AI Daily — 2026-06-02

English 中文

MAI Unveils Seven New World-Class Models Led by MAI-Thinking-1 · Claude Code Gets Workflows, Bigg...

Covering 32 AI news items

🔥 Top Stories

1. MAI Unveils Seven New World-Class Models Led by MAI-Thinking-1

MAI rolled out seven new world-class models, led by the 35B active parameter MAI-Thinking-1 MoE with a 256K context window, tuned for the MAIA 200 chip with 30% better performance per dollar and 1.4x better performance-per-watt versus GB200. The lineup signals a hardware-optimized, broad AI portfolio aimed at improving efficiency at scale and broadening MAI’s competitive footprint in the model ecosystem. Source-x

2. Claude Code Gets Workflows, Biggest Upgrade Since Skills

Claude Code gains workflows, the most significant upgrade since Skills and subagents, enabling automations for non-technical tasks and end-to-end processes. This expansion widens practical use cases for Claude Code beyond coding, potentially boosting enterprise adoption while inviting scrutiny over reliability and cost. Source-x

3. OpenAI Launches Sites into Codex for End-to-End Software

OpenAI introduces Sites into Codex to enable end-to-end software creation for users with varying technical fluency. Sites deploy to a URL, are private to workspaces, include authentication, host static files, and store dynamic data, with a preview rollout for business and enterprise teams before full workspace-wide release. Source-x

📰 Featured

LLM

MAI Unveils Seven New World-Class Models Led by MAI-Thinking-1 — MAI rolls out a hardware-optimized, multi-model lineup anchored by a 35B MoE with a 256K context window; efficiency gains are a key differentiator. Source-x
Claude Code Gets Workflows, Biggest Upgrade Since Skills — Workflows broaden Claude Code’s automation capabilities beyond skills, enabling end-to-end task execution. Source-x
OpenAI Launches Sites into Codex for End-to-End Software — End-to-end software creation via Codex Sites, with workspace-private URLs, authentication, and data hosting; preview for business/enterprise. Source-x
Anthropic Expands Project Glasswing, Extends Claude Mythos Preview — Glasswing access expanded to ~150 organizations across 15+ countries, broadening Mythos Preview reach. Source-x
Microsoft Announces Seven New AI Models at Build — A set of seven models spanning reasoning, coding, image processing, transcription, and voice, designed as a cohesive tool family. Source-x
Multi-Agent RL Improves LLM Workflows: Shared vs Isolated Policies — Comparative study on Shared-Policy vs Isolated-Policy RL for end-to-end LLM workflows, highlighting stability and tradeoffs. Source-huggingface
75M KeyLM Beats 135M Instruct Models on IFEval — A compact 75M-parameter model trained on 18B tokens outperforms a 135M model on an instruction-following benchmark, underscoring efficiency in small models. Source-reddit
TASTE Task Synthesis Improves Agent Benchmark Coverage — Proposes Task Synthesis to broaden benchmark coverage beyond NL-to-tool mappings, addressing saturation and cost. Source-huggingface
Domino: Decoupling Causal Modeling from Speculative Drafting — Domino separates drafting from causal modeling in speculative decoding to improve efficiency, balancing draft quality and drafting cost. Source-huggingface

⚡ Quick Bites

Linear Ensembles Erases LLM Watermarks — Proposes methods to erase watermarks in LLM outputs, signaling potential misuse risks. Source-huggingface
TradingAgents v0.2.5 Adds Grounded Sentiment Analyst and Dual-Region Support — Adds sentiment analysis and dual-region support for trading agents. Source-github
Machine Learning for Trading, 2nd Edition – GitHub Code — GitHub release of code for the second edition of the ML for Trading book. Source-github
Replaced Claude with Local Qwen3.6-27B in Multi-Agent Orchestrator — User swaps Claude for Qwen-3.6B in multi-agent orchestrator, testing local LLMs. Source-reddit
Benchmarks of 20 Small LLMs on 6GB RTX 4050 — Side-by-side benchmarks of small LLMs on a 6GB RTX 4050. Source-reddit
Gemma 4 E4B with LiteRT ~2.4x faster in text generation — Gemma 4 E4B with LiteRT delivers ~2.4x speedup in generation. Source-reddit
1-bit and Ternary Bonsai Image 4B: tiny diffusion models for local devices — Lightweight 1-bit and ternary diffusion models for local use. Source-reddit
Simple Coding Benchmark: Step 3.7 vs Qwen 3.5/3.6 — Quick benchmark comparing Step 3.7 to Qwen 3.5/3.6. Source-reddit
Claude Opus 4.8 Max Criticizes Its Content Claim — Claude Opus 4.8 questions its own content claims. Source-x
Natol Lambert departs Ai2 after 2.5 years — Longtime Ai2 researcher Natol Lambert departs after 2.5 years. Source-x
Perplexity Computer introduces hybrid agentic inference — Perplexity introduces hybrid agentic inference for agents. Source-x
Fork of OpenCode routes through Chipotle’s unsecured AI endpoints — Security concern as OpenCode routes pass Chipotle’s unsecured endpoints. Source-x
Hermes WebUI Launches Web Interface for Hermes Agent — WebUI provides a web interface for Hermes Agent. Source-github
Minimax M3 Appears to Have No Political Censorship — Claims no political censorship in Minimax M3. Source-reddit
Datacenter GPU in a Gaming PC for £200 — A datacenter-class GPU mounted in a consumer PC for 200 pounds. Source-reddit
What memory systems are you using for your agents? — Community discussion on memory architectures for agents. Source-reddit
Which Web Search API Delivers Clean Markdown for Local RAG? — Debate on search APIs for clean Markdown in Local RAG. Source-reddit
LLaMA.cpp adds Thinking mode toggle with reasoning levels — LLaMA.cpp now has a Thinking mode toggle with reasoning levels. Source-reddit
Ollie: World’s first AI family assistant that manages life — Introduction of Ollie, a household AI assistant. Source-x
Harness-1: State-Externalizing Harnesses for RL Search Agents — Harness-1 introduces state-externalizing capabilities for RL search. Source-huggingface
StepFun 3.5 MTP for Llama.cpp PR — PR for StepFun 3.5 MTP in Llama.cpp. Source-reddit
Imagine LLMs in the 80s, enable HLS playback — Nostalgic concept: enabling high-level synthesis-like playback for LLMs. Source-x

Generated by AI News Agent | 2026-06-02