# Have you heard these exciting AI news? - January 09, 2026 AI Updates Weekly

Source: https://www.youtube.com/watch?v=_y6GP23mK7A
Recap page: https://rapidrecap.app/video/_y6GP23mK7A
Generated: 2026-01-09T20:05:53.03+00:00

---
## Quick Overview

The AI Updates for January 9, 2026, covered a diverse range of topics including the launch of Yann LeCun's AMI startup, rapid advancements in LLM benchmarks showing new doubling rates for coding tasks, significant Gemini AI updates in Gmail, new NVIDIA hardware/software platforms, significant security vulnerabilities in a humanoid robot, the open-source Oh My Open Code Framework, the Rocket.new text-to-app framework, Julia McCoy's AI-driven YouTube business model, and the introduction of DeepSeek's MHC architecture, all while noting economic impacts like GPU price increases due to memory shortages.

**Key Points:**
- Yann LeCun launched his startup, Advanced Machine Intelligence Labs (AMI), in Paris, seeking to raise €500 million ($586 million) at a €3.5 billion valuation, focusing on World Models trained on video and spatial data.
- AI progress shows time horizons for coding tasks doubling every 4 months (down from every 7 months over 2019-2025), indicating an accelerated 'New Moore's Law' for AI capability.
- Google announced major Gemini AI updates for Gmail, expanding AI features like Email Thread Summaries, AI Overviews in Search, Enhanced Writing Tools, and the coming AI Inbox view.
- NVIDIA unveiled the Vera Rubin Supercomputer Platform, succeeding Blackwell, featuring 1,152 GPUs across 16 racks with 72 Rubins per rack, emphasizing 45°C hot water cooling for efficiency.
- The Unitree G1 humanoid robot was shown to have severe security issues, including RF hijacking, LLM attacks via prompt injection, and physical access vulnerabilities, leading to the conclusion that 'security is bad'.
- The Open Code Framework 'Oh My Open Code' enhances the OpenCode AI coding agent by adding advanced multi-agent orchestration capabilities.
- Anthropic raised $10 billion at a $350 billion valuation, nearly doubling its valuation from four months prior, while memory shortages cause the price of RTX 5090 GPUs to potentially rise from $2K to $5K.

![Screenshot at 00:06: The main slide listing the key AI updates for January 09, 2026, including the Nick Bostrom quote: "Machine intelligence is the last invention that humanity will ever need to make."](https://ss.rapidrecap.app/screens/_y6GP23mK7A/00-00-06.jpg)

**Context:** This presentation delivers a comprehensive update on the AI landscape as of January 9, 2026, structured as a list of diverse news items and technical deep dives. It covers significant corporate moves, such as Yann LeCun's new venture and Anthropic's massive valuation jump, alongside critical research findings regarding the limits of current scaling laws in pattern matching versus true causality. The content also details product releases like Gemini in Gmail and new NVIDIA hardware, alongside warnings about the security of emerging AI-powered robotics.

## Detailed Analysis

The AI Updates for January 9, 2026, began with the launch of Yann LeCun's startup, AMI, in Paris, which is seeking €500 million at a $3.5 billion valuation. AMI focuses on World Models using video/spatial data, building on Meta's V-JEPA architecture, with a goal to develop FDA-certifiable AI agents for healthcare. Following this, the acceleration of AI progress was highlighted: the doubling time for coding tasks has dropped to 4 months, suggesting a 'New Moore's Law' that vastly outpaces traditional hardware scaling. In product news, Google rolled out major Gemini AI updates to Gmail, introducing features like AI-powered summaries, search overviews, and the forthcoming AI Inbox view. NVIDIA also made major announcements at CES 2026, including the Vera Rubin Supercomputer Platform, which uses 72 Rubins per rack and employs advanced 45°C hot water cooling, and the Cosmos platform for physical AI. On the open-source front, the 'Oh My Open Code' framework enhances multi-agent orchestration for OpenCode. In a security segment, significant vulnerabilities were revealed in the purchasable Unitree G1 humanoid robot, including LLM prompt injection and internal Ethernet access, leading to the conclusion that security is poor. Economic news noted that memory shortages are driving up GPU prices, with RTX 5090 potentially rising to $5K, and Anthropic achieved a $350 billion valuation after raising $10 billion. Technical updates included Claude 2.1.0 with better agent tooling, the MIT RLM demonstrating long-context handling via recursive processing, and a discussion on Correlation versus Causality, arguing that current AI lacks true causal reasoning despite excelling at pattern matching. Finally, the structure of startups is changing, moving from pyramid to more flattened/networked structures to maximize speed, and recent tech layoffs in the US for 2026 were reported at 454 people so far.

### AI Updates - Jan 09, 2026

- Crowd-sourced "LM Arena" Leaderboard shows competition among models like Gemini, Claude, and OpenAI
- Best 14-40B Params Open LLMs list, showing Qwen2-235B-A22B leading in quality score (9.2/10)
- New Moore's Law: AI capability doubling every 4 months, coding tasks doubling every 7 months
- Yann LeCun AMI Startup launched in Paris, seeking €500M at $3.5B valuation, focusing on World Models
- Gemini AI in Gmail expanding features like Thread Summaries, AI Overviews, Enhanced Writing Tools, and AI Inbox view
- Lightricks LTX-2 Open-Source Video Model generating 4K videos with synchronized audio/lip sync
- Memory Shortage Causes GPU Price Increase: RTX 5090 price may rise from $2K to $5K
- Anthropic raised $10B at $350B Valuation, doubling its prior valuation
- Claude Code 2.1.0 Released January 7 with automatic skill hot-reload and improved agent workflows
- Boris Cherny Work Setup details parallel coding sessions and using a 'Chairman' model for synthesis
- Google ADK - Agent Development Kit introduced for efficient context-aware multi-agent frameworks
- DeepSeek MHC introduced Manifold-Constrained Hyperconnections for better transformer performance
- Correlation versus Causality discussion highlighting that AI still lacks causal reasoning
- Jobs: Tech layoffs in 2026 at 454 people (674/day in 2025)
- Julia McCoy AI Business Model details her YouTube content automation strategy
- Oh My Open Code Framework enhances OpenCode agent with multi-agent orchestration
- Rocket.new text-to-app Framework offers backend infrastructure for app development

### Yann LeCun AMI Startup

- Launched in 2026 in Paris, France
- Seeking to raise €500 million ($586 million) at a $3.5 billion valuation
- Executive Chairman: Alex LeBrun
- Technology Focus: World Models, training on video/spatial data
- Technology builds on V-JEPA architecture created at Meta
- LeCun vocal that scaling LLMs only by text prediction isn't enough for human-level AI
- Strategic partnership formed with health-tech company Nabla
- Will develop FDA-certifiable AI agents for healthcare

### CES 2026 Consumer Electronics Show

- Nvidia CEO Jensen Huang dominated the show with major announcements including Cosmos, an AI foundation model trained to replicate real-world physics, and Alpama yo, a specialized AI model for autonomous vehicles
- He also revealed Nvidia's next-generation AI superchip platform, Vera Rubin, has entered full-scale production
- AMD CEO Lisa Su introduced new Ryzen AI processors for AI-driven personal computing and the Ryzen 9 85X3D for gaming enthusiasts
- She also unveiled the Helios rack-scale system for the Instinct MI445X GPU and previewed next-generation MI500 chips
- Intel announced Core Ultra 3 PC processors (Panther Lake) and plans for a handheld gaming device platform
- Qualcomm showcased the Snapdragon X 2 for AI applications
- Lenovo unveiled Qira AI platform, new ThinkPads in the Aura edition, and Motorola's first book-style folding phone, the Razr Fold
- Google presented Gemini 3 integration with Google TV for natural language searches and image editing
- Samsung showcased the Galaxy Zfold in the United States for the first time, featuring built-in AI functionalities
- TCL announced the RayNeo Air 4 Pro smart glasses with HDR-capable Micro-OLED displays
- Siemens unveiled the Digital Twin Composer, launching mid-year on the Siemens Xcelerator Marketplace
- Caterpillar CEO Joe Cread introduced the Cat AI Assistant, bringing the company's digital ecosystem into one conversational interface for construction sites

### NVIDIA Alpama yo & Cosmos

- Alpama yo is Open AI for Driving, supporting full stack autonomous driving framework
- Cosmos is a platform for physical AI featuring state-of-the-art generative world foundation models (WFMs)
- Nvidia Rubin Supercomputer Platform announced, succeeding Blackwell, delivering 10x cost reduction in inference and 4x reduction in GPU usage for MoE models
- Rubin comprises six chips (Vera CPU and Rubin GPUs) and uses 45°C hot water cooling

### Humanoid Robots Security Problems

- Unitree G1 robot purchasable online has security issues
- Researchers demonstrated remote takeover and jailbreak to unlock premium features
- Key Vulnerabilities include RF Layer hijacking, Network Service Exploits, Authentication Bypass, Cloud Token Theft, and LLM Attacks (Prompt Injection, Unsafe Code Execution via eval[])
- Conclusion: security is bad

### Google's Gemini CLI - Agent Skills

- Version 0.23.0 introduced experimental support for Agent Skills, following Anthropic's earlier implementation for Claude
- Skills allow extending Gemini CLI with modular, task-specific capabilities including specialized expertise, procedural workflows, and domain-specific resources
- Skills discovered from three tiers: project-level, user-level, and extension-bundled
- Skills managed via the /skills command; release also includes full integration of Gemini Conductor for context-driven development

### MIT RLM - Recursive Language Model

- MIT researchers developed a Recursive Language Model (RLM)
- Extends GPT-5's effective context window from 16-33K tokens to 1 million+ tokens
- RLM writes Python code to programmatically chunk, inspect, and recursively process data through spawned sub-instances of itself, similar to an operating system using virtual memory
- Architecture has three components: a root LLM for orchestration code, deterministic Python execution handling loops and logic, and recursive sub-LLM instances for actual text processing and summarization
- RLM-wrapped GPT-5 achieves 58% accuracy for $0.33, delivering over 1000x performance improvement over standard GPT-5

### Correlation versus Causality

- Citing Martin Milani, causality is the missing foundation for AI, which excels at pattern recognition but lacks causal models and logical reasoning
- Two 2025 studies show simply scaling pattern-matchers hits a wall, with one from Apple showing model accuracy collapses as tasks grow complex
- Correlation can predict, but cannot explain; intelligence requires causal models
- Until AI can represent causes, test counterfactuals, and justify inferences, it will remain powerful pattern recognition masking as cognition

### Jobs

- Tech layoffs in 2024-2025 shown in bar chart
- US Tech layoffs in 2026 (as of Jan 8) at 454 people
- 2025 saw 245,953 layoffs (674/day), down from 264K in 2023 and 165K in 2022

![Screenshot at 00:20: The slide detailing the Crowd-sourced "LM Arena" Leaderboard for English and Coding tasks, showing the performance of various LLMs like Gemini, Claude, and Open Source models.](https://ss.rapidrecap.app/screens/_y6GP23mK7A/00-00-20.jpg)
![Screenshot at 01:47: The slide detailing the 'New Moore's Law' acceleration, showing coding tasks complexity doubling every 7 months, recently accelerating to every 4 months.](https://ss.rapidrecap.app/screens/_y6GP23mK7A/00-01-47.jpg)
![Screenshot at 03:55: The slide summarizing updates on Gemini AI in Gmail, highlighting features like thread summaries and the coming AI Inbox view, alongside the Gemini and Gmail logos.](https://ss.rapidrecap.app/screens/_y6GP23mK7A/00-03-55.jpg)
![Screenshot at 06:56: The slide detailing NVIDIA's Alpama yo & Cosmos for Open AI for Driving and the Vera Rubin Supercomputer Platform, with corresponding visual hardware diagrams.](https://ss.rapidrecap.app/screens/_y6GP23mK7A/00-06-56.jpg)
![Screenshot at 07:52: The slide detailing security problems found in the Unitree G1 humanoid robot, listing specific vulnerabilities like RF hijacking and LLM attacks.](https://ss.rapidrecap.app/screens/_y6GP23mK7A/00-07-52.jpg)
![Screenshot at 08:28: The slide covering Google's ADK \(Agent Development Kit\) focusing on context management for multi-agent frameworks.](https://ss.rapidrecap.app/screens/_y6GP23mK7A/00-08-28.jpg)
![Screenshot at 09:51: The slide detailing MIT RLM, the Recursive Language Model, and its performance charts comparing GPT-5 against RLM\(GPT-5\) on context length.](https://ss.rapidrecap.app/screens/_y6GP23mK7A/00-09-51.jpg)
![Screenshot at 23:25: The slide discussing Correlation versus Causality, featuring an image of Albert Einstein and a linear regression graph illustrating correlation.](https://ss.rapidrecap.app/screens/_y6GP23mK7A/00-23-25.jpg)
![Screenshot at 25:28: The slide detailing Tech Layoffs in 2024-2025 and the Tech Layoff Tracker data.](https://ss.rapidrecap.app/screens/_y6GP23mK7A/00-25-28.jpg)
![Screenshot at 25:42: The slide about the speaker, Lev Selector, Ph.D., listing his experience and interests, including Generative AI and Local AI.](https://ss.rapidrecap.app/screens/_y6GP23mK7A/00-25-42.jpg)