SketchAgentBench – evaluate LLMs on 3D spatial reasoning

AI is dramatically accelerating AI research

Show HN: Agents infer Lego mechanisms from sources, then animate them

Replit rebuilt its sandbox engine with microVMs

Devvm.dev: One VM for every project

China's CO2 emissions fell by 3% in the third quarter.

According to an analysis by Carbon Brief, China's CO2 emissions for the third quarter (July–September) fell by 3.1% year-on-year, marking the largest decline since 2022. Year-to-date, China's carbon emissions have cumulatively decreased by 0.5%. Data shows that China's oil consumption in the third quarter dropped by 4% year-on-year, while coal usage fell by 3%. One key factor behind the decline in third-quarter CO2 emissions is a 3.3% reduction in coal-fired power generation. This is due to a slowdown in electricity demand growth, coupled with a decrease in unused generation from wind and solar energy. Solar, wind, and nuclear...

DigitalOcean launches learning and data services to enable AI agents to self-evolve.

DigitalOcean has released the public preview of Learning Services and Agent Data Services, enabling developers to build, evaluate, and continuously improve AI agents on a unified AI-native cloud platform. Built on DigitalOcean's Inference Engine and Managed Agents, these services allow agents to learn from real-world interactions, identify weaknesses, test improvements, and utilize enterprise-owned data: we built the Inference Engine to enable agents to think, and Managed Agents to enable them to take actions, and...

AI-Native Enterprise R&D Toolchain: Architectural Refactoring and Continuous Evolution | AICon Beijing

When AI Agents truly integrate into enterprise workflows, which tasks can be delegated to AI? Discover the answers through 50+ real-world case studies! AI is evolving from a supportive tool for answering questions into an autonomous executor capable of understanding goals, invoking systems, and completing tasks. As Agents enter enterprise processes, the focus shifts from "how to build AI" to critical questions: which tasks can be delegated, how to validate outcomes, and how to manage permissions, costs, and risks. With large language models expanding into reasoning, tool invocation, multimodal processing, and long-horizon tasks, the development focus is transitioning from single calls to...

OpenAI launches GPT-6 Cyber with stronger exploit research and higher API price than Anthropic and Google

OpenAI released GPT-6 Cyber today, a specialized security model that the company says can uncover and exploit more software vulnerabilities while more reliably respecting the limits researchers impose.Read more

When the Coding Agent enters the R&D team: We have built and open-sourced the self-evolving R&D platform, LoopForge.

Introduction: While standalone coding agents can rapidly generate code, stepping into a real-world R&D team reveals challenges such as interrupted workflows, fragmented multi-user contexts, uncontrolled quality in free-form "vibe coding," time-consuming manual monitoring, and repeated pitfalls from similar issues. These quickly trap teams in the bottleneck of "individual efficiency gains at the cost of overall internal friction." To address this, Tencent's Lightweight Cloud team has developed and open-sourced LoopForge, a cloud-based self-evolving AI R&D platform featuring 7×24-hour cloud workspaces, multi-role shared sessions, engineering-grade SOP pipelines, unattended automation, and...

Tencent Launches Hy4, One of the Leading Open-Source AI Models

Vera Delivers the Best of Both Worlds: Highest Throughput and Fastest Agents

AI agents do more than generate answers. They search for information, query databases, write and test code and take action.  Each of those steps relies on a CPU, with agents moving between GPU reasoning and CPU execution throughout a task.  As businesses put more agents to work, AI factories need CPUs that can execute those […]

New NVIDIA Cosmos 3 Open Models for Interactive World Simulation

Axiom Math wants to keep AI slop out of mathematics

As AI floods math with unchecked proofs, Axiom Math is betting on verification. The startup’s AI agents just formally verified the Stacks Project, one of the field’s biggest libraries.

The Wizard of Open Weights: Chris Alexiuk Makes AI Easy With Open Models

Chris Alexiuk has a simple rule about open source AI: If you believe in it, be loud about it. “We should be shouting about this technology from the rooftops,” said Alexiuk, a Toronto-based senior product research engineer in deep learning at NVIDIA who specializes in making open large language models (LLMs) practical for working developers.  […]

Embodied AI: The Game Has Changed

Data collection orders are exploding, and Embodied Brain Company is panicking.

91% ≈ 0%? Google's latest research bursts the bubble of high accuracy in engineering agents.

41.3% of design items failed to meet safety thresholds, while complex engineering remains trapped in constraint coupling. Authors: Zheng Jia Mei, Cen Feng Editor: Cen Feng In the engineering world, 91% correctness is sometimes indistinguishable from 0% delivery capability. A single cable tray running overhead in a server room through fire protection piping can render dozens of pages of correct calculations worthless for delivery. Even if current-carrying capacity is calculated correctly, voltage drop meets requirements, and equipment selection is well-founded, once introduced into 3D space, multiple systems compete for the same clear space, forcing engineers to withdraw the proposal. AI Agents can quickly complete...

Ex-Googlers' new Apache 2.0 model Ripple-65B Preview claims the U.S. open coding lead on two major benchmarks

There's another American victory worth paying attention to today: a new, open weights AI model that its creators claim scores the higher on coding performance on two third-party benchmarks compared to any other open model from the U.S.Read more

Morning Brief: Tencent launches the official version of Hunyuan 4.0, Google adjusts its AI policy for K-12 education, among other updates.

New designs for the RetroPad mousepad from the Minority Club are now available for purchase, redeemable with membership points. Explore the latest co-created iPhone accessories for the Minority Club, featuring customizable combinations of protection, charging, cables, and stands. Don't miss out on our featured articles: Getting Started with Apple Intelligence (China Version): These Features Are Already... Read More

Towards Foundation Decision Models: Drex

Authors: Nischay Dhankhar, Alisher Amantay, Zhanibek Datbayev, Darkhan Akhmetov, Yash Shah, Nurmukhamed Ubaidullayev, Raman Saparkhan, Dos Baha, Abulhair SaparovUpvotes: 7Summary: Autoregressive approaches with large language models, such as long chain-of-thought reasoning, have become the default for new AI applications because they reach higher accuracy on a wide range of tasks. In production, however, they give inconsistent answers, require parsing of outputs, and add cost and latency. Dec...

Almost no one uses AI agents to book their Uber, company says

During its busiest week last quarter, Uber said only 4 people booked a ride using an AI agent.

How NVIDIA Factory Operations Blueprint Coordinates AI Agents to Investigate Production Issues

A Siemens report estimates that unplanned downtime costs an average large plant in the surveyed sectors $253 million per year. Whether investigating a failed...

Inkitt Builds a One-Shot PR Creator with Resumable Agent Orchestration

Inkitt built a one-shot pull request creator that automates small, low-risk code changes from task intake to draft PR creation. Its event-driven orchestrator persists workflow state, coordinates short-lived coding agent containers, bounds retries, and enforces permission restrictions while retaining human review and approval before changes can be merged. By Leela Kumili

Zero-Shot Coordination among LLM Agents

We study zero-shot coordination (ZSC), where independently developed agents must coordinate at test time. While ZSC has been well studied in the RL literature, far less is known about the performance of LLM agents despite their increasing deployment in such settings. Existing work on LLM coordination often relies on specialised scaffolding that independently developed agents are unlikely to share in practice. Moreover, evaluations on complex environments (e.g., Hanabi) make it difficult to pi...

Presentation: Panel: Who Builds the Frontend Now? (And What Breaks When They Do)

The panelists explain how AI tools and vibe coding reshape frontend and mobile engineering. They share insights on evolving junior developer roles, managing AI-generated code review fatigue, designing systems for autonomous agents, enforcing architectural guardrails, and navigating economic shifts in open-source software sustainability. By Luca Mezzalira, James Hall, Danielle An, Ivan Zarea

In the era of large models, the "calculation power victory" of intelligent driving chips has become obsolete.

If you can't compete with general-purpose chips, go for custom chips.

Anthropic is making a slight push; is the sky about to fall for AI relay stations?

Just like sending an email with a digital signature, #Follow our official WeChat account iFANR (WeChat ID: ifanr) for more exciting content delivered right to you.

A.S.E 2.0: Understanding How Repository Context Impacts the Security of LLM-Generated Code

Large language model (LLM) coding assistants and agents construct a repository context before generating code. For security-sensitive code, this context decides whether the model sees the project's safeguards (e.g., authorization checks) or insecure code to imitate. However, context construction is designed almost exclusively for functional correctness, and secure-code benchmarks treat the context as fixed, so how repository context affects the security of generated code remains unclear. We c...

Personal Agent: Ant wants to make inspiration a constant spark, not just a fleeting one.

How far is Ant from having a moment of inspiration to having inspiration become a permanent fixture?

Blog Review: Nov. 4

Chiplets for physical AI; rack-level considerations; verification complexity; sparse matrix workloads; AI data centers. The post Blog Review: Nov. 4 appeared first on Semiconductor Engineering.

The Little Book of Generative AI Foundations: An Intuitive Mathematical Primer

Reddit tests ads in search results as traffic declines and growth concerns mount

The new ad placements could help Reddit squeeze more revenue from existing users as Google's AI search changes how people discover its content.

How I use AI (Oct 2026)

Qwen 3.8 27B on my laptop can interpret model weights, too

USDB: Scaling Storage for Next-Generation Recommender Systems at ByteDance

Google will shut down search for U.S. schools that refuse AI.

Google has announced that AI features will be disabled by default for software used in U.S. K-12 schools, unless school administrators opt to enable them. This move aims to address the concerns of educators and parents, who seek greater control over their children's interactions with AI. Today, AI features are standard in Google's commercial search engine. Chris Phillips, Google's vice president of education, stated that search and AI are deeply integrated; if schools disable AI, students will no longer be able to use search. Google is widely used in U.S. K-12 schools...

The Download: controlling AI agents and “Super Intelligence” loyalty tests

This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology. How “reverse game theory” could keep AI agents in check A new approach known as “reverse game theory” aims to control powerful AI models by designing systems that incentivize them to…

Accelerating engineering for the physical world

Discover how ChatGPT and Codex are helping engineers accelerate CAD, PCB design, simulation, and other workflows across the physical world.

Building the Next-Generation LLM Inference Engine for Agentic Reasoning

Against the backdrop of rapidly growing inference workloads for agents, traditional inference engines designed for chat scenarios face a series of structural challenges, including the failure of throughput optimization goals, insufficient cache utilization, and lagging hardware adaptation. This article is compiled from the talk "Building the Next-Generation LLM Inference Engine for Agentic Inference" by Shang Xuchun, an independent contributor to TokenSpeed, at AICon 2026 Shenzhen. He systematically introduces the complete technical path of the project from its inception to version iteration, focusing on workload characteristics in agent scenarios, a layered decoupled system architecture, ...

Apple Watch’s AI features for rewinding and recapping conversations begin beta testing

Now in beta: Apple's new Audio Intelligence features can help you remember what someone said or take notes on conversations.

The developer behind the AI-powered replication of Adobe's software suite admits that the project is far from ready.

Brandon Thomas posted his open-source alternatives for seven Adobe applications built with AI to Hacker News, but quickly tempered expectations. "I'm not quite ready to post this to HN," he wrote, "it's still far from being readable." He indicated he would post again once the project exits beta. A few days prior, five Reddit subreddits had posted teasers conveying a different message, stating the project was about to enter beta and expected to achieve near 100% feature parity within a month. Meanwhile, the codebase's...

Temperatures of 40°C could reach the Nordic region.

Heatwaves in Nordic countries such as Norway, Sweden, and Finland have traditionally been less intense than those in other parts of continental Europe. However, in recent extreme weather events, such as the 2026 heatwave, record-breaking high temperatures have been observed across much of Europe, with 40°C heatwaves also occurring in regions farther north than previously seen. According to a study published in the journal npj Natural Hazards, Finnish scientists report that 40°C heatwaves in the Nordic region are no longer considered impossible as once thought. Researchers estimate that in 2025, a certain...

Voluntary Servitude: Weaponized Safety Warning Mechanisms for Controlling AI Agents

LLM agents increasingly act on information retrieved from public sources they treat as credible, which turns a manipulated recommendation into a transaction, a dependency, or a decision with physical consequences. The security literature models this threat as instruction injection, in which an adversary writes a command into retrieved content and deployed defenses learn to recognize command syntax. We show that an attacker can instead steer an agent through the agent's own harm-avoidance. A s...

Twice as Fast, Until It Isn't: Quantized Serving of Mid-Size Open LLMs

A team serving a 9B-20B open-weight language model on vLLM or SGLang must choose a quantization recipe, and needs two numbers: the speedup over BF16 and the accuracy given up. We assemble every public measurement of these numbers that we could verify and find that the record supplies them only in pieces. Only two checkpoints, both INT4 builds of Gemma 3 12B, have both a vLLM timing and an accuracy score, and that score covers a single MMLU subject, one that the AWQ build was calibrated on. Th...

The Anthropic team, starting with zero Rust experience, leveraged Claude to perform a cross-language refactoring of the core API, achieving a 40% reduction in CPU usage and a 20% decrease in latency.

Authors: Lin Qibei, Editors: Cai Fangfang. Anthropic recently completed a highly representative engineering transformation: a team with no prior Rust developers used Claude to migrate the company's Messages API from Python to Rust. According to Katelyn Lesse, Anthropic's Head of Platform Engineering, the transformation reduced CPU resource consumption for the relevant services by approximately 40%–50% and improved first-token latency by about 10%–20%. In this project, Claude handled a significant portion of the code generation, while human engineers primarily focused on system architecture...

BYD sold more EVs last month than the US did all quarter, topping 300,000

SLATE: A Large-Scale Benchmark for Long-Horizon API Selection in E-commerce Tool Use

Enterprises increasingly want LLM agents to automate standard operating procedures over their software stacks -- mapping each step of a procedure to the right call among thousands of internal APIs, without hand-authored step-to-API mappings. The core difficulty is API selection at scale: choosing and composing the correct tools over a long horizon, judged only by whether the overall task succeeds. Current tool-use benchmarks do not test this regime -- they expose only a handful of tools, eval...

Astra can read and write weights to improve Qwen

Can LLM Agents Improve Imperfect Earth System Model Parameterizations?

Earth system models (ESMs) rely on parameterizations to represent unresolved processes such as convection, cloud microphysics, and turbulence. However, translating advances in physical understanding into stable, validated parameterizations and integrating them into ESMs often takes years. Agentic AI could accelerate this process and enable more systematic exploration of design choices. Using warm-rain collision-coalescence as a test case, we evaluate agents' ability to synthesize scientific l...

Playing Dumb: How a Chess Transformer Simulates Users Across Skill Levels

Generative models increasingly simulate users at different skill levels in education and AI evaluation. Yet it remains unclear whether lower-skill behavior reflects less internal knowledge or a different use of retained knowledge. This distinction is difficult to study in natural language, where proficiency and relevant knowledge are hard to measure. In this work, we study Maia-2, a chess transformer that predicts human moves conditioned on player skill, using quantitative skill ratings and d...

Anthropic's internal AI use, at retail prices, now exceeds its employee pay

Elon Musk throttles Grok Bot limits due to ‘unsustainably high’ costs

Musk recently referred to Grok Bot as having "the fastest growth of a product that I've ever seen in my life."

Support New AI Frameworks Faster Using NVIDIA Workload Map

NVIDIA DSX AI Factory Platform brings together technologies to help the ecosystem design, build, and operate more efficient AI factories, including open source,...

OpenAI is finally merging ChatGPT’s Chat and Work tabs

OpenAI said that ChatGPT and Codex will draw usage from the same allocation pool after this change.

From Local Skills to Long-Horizon Web Tasks: Progressive Skill Exploration for LLM Web Agents

Large language models (LLMs) are driving autonomous agents toward real-world environments, with the Web emerging as one of the most representative interaction scenarios. However, training Web agents with naive exploration remains challenging: successful trajectories are rare, and the sparse terminal rewards make it difficult for agents to obtain useful feedback from failed trajectories. Inspired by skill acquisition theory, we propose Skill-Composed Progressive Exploration (SCOPE), a short-to...

How to play the second half of embodied AI? Agibot gave its own answer at a press conference.

After the arrival of GPT-6, do robotics companies still hold value? Agibot responds with the "Worken" compute ledger behind its 20,000-unit shipment. By Author: Qi Chengyong; Edited by Cen Feng. In 2026, the embodied intelligence sector is sinking into an unprecedented wave of emotional anxiety. The embodied control capabilities demonstrated by GPT-6 Astra have fueled the prevailing narrative that "physical AI is about to be breached by AI"; meanwhile, the S2 layer, responsible for perception and decision-making, is being overtaken by general intelligence at a visibly accelerating pace. The industry has even seen a clear divergence in strategic pathways: some companies focused on models are beginning to pivot...

ChatGPT is secretly profiling you; use this prompt to make it reveal it.

Perhaps the world depicted in *Her* has already arrived. #Follow our official WeChat account iFANR (WeChat ID: ifanr) for more exciting content delivered right to you.

Linux 7.4-rc1 released

Linus Torvalds announced on the kernel mailing list that Linux 7.4-rc1 has been released, closing the merge window, with the official release expected before Christmas. Linux 7.4 includes a record 18,591 non-merge commits. A total of 2,678 developers contributed code to v7.4, with 625 making their first contribution. Torvalds noted that he was pleased to see a significant influx of AI-related work during the merge window; the patch size merged for Linux 7.4 exceeds that of typical releases, indicating that AI fixes are gradually becoming integrated into the regular development workflow...

Stack Overflow says it's still alive.

Stack Overflow's own statistics show that its monthly new question count has dropped below 1,000. This has once again sparked discussions on social media about the impending death of the programming Q&A community. The Stack Overflow blog responded by stating that death predictions are exaggerated, conflating the changing ways of acquiring knowledge with the disappearance of knowledge itself—and the company behind it. AI has altered developer behavior. In the recently released "2026 Developer Survey" from Stack Overflow, 64% of respondents stated that for simple questions...

Overnight, both GPT-6.2 Astra and Fable 5.5 were revealed, signaling that November is going to be a wild month for the AI community.

Global LLM Showdown: The Mad November Launches Officially! #Follow our official WeChat account iFANR (WeChat ID: ifanr) for more exciting content delivered straight to you.