02-22-Daily AI News Daily

AI News Daily 2026/2/22

AI News | Daily Briefing | Web Data Aggregation | Frontier Science Exploration | Industry Voice | Open Source Innovation | AI & Human Future | Visit Web Version | Join Group Chat

Today’s Digest

Pika launches AI Selves, Replit integrates Gemini 3.1, Qwen3-Coder-Next API goes live
Claude Code desktop adds server preview, code review, PR auto-fix and more
Taalas unveils HC1 chip with 10x faster inference than Cerebras, models directly etched onto silicon
S2Q multi-agent retains alternative strategies, MolmoSpaces builds 230k indoor environment robot benchmark
AI customer service satisfaction drops, community reflects on code cognitive debt and proliferation of low-cost projects

Product & Feature Updates

1. Pika is stepping up its game, moving beyond just videos to ✨ create “AI Selves” — essentially, a “second you.” Users can simply upload a selfie and record some audio, and a Digital Twin Born Instantly (AI News) is generated. This digital self boasts incredible long-term memory and personality, so much so that you can even set it to have a peanut allergy! 🥜 It can post for you, make calls, and stay active across various platforms. Currently, this cool feature is invite-only, so you’ll need a code to experience it.

2. Replit has announced a major upgrade to its Design Mode. 🚀 It’s been completely overhauled, now powered by Gemini 3.1 underneath. The result? Sharper Visuals, Smarter Layouts (AI News) . The official word is it “has better taste now.” 😉 Front-end developers are in for a significantly enhanced experience.

3. The Qwen3-Coder-Next API has officially launched on Alibaba Cloud, following the open-sourcing of Qwen3-Coder-Next by Alibaba Tongyi Qianwen. You can find the API Now Live on Alibaba Cloud (AI News) . It’s also integrated into the Coding Plan solution. Development teams can now flexibly call the API, making it perfect for scenarios requiring scalable endpoints. The official team is super keen on feedback! 👍

4. Claude Code Desktop by Anthropic just dropped four major updates! 🚀 First up, there’s a new server preview feature, letting you check out your work directly within the interface. They’ve also added local code review, so you can Get Inline Suggestions Before Push (AI News) . Plus, PR monitoring now supports auto-fix and auto-merge. And for ultimate convenience, your sessions can move across devices – think seamless switching from CLI to desktop to mobile. How cool is that? ✨

Frontier Research

1. S2Q is making waves in multi-agent reinforcement learning. Typically, value decomposition methods often get stuck in sub-optimal strategies. But a new paper introduces Successive Sub-value Q-learning, which learns multiple sub-value functions to retain high-value alternative actions. Its Softmax behavior policy encourages continuous exploration, helping agents to Quickly Adapting to Changing Optimal Solutions (AI News) . Experiments have shown that S2Q significantly outperforms benchmarks across the board. 🏆

2. MolmoSpaces has built a massive robotics benchmark ecosystem. 🤖 The new paper unveils over 230,000 indoor environments and 130,000 annotated object assets. It supports various simulators like MuJoCo and Isaac, Covering All Navigation and Manipulation Tasks (AI News) . What’s really impressive is the simulation-to-reality correlation, hitting a high R=0.96! This provides a truly scalable data and training foundation for robotics learning research.

3. M-Attack-V2 is a game-changer for black-box attacks on large vision-language models. Facing the challenge of attacking these models in a black-box setting, researchers introduced multi-crop alignment and auxiliary target alignment strategies. This led to a massive jump in attack success rates on Claude-4.0, from 8% to a whopping 30%, and for GPT-5 Reaching 100% (AI News) ! The method sports a modular design, allowing for flexible layering. Plus, the code is open-source! 🔓

Industry Outlook & Social Impact

1. Taalas, founded by a dream team of former AMD executives, is taking on Nvidia with its new HC1 chip. 🔥 This bad boy boasts a peak inference speed of 17,000 tokens per second. That’s a mind-blowing 10 times faster than Cerebras, and it Costs Only One-Twentieth (AI News) ! The core idea is to etch models directly onto the silicon, making the chip the model itself. However, this hard-coding approach has sparked controversy over potential obsolescence, with netizens’ opinions being quite polarized. 🤷‍♀️

2. Karpathy has stirred up a debate with his “Claw” concept: resident, schedulable AI agents designed for long-term operation, triggered by plans or events. The community is fiercely debating: Is it really safe to hand over your passwords to a Claw? 😬 Container Sandbox Cannot Fundamentally Eliminate Risks (AI News) . Self-hosting is seen as a viable but high-cost solution. Even the naming and commercialization have drawn some cultural snark.

3. Anthropic Opus 4.6 is sparking controversy over its vulnerability detection. Anthropic claims Opus 4.6 discovered around 500 high-severity vulnerabilities, vastly outdoing Google BigSleep’s 20. The community is questioning the false positive rate and the definition of “high-severity,” Demanding Disclosure of Cost/Vuln Confusion Matrix (AI News) . Some even worry that low-barrier scanning could be abused to unearth open-source zero-days. Nevertheless, the virtual security engineer path seems to be gaining practical consensus. 🛡️

4. The proliferation of AI side projects is causing the community to reflect. “Just write a prompt and get on Show HN” has become the new normal, with tons of low-effort AI projects drowning out quality work. Community Calls for Adjusting Curation Mechanisms (AI News) , with some suggesting a requirement to showcase effort and technical details. However, others argue that AI lowers the barrier, empowering more people to build projects. Still, architectural risks and cognitive debt are not to be ignored. 🤔

5. AI prediction markets are stirring up ethical debates concerning insider trading. Some folks are using AI on Polymarket to identify insider trading and bag some alpha. Supporters argue that insider bets can boost prediction accuracy, but Opponents Frankly State Bribery for Information is Illegal (AI News) . While on-chain transactions are traceable, the advantage of copy trading can easily be wiped out by competition. Let’s be real, most regular users are just in it for the thrill of gambling. 🎲

6. Big Tech AI products are facing a harsh reality: rapid user churn after massive investments. One big tech company reportedly poured billions into training models, marketing, and handing out red packets. On Lunar New Year’s Eve, user numbers hit a historic peak. Sadly, half of them vanished the next day, and another half the day after that. 📉 Short-Term Gain Driven, Others Take Over the Mess (AI News) . This has sparked deep reflection across the industry on the importance of long-term retention.

7. The truth about AI customer service is that satisfaction rates are actually dropping. Many companies have rolled out AI customer service, only to find that Customers Feel They Are Being Brushed Off (AI News) . The smart move? Keep human agents on the front lines, with AI as their behind-the-scenes assistant. Customers don’t just need professional answers; they need emotional support too. AI should be a tool for human customer service, not a replacement. 🗣️

Open Source TOP Projects

1. PentAGI, an AI penetration testing tool written in Go, is blowing up! 🔥 It gained 2107 new stars today, pushing its Total Stars Reaching 4909 (AI News) . This project aims for automated security assessment. The open-source community is absolutely buzzing, with its fork count hitting 648. ✨

2. FossFLOW, a workflow engine written in TypeScript, continues to be a hot topic! 🔥 It has Accumulating ⭐17634 Stars (AI News) , with an additional 101 stars today. This project is all about open-source workflow orchestration. Its fork count has already topped 1151, and the ecosystem keeps growing. 🌱

3. GitNexus, a TypeScript-developed Git enhancement tool, is rapidly gaining traction. 🚀 It Today Added 133 Stars to ⭐888 (AI News) . The project specifically targets the pain points of multi-repository management. With 82 forks, it’s definitely in a fast-growth early stage!

4. Hiddify, developed in Dart, consistently ranks high. 🌟 It has accumulated Total ⭐26271 Stars (AI News) , with an additional 121 stars today. This tool supports multi-platform network proxy functionalities. With 2271 forks, it’s one of the most popular open-source projects in its category.

Social Media Buzz

1. OpenClaw can even find movie and TV resources! Blogger Fánrén Xiǎoběi tested it and discovered that OpenClaw could scour the web for streaming content. “Checked it out, and the Link Not Expired (AI News) ,” they exclaimed. The blogger also griped about platforms like iQiyi, Youku, and Tencent Video constantly piling on member tiers and locking screen mirroring to 720p. It seems even legitimate supporters are being pushed to the wall. 😤

2. The 1983 Wall Street simulation game has been reverse-engineered and brought back to life! Someone managed to reverse-engineer the 115,000 lines of BASIC code for “Wall Street Raiders.” Believe it or not, This Game Once Influenced Real CEO Investment Decisions (AI News) . Even Disney couldn’t crack the source code back in the day! Now, you can actually play it online. 🕹️

3. AI programming can easily pull you into a “joy of creation” trap! Blogger Nazha recounts how they originally just wanted to publish a post. Instead, they casually let Claude Code write an Obsidian plugin. Hours Later Remembered the Main Task (AI News) ! The flow state induced by AI programming is just too powerful. Their attention got redirected, but thankfully, they eventually made the post. Phew! 😅

4. Remotion is liberating content creators by generating videos with just a single sentence! Blogger Huang Yun demonstrated how it works: you can create a video using Remotion just by typing a sentence. Back in the day, they’d spend ages hunting for material, often getting stuck in a “three days fishing, two days drying nets” cycle. 🎣 Now, AI Now Handles 10 Pending Posts in Minutes (AI News) ! But if creation is no longer hands-on, where’s the fun? 🤔

5. An article on OpenClaw multi-agent configuration is receiving rave reviews! Blogger Huang Yun recommended a piece explaining how to equip agents with different large models. They argue that Multi-Agent Team Play is the Prototype of a One-Person Company (AI News) . You get assistants who might earn 5K a month and also “god-tier” experts making 50K. 💸 The possibilities for pay-per-use outsourcing are endless!

6. Xiaohongshu Diandian is making its breakthrough with a “strategy guide” mode. Blogger rosicky analyzes that data importance is being re-prioritized. Xiaohongshu Diandian chose “Strategy Mode” as its core feature, Based on Community UGC Real Long-Tail Data (AI News) . Real-world tests of its Hokkaido ski guide showed impressive results. ⛷️ There’s huge potential for expansion in the lifestyle services sector!
   
   

7. How is your team managing the cognitive debt from AI-generated code? A Reddit user brought up the concept of “cognitive debt”: teams simply don’t understand the code churned out by AI. On-Call Engineers Cannot Debug, Newcomers Cannot Get Started (AI News) . Suggestions include setting up comprehension checkpoints and PR explanation templates. This debt is far more insidious than technical debt, only revealing itself when things go sideways. 😬

AI News Daily: Multi-Channel

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