01-23-Daily AI News Daily

AI Daily News Digest 2026/1/23

AI News | Daily Briefing | Aggregated Web Data | Cutting-Edge Science | Industry Voice | Open-Source Innovation | AI & Human Future | Visit Web Version | Join Group Chat

Today’s Highlights

Qwen3-TTS open-source supports ten-language end-to-end synthesis.
Grok video extends to 10 seconds; Fudan releases model safety evaluation.
Nadella says AI needs useful scenarios, raising bubble concerns.
Kimi claims 1% compute power surpasses closed-source models; in-house compute trend rises.
Open-source projects like Remotion, Goose, FlashMLA gain thousands of stars.

Product & Feature Updates

1. Qwen3-TTS Speech Synthesis Officially Goes Open-Source! The Tongyi team just dropped their entire Qwen3-TTS Family (AI News) , featuring VoiceDesign, CustomVoice, and Base modules. You can grab it with 0.6B or 1.8B parameters, and it supports 10 languages. This bad boy offers native end-to-end multimodal synthesis, waving goodbye to those robotic tones 👋. Devs can fine-tune all parameters and deploy it privately with zero hassle.
   

2. xAI Grok’s Video Generation Now Supports 10 Seconds! Grok Imagine just officially launched its 10-second video generation capability (AI News) . The video stability and detail have seriously leveled up. Audio sync is natural, and noise interference issues are a thing of the past. Musk himself confirmed the update, calling it “fantastic” 😍. While precise timing options are missing for now, they’ll be added soon. The community’s already diving into creative short film tests.
   

3. Perfect Corp. Expands Virtual Try-On to Nine Categories. Perfect Corp. has just added new virtual try-on features for nine categories including watches, rings (AI News) to its fashion API. Leveraging generative AI and computer vision, it accurately identifies human bodies and lighting. The result? Highly realistic outfit mock-ups that boost purchasing confidence ✨. It also supports MCP protocol, making it compatible with all e-commerce channels. This helps brands crank up conversion rates and slash returns.

Cutting-Edge Research

1. Fudan University and Others Release Frontier Large Model Security Report. Fudan University, in collaboration with Shanghai Institute for Intelligent Interaction and other institutions, just dropped a security evaluation report for six leading models (AI News) covering six leading models. The assessment includes 30 types of jailbreak attacks and 18 language scenarios. GPT-5.2 scored an average safety rate of 78.39%, leading the pack! 🏆 Multi-round adaptive attacks emerged as the biggest threat, with cross-language safety gaps reaching 20-40%. The report advocates for building a dynamically evolving security evaluation system.

2. DrivIng Dataset, Including Digital Twins, Has Been Released! A research team unveiled the DrivIng large-scale multimodal driving dataset (AI News) , covering approximately 18 kilometers of urban to highway segments. It provides continuous recording from six RGB cameras and LiDAR, including day and night scenarios. All sequences are annotated with 3D bounding boxes at 10Hz, totaling 1.2 million instances. This baby supports 1:1 real-world traffic simulation transfer and flexible scenario testing—pretty sweet! 😎 The dataset and code are now publicly available.

3. Text-to-Video Survey: Can Sora Model the World? Researchers have compiled 250+ text-to-video and world modeling research papers (AI News) , systematically evaluating the current state of the tech. Recent models have shown continuous progress in spatial understanding, actions, and strategic intelligence. Support for completeness and consistency is good, with creativity and interactive control steadily improving. The conclusion? Text-to-video already possesses world-modeling capabilities! ❤️ However, balancing diversity and consistency remains a challenge.

4. CityCube Tests VLM Urban Spatial Reasoning. The new benchmark, CityCube, has been released to focus on VLM cross-view urban spatial reasoning (AI News) evaluation. It integrates multiple platform perspectives from vehicles, drones, and satellites. Featuring 5022 pairs of multi-view QA, it covers five cognitive dimensions. Large-scale VLMs only achieve a maximum accuracy of 54.1%, trailing human performance by 34.2%—whoa! 🤯 Interestingly, smaller fine-tuned models actually surpassed 60%, highlighting the benchmark’s value.

Industry Outlook & Social Impact

1. Nadella Says AI Needs Useful Scenarios. Microsoft CEO Satya Nadella stated that we need to find useful scenarios for AI (AI News) , sparking a hot debate in the community. Critics argue that LLMs offer limited productivity boosts for average users, claiming “96% of people won’t significantly benefit.” AI training is driving up GPU and flash memory prices, with energy costs becoming the next bottleneck 💸. Under this VC-driven bubble, the pressure for Product-Market Fit (PMF) validation is immense. Corporate rhetoric is being criticized as vying for “social license” rather than user value.

2. Kimi from Moonshot AI Makes a Splash at Davos. Kimi President Zhang Yutong announced at Davos that, using just 1% of US computing power, they have surpassed closed-source models (AI News) . An engineering-first mindset is becoming the key path for China’s AI breakthrough. Kimi K2 Thinking shows outstanding performance in handling complex task chains—nice! ✨ Their open-source strategy accelerates community feedback and iteration cycles. A new generation of models is on the horizon, enhancing multimodal and agent capabilities.

3. A Trend of Enterprises Building Their Own AI Compute is Emerging. More and more companies are opting to build their own local AI workstations (AI News) instead of relying on cloud APIs. Hardware investments typically see a 1.5 to 2.5-year ROI, highlighting the economic benefits. Different GPU and memory configurations are needed depending on task complexity. Kingston has rolled out a full-stack solution 🛠️, covering DDR5 and enterprise-grade NVMe. Local deployment offers both data security and supply chain resilience.

4. eBay Bans AI Proxy Buying, Sparking Double Standard Questions. eBay has updated its user agreement 🤖, explicitly prohibiting AI proxy buying agents (AI News) from placing automatic orders. Critics are slamming them for a double standard against sniper bots, which are still tolerated. LLM agents could lead to mistaken purchases and increased refund costs 💸. Detection methods like browser fingerprinting create an ongoing cat-and-mouse game. The new terms might be paving the way for post-facto accountability and monetization.

Top Open-Source Projects

1. Remotion: A Programmatic Video Production Framework. Remotion is an open-source framework that lets you create videos programmatically using React 🎬, and it’s already bagged ⭐26.4k (AI News) stars. Developers can control every frame of a video with code, making it perfect for automated marketing videos and data visualization scenarios—how cool is that?! ✨

2. Goose: An Open-Source, Scalable AI Agent. The Block team has built Goose, a scalable AI agent 🚀 that’s already snagged ⭐27.0k (AI News) stars. It goes beyond just code suggestions, supporting installation, execution, editing, and testing. Plus, you can drive it with any LLM, giving it super high flexibility! 💡

3. FlashMLA: An Efficient Attention Kernel. DeepSeek open-sourced FlashMLA 🔥, designed specifically for efficient multi-head latent attention (AI News) , and it’s already hit ⭐12.1k stars. It optimizes MLA inference performance, significantly slashing computational overhead—score! ✨

4. Mastra: A Modern TypeScript AI Framework. Crafted by the Gatsby team 💡, Mastra is a TypeScript tech stack AI application framework (AI News) that’s already reached ⭐20.0k stars. It supports building AI-driven applications and agents with a high degree of engineering sophistication! 🚀

5. VidBee: A Global Video Download Tool. VidBee lets you download videos from 🌐 pretty much any website out there, and it’s already at ⭐4.3k (AI News) stars. It covers all the major platforms and is super easy to use! 😍

6. Dexter: An Autonomous Agent for Deep Financial Research. Dexter is an autonomous agent (AI News) specially designed for deep financial research 💰, already boasting ⭐8.2k stars. It handles complex investment research analysis and data processing like a pro! 🔍

Social Media Shares

1. Claude Code Skills Mechanism: A Deep Dive. @shao__meng re-read the official docs 💡, breaking down the differences between Skills and CLAUDE.md (AI News) . Skills load on demand, saving context and offering strong reusability—pretty neat! ✨ Paired with Subagents, they can even hatch sub-agents. MCP provides external connections 🔗, while Skills offer usage guidance, making them complementary.
   
   

2. Pencil: A Design Canvas Based on Claude Code. @Gorden_Sun shared 🎨 Pencil, an infinite canvas tool (AI News) that requires Claude Code login. It comes with multiple built-in design component libraries and offers interactive visual operations—super intuitive! ✨ It supports importing Figma files and manually adjusting text colors. Compared to Stitch, you don’t have to manually tweak code, but it offers fewer alternative solutions.
   
   

3. AI Will Force Cognitive Upgrades in Eight Professions. @huangyun_122 reposted a hot take 🔥, stating that AI will inevitably bring two changes (AI News) . People will either be forced to transform and upgrade their cognition, or they’ll be marginalized and out of a job 😰. Professionals in eight specific fields will be the first to face this compelled growth. The awakening of personal brand awareness 💡 might just be the chance to break through the old “involution” system.

4. Tech Cycles: From DOS to the Skill Era. After messing with skills for a few days, @frxiaobei muses 🌀 that tech really does go in circles (AI News) . From DOS commands to Windows GUIs and then back to the terminal. Now, everyone’s collectively returning to the command line 💻, just with natural language instead—pretty wild! ✨ What truly creates the gap always comes down to foundational capabilities.

5. Selling HTML Files and Skills: The Next Big Business. @dotey believes that selling HTML files is totally fine (AI News) because browsers can do so much 🚀. Once Agent OS becomes widespread, selling Skills will be a huge business 💰. The trick is to find customers and actually sell ’em! ✨

6. Anthropic Releases Claude Constitution Document. @emollick points out that the 📜 Claude Constitution showcases Anthropic’s thoughts on the future of AI (AI News) . This is a massive document covering a ton of philosophical issues 🤔. It deserves serious attention beyond just the AI community. Other labs should totally be this transparent too!
   

AI News Daily Voice Edition

🎙️ Little Universe	📹 Douyin
Afterlife Tavern	Self-Media Account