02-20-Daily AI News Daily

AI News Daily 2026/2/20

AI News | Daily Morning Read | All-Network Data Aggregation | Frontier Science Exploration | Industry Free Expression | Open Source Innovation Power | AI and Human Future | Access Web Version↗️ | Join Group Chat🤙

Today’s Summary

Google Lyria 3 enables music generation from text, images, and video; NetEase LobsterAI open-sources its all-scenario Agent.
Re-TRAC recursive trajectory compression framework with 4B model achieves SOTA; Co-rewarding boosts unlabelled RL by 12.9%.
OpenAI's removal of its safety mission sparks controversy; Fei-Fei Li's World Labs valued at $5 billion.
Fake web pages trick ChatGPT, exposing RAG's trust gap; AI acceleration might worsen technical debt.
AI Agent long-task success rate less than 20%; Anthropic releases Claude skill-building guide.

Product and Feature Updates

1. Google Lyria 3 music model has officially dropped! 🎉 DeepMind rolled out this next-gen music generation model, allowing anyone to create tunes without any musical background. It supports text, image, and video inputs to pump out 30-second full songs, complete with vocal accompaniment. Currently, it’s just for non-commercial use. Find out more in the official intro (AI News) .

2. Anthropic has put a stop to third-party subscription calls. 🚫 The company explicitly limited OAuth tokens in its documentation, stating that subscriptions are only for Claude Code and the official website. Any third-party tool integration will be considered a violation, which has understandably irked a ton of independent developers. 😠 Now, the community is already looking at alternative solutions like Kimi (AI News) .

3. NetEase LobsterAI has officially hit open source! 🚀 NetEase just dropped this all-scenario personal agent tool, ready to execute automated tasks 24/7. Its local-first design ensures both privacy and low latency, and it can hook up with platforms like Feishu, DingTalk, and Telegram. It’s officially open source, so check out the experience address here (AI News) .
   
   

Cutting-Edge Research

1. Re-TRAC helps agents learn from their flops! 💡 Microsoft, teaming up with Southeast University, proposed this Recursive Trajectory Compression framework, letting agents pass on search experiences across turns. The 4B model nails SOTA performance on multiple benchmarks, and the 30B version even outdoes the beefier 358B-parameter GLM. Impressive! The paper is out (AI News) , and the code’s on GitHub (AI News) .

2. Co-rewarding is here to fix the unlabeled RL crash! 🔥 A team from Hong Kong Baptist University and Shanghai Jiao Tong University unveiled a new self-supervised RL framework. Their complementary perspective supervision effectively prevents reward speculation, allowing stable training without needing real answer labels. This work boosted performance on math benchmarks by an average of +12.90% and has already been accepted by ICLR 2026 (AI News) .

3. CAFE uses causal discovery to guide feature engineering. ☕ A fresh paper introduces this causal-guided automated feature engineering framework, combining causal graphs with multi-agent deep Q-learning. It boosted performance by up to 7% across 15 benchmarks and slashed performance degradation under distribution shift by roughly 4x. Pretty neat! Check out the full paper (AI News) .

4. Researchers are tackling the “forgetting” problem in continuous learning! 🧠 They’re challenging the old paradigms, noting that when memory isn’t an issue, the core problem shifts to lack of plasticity. They’ve come up with a weight space merging method to restore learning capacity, which has shown effectiveness in both image classification and LLM fine-tuning. The paper’s details (AI News) have been updated to v5.

5. Researchers are digging into what knowledge generative social robots really need! 🧑‍🏫 They’re exploring the design of social robots in university settings, and through 12 interviews, they’ve pinpointed three key knowledge requirements: self-knowledge, user knowledge, and situational knowledge. The goal? To make LLM-driven robots more responsible. The paper’s address (AI News) is now public.

Industry Outlook & Social Impact

1. OpenAI quietly dropped its safety mission. 🤫 Tax documents reveal that OpenAI tweaked its mission statement, with “safety” and “not constrained by profit” notably absent. Employee Peter Girnus even posted a scathing critique, dubbing it the company’s “seven deadly sins.” The mission alignment team has been disbanded, and its head? Reassigned as “Chief Futurist.” Hmm. 🤔 The controversy keeps brewing (AI News) .

2. Fei-Fei Li’s World Labs just snagged a massive $1 billion in funding! 🤯 This spatial intelligence company’s valuation has rocketed to an incredible $5 billion, with tech giants like AMD, Nvidia, and Fidelity all jumping in. Its valuation quintupled in less than two years – talk about speed! 🚀 Their first product, Marble, is already the most advanced 3D world model out there. Li Feifei’s interview details (AI News) are packed with info.

3. Fake web pages managed to trick both ChatGPT and Google! 😲 An author cooked up some bogus hot dog championship info on their personal website, and both ChatGPT and Google spat it out as fact. This really exposes a source trust gap in RAG retrieval. The community is now calling for a mechanism to verify claims item by item. The discussion details (AI News) have stirred up a lot of debate.

4. Research reveals how X platform’s algorithms are reshaping the political landscape. 📈 Apparently, recommendation algorithms are influencing users’ political views. After Musk’s takeover, a mass exodus of left-leaning users occurred. Now, paid blue-check replies are drowning out valuable discussions. Critics are demanding open protocols and algorithm transparency. Nature-related research (AI News) on this topic has been cited multiple times.

5. AI acceleration could totally supercharge tech debt. 😬 Martin Fowler and others are sounding the alarm about a “speed trap”: unstandardized AI outputs are becoming debt accelerators. The community is floating ideas like risk layering and TDD (Test-Driven Development) guard strategies. Plus, agent pipelines are introducing a whole new set of failure modes. Yikes! This in-depth discussion (AI News) covers both costs and safety.

6. AI might just be humanity’s compass! 🧭 A trending Reddit post throws out a counter-intuitive idea: AI taking over knowledge work is a necessary course correction. It suggests humans will circle back to fusion energy and space exploration, trading digital comfort for physical world progress. Wild, right? This complete argument (AI News) has sparked a lot of buzz.

7. OpenClaw has some serious security flaws. 🚨 This agent tool hands LLMs local system permissions, meaning prompt injection could lead to account and data breaches. The community is suggesting mitigations like physical isolation or layered execution, but regulation and accountability are seriously lagging behind this tech’s spread. Developers, beware! This security analysis (AI News) is worth a close look.

Top Open Source Projects

1. pyrite64: A new N64 emulator star is born! ✨ This N64 emulator project, coded in C++, is blowing up! It racked up 605 new stars in a single day, hitting a total of 1539 stars. Community attention is absolutely skyrocketing! 🚀 The project address (AI News) is now open.

2. cs249r_book: Harvard’s Edge AI textbook is making waves. 📖 This open-source embedded AI textbook project from Harvard University has collected a whopping 19964 stars, with 660 added just today. It covers the entire knowledge system for edge computing and AI deployment. The GitHub repository (AI News) is continuously being updated.

3. open-mercato: An open-source transfer market built with TypeScript. ⚽ This open-source transfer platform project has already snagged 563 stars, and its community is steadily growing. It offers an open sports data trading framework. Project details (AI News) – contributions are welcome!

4. Composio: Your go-to AI Agent tool integration platform. 🛠️ This open-source framework helps agents connect with external tools. It’s already racked up a massive 26817 stars, boasting a mature ecosystem. It supports TypeScript development and multi-platform access. Check out the repository (AI News) to get started quickly!

Social Media Buzz

1. Self-hosting is on the cusp of an AI-driven renaissance! 🚀 Jike user cosformula dropped a bold prediction: AI agents could zero out self-host operations costs. Tokens might just become a monthly utility bill, like electricity. This could mean personal data flowing back to local storage, cutting ties with cloud dependence. Food for thought, right? Check out the original post’s viewpoint (AI News) that sparked all this buzz.

2. A K12 learning machine demo with Claude was whipped up by Jike user dangjin! 🤩 The user’s inner K12 student was awoken by a learning machine product, prompting them to build a test paper analysis tool using DeepSeek + Next.js. It supports uploading test papers for problem analysis and extending learning with similar examples. Alibaba Cloud’s ReadLight handles the intelligent splitting of test papers. The demo video (AI News) looks pretty solid!

3. Anthropic has dropped a Claude skill-building guide! 📚 This 32-page comprehensive guide covers everything from planning to distribution. Five core design patterns are highlighted as essential learning points. The idea is to package common workflows into a Skill for long-term convenience. You can download the full PDF now (AI News) .
   
   

4. An AI toolkit was installed for a cousin! 🎁 Xiangyang Qiaomu went home for Chinese New Year and put AI popularization into practice. He loaded up a Windows PC with Claude Code and CC Switch, plus tools like Listenhub and Happycapy. The goal? To get his family using advanced productivity tools ASAP! 💪 This tool list (AI News) is worth checking out.

5. Suno v5 vs. Gemini music generation: Xiangyang Qiaomu put them head-to-head using identical prompts! 🎶 Suno v5 absolutely crushed Gemini in completeness and vocal style. Gemini could only pump out 30 seconds of oddly Chinese-sounding audio. 😅 The gap was way bigger than expected, LOL! The comparison video (AI News) speaks for itself.
   
   

6. AI Agents are clocking less than 20% success rate on long tasks. 📉 The LongCLI-Bench benchmark results are a bit of a reality check: leading agents achieved a success rate below 20% in complex CLI tasks. Self-correction barely helped at all. 🤷‍♀️ Looks like human-computer collaboration is the way to go. This paper’s address (AI News) is definitely one to watch.
   
   

AI News Daily: Multi-Channel

WeChat Official Account	Douyin
Official Account: Hexi 2077	Self-Media Account