AI Daily — 2026-03-30

English 中文

Qwen3.5-Omni Unveils Native Omni-Modal AGI · ARC-AGI-3 Benchmark: AI Scores Near Zero in Novel En...

Covering 39 AI news items

🔥 Top Stories

Alibaba’s Qwen3.5-Omni launches a native omni-modal AI capable of text, image, audio, and video understanding with real-time interaction. It introduces Audio-Visual Vibe Coding to generate websites or games from vision, claims SOTA audio-visual performance, and supports long-form media, many languages, and fine-grained voice control. This marks a shift toward fully integrated multimodal stacks that could alter competitive dynamics in enterprise AI adoption. Source-x

2. ARC-AGI-3 Benchmark: AI Scores Near Zero in Novel Environments

François Chollet’s ARC-AGI-3 introduces 135 novel game environments with no instructions or goals, requiring real-time exploration and adaptation. Untrained humans solve all environments, while top AI models score below 1%, underscoring a severe generalization gap and the ongoing challenge of robust, instruction-free general intelligence. The reset to ARC-AGI-3 underscores how far current systems are from flexible, real-world adaptation. Source-x

3. 397B Qwen3.5 on MacBook: No Cloud, Pure C/Metal

An AI engineer demonstrates running a 397B-parameter model (Qwen3.5-397B) on a MacBook Pro with 48GB RAM using a pure C/Metal inference engine (flash-moe) with no cloud, GPUs, or Python. The system streams a 209GB model from SSD, loads only 4 of 512 experts per token, uses ~5.5GB active memory, and achieves 4.4 tokens/sec with full tool calling. The feat highlights striking consumer-hardware capabilities for locally hosted AI workloads. Source-x

📰 Featured

Open Source & Embeddings

Microsoft Harrier OSS Multilingual Embedding Models (27B/0.6B/270M) — Harrier OSS v1 provides multilingual decoder-only text embeddings with last-token pooling and L2 normalization for retrieval, clustering, semantic similarity, classification, and more; claims state-of-the-art on Multilingual MTEB v2 at release. Source-reddit
Qwen3.5-27B Fine-Tuned on Claude Distillations, Local Inference — Qwen3.5-27B fine-tuned on Claude-4.6-Opus distillations improves reasoning; runs locally on 16GB (4-bit) or 32GB (8-bit) hardware and is available on HuggingFace. Source-x

Tools & Interfaces

Claude Code Adds Computer Use for UI Automation — Claude Code now supports computer use to open apps, click through UIs, and test builds from the CLI; availability in research preview on Pro and Max plans. Source-x
Codex now usable in Claude Code via new plugin — Anthropic’s Claude Code gains OpenAI Codex support through a new plugin marketplace entry (openai/codex-plugin-cc) enabling commands like /codex:review and /codex:adversarial-review. Source-x

AI Safety & Research

Stanford and Harvard Release Disturbing AI Paper — A Reddit post highlights an arXiv preprint (2602.20021) by Stanford and Harvard described as “the most disturbing AI paper of the year,” with discussions on broader implications. Source-reddit

Benchmarks & Frontier Models

Opus Beats Claude Code by 20% in Cursor Benchmark — In a frontier-model benchmark, Cursor outperformed Claude Code and other harnesses; Opus improves from 77% to 93%, GPT-5.4 from 82% to 88%, and Gemini from 52% to 57% on a 100-feature PRD measure. Source-x

Multimodal & Vision

Hybrid Memory Enables Dynamic Video World Models for Hidden Subjects — Introducing Hybrid Memory, a memory architecture for video world models that preserves static backgrounds while tracking dynamic subjects to avoid disappearing visuals, aiming for better long-term video understanding. Source-huggingface

(Note: The 39-item count reflects both Featured Items and Brief Items combined in this issue.)

⚡ Quick Bites

ShotStream Enables Real-Time Interactive Multi-Shot Video Generation — Real-time interactive video generation via ShotStream. Source-huggingface
Open-source Claude code best-practice guide released — A best-practice guide for Claude code is now available openly. Source-github
HJB Equation Connects RL and Diffusion Models — A discussion linking reinforcement learning and diffusion models via HJB. Source-rss
AI unbundles jobs into lower-paid chunks, not killing them — Analysis suggests AI reconfigures job tasks rather than eliminating roles. Source-rss
Police used AI facial recognition to wrongly arrest TN woman in ND — Case highlights risks of biased or erroneous facial recognition. Source-rss
AI data center boom could become a $9T bust — Industry-level risk assessment of AI infrastructure expansion. Source-rss
Local Qwen3-VL Enables Transcription-Free Semantic Video Search — Local Qwen3-VL enables semantic search without transcription. Source-reddit
Small Local and OpenRouter Models Benchmark via Text-to-SQL — Benchmarking local models on Text-to-SQL tasks. Source-reddit
Apple ANE backend for llama.cpp on Apple Silicon — Apple Neural Engine backend brings improved performance to llama.cpp on Apple Silicon. Source-reddit
llama.cpp reaches 100k stars — Community uptake milestone for llama.cpp. Source-reddit
Qwen 3.6 Plus Preview Spotted on OpenRouter — Preview release signals upcoming improvements. Source-reddit
Sakana AI apologizes over Namazu naming for new LLM series — Naming controversy in LLM lineup. Source-x
Sharing Hidden, Useful Claude Code Features — Reveals lesser-known Claude Code features. Source-x
Hermes Agent v0.6 Adds HLS Playback — Hermes Agent adds HTTP Live Streaming playback support. Source-x
Boaz Barak Publishes Blog: State of AI Safety in Four Fake Graphs — Critical perspective on AI safety narratives. Source-x
Learn Claude Code by Doing, Not Reading — Practical approach to learning Claude Code. Source-rss
AI Rewrites the Engineering Career Ladder — Examines how AI reshapes professional progression. Source-rss
How the AI Bubble Bursts — Analysis of AI market dynamics and over-optimism. Source-rss
Visual Guide to Claude Code on GitHub — Visual guide to Claude Code usage. Source-github
Mathematical Methods and Human Thought in the AI Era — ArXiv preprint exploring math and cognition. Source-arxiv
Claude Code auto-resets repo with git reset —hard origin/main every 10 minutes — Repo resets to origin periodically. Source-github
Miasma traps AI web scrapers in endless poison pit — Project demonstrates data-scraping pitfalls. Source-github
What if AI doesn’t need more RAM but better math? — Discussion on efficiency and math-centric AI improvements. Source-rss
The First 40 Months of the AI Era: A Retrospective — Retrospective look at the AI era so far. Source-rss
Technical clarification on TurboQuant and RaBitQ discussions — Clarifications on model quantization tools. Source-reddit
PSA: Claude Code breaks local KV cache; fix via settings — Practical fix for Claude Code cache issue. Source-reddit
llamafile v0.10.0 adopts new build system for llama.cpp compatibility — Build system update for compatibility. Source-reddit
First language model traced to Markov in 1913 — Historical note on language modeling origins. Source-x
What is Claude’s secret sauce and why can’t it be replicated? — Discussion on Claude’s design features. Source-reddit

Generated by AI News Agent | 2026-03-30