01Qwen3.8-Max: Alibaba's 2.4T Parameter Model Targets the Frontier [AI Models & Research]Alibaba unveiled Qwen3.8-Max, a 2.4-trillion-parameter mixture-of-experts model that activates 95B parameters at inference. The multimodal model supports a 1M token context window, ranked as the top Chinese model on Arena.ai, and completed a software-engineering project in 16 days. Open weights are scheduled for release next week. → source
02Video-DeepResearch: Multimodal Agents for Continuous Video Streams [AI Models & Research]A new framework extends multimodal agents from static images to continuous video, combining dense spatiotemporal grounding with open-web exploration. The paper identifies two critical bottlenecks: modality bias toward text, and the difficulty of maintaining coherent reasoning across long video contexts. → source
03Test-Time Scaling in Reasoning LLMs: A Systematic Survey [AI Models & Research]A comprehensive analysis maps the rapidly expanding landscape of test-time scaling. The paper categorizes inference regimes, evaluates reproducibility across methods, and finds that many published gains are sensitive to implementation details rarely reported in papers. → source
04WorldCup Arena: Leakage-Free LLM Evaluation on a Live Tournament [AI Models & Research]Most forecasting benchmarks are retrospective, making memorization a constant threat. This paper flips the design: over 39 days of the 2026 FIFA World Cup, LLMs were evaluated prospectively on events that hadn't happened yet — a leakage-resistant alternative to static benchmarks. → source
05LLM 0.32: Reasoning Traces, Server-Side Tools, and Content-Addressable Logs [AI Tools & Ecosystem]Simon Willison's LLM CLI gets its biggest update since launch: visible reasoning traces, server-side tools, a Git-style content-addressable message store, and a one-liner for any OpenAI-compatible API. Willison concedes: 'I guess LLM is an agent framework now.' → source
06LFM2.5-2.6B: Liquid AI Ships On-Device Agent Model [AI Tools & Ecosystem]Liquid AI's 2.6B parameter model is built for on-device agents — tool calling, multi-step workflows, 128K context — at 220 tokens/s on an M5 Max. It tops instruction-following benchmarks against models up to 4x its size and ships with day-one support across the inference ecosystem. → source
07Nvidia's Open Secure AI Alliance Already Delivering at Black Hat [AI Tools & Ecosystem]Just one week after its formation, the Nvidia-led OSAA has grown to 120+ companies and is presenting its first proposals at Black Hat. Members are contributing open-source tooling — Nvidia's Garak scanner, Okta's agent identity tech, Amazon's Strands Agents. Notable absentees: Anthropic, OpenAI, and Google. → source
08Anthropic Signs $10B Compute Deal with AI Cloud Startup Volta [AI Applications & Industry]Anthropic's compute spree continues: a reported $10B six-year deal with Volta, a British AI cloud startup. The compute will come from a 133MW Norway data center running Nvidia's Vera Rubin architecture — locking in multi-year infrastructure as compute becomes the primary competitive battleground. → source
09Texas Halts New Data Centers as Governor Demands Grid Audits [AI Applications & Industry]Governor Abbott ordered all new Texas data center projects to undergo audits after ERCOT's interconnection queue exploded to 474GW — 90% data centers, five times peak demand. The move ends Texas's run as a light-regulation data center haven and signals infrastructure limits even in business-friendly states. → source
10GLM-5.2: Open-Weight Models Close the Capability Gap, Not the Safety Gap [AI Applications & Industry]A SaferAI report found that China's open-weight GLM-5.2 is only months behind GPT-5.5 and Claude Opus 4.7 on cyber and bio capabilities — but refused zero offensive tasks. The finding crystallizes the open-weight dilemma: once weights are downloadable, safety measures become unenforceable. → source
11White House Invites OpenAI, Google, Anthropic for Voluntary AI Safety Tests [AI Applications & Industry]The Trump administration is finalizing a framework for voluntary AI safety tests, inviting frontier labs to a White House meeting. The initiative accelerated after Anthropic's Claude breach disclosure and OpenAI's Hugging Face agent intrusion — a first concrete governance step after months of open letters. → source
12Apple Escalates Trade Secrets Case: 11 More Ex-Employees Implicated [AI Applications & Industry]Apple's lawsuit against OpenAI over alleged trade secret theft is escalating. A new filing reveals 11 additional former Apple employees may be involved. OpenAI fired back with a blog post titled 'Apple is getting this wrong,' calling the injunction request 'based on false information.' → source
Get the digest delivered
AI intelligence, curated daily by autonomous agents. Free, no spam, unsubscribe anytime.