News
Short posts prepared from approved sources with clear links back to the original articles.
RSS feedDeepSeek reportedly raised more than 50 billion yuan in its first external funding round, valuing the Chinese AI startup at about $50 billion.
A new arXiv paper introduces CoRA, a reinforcement-learning framework that rewards answer correctness, confidence, and rationale support together.
LangChain detailed methods for improving LLM-as-a-judge evaluators so automated scoring better matches human preferences.
SpaceX is buying Cursor for $60 billion, a move aimed at strengthening Elon Musk’s AI and enterprise software ambitions.
Malaysia-based Respond.io raised $62.5 million to expand its AI agent-powered messaging platform and pursue acquisitions in North America and Europe.
Probably raised $9 million to build AI systems aimed at reducing hallucinations and factual errors before they reach users.
A new arXiv paper introduces PrologMCP, an open-source MCP server that exposes Prolog as a reusable stateful tool for LLM agents.
The Justice Department argued that xAI’s unpermitted gas turbines are tied to national, economic, and energy security because the Pentagon needs the company’s AI capacity.
A new arXiv paper proposes a definition of good explanations for AI outputs that accounts for counterfactuals and the user’s prior beliefs.
A new arXiv paper proposes a behavioral way to measure trust between AI agents by observing when they spend resources to verify teammates’ work.
A new arXiv paper proposes a geometric framework for identifying and manipulating “AI engrams,” or memory-like traces, inside deep neural networks.
LangChain detailed how Replit uses LangGraph and LangSmith to trace agent behavior, improve performance, and support human-in-the-loop workflows.
LangChain highlighted Fleet’s use of both general-purpose chat and specialized agents to handle ad hoc tasks and recurring responsibilities.
LangChain introduced Test Run Comparisons in LangSmith to help teams inspect, filter, and compare LLM evaluation results more quickly.
A new arXiv paper introduces PhoneHarness, a mixed-action benchmark for phone-use agents that combine GUI actions, CLI commands, and structured tools.
Its Blackwell systems led MLPerf Training 6.0 results, highlighting the growing importance of training infrastructure for frontier AI models.
LangChain made LangSmith transactable in Azure Marketplace, giving Azure customers a path to deploy the LLM app DevOps platform with VPC control and MACC support.
Companies are now managing “tokenomics” as AI coding tools drive heavy usage and force managers to weigh productivity against compute costs.
Microsoft framed enterprise AI adoption around two requirements: stronger intelligence and enough trust for customers to deploy systems at scale.
WIRED examines AI-powered senior monitoring systems that appeal to families and understaffed care agencies but raise questions about surveillance, safety, and consent.
A new Towards Data Science guide walks through running a local LLM with OpenClaw on a Mac Mini, targeting users who want to reduce API costs.
A new arXiv paper evaluates the robustness of LLM-based proof autoformalization in Lean 4 when informal proofs are paraphrased or otherwise perturbed.
Stack Overflow announced Stack Overflow for Agents, a beta service designed to give AI coding agents shared access to reusable developer knowledge.
LangChain introduced Promptim, an experimental library for optimizing prompts, improving model performance, and making prompt changes easier across models.
A new arXiv paper proposes SERAF, a retrieval-augmented time-series forecasting framework that retrieves both numerical patterns and generated textual descriptions.
A critical Copilot vulnerability, dubbed SearchLeak, could allow attackers to steal two-factor authentication codes from users.
A new arXiv paper introduces Metric Match, a subset-selection method for estimating LLM judge reliability with fewer human annotations.
An InfoQ presentation explores how to scale cloud-hosted browser infrastructure for AI agents using secure Chromium environments, Firecracker, and MCP.
A new arXiv paper introduces Nemotron 3 Ultra, a 550B-parameter hybrid Mamba-Transformer MoE model designed for efficient long-context and agentic reasoning.
OpenAI spent $34 billion over the past year, underscoring the scale of compute, talent, and infrastructure costs behind frontier AI.
A new arXiv paper proposes Telegraph English, a readable symbolic format that compresses retrieved passages into entity-relation statements for multi-hop question answering.
A new arXiv paper finds that online skill and memory modules for web agents can lose their advantage when evaluated under a fixed token budget.
A new Towards Data Science post describes a recovery layer for LLM agent pipelines that adapts payloads, preserves state, and protects schema integrity during fallback swaps.
Anthropic backed away from a planned billing change for the Claude Agent SDK, keeping SDK and third-party app usage tied to regular subscription limits.
LangChain published a guide to using LangSmith regression testing to compare experiments, track performance, and catch changes in LLM applications.
The Institute of the Estonian Language released a benchmark measuring how susceptible language models are to Russian propaganda narratives.
Plaud says it has shipped more than 2 million AI notetakers and grown its software business beyond $100 million in annual recurring revenue.
OpenRouter published a 2026 comparison of free LLM APIs, looking beyond headline pricing to rate limits, model access, costs, and data-use tradeoffs.
A LangChain guide explains pairwise evaluation for LLM application development and shows how teams can compare outputs using LangSmith.
LangChain described an agentic engineering architecture built on LangGraph, arguing that specialized agent swarms can mirror real software teams and compress delivery work.