Author: RadarAI Editorial
Editor: RadarAI Editorial
Last updated: 2026-05-31
Review status: Editorial review pending
Brief
速报
官方
AI动态
开源
Microsoft releases 45-year-old MS-DOS source code, restored via OCR from paper archives; Xiaomi unveils MiMo-V2.5—its full-stack inference optimization architecture with Hybrid SWA, MoE, and multimodal synergy; Fireworks AI's valuation hits $1.5B, signaling rapid capitalization of AI inference infrastructure.
Editorial standards and source policy: Editorial standards, Team. Content links to primary sources; see Methodology.
## 🔍 Key Insights
Microsoft has publicly released the **source code for DOS**, originally written **45 years ago**, reconstructing this pivotal piece of computing history from scanned paper archives using OCR technology [0]. Xiaomi has unveiled, for the first time, the full technical details of its **MiMo-V2.5 end-to-end inference optimization architecture**, centered on Hybrid SWA + MoE + multimodal coordination [2]. Fireworks AI’s valuation has surged to **$1.5 billion**, signaling accelerating capitalization in the global AI inference infrastructure race [6].
## 🚀 Top Updates
- **Microsoft releases 45-year-old DOS source code — back when code was printed on paper** [0]: Recovered via OCR scanning and meticulous human verification from physical archives, the early DOS 1.1/2.0 source is now open-sourced on GitHub.
- **ego lite: A browser purpose-built for AI agents launches** [1]: Features user login-state inheritance and strict background process isolation; claims at least **20% faster inference execution** versus mainstream alternatives.
- **Xiaomi MiMo reveals full-stack optimization for model inference** [2]: Overhauls KVCache management, prefix caching, and Prefill/Decode scheduling — enabling efficient on-device deployment of large language models.
- **Google’s “Big Four” gather: Model distillation is like “squeezing lemons”!** [3]: Jeff Dean and other DeepMind leaders unpack Gemini 3.5 Flash’s self-iterative training paradigm — where the model *writes code to improve itself*.
- **Morning Brief: Apple uses Google AI to train its own models / Huang Renxun: “I’ll explode if engineers stop burning tokens!”** [4]: Confirmed: Apple leverages Gemini for on-device model pretraining; Huang stresses that sustained, high-intensity compute investment remains non-negotiable for AI R&D.
- **Fudan alumna builds a $15B company** [6]: Fireworks AI — co-founded by Fudan University graduate Jolin Qiao — now valued at **$1.5 billion**, ranking among the world’s highest-valued AI inference platforms.
- **AnySearch Lens demo: Real-world applications in tech and finance** [7]: Using the hypothetical SpaceX IPO as a case study, it demonstrates query matrix expansion, multi-source evidence clustering (official docs / news / market data), and reliability scoring.
- **Try AnySearch: Turn search into an observable, reusable workflow** [8]: Supports custom MCP/REST/SKILL integrations via Codex — tailored for academic research, market analysis, and information verification.
## 🔗 Sources
[0] Microsoft releases 45-year-old DOS source code — back when code was printed on paper — https://www.bestblogs.dev/article/9891baee?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
[1] ego lite: A browser purpose-built for AI agents launches — https://www.bestblogs.dev/status/2060696809400619450?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
[2] Xiaomi MiMo reveals full-stack optimization for model inference — https://www.bestblogs.dev/article/c63fbd9f?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
[3] Google’s “Big Four” gather: Model distillation is like “squeezing lemons”! Gemini will use its own generated code to “improve” itself! In five years, Google may ship just one product — https://www.bestblogs.dev/article/0a87fe
Microsoft has publicly released the source code for DOS, originally written 45 years ago, reconstructing this pivotal piece of computing history from scanned paper archives using OCR technology [0]. Xiaomi has unveiled, for the first time, the full technical details of its MiMo-V2.5 end-to-end inference optimization architecture, centered on Hybrid SWA + MoE + multimodal coordination [2]. Fireworks AI’s valuation has surged to $1.5 billion, signaling accelerating capitalization in the global AI inference infrastructure race [6].
🚀 Top Updates
- Microsoft releases 45-year-old DOS source code — back when code was printed on paper [0]: Recovered via OCR scanning and meticulous human verification from physical archives, the early DOS 1.1/2.0 source is now open-sourced on GitHub.
- ego lite: A browser purpose-built for AI agents launches [1]: Features user login-state inheritance and strict background process isolation; claims at least 20% faster inference execution versus mainstream alternatives.
- Xiaomi MiMo reveals full-stack optimization for model inference [2]: Overhauls KVCache management, prefix caching, and Prefill/Decode scheduling — enabling efficient on-device deployment of large language models.
- Google’s “Big Four” gather: Model distillation is like “squeezing lemons”! [3]: Jeff Dean and other DeepMind leaders unpack Gemini 3.5 Flash’s self-iterative training paradigm — where the model writes code to improve itself.
- Morning Brief: Apple uses Google AI to train its own models / Huang Renxun: “I’ll explode if engineers stop burning tokens!” [4]: Confirmed: Apple leverages Gemini for on-device model pretraining; Huang stresses that sustained, high-intensity compute investment remains non-negotiable for AI R&D.
- Fudan alumna builds a $15B company [6]: Fireworks AI — co-founded by Fudan University graduate Jolin Qiao — now valued at $1.5 billion, ranking among the world’s highest-valued AI inference platforms.
- AnySearch Lens demo: Real-world applications in tech and finance [7]: Using the hypothetical SpaceX IPO as a case study, it demonstrates query matrix expansion, multi-source evidence clustering (official docs / news / market data), and reliability scoring.
- Try AnySearch: Turn search into an observable, reusable workflow [8]: Supports custom MCP/REST/SKILL integrations via Codex — tailored for academic research, market analysis, and information verification.
🔗 Sources
[0] Microsoft releases 45-year-old DOS source code — back when code was printed on paper — https://www.bestblogs.dev/article/9891baee?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
[1] ego lite: A browser purpose-built for AI agents launches — https://www.bestblogs.dev/status/2060696809400619450?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
[2] Xiaomi MiMo reveals full-stack optimization for model inference — https://www.bestblogs.dev/article/c63fbd9f?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
[3] Google’s “Big Four” gather: Model distillation is like “squeezing lemons”! Gemini will use its own generated code to “improve” itself! In five years, Google may ship just one product — https://www.bestblogs.dev/article/0a87fe
← Back to Updates