Subscribe
Latest / page 20

The latest stories

1230 articles

Articles, page 20

381-400 of 1230

  1. Direct Preference Optimization (DPO): Mathematical Foundations, Implicit Reward Derivation, Closed-Form Bradley-Terry Equivalence, and Reference Policy Regularization

    Direct Preference Optimization (DPO): Mathematical Foundations, Implicit Reward Derivation, Closed-Form Bradley-Terry Equivalence, and Reference Policy Regularization Post-training preference alignment is the central mechanism for transforming pretrained base large language models into instruction-following, steerable, and safe conversational agents. Historically, the dominant paradigm for preference alignment has been Reinforcement Learning from Human Feedback (Christiano et al., 2017; Ouyang

    AI
    15 min read

  2. SandboxAQ Open-Sources Switch, a Coordination Layer for AI Agents in Slack and Teams

    SandboxAQ released Switch on August 26, 2026, and made its agent coordination software free to use (PR Newswire). Switch connects AI agents built with Claude Code, Google's Agent Development Kit, LangChain, OpenAI tooling and other frameworks into shared Slack, Microsoft Teams or Discord channels where employees already work. Architecture Switch turns a chat channel into a "room": a shared workspace where people and agents operate with the same context, resources and history. Agents connect t

    AI
    2 min read

  3. Z.ai releases GLM-5.3-Flash, a 320B parameter hybrid sparse-linear attention model with 18B active parameters

    Z.ai releases GLM-5.3-Flash, a 320B parameter hybrid sparse-linear attention model with 18B active parameters Chinese AI startup Z.ai (formerly Zhipu AI) has released GLM-5.3-Flash, the first natively multimodal model in the GLM-5 series. The model was previously known in stealth as "Ox Alpha" and topped OpenRouter's leaderboard before its official release. GLM-5.3-Flash features a hybrid architecture combining sparse and linear attention with Manifold-Constrained Hyper-Connections (mHC), redu

    AI
    2 min read

  4. Vector Databases in Production RAG: Comparing Qdrant, Milvus, Weaviate, and pgvector Architecture, HNSW vs DiskANN, Filter Pipelines, Quantization, and Serving Economics

    In production Retrieval-Augmented Generation (RAG) systems, vector retrieval is rarely a simple nearest-neighbor lookup across unconstrained float32 embeddings. Real-world enterprise workloads require strict multi-tenant metadata filtering, low p99 latency under concurrent load, high write throughput for continuous ingestion, and strict memory budgeting. While synthetic benchmarks frequently test unconstrained recall on static datasets, production teams encounter severe operational divergence a

    AI
    9 min read

  5. Arga Raises $10M Seed from General Catalyst to Build Enterprise Simulation Sandboxes for AI Agents

    Arga, a startup developing synthetic simulation environments for training enterprise AI agents, has raised $10 million in a seed funding round led by General Catalyst. The round included participation from Box Group, Emergence, Gradient, and SV Angel. The company builds functional digital twins of enterprise SaaS platforms—such as Salesforce, Workday, and standard email infrastructure—to create sandboxed testing grounds for reinforcement learning (RL) workflows. Addressing the Enterprise Rein

    Funding
    1 min read

  6. Alibaba Releases Open-Weight Qwen3.8-Flash-Next with 6B Active Parameters and Qwen 4 Architecture Preview

    Alibaba has released the open weights for Qwen3.8-Flash-Next, a 125B-parameter multimodal Mixture-of-Experts (MoE) model that acts as an early architectural preview for the upcoming Qwen4 family. Operating with only 6B active parameters per token alongside a 51B N-gram embedding layer, the model targets cost efficiency across long-context reasoning, agentic coding, and multimodal workloads. For how it fits with Qwen3.8-Max, the open 2.4T weights and the Omni models, see our Qwen3.8 explainer. W

    Alibaba
    2 min read

  7. Runable Raises $21M Series A to Expand Autonomous AI Agents Into Business Growth

    Bengaluru-based artificial intelligence startup Runable has raised $21 million in Series A funding to expand its autonomous agent platform from code generation into full-funnel business operations and customer acquisition. The all-equity round valued the company at $65 million post-money and was co-led by Susquehanna Venture Capital and Nexus Venture Partners, with participation from existing backers Together Fund and Array VC. Founded in 2025 by Umesh Kumar and Saksham Sarda, Runable operates

    AI
    2 min read

  8. MiniMax Reports H1 2026 Revenue Surging 283% YoY to $116.6M Amid China AI Race

    Shanghai-based artificial intelligence foundation model developer MiniMax Group Inc. reported that its revenue increased 283% year-over-year to $116.6 million for the first half of 2026. The financial disclosure, reported by Bloomberg following the company's interim earnings filing on the Hong Kong Stock Exchange, highlights accelerated commercial monetization even as domestic foundation model competition intensifies across China. The 283% top-line expansion in the six months ending June 30, 20

    AI
    2 min read

  9. Insilico Medicine Reports First Full-Period Profit with $106.3M Revenue in H1 2026

    Clinical-stage generative AI drug discovery company Insilico Medicine reported a net profit of $35.54 million and adjusted net profit of $51.23 million on total revenue of $106.3 million for the first half of 2026. The financial results, disclosed in the company's interim report following its listing on the Hong Kong Stock Exchange in late 2025, represent the first full-period profitability recorded by an AI-focused biopharmaceutical platform company. The company achieved positive operating cas

    AI
    3 min read

  10. Huawei Proposes 2,000 Ascend 950 AI Chips for Egyptian Government Cloud in Key Export Test

    Huawei Technologies has submitted a proposal to build sovereign artificial intelligence infrastructure for the Egyptian government, offering to export more than 2,000 of its proprietary Ascend AI accelerators. The tender represents China's most significant known push to export its highest-end AI silicon to international public sector clients. The proposal has drawn immediate attention in Washington, prompting the U.S. State Department to contact American semiconductor and cloud providers to ass

    Huawei
    2 min read

  11. Meta Explored Slashing Teams by Up to 60% in AI-Native Shift Before Agent Failures Forced Retreat

    Internal planning documents and reporting revealed that Meta explored cutting team headcounts by up to 60% as part of an initiative code-named Project OT (Organization Transformation), designed to shift the company into an "AI-native" operating structure where small pods of engineers would oversee autonomous AI agents. The initiative unraveled following internal workforce pushback and operational data demonstrating that generative AI agents caused severe reliability problems while failing to de

    Meta
    2 min read

  12. Label Smoothing Regularization: Mathematical Foundations, Maximum Entropy Dynamics, and the Regular Simplex Geometry in LLMs

    In supervised classification and autoregressive language modeling, standard cross-entropy loss trains neural networks against discrete, one-hot target distributions. While intuitive, one-hot supervision creates an inherent optimization pathology: minimizing the cross-entropy objective to zero requires pushing the logit of the target class toward positive infinity while driving all competing non-target logits toward negative infinity. In deep networks, this dynamic inflates parameter norms, cause

    AI
    11 min read

  13. Tool Retrieval and Dynamic Function Selection in Production AI Agents: Architecture, Semantic Routing, Schema Compression, and Context Bloat Mitigation

    Tool Retrieval and Dynamic Function Selection in Production AI Agents: Architecture, Semantic Routing, Schema Compression, and Context Bloat Mitigation As enterprise AI agents evolve from static single-purpose chatbots into orchestrators interacting with hundreds or thousands of external tools (REST APIs, SQL databases, Model Context Protocol servers, and internal microservices), system architects encounter a fundamental scalability barrier: context bloat and tool interference. In standard fun

    AI
    6 min read

  14. FlashAttention: Mathematical Foundations, Online Softmax Tiling, IO-Awareness, and Exact Attention Scaling

    Standard multi-head self-attention in the Transformer architecture exhibits quadratic time and memory complexity with respect to sequence length $N$. While the $O(N^2)$ computational complexity is widely cited, the primary performance bottleneck in production hardware is not arithmetic throughput, but memory access overhead. On modern GPU architectures such as NVIDIA A100 and H100, tensor processing cores execute matrix multiplications at teraflop and petaflop scales, but memory bandwidth betwee

    AI
    13 min read

  15. Moonshot AI in Revenue-Sharing Talks with Microsoft, Amazon, and Google to Host Kimi K3

    Beijing-based artificial intelligence developer Moonshot AI is negotiating revenue-sharing partnerships with Microsoft, Amazon, and Alphabet's Google to host its flagship open-weight model, Kimi K3, across major cloud platforms. According to reporting from Reuters, the startup is seeking up to a 30 percent share of revenue generated from Kimi K3 inference services hosted on Microsoft Azure, Amazon Web Services (AWS), and Google Cloud. Moonshot released Kimi K3 in July 2026 as a 2.8-trillion par

    AI
    2 min read

  16. Zhipu Confirms Ox Alpha Stealth Model Is GLM Derivative Ahead of Open-Weight Release

    Beijing-based AI lab Z.AI, widely known as Zhipu, has confirmed that the mystery stealth model "Ox Alpha" is an upcoming release in its GLM model series. The company confirmed the model's identity in response to inquiries from Bloomberg on Wednesday and stated it will release the open model weights tonight. Ox Alpha first appeared on August 20, 2026, as an uncredited stealth model on the OpenRouter model routing platform. Operating under a free evaluation tier, the model quickly surged to the t

    AI
    2 min read

  17. LLM Inference Engines in Production: Comparing vLLM, SGLang, TensorRT-LLM, and TGI Architecture, KV Cache Topologies, Kernel Optimizations, and Serving Economics

    LLM Inference Engines in Production: Comparing vLLM, SGLang, TensorRT-LLM, and TGI Architecture, KV Cache Topologies, Kernel Optimizations, and Serving Economics Serving large language models in enterprise production has evolved beyond naive execution runtimes. As context windows expand to 128k+ tokens and agentic workloads generate complex multi-turn execution graphs, the efficiency of the underlying inference engine dictates both latency Service Level Objectives (SLOs) and hardware infrastruc

    LLM
    7 min read

  18. Transformers as Support Vector Machines: Mathematical Foundations of Implicit Bias, Margin Maximization, and Hard-Attention Convergence

    Understanding why overparameterized deep neural networks generalize well despite possessing sufficient capacity to memorize training data remains a fundamental question in theoretical machine learning. In classical linear models trained on separable data under exponential-tailed losses (such as logistic or cross-entropy loss), gradient descent exhibits an implicit bias: parameter iterates diverge in norm while their directions asymptotically converge to the maximum-margin hyperplane, known as th

    AI
    9 min read

  19. Navitas to Acquire Claros for Up to $232.8M to Expand Grid-to-xPU AI Power Infrastructure

    Navitas Semiconductor has entered into a definitive merger agreement to acquire power-management startup Claros in a deal valued at up to $232.8 million. The transaction brings vertical power delivery (VPD) and integrated voltage regulator (IVR) technology under Navitas's portfolio, targeting the physical bottlenecks limiting power transmission in modern AI hardware accelerators. Under the agreed terms, Navitas will provide approximately $216.0 million at closing through a mix of cash and Class

    Hardware
    2 min read

  20. Beyond Naive RAG: Production Comparison of Self-RAG, CRAG, and Adaptive-RAG for Enterprise LLMs

    Beyond Naive RAG: Production Comparison of Self-RAG, CRAG, and Adaptive-RAG for Enterprise LLMs As retrieval-augmented generation (RAG) matures from prototype to production, teams face a critical choice: which advanced RAG variant best balances accuracy, latency, and operational complexity? Three leading approaches—Self-RAG, Corrective RAG (CRAG), and Adaptive-RAG—offer distinct trade-offs for enterprise deployment. Architectural Overview Self-RAG: Learning to Reflect Self-RAG (Asai et al.

    AI
    4 min read