GLM-5.3 Tops Open-Source Leaderboard, Anthropic Pauses RL Training, Google Launches AI Study Tools

Model Releases & Updates

Liquid AI Releases LFM2.5 QAD Q4_0 Quantized Checkpoints, Recovering 97% of Precision Loss

Liquid AI released four Q4_0 GGUF checkpoints (LFM2.5-230M, 350M, 1.2B-Instruct, 2.6B) trained with Quantization-Aware Distillation (QAD), recovering 97% of BF16 average precision loss while preserving native Q4_0 memory and speed advantages.

Source: Hugging Face Blog

---

Product Releases & Updates

GLM-5.3 Launches: 60 on AA Benchmark, Tied for #1 Open-Source, Lower Cost

GLM-5.3 API is now live, excelling at complex coding, defensive cybersecurity, and long-horizon tasks. It scores 60 on the AA comprehensive intelligence benchmark, on par with closed-source flagships like Claude Fable 5 and GPT-5.6 Sol, and tied with Kimi K3 for open-source #1. With a smaller parameter footprint and lower inference cost, it lowers the barrier to frontier intelligence. API pricing matches GLM-5.2; weights open next Friday.

Source: Zhipu (GLM) WeChat Official Account

FastMetal Generates 5-Second 480P Video on Mac in 30 Seconds Flat

A 5-second 480P video generated entirely on a Mac in 30 seconds. No CUDA, no cloud, just 3.9 GiB of memory. FastMetal brings FastWan-QAD to Apple Silicon with DiT, DMD samplers and decoders running on Metal via MLX, INT8 by default. Three model sizes: 1.3B (480P), 5B (720P), 14B (quality-focused).

Source: X: Sky Computing Lab (@haoailab)

Google Search Launches 5 AI Learning Tools

Google rolled out 5 AI learning tools in Search. AI Mode's generative UI is now live globally in English with interactive visualizations and custom simulations. Practice quizzes covering ACT, SAT and other standardized tests are free in AI Overviews and AI Mode.

Source: Google Blog: AI (RSS)

Replit Launches Free Mode Powered by GPT-5.6 Luna, Build Software for Free

Replit's Free Mode, powered by GPT-5.6 Luna, lets users turn ideas into running software without worrying about token costs. Leveraging GPT-5.6's price-performance advantage and OpenAI's recent price cuts, it serves millions with fast answers, suggestions, and project analysis—and can seamlessly switch to GPT-5.6 Sol for higher reasoning while preserving project context.

Source: OpenAI: News (RSS, excluding enterprise/customer stories)

Claude Code v2.1.236: New Default Model Env Var and Cross-Session Idle Notifications

Claude Code v2.1.236 introduces the ANTHROPIC_DEFAULT_MODEL environment variable to set a default model for new sessions, while /model overrides still apply and persist across restarts.

Source: Claude Code: GitHub Releases (RSS)

---

Industry News

Anthropic Slows Model Development in the Era of Critical Cyber Capabilities: Pauses RL Training, Strengthens Safety

Anthropic is temporarily slowing model scaling—including pausing RL training on its latest model for two weeks—due to the OpenAI-Hugging Face incident and concerns that the upcoming Astra model may reach a "critical cybersecurity capability threshold." The company has strengthened research environment security: workload isolation, network isolation, continuous safety testing, and expanded chain-of-thought monitoring. Workloads involving Astra or cyber models must meet the strictest safety standards; some training and evaluation remains paused.

Source: Hacker News Top (buzzing.cc Chinese translation)

OpenRouter Announces Merger with Stripe

OpenRouter is merging with Stripe to accelerate global economic growth. Currently processing 10+ trillion tokens daily from 400+ AI models for 10M+ developers and companies, with inference volume growing at least 10x year-over-year since founding. Post-merger, OpenRouter continues independently under its original name and mission with unchanged products and roadmap; routing decisions remain user-first. Deal expected to close in coming weeks.

Source: OpenRouter: Announcements (RSS)

---

Research Papers

Breaking DeepSeek-V4-Pro Service Limits: Multi-Scenario Optimization on H20 GPUs

The LMSYS team achieved near-B300 performance for the 1.6-trillion-parameter MoE model DeepSeek-V4-Pro on H20 GPUs through scenario-specific serving configurations. A single H20-141GB reference implementation reaches 271 output tokens/s, narrowing the gap with B300's 383.7 tokens/s to just 1.42×.

Source: LMSYS: Blog (Chatbot Arena Team)

Apple Research: Multi-Dimensional Analysis of LLM Human-Like Behavior

Apple's machine learning research team conducted a multi-dimensional analysis of LLM human-like behaviors (expressing thoughts and emotions, building relationships with users, refusing requests while maintaining boundaries), covering their prevalence, implications, and controllability. The study combined LLM-as-a-judge with human evaluation across 21,000+ data points, aiming to provide researchers and practitioners with guidance on when and what types of human-like behaviors to deploy.

Source: Apple Machine Learning Research (RSS)

The P-Completeness of Inverted Index Traversal: Complexity of Boolean Query DAGs

Modern AI agents rely on search infrastructure for neurosymbolic reasoning, often compiling to deeply nested non-monotonic Boolean queries. Standard inverted index query evaluation strategies face fundamental theoretical limits: stateful iterator models (Document-at-a-Time) are constrained by NC^1 formula evaluation structure, with worst-case query complexity exploding exponentially at O(2^|Q|) during unraveling and re-convergence.

Source: Apple Machine Learning Research (RSS)

---

Tips & Perspectives

GitHub Copilot App Beginner Tutorial: Managing Your Work with the My Work Panel

GitHub Copilot app's My Work panel centralizes pull requests and issues in one view, with built-in All, Active, Review requests, and Done default views plus custom views and filters. Users can launch agent sessions directly from issues or PRs, batch-select multiple items for sessions, toggle between list and table views, and adjust repo scope or create new issues.

Source: GitHub Blog

How Slack Builds Human-Agent Teams: Conversation as Knowledge

Slack CPO Jaime DeLanghe shared best practices for turning conversation into organizational knowledge: default to public channels so agents learn from visible dialogue, connect meetings, email, and calendar context to reduce repetition. In Slack, Claude-powered agents handle drafting, summarizing, and monitoring while humans review, decide, and hand off—forming a continuous collaboration loop.

Source: Claude: Blog

Designing Effective Genie Agents from a Single Prompt

Databricks' guide shows how targeted prompt design helps Genie Agents outperform general agents on queries like revenue analysis, which often only fetch the first relevant table. The guide helps developers optimize agent behavior for greater accuracy and relevance.

Source: Databricks: Blog (RSS)

---

【Hacker News Top Posts】

Keywords: AI OR GPT OR LLM OR Claude OR OpenAI OR "machine learning"...
Source: hnrss.org | Filtered from Hacker News

1. Show HN: Zero () friction local AI for Mac

Base Compute launches Zero friction local AI for Mac—automatically analyzes hardware, optimizes model parameters, and recommends the best models for your device. Supports PDF chat, meeting recording and summarization, coding agents, and Office Mode where the strongest machine in the office runs AI for the whole team with zero data leaving the premises.

• View on HN →

2. Show HN: Huzzah – a novel approach to coding with AI

An experimental editor using a new human-AI collaboration paradigm: write pseudocode your way, and the editor automatically syncs it to real source code on save. The pseudocode is persisted alongside generated code as a permanent record of intent—combining the benefits of AI assistance with the clarity of handwritten code.

• View on HN →

---

Originally published on WeChat Official Account 「比特财商」.