AI News

TrackMCP shows who is using your MCP server, what they are trying to do, whether the work gets done, and where to improve. All from one line of code.

Airtop is a conversational web agent builder.
It lets you create AI agents that can log in, browse any website, extract data, and complete tasks - just by describing what you want in plain English. No code, no workflows, no APIs. Just words.

At IFA 2026, NVIDIA and partners are providing faster inference and new tools that make agents easier to set up and run locally.

Omarchy is an omakase Linux desktop on Arch, Hyprland, and Quickshell. It ships a complete keyboard-first workstation with system-wide themes, packaged updates, and coding agents that can reshape the setup instead of making you assemble it yourself.

Alibaba's 27B dense multimodal model for agentic coding, tool use, research, and long-running workflows. It accepts text and image inputs and supports configurable reasoning.

Banger paper from Google DeepMind and colleagues.

(bookmark it)

A model reads its entire KV cache on every generated token, even though it ends up attending to a tiny slice of it.

In other words, if you ask about one detail from a 1M-token conversation the global attention layers re-read all of …

Join the discussion on this paper page

Join the discussion on this paper page

WeatherNext 3 is a major breakthrough in how we forecast global weather. ⛅

Developed with @GoogleResearch, the model learns directly from real-world, real-time observations to give more localized highly accurate predictions faster. 🧵

A Blog post by IBM Research on Hugging Face

Qualcomm and ASUS partner to launch a local-inference Pharmacy AI Agent for community pharmacists in southern Taiwan, donating AI laptops to 50+ pilot pharmacies.

The DLL leaked inside this title just days ago

Audacity 4.0
Audacity 4 rebuilds the application interface on Qt and introduces many new quality-of-life improvements, including a new clip-editing model. Most Audacity 3 workflows remai

Atlas is World Labs' omni world model. It takes text, images, video, and 3D, then generates camera-controlled 1440p video up to a minute, reconstructs scenes from a few photos, and simulates space-time for robotics. Early access.

AMD Threadripper Halo Station debuts at IFA 2026, offering 576GB of HBM3E memory and the ability to run trillion-parameter AI models locally without cloud compute. AMD also launched Kraken Halo (Ryzen AI Max Pro 400) with 192GB unified memory for 300-billion-parameter models, with Lenovo and HP

Hermes Desktop now sets up local models in one click.

It automatically reads your hardware, picks the best model for you, then downloads it and configures the runtime.

Explore K2 Horizon, IFM’s open-source fleet of six frontier AI models for reasoning, coding, agentic workflows, edge devices, and enterprise deployment.

Reasoning traces of large language models are widely read as containing "breakthrough" moments and early-legible fates. Both readings rest on measurements missing a counterfactual control at the level of the claim; we supply both controls. First, a restart-controlled truncation probe separates when a solution fits the continuation budget from when a prefix carries value that fresh computation cannot buy, comparing per-anchor continuation solve rates against from-scratch restart curves at matched total generated-token budget. Applied to 178 problem-model cells (89 MATH problems x two small open models, an outcome-blind but difficulty-targeted cohort), exactly 1 of 178 cells survives as prefix-limited; restart dose-response separates a compute-starved model from a capability-limited one; and wherever the matched budget lies inside the restart grid, continuing the model's own prefix beats restarting (9 of 9) -- predominantly compute compression rather than expanded reachability. Second, a pre-registered, difficulty-controlled test finds no detectable outcome information in early-window internal signals beyond a problem-difficulty baseline, and two generation-free analyses of public corpora show why this control is needed: a trace-blind difficulty proxy reaches AUROC 0.873 on 192K DeepSeek-R1 generations -- inside the published probe range -- and a closely matched reconstruction of the closest published early-window positive recovers a comparable pooled result (0.849) while within problem it is statistically indistinguishable from chance at all ten anchors (0.496 at t=4); a post-hoc within-targeted probe finds only a small average residual, concentrated in three low-failure problems. High pooled probe AUROCs cannot by themselves establish within-attempt information; a question-only baseline or within-problem evaluation is required.

Join the discussion on this paper page

Grok Bot is now available for enterprises. Grok and Cursor Enterprise customers have free usage for the next two weeks, and can invite their whole organization, including people without an existing seat.

Nvidia said Hugging Face hosts over 3 million models and is used by over 18 million developers.

worlds via code, from fable 5.1. Contribute to PhiloLabs/fable51-worlds development by creating an account on GitHub.

HTX showcased OceanOneK at Milipol TechX 2026 and plans a humanoid robotics centre by September 2026 for hazmat handling and firefighting.

Acer showcased the Acer SFF RTX Spark design for a compact desktop powered by the NVIDIA RTX Spark superchip at Acer's next@Acer global press conference on September 2 during IFA 2026. NVIDIA RTX Spark redefines personal computing, combining personal agents, advanced content creatio