# Have you heard these exciting AI news? - January 23, 2026 AI Updates Weekly Source: https://www.youtube.com/watch?v=cZYFF1zvI8k Recap page: https://rapidrecap.app/video/cZYFF1zvI8k Generated: 2026-01-23T21:34:53.242+00:00 --- ## Quick Overview The January 23, 2026 AI Updates cover significant progress in AI agent orchestration via skills.sh (now with 6574 skills), the continued rise of Claude models in the LM Arena, the decline of Stack Overflow as a public coding knowledge source, the release of GLM-4.7 Flash, Baidu ERNIE 5.0, and numerous product/workflow updates across agents, video generation (PixVerse V5), and speech processing (Nvidia personaplex-7b-v1). **Key Points:** - Claude Opus models (like Claude-opus-20251101-thinking) are strongly positioned in the top ranks of the crowd-sourced LM Arena for both English and Coding benchmarks as of January 16. - Stack Overflow is experiencing a rapid decline in traffic and value as developers pivot to using AI agents and LLMs for instant coding help, leading to its platform being sold for $1.8 Billion in 2021. - Google released TranslateGemma, an open-weight LLM built on Gemma 3 with 48, 128, and 278 parameters, supporting 55 languages and multimodal tasks, though image translation quality is noted as 'not very good' due to hallucinations. - PixVerse, an Alibaba-backed startup, launched its real-time interactive AI video generation tool on January 13, 2026, having raised over $60 million and serving 16 million monthly active users. - Nvidia released personaplex-7b-v1, a full duplex speech-to-speech conversational model that replaces the previous ASR->LLM->TTS cascade with a single Transformer model for streaming speech understanding and generation. - The MyhroThinker v1.5 open-source web research agent ranks at the top of benchmarks like HLE and GAIA, supports 256K-token context windows, and handles up to 400 tool calls per task. - Microsoft Research predicts significant job market shifts, with AI skills earning roughly 56% more than those without by 2027, emphasizing strategies like collaboration, human skills (ethics, strategy, emotional intelligence), and gaining AI literacy. ![Screenshot at 00:20: The slide displays the 'Crowd-sourced "LM Arena" Leaderboard' data from January 16, highlighting Claude Opus models ranking highly against competitors like Gemini Pro and other open-source models in both English and Coding benchmarks.](https://ss.rapidrecap.app/screens/cZYFF1zvI8k/00-00-20.jpg) **Context:** This weekly AI updates presentation from January 23, 2026, summarizes recent developments across various domains including LLM benchmarking, agent frameworks, specialized models (speech, video, image generation), company news (layoffs, investments, leadership changes), and job market predictions based on Microsoft Research findings. ## Detailed Analysis The updates begin with the crowd-sourced LM Arena leaderboard (data for January 16), showing Claude Opus models maintaining strong positions, particularly in coding tasks, suggesting their superior performance despite the cost of use. The presentation highlights the rapid decline of Stack Overflow as developers increasingly rely on AI coding assistants like Cursor and Claude Code, which now also has a VS Code extension. New models and tools are featured: Google's TranslateGemma, an open-weight, multimodal LLM supporting 55 languages; PixVerse V5, an Alibaba-backed real-time AI video generation tool that has raised $60 million; and FLUX.2 Klein for rapid 0.5-second image generation. In speech technology, Nvidia announced personaplex-7b-v1, a full-duplex speech model that replaces the traditional ASR->LLM->TTS pipeline with a single Transformer stack. Other developments include Aion UI, an open-source 'co-work' alternative restricted to Anthropic Pro/Max users, and Manus AI (Meta) for developing mobile apps. Claude Code extension for VS Code enhances the terminal experience with features like context window management. Deepseek Engram introduced a conditional memory module for MoE models, enabling O(1) lookups for massive context. Microsoft Research job market predictions suggest AI skills will earn roughly 56% more by 2027, advocating for collaboration, human skills, and AI literacy. Finally, the presentation notes MyhroThinker v1.5 is leading benchmarks and can run locally via GitHub, while Wikipedia is signing AI training deals with Microsoft, Meta, and Amazon. ### LM Arena Leaderboard - Claude Opus models ranking highly in English and Coding contests (e.g., Claude-opus-20251101-thinking leading in coding); Starting Elo rating is 1000. ### Coding Agents & Tools - Skills.sh is now a directory of 6574 skills, installable via 'npx skills add'; Claude Code extension for VS Code offers user-friendly UI with custom commands and context management. ### Model & Tool Updates (General) - Cursors ran 100s of AI coding agents for weeks, building a 3M+ line web browser on GPT-5.2; Baidu released Ernie 5.0, a native omni-modal large model with 2.4T MoE parameters (3% active). ### Model & Tool Updates (Vision/Speech) - Nvidia personaplex-7b-v1 is a full duplex speech model replacing ASR->LLM->TTS; PixVerse V5 offers real-time video generation with V3.5 features; Google Studio upgraded with VEO 3.1 integration; FLUX.2 Klein generates images in 0.5 sec. ### AI Research & Frameworks - Deepseek Engram introduced conditional memory via scalable lookup for MoE models, storing up to 27B parameters; MyhroThinker v1.5 is an open-source web research agent outperforming competitors like HLE and GAIA. ### Industry & Career News - Taiwan invests $250B into US semiconductor/energy/AI production; Thinking Machines Lab co-founders left for OpenAI; OpenAI launched a $8/mo 'ChatGPT Go' plan; Microsoft Research predicts AI skills will earn 56% more by 2027, emphasizing human skills and AI literacy. ![Screenshot at 00:20: The slide displays the 'Crowd-sourced "LM Arena" Leaderboard' data from January 16, highlighting Claude Opus models ranking highly against competitors like Gemini Pro and other open-source models in both English and Coding benchmarks.](https://ss.rapidrecap.app/screens/cZYFF1zvI8k/00-00-20.jpg) ![Screenshot at 01:14: Detailed explanation of skills.sh directory structure showing how skills are organized in Markdown files, with an example of 'bigquery-skill'.](https://ss.rapidrecap.app/screens/cZYFF1zvI8k/00-01-14.jpg) ![Screenshot at 03:42: Comparison slide detailing Cline vs Claude Code AI coding assistants, showing Claude Code has over 3M downloads for its VS Code extension, more than Cline.](https://ss.rapidrecap.app/screens/cZYFF1zvI8k/00-03-42.jpg) ![Screenshot at 06:17: Overview of Nvidia personaplex-7b-v1, a full-duplex speech model that uses a single Transformer model for streaming speech understanding and generation.](https://ss.rapidrecap.app/screens/cZYFF1zvI8k/00-06-17.jpg) ![Screenshot at 27:37: Slide detailing Nvidia's personaplex-7b-v1, which uses ASR-\>LLM-\>TTS conversion through a single full duplex model, achieving high accuracy \(75%\) on emotion detection using bodily signals.](https://ss.rapidrecap.app/screens/cZYFF1zvI8k/00-27-37.jpg)