AI Daily Brief — 06 March 2025
Thursday delivered the year’s biggest Chinese agent debut, the most aggressive Western OCR drop, and the year’s first credible RL-reasoning open-weights challenger to DeepSeek-R1. Butterfly Effect (Singapore-based, China-founded) launched Manus in invite-only beta as “the world’s first general AI agent” — a multi-agent autonomous system built atop Anthropic’s Claude 3.5 Sonnet and fine-tuned Alibaba Qwen models. The launch demo (autonomous resume screening, stock analysis) drew 1 million-plus views in 20 hours, with over 2 million people waitlisted within a week. Mistral released its OCR API at 1,000 pages per dollar (≈double with batch inference), processing up to 2,000 pages per minute on a single node, 98.96% accuracy on scanned documents — outperforming Google Document AI, Azure OCR, and GPT-4o on Mistral’s eval. Alibaba’s Qwen Team open-sourced QwQ-32B under Apache 2.0: a 32-billion-parameter RL-trained reasoning model with 131K context, claimed to rival DeepSeek-R1 and outperform OpenAI o1-mini on AIME 24, LiveCodeBench, IFEval. Alibaba’s stock jumped 8% on the news. Hugging Face co-founder Thomas Wolf published a sharp X essay directly challenging Anthropic CEO Dario Amodei’s “compressed 21st century” vision, arguing current paradigms produce “a country of yes-men on servers” rather than “a country of geniuses.” Manus invite codes hit resale prices of ¥50,000-¥100,000 ($7,000-$13,800) on Xianyu, with some individual codes reportedly trading up to ¥1 million.
Top stories
- Manus AI launches invite-only beta — “world’s first general AI agent.” Multi-agent autonomous system built atop Claude 3.5 Sonnet and fine-tuned Qwen models. Launch demo: 1M+ views in 20 hours, 2M+ waitlisted within a week. Jack Dorsey and Hugging Face’s Victor Mustar amplified. Dubbed “the second DeepSeek.” via Wikipedia
- Mistral releases OCR API. 1,000 pages per dollar (≈double with batch), 2,000 pages/min/node, 98.96% accuracy on scanned documents / 89.55% multilingual. Outperforms Google Document AI, Azure OCR, GPT-4o on Mistral’s eval. Available on la Plateforme with on-prem option pitched at regulated industries. via Mistral
- Alibaba open-sources QwQ-32B under Apache 2.0. 32B-parameter RL-trained reasoning model with 131K context. Claimed to rival DeepSeek-R1 and outperform OpenAI o1-mini on AIME 24, LiveCodeBench, IFEval. Alibaba’s stock jumped 8%+ on the news. via SiliconANGLE · via Hugging Face
- Thomas Wolf publishes “yes-men on servers” essay. Hugging Face chief science officer’s X essay directly challenges Anthropic CEO Dario Amodei’s “compressed 21st century” vision. Argues current paradigms produce “a country of yes-men on servers” rather than “a country of geniuses” capable of Nobel-level discovery; calls for new benchmarks measuring counterfactual reasoning and ability to ask non-obvious questions. via TechCrunch
- Manus invite codes hit five-figure resale prices. Demand drives a secondary market on Xianyu: ¥50,000-¥100,000 ($7,000-$13,800) per code, some individual codes reportedly trading up to ¥1M. The Manus team later said they never created paid channels or marketing budget. via AIbase
Who shipped
Butterfly Effect shipped Manus. Mistral shipped OCR. Alibaba shipped QwQ-32B. The single densest Western+Chinese release day of the month. OpenAI, Anthropic, Google DeepMind, Meta, and xAI made no dated launches.
Open-source pulse
QwQ-32B and the broader Qwen family continued the open-weights surge from Wan 2.1 (Feb 25). Cohere’s Command R7B Arabic (Feb 27 release) continued enterprise rollout for MENA — 8B Arabic-optimized open weights, 128K context, runs on low-end GPUs.
Quiet corners
Anthropic Series E day-3 coverage continued, with the $61.5B valuation positioning Anthropic among the most valuable private AI companies. The 2024 Turing Award announcement from the prior day kept landing — particularly relevant given QwQ-32B’s RL-on-outcome-rewards approach echoing Sutton’s foundational work.
By the numbers
- 1M+ / 2M+ — Manus demo views (20 hours) / waitlist (1 week)
- 1,000 pages / $1 — Mistral OCR throughput pricing
- 32B / Apache 2.0 — QwQ-32B size / license
- +8% — Alibaba stock move on QwQ-32B
- $7K-$13.8K — Manus invite-code resale band
- Most-mentioned country: China
Compiled by AI Feed’s editor from verified web sources for 6 March 2025.