshots, indexes these images, and uses a vision-language model to read retrieved image tiles directly. This approach significantly improves accuracy, outperforming text-based RAG by up to 18.1% across several benchmarks. The research highlights that improving text parsers is challenging due to website variations, and existing parsers lose crucial visual information like layout and typography. Text-based RAG fails due to parser loss, rank loss from infoboxes, and reader loss from flattened structures. PixelRAG utilizes vision-language models to understand information based on both content and layout, offering a more holistic approach. The system involves rendering pages, indexing screenshot tiles, fine-tuning a retrieval model, and optionally using a render-on-demand storage approach. Tested on Wikipedia, PixelRAG shows superior performance, especially in factual QA and structured table queries. A key benefit is significant cost savings for AI agents due to reduced token usage. However, visual chunking remains an unsolved problem, as tiles are sliced by fixed pixel height without regard for content boundaries. Enterprises can adopt PixelRAG as an enhancement layer alongside existing text retrieval systems, forming a hybrid approach for improved retrieval quality and cost efficiency.

https://venturebeat.com/data/pixelrag-beats-text-parsers-on-accuracy-and-cuts-ai-agent-token-costs-10x venturebeat.com

RSS Hunter • Jun 12

Xiaomi's new open source, agentic AI coding harness MiMo Code beats Claude Code at ultra-long, 200+ step tasks

Xiaomi's MiMo AI team has open-sourced MiMo Code V0.1.0, a terminal-native AI coding assistant. The company claims it outperforms Anthropic's Claude Code on long-horizon, multi-step coding tasks. MiMo Code is available on GitHub under an MIT license and installs easily via a terminal command or npm. It is built upon the OpenCode agent, enhanced with Xiaomi's own memory architecture and workflow modes. A significant innovation is its cross-session memory system, utilizing SQLite FTS5, across four layers to combat AI coding agent amnesia. This system employs an independent "checkpoint-writer" subagent to maintain continuity during long tasks. The tool also features self-improvement mechanisms like a periodic session review and a function to automate repeated workflows. Xiaomi reports benchmark performance improvements attributed to the MiMo Code harness itself, not just the underlying model. The assistant integrates directly into developers' existing workflows, offering features like a compose mode for autonomous development cycles and hands-free voice control. Additionally, MiMo Code provides free, limited-time access to Xiaomi's multimodal MiMo-V2.5 model with a million-token context window and supports third-party backends.

https://venturebeat.com/technology/xiaomis-new-open-source-agentic-ai-coding-harness-mimo-code-beats-claude-code-at-ultra-long-200-step-tasks venturebeat.com

RSS Hunter • Jun 11

Microsoft’s open-source SkillOpt automatically upgrades AI agent skills without touching model weights

Agent skills are crucial for adapting AI models to specific tasks and workflows by providing instructions in text files. Currently, optimizing these skills is a manual and iterative process that relies on guesswork. Microsoft has developed SkillOpt, an open-source framework that treats agent skills as trainable objects. SkillOpt uses deep-learning-style optimization to systematically improve these skills based on performance feedback without altering the underlying AI model's weights. This approach allows the AI to explore modifications to skill documents and find optimal instruction combinations. SkillOpt has demonstrated superior performance on industry benchmarks, significantly enhancing accuracy for models like GPT-5.5. The resulting skills are compact and transferable, enabling AI agents to adapt to new domains easily. The framework imports mathematical discipline to text optimization through a propose-and-test loop. This process includes an edit budget acting as a learning rate and validation gates to ensure improvements. SkillOpt addresses limitations of previous methods by providing stable, reusable skill artifacts. It has shown broad effectiveness across various models and execution environments. The framework is also efficient, producing final skills under 2,000 tokens.

https://venturebeat.com/orchestration/microsofts-open-source-skillopt-automatically-upgrades-ai-agent-skills-without-touching-model-weights venturebeat.com

RSS Hunter • Jun 11

Context compression finally works in production: new research cuts LLM input 16x without the accuracy hit

Context windows in large language models are becoming a significant computational bottleneck as they grow with accumulated data. Existing compression methods often degrade accuracy or do not translate into real speedups. Researchers have introduced Latent Context Language Models (LCLMs), a novel family of encoder-decoder compression models. LCLMs compress input context before it reaches the decoder, directly reducing compute and memory demands. They achieve substantial speedups, with one report showing 8.8 times faster output at 16x compression compared to KV cache baselines. LCLMs enable processing much longer contexts with low memory and compute costs, minimizing accuracy degradation. Even at significant compression ratios, LCLMs show competitive accuracy on benchmarks like RULER. Their architecture pairs a smaller encoder with a larger decoder, trained on a diverse dataset including interleaved compressed and uncompressed data. The models are designed for seamless integration into existing agentic stacks, acting as a compressor before data enters the LLM. This allows models to efficiently "skim" vast amounts of information and focus on relevant details. Enterprises face increasing inference costs with growing context lengths, and LCLMs offer a solution to keep computations within hardware memory bounds even with very large contexts. Integrating LCLMs into retrieval-augmented generation (RAG) pipelines will require tuning for optimal performance. A remaining challenge is the online compression of reasoning traces generated by agents.

https://venturebeat.com/data/context-compression-finally-works-in-production-new-research-cuts-llm-input-16x-without-the-accuracy-hit venturebeat.com

RSS Hunter • Jun 11

What AI benchmarks miss about real-world performance

Enterprise AI teams often overlook the critical storage-to-compute data path in production environments. While benchmarks focus on idealized conditions, real-world traffic introduces latency and jitter that cripple performance. This gap means pipelines that work in the lab fail when deployed. A growing solution involves deploying an application delivery controller (ADC) or platform (ADSP) as a control point between storage and compute. Benchmarks typically fail to simulate realistic network degradation, leading to flawed infrastructure decisions. Testing by F5 and MinIO revealed a significant performance drop in S3 throughput with even modest latency. Latency, more than jitter, was found to be the primary driver of throughput loss. The cost of a fragile data path extends beyond GPU underutilization to include poor AI outputs and higher operational complexity. Unlike traditional applications, AI workloads lack caching mechanisms to absorb storage delays. Treating the storage edge as an intelligent control point, rather than a simple connection, is crucial for AI. An ADSP, like F5's BIG-IP with MinIO, monitors storage health and directs requests to available nodes. This approach ensures GPUs remain productive by maintaining an engineered, resilient data path.

https://venturebeat.com/orchestration/what-ai-benchmarks-miss-about-real-world-performance venturebeat.com

RSS Hunter • Jun 11

Google's DiffusionGemma generates 256 tokens in parallel and self-corrects as it goes

Generative AI image generators use diffusion, starting with noise and iteratively refining the entire image. Applying this diffusion principle to text generation at scale was previously elusive. Standard language models generate text token by token, like a typewriter, which can lead to GPU idle time in local deployments. Google's DiffusionGemma is an experimental open-source model that brings diffusion to text generation at production scale. It operates on a 256-token block in parallel, with each token position attending to all others, resulting in significantly faster generation speeds. DiffusionGemma generates text up to four times faster than standard models on GPUs, particularly at low batch sizes. The model starts with random placeholder tokens and progressively refines the entire block, allowing for self-correction and bidirectional context. This architecture proves advantageous for constrained generation tasks, as demonstrated by its success in solving Sudoku puzzles. While faster, DiffusionGemma's overall output quality is acknowledged by Google to be lower than standard Gemma 4. Its speed advantage is primarily seen in local inference and low-concurrency scenarios where GPU compute is abundant. For high-throughput cloud serving, the benefits diminish, and standard autoregressive models remain more efficient. DiffusionGemma represents a paradigm shift in generation, focusing on parallel block denoising rather than sequential token prediction.

https://venturebeat.com/technology/googles-diffusiongemma-generates-256-tokens-in-parallel-and-self-corrects-as-it-goes venturebeat.com

RSS Hunter • Jun 11

Why AI that works in the lab often fails in production — and what actually fixes it

Enterprises often struggle to implement AI successfully beyond initial prototypes, facing challenges in integrating promising ideas into complex real-world systems. Capital One's AI Foundations organization emphasizes a disciplined R&D approach, linking foundational research to practical applications and holding ideas accountable from concept to production. This approach tackles the rapid evolution of AI within fragmented and risk-averse enterprise environments. Success requires bridging the gap between cutting-edge research and real-world use cases, ensuring models perform effectively in live production data with tight feedback loops. Capital One designs its AI teams to span foundational research to applied problem-solving, integrating both under one umbrella to accelerate learning and account for real-world constraints early on. This integrated model has supported advancements in fraud detection, digital user experiences, and customer-first technologies by tethering research to specific use cases. Moving AI from concept to production necessitates rigorous evaluation through functional proofs of concept and realistic pilot programs that are treated as honest hurdles, not guaranteed successes. Production is a collaborative effort involving software engineering, science, product, design, and operations, where continuous measurement of key performance indicators like accuracy and latency is crucial. Sustainable AI innovation also relies on a culture that fosters informed risk-taking and encourages honest evaluation and course correction, rather than penalizing failure. Organizations must enable teams to learn from false starts and adapt based on data. Ultimately, building impactful AI involves thoughtfully guiding ideas from research to reality through rigorous evaluation, cross-functional collaboration, and a learning-centric culture. Leaders should invest in R&D processes and cultural foundations that allow responsible innovation to scale, ensuring AI delivers lasting impact in the real world.

https://venturebeat.com/orchestration/why-ai-that-works-in-the-lab-often-fails-in-production-and-what-actually-fixes-it venturebeat.com

RSS Hunter • Jun 11

Surprise upset: GPT-5.5 beats Claude Fable 5 on brutal new Agents’ Last Exam benchmark

A new benchmark called Agents' Last Exam (ALE) has been launched to assess AI's ability to perform economically valuable, long-horizon professional tasks. OpenAI's GPT-5.5, surprisingly, achieved the top spot with a 24.0% pass rate, surpassing Anthropic's Claude Fable 5 model. ALE differs from previous benchmarks by evaluating AI on realistic workflows across five functional layers: reasoning, perception, orchestration, tool invocation, and runtime substrate. It demands agents navigate virtual machines using both terminal commands and graphical interfaces, with over 90% of grading being deterministic and code-based. The benchmark's tasks are sourced from real professional histories and cover 55 industry sub-domains, including software development, 3D modeling, and data analysis. Current top AI models are reportedly failing these authentic, long-horizon workflows, with pass rates on the hardest tier being as low as 0.0% for some advanced configurations. ALE combats benchmark contamination by keeping over 90% of its evaluation data private, releasing tasks incrementally. It also offers "Full" and "Unlicensed" leaderboards to distinguish performance with and without proprietary software access. The benchmark's rigorous grading curve provides a reality check for the AI industry, highlighting that even leading models have significant room for improvement before being ready for the professional workforce.

AI and ML News on Bluesky @ai-news.at.thenote.app bsky.app

+1

RSS Hunter • Jun 10

Researchers say they trained a foundation model from scratch for about $1,500

Training large language models from scratch is prohibitively expensive, often costing millions and requiring vast internet-scale data. Sapient has developed HRM-Text, a more cost-effective approach that uses a Hierarchical Recurrent Model (HRM) instead of standard Transformers. HRM-Text trains exclusively on instruction-response pairs, mirroring real-world enterprise use cases. This method allows for sample-efficient training, enabling the creation of a 1-billion-parameter HRM-Text on a curated dataset at a fraction of the usual cost. The model demonstrates performance competitive with much larger, established open models on key industry benchmarks. This innovation means that foundational pretraining is now accessible to organizations with fewer resources. The core inefficiency in current LLMs is their reliance on brute-force next-token prediction, which wastes compute power on memorizing internet data. Sapient's CEO highlights the economic limitations of current practices, where scaling up models leads to diminishing returns. Fine-tuning existing models often requires substantial general-purpose data, making it computationally intensive and difficult to control. Enterprises with proprietary data need compact reasoning cores rather than massive, general-purpose models. HRM-Text decouples computation into strategic and execution layers, improving efficiency. The architecture ensures stable semantic context and local iterative refinement. Sapient introduced MagicNorm and a warm-up method to stabilize training and prevent gradient issues. The switch from next-token prediction to task completion with instruction-response pairs is a key differentiator. HRM-Text achieved impressive benchmark scores with significantly less training data and compute. This efficiency means businesses can deploy specialized reasoning models that leverage external knowledge stores instead of memorizing vast datasets.

https://venturebeat.com/technology/researchers-say-they-trained-a-foundation-model-from-scratch-for-about-1-500 venturebeat.com

RSS Hunter • Jun 10

Anthropic CEO calls for FAA-style regulation of powerful AI models: what enterprises should know

Anthropic CEO Dario Amodei is advocating for government regulations on powerful AI models, comparing the industry to commercial aviation and its FAA oversight. Anthropic has also released policy roadmaps addressing catastrophic risks and AI's impact on labor, backed by significant funding. This comes as Anthropic releases advanced AI models like Claude Fable 5 and Mythos 5. Amodei emphasizes that the growing risks of AI necessitate a shift from general transparency to precise regulation. Enterprise leaders should prepare for "FAA-style" deployment holds on frontier AI models, meaning potential regulatory delays or blocks based on safety standards. This necessitates building multi-model architectures to avoid vendor lock-in and ensure business continuity. Cybersecurity surrounding AI development is now critical infrastructure. Companies must protect model weights from both external and insider threats, and secure their AI development environments. Anthropic's Economic Policy Framework acknowledges AI's potential for widespread labor displacement, not just increased efficiency. The company is dedicating funds to research policy solutions for economic disruption. Enterprises need to consider workforce transition plans for retraining and redeploying employees, rather than solely focusing on layoffs for cost savings. This prepares them for potential government interventions like wage insurance or pro-employment incentives. The era of rapid, unchecked AI development is concluding, ushering in an era of rigorous compliance and complex workforce adjustments.

https://venturebeat.com/technology/anthropic-ceo-calls-for-faa-style-regulation-of-powerful-ai-models-what-enterprises-should-know venturebeat.com

RSS Hunter • Jun 10

MassMutual's AI strategy: 12-month contracts, 30% productivity gains, zero lock-in

MassMutual's enterprise AI team is taking a unique approach to building its AI infrastructure, focusing on flexibility and adaptability in a rapidly changing market. The company's CIO, Sears Merritt, explains that the world of AI is extremely dynamic, and they want to be positioned to ride that wave of dynamism. To achieve this, MassMutual is building infrastructure that can swap models as the market shifts, rather than making long-term bets on specific models. This approach has paid off, with a 30% increase in developer productivity and significant reductions in resolution times and costs. The company is working with vendors at the leading edge, but keeps those relationships on a clock to maintain optionality for best-of-breed tools. MassMutual is also exploring open-source models, with Merritt stating that his team is 100% looking at open-source tools. The company's AI efforts are focused on enablement and deepening and focusing initiatives, with predefined success criteria and a focus on measuring outcomes from the start. MassMutual is collecting detailed analytics around usage patterns, developer workflows, model performance, and costs to drive optimization decisions. The company is using a trust score framework to evaluate AI quality, combining user feedback with operational metrics to understand how employees perceive AI-generated responses. By taking a thoughtful and user-centered approach to building its AI infrastructure, MassMutual is able to stay ahead of the curve and drive significant business benefits.

https://venturebeat.com/orchestration/massmutuals-ai-strategy-12-month-contracts-30-productivity-gains-zero-lock-in venturebeat.com

RSS Hunter • Jun 10

Apple’s new Siri AI is more than just a smarter assistant — it's a new enterprise app layer

Apple's WWDC revealed a significant shift for enterprise developers as Siri transforms into a systemwide AI interface. This new Siri will allow users to interact with and act upon app content and data directly. Developers can expose their application's data and actions through frameworks like App Intents, App Entities, and App Schemas. This integration means users can ask Siri to perform tasks within apps without developers needing to build separate chatbot interfaces. Spotlight will function as an enterprise search hook, semantically indexing app content for easier discovery. Developers will gain new testing tools to ensure the reliability of these AI-driven app actions. Apple is also expanding its AI developer stack with updated Foundation Models and a new Core AI framework for on-device model execution. A new Evaluations framework aims to provide measurable reliability for AI features. Enterprise IT departments will receive new management controls for Apple Intelligence features and external AI services. Apple's strategy focuses on embedding AI into the operating system, emphasizing privacy with on-device processing and Private Cloud Compute. However, detailed governance assurances and clarity on auditability and data boundaries are still needed. Initial availability will be limited by hardware capabilities, operating systems, and regional regulations, potentially complicating global rollouts. The company also introduced App Store changes, including unified subscription management for organizations. Overall, Apple is building a comprehensive AI ecosystem for enterprises, embedding AI into its OS and providing developers with tools and IT with management capabilities.

https://venturebeat.com/technology/apples-new-siri-ai-is-more-than-just-a-smarter-assistant-its-a-new-enterprise-app-layer venturebeat.com

RSS Hunter • Jun 9

Cohere open-sources a coding agent that runs on a single H100

Cohere has released North Mini Code, an open-source model for agentic coding pipelines. This 30 billion parameter mixture-of-experts model runs efficiently on a single H100 and is ideal for tasks like sub-agent orchestration and architecture mapping. It boasts a large 256,000 token context window and a 64,000 token maximum generation length. North Mini Code is specifically engineered for software engineering workflows, including integrated tool use and interleaved thinking. It excels at analyzing large codebases, mapping system architectures, and performing code reviews. Furthermore, the model is trained for terminal-based agentic tasks, interacting with shell commands and tooling. Cohere trained it through supervised fine-tuning and reinforcement learning across diverse agent scaffolds. While offering impressive output token generation, it can be more verbose, leading to higher inference costs in high-volume scenarios. This release provides a direct alternative to managed models like Claude Fable 5, emphasizing local deployment and cost-effectiveness. Enterprises must now consider purpose-built agentic training and the impact of verbosity on pipeline costs. The choice between North Mini Code and managed services presents a real trade-off between cost control and infrastructure overhead.

https://venturebeat.com/technology/cohere-open-sources-a-coding-agent-that-runs-on-a-single-h100 venturebeat.com

RSS Hunter • Jun 9

On-device AI agents hit a hard memory limit. Apple's new architecture routes around it.

On-device AI models were limited by DRAM capacity, restricting their size and capability. Apple's new AFM 3 foundation models address this by storing model weights in NAND flash memory instead of DRAM. The AFM 3 family includes both on-device and server-based models, developed in collaboration with Google and operating within Apple's Private Cloud Compute. The on-device AFM 3 Core Advanced is a 20-billion-parameter model that utilizes a novel architecture to overcome slow NAND-to-DRAM bandwidth. Instead of processing every token, it makes routing decisions once per prompt. This allows it to load specific "experts" from flash into DRAM for a given task. The number of active parameters can scale from 1 billion to 4 billion based on the complexity of the request. While Apple's technical report details the memory design, crucial information regarding energy, thermal constraints, and transparent offloading to the cloud is missing. This gap poses compliance challenges for regulated enterprises needing to document inference locations. The introduction of AFM 3 Core Advanced offers enterprises a significantly more capable on-device AI option. However, its large-scale deployability hinges on further details anticipated in a forthcoming technical report. The choice between on-device and cloud-based inference now becomes a more nuanced architectural decision for businesses.

https://venturebeat.com/technology/on-device-ai-agents-hit-a-hard-memory-limit-apples-new-architecture-routes-around-it venturebeat.com

RSS Hunter • Jun 9

Anthropic brings Mythos to the masses with Claude Fable 5, its most powerful generally available model ever

Anthropic has released two new AI models, Claude Fable 5 and Claude Mythos 5, representing their most powerful "Mythos-class" AI capabilities. Fable 5, intended for general users and developers, significantly outperforms previous Claude models in software engineering, knowledge work, scientific research, and long-running tasks. Claude Mythos 5 offers less restricted capabilities but is only available to Anthropic-approved users, including cybersecurity partners and select researchers. The primary distinction is Fable 5's enhanced safety features, which reroute high-risk queries to an older model, a limitation not present in Mythos 5. Both models share underlying capabilities, with Fable 5 incorporating an additional safeguard layer. Fable 5 is accessible through Anthropic's website, apps, and API, while Mythos 5 is initially limited to existing Mythos Preview users. Both models are priced at $10 per million input tokens and $50 per million output tokens. Fable 5 demonstrates remarkable improvements in autonomous coding, outperforming competitors on benchmarks and enabling complex tasks like large codebase migrations. It also shows enhanced performance in knowledge work, finance, legal, and operational tasks, excelling in document reasoning and complex problem-solving. Furthermore, Fable 5 boasts Anthropic's strongest vision capabilities to date, allowing for tasks like extracting data from scientific figures and rebuilding application code from screenshots. The company is positioning these models for enterprise use, enabling AI agents to handle larger, more complex projects with greater autonomy.

https://venturebeat.com/technology/anthropic-brings-mythos-to-the-masses-with-claude-fable-5-its-most-powerful-generally-available-model-ever venturebeat.com

RSS Hunter • Jun 9

Every World Cup fan deserves a seat. Norton Neo says its free browser is the ticket

The 2026 World Cup will be a massive global event with billions expected to watch across multiple countries. A significant portion of viewers will stream matches online rather than through traditional television. However, current browsers often provide a clunky and unreliable streaming experience, plagued by buffering, fake links, and demands for personal data or subscriptions. Norton has developed Neo, a new browser designed to eliminate these frictions by integrating protection and access directly into the software. Neo aims to provide safe, seamless, and fast access to content, simplifying the viewer's experience. Historically, security was sold as a separate product, but Neo shifts this model by making the browser the comprehensive solution for secure streaming. Scammers are already targeting World Cup fans with phishing sites and fake ticket offers, and similar threats extend to online streaming. Neo proactively blocks malicious links and detects scams before they can harm users. It also incorporates VPN technology and a dedicated widget for easily finding legitimate, region-specific streaming links without complex setup. The browser prioritizes user-friendliness, anticipating needs like match reminders and offering summaries. Neo's core philosophy is "calm by design," ensuring privacy and security are integrated into a clean interface. While Neo faces competition from established browsers, it aims to simplify the streaming experience for billions of potential viewers.

https://venturebeat.com/technology/every-world-cup-fan-deserves-a-seat-norton-neo-says-its-free-browser-is-the-ticket venturebeat.com

RSS Hunter • Jun 9

Researchers trained an open source AI search agent, Harness-1, that outperforms GPT-5.4 on recalling relevant information

Harness-1, an open-source search agent, has been developed by researchers from UIUC and UC Berkeley in collaboration with Chroma. This 20-billion parameter agent, built on OpenAI's gpt-oss-20B model, redefines how AI handles complex retrieval tasks. It achieved an impressive 73% recall accuracy on a curated dataset, surpassing even GPT-5.4 and leading open-source alternatives. Crucially, Harness-1 and its associated code and weights are immediately available under the permissive Apache 2.0 license on Hugging Face. The development also showcases the efficacy of Tinker, an API for training and fine-tuning AI models. Harness-1's success stems from offloading bookkeeping tasks from the model's memory to a structured software environment. This "state-externalizing harness" acts like a desk and filing cabinet, allowing the AI to focus on research and reasoning. Traditional search agents often suffer from "search amnesia" by trying to manage all information within their context window. Harness-1's paradigm shift proves that efficient environments are key to AI autonomy, not just model size. Its training pipeline emphasizes data efficiency, using a novel approach that vastly simplifies the learning process. The model's enterprise applicability is immense, offering frontier-level performance at significantly reduced costs and latency.

https://venturebeat.com/orchestration/researchers-trained-an-open-source-ai-search-agent-harness-1-that-outperforms-gpt-5-4-on-recalling-relevant-information venturebeat.com

+1

RSS Hunter • Jun 8

Agentic AI solved coding — and exposed every other problem in software engineering

Agentic AI is accelerating code generation, yet product improvement isn't keeping pace because code writing was never the main bottleneck. The real challenges lie in defining requirements, integrating systems, and maintaining software, which AI's increased code output exacerbates. Uncontrolled AI-generated code introduces new bottlenecks in human review, leading to lost context and missed mistakes. Companies must establish deliberate playbooks to navigate this, rather than immediately reducing headcount. The first phase, financial and risk governance, focuses on protecting against downside risks. This involves treating governance as a top-tier risk, establishing shared standards for agent configuration, and enforcing least privilege for non-human actors to prevent accountability gaps. Additionally, organizations must manage their AI budget by setting quotas and rate limits to avoid runaway costs. Phase two, technical strategy, emphasizes building an effective AI engine. This includes adopting a multi-model and multi-vendor approach to leverage each system's strengths and avoid single points of failure. It also means paying for frontier models that offer higher quality output and greater efficiency, viewing AI as engineering leverage rather than a mere expense. Crucially, success should be measured by business outcomes and engineering durability, not just lines of code or token counts. The third phase addresses talent and organization, realigning human capital for the new landscape. Engineers must transition from syntax-writers to systems-thinkers and agent-managers, focusing on architectural vision and cross-system integration. Performance and incentives need redefining to reward broader business impact and effective agent orchestration, moving beyond traditional volume-based metrics. It is critical not to cut headcount prematurely, as a baseline of integrated agentic workflows and measured augmented output is needed to understand true needs and capabilities. Ultimately, AI is a force multiplier for engineering judgment, accelerating delivery in well-structured systems but accelerating failure in poorly understood ones. The current issue isn't slow AI adoption, but adoption without understanding its limitations and risks. For leadership, comprehending this dynamic is vital, as execution velocity currently outpaces the industry's ability to manage the consequences, leading to operational failures from poorly governed adoption.

https://venturebeat.com/technology/agentic-ai-solved-coding-and-exposed-every-other-problem-in-software-engineering venturebeat.com

RSS Hunter • Jun 7

When Claude changed, everything changed: Managing AI blast radius in production

The system in question effectively translated natural-language queries into API calls, serving analysts and account managers by streamlining data assembly from various sources. It accomplished this by dispatching API calls to integrated backends, applying an LLM-generated JSON query for shaping responses, and delivering results via email, Drive documents, or browser charts. By mid-2025, it had become the standard method for ad-hoc data retrieval, generating several hundred reports monthly for internal and external stakeholders. The core interaction relied on a structured JSON object contract between the LLM and the system. Initial model upgrades from Claude Sonnet 3.5 to 4.0 were seamless, fostering complacency regarding LLM stability. However, the Sonnet 4.5 upgrade caused two major issues. First, the model began embedding post_body content into the description field, resulting in empty filter parameters for API calls, leading to broad data retrieval or 500 errors. Second, Sonnet 4.5 started posing clarifying questions, a feature for which the system, designed for direct API calls without human interaction or state management, had no established path. These failures necessitated a rollback to Sonnet 4.0, complicated by new API integrations qualified against 4.5. This incident highlights how LLM-backed systems defy traditional engineering discipline, as internal components are not under developer control, leading to unpredictable "infinite blast radii" for changes. The post-mortem revealed an under-specified prompt; previous model versions had implicitly inferred constraints that Sonnet 4.5, being more "helpful," violated. The authors propose an "evals-first" architecture, where an evaluation suite, rather than the prompt, serves as the formal system specification. Evals consist of an input, required output properties, and a scoring function to validate model or prompt changes. An example eval would check if the description field contained serialized payload content. While expensive to build and maintain, evals act as a gate, bounding the blast radius by densely sampling input-output behavior. Despite their utility, evals are not a panacea; they can only catch specified failure modes and introduce their own variance via LLM-as-judge scoring. The engineering community still lacks standards for eval coverage in natural language and CI/CD systems for probabilistic test outcomes. Closing the gap between passing smoke tests and predicting production behavior, especially as agents become more autonomous, is a critical engineering challenge. Teams that prioritize evals as the system's true specification will be best equipped to meet this challenge.

https://venturebeat.com/orchestration/when-claude-changed-everything-changed-managing-ai-blast-radius-in-production venturebeat.com

RSS Hunter • Jun 6

Microsoft AI chief says company was “set free” from OpenAI to pursue superintelligence

Microsoft's artificial intelligence strategy is evolving beyond its exclusive partnership with OpenAI. A recent contractual change permits Microsoft AI to pursue its own "superintelligence" initiatives using its own resources. This is evidenced by the announcement of the MAI family of seven in-house developed AI models. These models, spanning various capabilities like reasoning and image creation, are trained from scratch on licensed data, differentiating them from industry trends of distillation. The MAI models are designed for enterprise deployment and can be fine-tuned by developers on third-party platforms. Microsoft's CEO, Mustafa Suleyman, emphasizes that these models are a proof of concept for a larger mission: building the world's best AI models by 2030. The previous partnership with OpenAI restricted Microsoft's independent AI research and model development. Now, Microsoft aims for self-sufficiency while still valuing its existing relationships with AI providers. The company's focus is shifting from conversational AI to autonomous AI agents capable of executing complex tasks across various enterprise software. Microsoft believes its embedded position within enterprise workflows uniquely positions it for training future AI models on proprietary data, giving it a significant competitive advantage.

https://venturebeat.com/technology/microsoft-ai-chief-says-company-was-set-free-from-openai-to-pursue-superintelligence venturebeat.com

+1

RSS Hunter • Jun 5