All AI news

AIByteDance Seed researchers found that language models compressing their KV cache in fixed-size chunks retrieve the same information unevenly depending on token position. In a 128K-token needle-in-a-haystack test, base DeepSeek-V4 checkpoints differed by up to 40.2 percentage points by phase, and post-training narrowed but did not eliminate the gaps. The authors urge evaluating such models across positional phases, since high average accuracy can hide systematic failures.

AIHuawei's OceanStor M900 and Nvidia's CMX move reusable inference KV cache into a shared storage tier, but no agreed SSD specification exists for it. A storage executive said endurance requirements for one design rose from 3 to 9 drive writes per day and could change again. Industry sources expect convergence to take 6 to 12 months, with another year for development and validation.

AISimplexity Robotics says it trained a single robot to load and unload two CNC lathes on its own, using 600 real-robot trajectories and reporting a 100% success rate on the precision CNC insertion task. The work, presented at IROS 2026 on September 29, combines the SimpleWAM world action model, a DRAM memory module and DPE action scoring, with force and torque feedback for insertion recovery. The company did not say how many trials the 100% figure covers, and it does not describe the yield of the whole cell.

AIHuawei has opened its DevEco Studio for HarmonyOS PCs to public beta, alongside first public betas of the AI tool DevEco Code and the agent toolkit DevEco CLI. The beta requires HarmonyOS 7.0.0.107 or later, at least 16 GB of memory and 100 GB of storage, and runs on several MateBook models and the MatePad Edge. DevEco Code ships with Zhipu AI's GLM-5.3 and GLM-5.1 models and supports third-party model connections.

AIHong Kong's Centre for Artificial Intelligence and Robotics (CAIR), Chinese Academy of Sciences, unveiled CARES 4.0, a multimodal clinical AI agent that carries out tasks rather than only answering questions about images or video. Built on the Harness agent framework and CAIR's own CT, MRI, ultrasound, endoscopy and EEG foundation models, it has been validated at several top-tier hospitals. CAIR says the system gives suggestions with reasoning paths and sources, and that the doctor remains the final decision-maker.

AIMirroS released AgentGarten, which pairs executable code environments with a real-time neural renderer running above 30 fps so agents can act, observe, and learn. In a one-on-one hide-and-seek setup, the hider learned to block passages by round 4 and the seeker learned to climb ramps by round 10, guided by notes the agents wrote after each round. The authors report applying the same loop to four other tasks, including a dog-companion game, a narrow-bridge car passing task, herding, and quarry loading.

AITencent Hunyuan, with Fudan and Tsinghua researchers, released ExplorationBench, a benchmark testing how AI systems explore through verifiable "Alien Worlds" with executable rules that conflict with familiar knowledge. Across 10 frontier systems, feedback mattered most: the best AlienCode run reached 89.0% after four rounds of probing, versus 0.5–11.0% without feedback. Answers are graded by an interpreter or proof checker rather than an LLM judge.

AIOpenClaw says it has added many new features and quality-of-life improvements over the past few months. A video covers new models, multiplayer, interactive dashboards, memory and skills, meetings and voice, and easier Mac setup. It also previews plans for the coming months.

AICoreWeave is layering managed services over its infrastructure to address AI inference bottlenecks, including a preview capability called CoreWeave RL Rollouts that improved model reload latency by 15x versus a baseline configuration in testing. The capability is built on Nvidia's Dynamo framework, and the features are packaged into CoreWeave Forge, a platform that is free to start with paid tiers offering additional capabilities.

AIPeter Thiel, Palantir and Founders Fund co-founder, argued in $100-per-ticket Nashville lectures that government attempts to regulate AI are evil, according to recordings obtained by Politico. He attacked former President Barack Obama and Pope Leo XIV over AI, and criticized effective altruism as a belief system that could form a one-world government to stop AI progress. The article notes Thiel's wealth is heavily invested in AI, including a $418 million AI-focused portfolio at Thiel Macro LLC.

AIAI systems are now solving math problems faster than humans can verify the answers, according to the source's headline. The source text is limited to a brief note that OpenAI's math breakthroughs may trigger an identity crisis for mathematicians.

AIThe University of Michigan's EECS 498 course Applied Agentic Software Engineering teaches a coding agent across three phases, from applying and analyzing agents to building one. Its Elephant-Goldfish Model packages a design-first workflow into five Skills, with human handoffs between each step, and the course materials are public on GitHub.

AITeknium announced that TinyFish, a new browser backend, is now available on the Hermes plugins catalog. According to a related post, TinyFish is added as a first-party Hermes plugin that lets agents search and fetch the live web for free, and it can be installed with hermes plugins install tinyfish.

AITeknium called a Hermes Desktop demo "pretty sick" after Jonathan Bylos reported that Hermes Agent produced an intelligent UI embed during a design discussion without being asked. Bylos said the feature has been running in Hermes Desktop for a few days.

AIswyx says the going compensation package for this role at frontier agent labs is between $5M and $50M. The quoted post describes it as among the hardest roles to hire for, citing a small pool who combine trend awareness, creative taste, AI knowledge, and execution.

AIMiles Brundage calls a question from Jordan Schneider about what individuals can do to help pass the MATCH Act a very good question. The post itself gives no further details on the bill or proposed actions.

AIOpenAI published 719 AI-generated math proofs covering 372 result families, after withdrawing 3 for a symbol error. Reports say the release falls short of the AGMAI advisory group's standards, since it uses proprietary models, includes reasoning chains for only 10 manuscripts, and leaves about 42% unformalized. Terence Tao argues that rapidly solving famous problems harms the mathematical community's understanding and collaboration.

AIGoogle Cloud introduced the Gemini agent, a general office agent that can search, write emails, build slides, analyze data, run code, and coordinate sub-agents. It can take on an enterprise identity with email, calendar, and account, and it selects underlying models automatically, including Anthropic's Claude. The article presents this alongside OpenAI's Dots and Meta's Muse as competing office and personal agents.

AIGoogle Cloud unveiled Gemini Agent at its Gemini at Work 2026 event, a general-purpose work agent for enterprise customers that works across Gemini Enterprise and Workspace. Currently the agent supports Gemini and Claude models, with an API planned so developers can integrate it into third-party applications.

AIFactory has appointed Connor Maloney as Head of Federal to lead its expansion into the public sector. Maloney previously served as Vice President of Federal at Rubrik and began his career as a Program Manager at the Department of War. Factory also recently announced a partnership with Carahsoft and says it can run in managed, private cloud, on-premises, or fully air-gapped deployments.

AIJerry Liu is preparing for the Agent Economy Gala, which has reached its 400-guest capacity and is organized by @imaanxsultan. The source says attendees include 203 founders with over $1B raised combined, 67 founding operators, 58 investors, 23 creators, and 295 guests building agent applications or infrastructure.

AIJeremy Howard says Jasmine Wang accessed an executive's email that OpenAI had given her access to, despite her request that the access be removed and IT's failure to remove it. He implies she was fired because of that IT failure, not misconduct on her part.

AISnyk moved its internal support agent, Snyk Assist, into the core Snyk product in September 2026, giving every paying customer access. Built on LangChain and LangGraph with observability in LangSmith, the agent answers questions in plain language and can open support cases or log feature requests. It runs as a single agent behind Slack, web and API surfaces, with tools attached per user permissions.

AIAnthropic has barred users from exhibiting "sustained and needless abusive or cruel behavior" toward its models, according to a policy change first reported by The Verge. The San Francisco-based company says the ban does not apply to common user frustrations, model testing, or "dark creative themes." The change follows an August feature that lets Claude end conversations when a user is persistently harmful, which Anthropic framed as a safeguard for AI welfare.

AIAustralian AI data center operator Firmus, backed by Nvidia, has withdrawn its planned initial public offering, citing market volatility and conditions. Its board concluded the proposed terms did not adequately reflect the company's business strength and long-term growth outlook. Firmus said it will now pursue private market capital and consider other public and private options.

AITessl's blog post argues that teams should record agent mistakes as attributed, signed diary entries, then curate them into reusable context packs rather than adding unverified rules to files like AGENTS.md. The author describes a REST API case where an agent regenerated the OpenAPI spec and TypeScript client but missed the Go client, and the same lesson had to be re-taught in a fresh session.

AITessl argues that an AI-native organization collapses the handoff between people who own outcomes and the work itself, so product managers and designers can execute changes through agents. It says PR count and token spend are insufficient measures, and that merge rate better shows whether the new workflow is working. The article also says the boundary should follow decision authority, with engineers still owning architecture and data models.

AIThe author argues that system use cases, with actors, preconditions, scenarios, and acceptance criteria, work better than user stories as the input for AI code generation in enterprise business applications. He describes a process that skips the plan-and-task phase, reverse-engineers legacy systems into use cases and entity models for modernization, and recommends self-contained system verticals and risk-based review.

AIAI DevCon New York, running November 2–4 at Industry City in Brooklyn, centers its program on software factories, the systems needed to make agentic development repeatable, trustworthy and scalable. The article argues that moving from one developer using an agent to an engineering organization requires layers covering context and skills, harnesses and tools, orchestration, verification and evaluation, and feedback.

AIThe article explains how a software factory combines agents, tools, and engineering practices to take work from intent to verified output. It argues that teams should enforce domain rules, workflow gates, and tool permissions independently, version shared guidance and checks, and review lessons before reuse.

AIOpenAI's DevDay 2026 session demonstrates Codex shifting from a single-user tool to a team-oriented agent. The session shows a persistent personal agent investigating a 2am outage, from the first Slack message through a reviewed fix, using voice, Appshots, plugins, and meeting notes to keep the team informed.

AIOpenAI's DevDay 2026 video presents what's new in Codex across the CLI, app, and web. Live demos show moving from delegating individual tasks to having Codex tackle entire problems and goals. The source is a short description only, so no specific features or figures are confirmed.

AIOpenAI's DevDay 2026 session shows how its teams use ChatGPT to manage launches, research competitors, and turn expertise into tools. The source provides no further details on specific features, metrics, or availability.

AIA DevDay 2026 session from Higgsfield explores how its agents choose models and provision their own compute. It covers when to use dedicated GPUs versus hosted models, how to deliver predictable throughput, and where the approach can fail.