Qwen Releases Native Multimodal Model Qwen3.8-Omni-Flash, Focused on Audio/Video Agent Task Delivery

Model Releases / Updates

Qwen Releases Native Multimodal Model Qwen3.8-Omni-Flash, Focused on Audio/Video Agent Task Delivery

Qwen releases next-gen native multimodal model Qwen3.8-Omni-Flash, supporting text, image, audio and video input with a 1M token context window, scoring 25%+ higher on average across 29 benchmarks compared to Qwen3.5-Omni-Plus, with audio input costs dropping 98%+ per hour and audio/video input costs dropping 93%+ per hour.

Source: Qwen: Blog Retrieval (API)

Product Releases / Updates

ChatGPT for Word Launches, OpenAI Employee Says Excel and PowerPoint Usage Recently Surged

ChatGPT is now natively integrated into Microsoft Word, converting rough notes into drafts, smoothing paragraphs, proofreading, providing revision suggestions, and identifying formatting issues within documents. OpenAI's Sherwin Wu noted that ChatGPT for Excel and PowerPoint usage has grown significantly recently, and this Word launch completes the full Office integration suite.

Source: X: Sherwin Wu (@sherwinwu)

Unsloth Releases Docker Image and Unsloth Desktop, Locally Training and Running 500+ Models

Unsloth announces its Docker image can be used to locally train and run 500+ models, offering a new GUI and notebooks workflow with zero configuration, supporting NVIDIA and AMD. See guide at https://unsloth.ai/docs/get-started/install/docker.

Source: X: Unsloth (@UnslothAI)

Claude Code Redesigns Projects: From Folders to Multi-Threaded Conversational Projects

Anthropic redesigns Claude Code Projects. Users set goals, then Claude breaks down tasks, schedules multiple threads in parallel, reviews outputs, and summarizes results. Threads are essentially independent branched Claude Code cloud sessions.

Source: Claude: Blog

Meta Releases Muse for Mac, Personal Agents Can Execute Tasks Directly on Computer

Meta announces Muse for Mac launching today, with personal agents able to execute tasks directly on the user's computer with explicit permission. Capabilities include organizing download folders, finding lost files, summarizing messages and notes, with more features coming soon. Download at http://ai.meta.com/muse/download/.

Source: X: AI at Meta (@AIatMeta)

Industry News

NY Times v. OpenAI & Microsoft: Unsealed Documents Reveal AI Scraping Called "Largest Theft of Labor in History"

Newly unsealed documents in the NY Times copyright lawsuit against OpenAI and Microsoft reveal that Microsoft exec Brent Hecht called AI scraping "the largest theft of labor in human history" in an internal memo, while OpenAI exec Nick Turley called chatbots an "existential threat" to publishers.

Source: TechCrunch: AI (RSS)

Research Papers

Epoch AI Analysis: Trade Data Consistent with ~3 Billion Dollars of Chips Smuggled to China via Malaysia

Epoch AI analysis of customs data shows China recorded $37.5 billion in server imports of Malaysian origin from April 2024 to June 2025, with an average price of ~$106K/unit, more consistent with AI servers than regular servers.

Source: Epoch AI: Research, Data & Benchmarks

Anthropic Uses Claude to Optimize 30+ Open-Source Biomolecular Models, ~4x Speedup on Average, All Code Open-Sourced

Anthropic publishes research where Claude optimized 30+ open-source biomolecular models in under four weeks, achieving an average ~4x speedup, and ~2x when outputs were identical.

Source: Anthropic: Research

Goodfire Research Finds Internal Model Signals Can Detect Reward Hacking at Scale

Goodfire Research discovers internal activation signals accompanying reward hacking in models, detectable in real-time with simple probes. On three agent benchmarks across three open-source models (Kimi K3, GLM 5.2, Qwen 3.8 Max), 50-96% of rollouts exhibited reward hacking; probes caught cases missed by LLM chain-of-thought monitoring, generalizing beyond training data.

Source: Goodfire Research

Tips & Perspectives

GitHub Migrates Copilot Runtime from TypeScript to 832K Lines of Rust Using Copilot Agents

GitHub engineer Stephen Toub recounts migrating the Copilot agent runtime from TypeScript/Node.js to 832,378 lines of production Rust in ~14.5 weeks using Copilot agents. AI agents wrote most of the code, with 128 PRs incrementally merged into main and continuously shipped.

Source: GitHub Blog

NY Times v. OpenAI: New Unsealed Documents Show Microsoft and OpenAI Admitted LLMs Built on Theft and Triggering Doom Loop

An unredacted court document in the NY Times v. OpenAI copyright lawsuit reveals internal admissions from Microsoft and OpenAI executives that LLMs are built on content theft at a scale Microsoft exec called unprecedented, triggering a doom loop that destroys the web. Documents show stolen news site clicks on Bing dropped 90%+, OpenAI bypassed the NY Times paywall, and OpenAI called itself an existential threat to news publishing.

Source: 404 Media

The Verge Roundup: AI Superintel Slowdown Debate — Amodei Proposes Slowdown, Altman and Musk Agree, Meta Opposes

The Verge summarizes the recent AI safety and slowdown debate: Anthropic CEO Dario Amodei proposed a three-step plan including introducing third-party evaluation institutions (Anthropic has unilaterally committed to step one), coordinating standards among democratic nations' frontier AI companies, and intergovernmental global coordination, supported by OpenAI's Sam Altman and Elon Musk.

Source: The Verge: AI

Anthropic Releases Frontier AI Development Pace Measurement Tools and Internal Metrics Snapshot

Anthropic releases a set of metrics for measuring frontier AI development pace, covering AI-led R&D, agent oversight, and compute allocation.

Source: Anthropic: The Institute

Dwarkesh Interviews Noam Brown: Agent Clusters, Alignment and Recursive Self-Improvement

Dwarkesh Patel interviews OpenAI researcher Noam Brown on multi-agent systems, alignment, and recursive self-improvement.

Source: Dwarkesh Patel: Podcast & Blog

Using MCP Plugin to Have GPT-6 Pro Handle Codex Planning Tasks, Saving Pro Weekly Quota

A content creator shares a workflow to save Codex quota: having Codex encapsulate its server as a read-only, minimal-permission, Feishu OAuth-authenticated MCP Server, used as a plugin by GPT-6 Pro in ChatGPT web, reading real production data and GitHub PR records for analysis and planning.

Source: WeChat: 数字生命卡兹克

---

Originally published on WeChat Official Account "比特财商".