# Have you heard these exciting AI news? - October 24, 2025 AI Updates Weekly

Source: https://www.youtube.com/watch?v=l1TIERzS-Zw
Recap page: https://rapidrecap.app/video/l1TIERzS-Zw
Generated: 2025-10-24T19:33:55.337+00:00

---
## Quick Overview

The AI Updates for October 24, 2025, highlight significant advancements across the AI landscape, featuring the launch of Google's new "vibe coding" product, Google Build, notable performance improvements in Claude's agent skills, Gemini's URL context grounding via API, the emergence of new terminals on macOS, and competitive shifts where ChatGPT is losing market share to Anthropic and DeepSeek, alongside a discussion on the energy demands of AI.

**Key Points:**
- Google launched 'Google Build,' a new 'vibe coding' product enabling direct AI application deployment from the browser with features like full-code visibility and instant deployment via GitHub integration.
- Claude Agent Skills received major updates, including executable code support, which allows for complex, locally replicated tasks, and support for multiple models like Gemini in the same context.
- Gemini introduced 'URL context grounding' via API, allowing users to extract full content (HTML, pages, PDFs, images) from URLs to ground LLM responses, featuring RAG and URL context grounding diagrams.
- The competition landscape shows ChatGPT losing dominance, with its market share falling to roughly 60% while Anthropic and DeepSeek gained ground, with Anthropic now leading enterprise market share at 32%.
- Nvidia unveiled the Blackwell 72GB GPU, which offers a 50% increase in memory over the original 48GB model and is manufactured with TSMC in the US (Phoenix).
- New local LLM execution methods on Android phones are emerging, including NN-Chat, MuC Chat, and running Llama via Termux, with performance varying based on hardware.
- The video covered Orchestration Frameworks (LangGraph, CrewAI, Microsoft Agent Framework) and new terminals for macOS (Term2, Kitty, Warp, Hyper, Core Shell, Tubby), noting the importance of these tools for managing complex multi-agent systems.

![Screenshot at 0:06: The main slide detailing the key AI updates for the week, organized into sections like Claude Agent Skills, Gemini URL context grounding, Google Build, and market competition metrics.](https://ss.rapidrecap.app/screens/l1TIERzS-Zw/00-00-06.png)

**Context:** This video provides the weekly AI updates for October 24, 2025, delivered by Lev Selector, Ph.D. The updates cover a broad range of topics including new product releases from major players like Google and Nvidia, significant feature updates for existing models like Claude, competitive analysis showing shifts in market dominance, and practical developments in local LLM deployment and agent orchestration.

## Detailed Analysis

The AI updates for October 24, 2025, covered several critical areas. Google introduced 'Google Build,' a new 'vibe coding' product that allows users to create and deploy AI applications directly in the browser using Gemini 2.5 Pro, featuring one-click deployment to Cloud/Run and version control via a pre-built app gallery. Claude Agent Skills have significantly evolved, now supporting executable code (like Python) run in secure environments via the Claude Execution Tool Beta, and supporting multiple models (including Gemini) simultaneously in the same context. Gemini added 'URL context grounding' via API, allowing extraction of full content from web pages to ground responses, a process illustrated with RAG and URL Context Grounding flowcharts. In terms of market dynamics, ChatGPT's dominance is rapidly declining, with its market share dropping to around 60% (down from 90% previously), while Anthropic and DeepSeek gained ground. Anthropic now leads the enterprise market with 32% share. Nvidia announced the Blackwell RTX Pro 5000 with 72GB of GPU memory, a 50% increase over the 48GB model, manufactured at TSMC in the US. The presentation also touched upon running LLMs locally on Android phones using tools like NN-Chat and Termux, noting that performance varies significantly based on hardware and model size. IBM's CUGA agent framework was highlighted for its evolutionary architecture and systematic approach to learning from failure. Finally, there was a brief mention of alternatives to JSON for data interchange (Protobuf, FlatBuffers, MessagePack, CBOR) and a slide on preventing loops in multi-agent systems, alongside a discussion on the growing energy demands for AI, citing Peter Diamandis's view that the AI race is an energy race, with China currently leading the US in chip manufacturing.

### Claude Agent Skills

- Skills are now uploaded as specialized knowledge; new 'h2_skills' endpoint offers versioning/management; skills require Code Execution Tool beta for secure execution environment; developers can create custom skills; supports multiple AImodels simultaneously (Anthropic, Claude, Gemini, etc.); supports code execution (e.g., Python) in a secure sandbox; allows using skills with agents in the same context (e.g., Claude Code agents and Sub-agents).

### Gemini 'URL context grounding' via API

- Asks model via API to extract data from URL; retrieves full content (HTML, pages, PDFs, images); can function standalone or combine with Google Search grounding; common applications include extracting key data points, synthesizing data, answering questions about page content, and analyzing content for particular purposes; feature available with added input tokens at standard rates.

### Google Build - new 'vibe coding' product

- Allows users to create/deploy AI applications directly in the browser using Gemini 2.5 Pro; features one-click deployment to Cloud/Run, GitHub integration for version control, pre-built app gallery with templates; includes full code visibility, checkpoint restoration, and API building features.

### LLM Competition & Market Share

- OpenAI's dominance is rapidly declining; ChatGPT market share fell from 90% to ~60% (losing approx 1.5% monthly); Anthropic now dominates enterprise market with 32% share; Anthropic generates $22/user vs. OpenAI's $25/user in efficiency.

### Hardware & Energy

- Nvidia unveils RTX 5000 Blackwell 72GB GPU (50% increase over original 48GB model) made with TSMC in the US (Phoenix); Google's Millow quantum chip achieved verifiable quantum advantage; AI energy race is on, with China leading the US.

### Orchestration Frameworks

- Top ready-to-use frameworks include LangGraph, CrewAI, Microsoft Agent Framework; all three coordinate a Cline instance; LangGraph is cellular function or MCP server; provides monitoring platforms like Galileo for real-time observability.

### Terminals on macOS

- Review covers Term2, Kitty, Warpy, Hyper, Core Shell, Tubby; Term2 is considered the best; AppleScript can be used to open a terminal window/tab via GitHub repo.

![Screenshot at 0:02: Title slide for the 'AI Updates - October 24, 2025' presentation, showing the agenda topics.](https://ss.rapidrecap.app/screens/l1TIERzS-Zw/00-00-02.png)
![Screenshot at 0:33: Detailed view of the 'Crowd-sourced LM Arena Leaderboard' comparing English and Coding performance, showing minimal changes from the previous week.](https://ss.rapidrecap.app/screens/l1TIERzS-Zw/00-00-33.png)
![Screenshot at 0:50: Slide detailing 'Claude Agent Skills' updates, highlighting the ability to upload specialized knowledge and the new 'h2\_skills' endpoint.](https://ss.rapidrecap.app/screens/l1TIERzS-Zw/00-00-50.png)
![Screenshot at 1:04: Slide comparing 'Claude Code for the Web' features \(session management, GitHub integration\) against 'OpenAI Atlas Web Browser' features \(persistent sidebar, agent nodes\).](https://ss.rapidrecap.app/screens/l1TIERzS-Zw/00-01-04.png)
![Screenshot at 1:35: Slide discussing the older 'Model Context Protocol' vs. the new 'Skills' approach for giving Claude access to external data.](https://ss.rapidrecap.app/screens/l1TIERzS-Zw/00-01-35.png)
![Screenshot at 1:47: Explanation of how Claude skills can be organized locally using a directory structure.](https://ss.rapidrecap.app/screens/l1TIERzS-Zw/00-01-47.png)
![Screenshot at 2:15: Slide detailing 'Emulate Skills in Google' using Gemini API in Python, showing code snippets for client initialization and function calling.](https://ss.rapidrecap.app/screens/l1TIERzS-Zw/00-02-15.png)
![Screenshot at 2:26: Slide comparing 'Claude Code for the Web' features against 'OpenAI Atlas Web Browser' features, including screenshots of both UIs.](https://ss.rapidrecap.app/screens/l1TIERzS-Zw/00-02-26.png)
![Screenshot at 2:34: Slide on 'Claude Skills vs Agents,' detailing architectural differences in autonomy and governance between the two approaches.](https://ss.rapidrecap.app/screens/l1TIERzS-Zw/00-02-34.png)
![Screenshot at 2:50: Slide on 'Run LLM on Android Phone' listing various methods like NN-Chat, MuC Chat, and using Termux, with a list of key performance bottlenecks \(key is I/O\). \(This slide is actually about running LLMs locally on Android, not the next slide\). Wait, this slide is about running LLMs on Android, not the next slide. The next visible slide is 24, 'Preventing Loops in Multi-Agent System'. Let's adjust the screenshot list to reflect the actual sequence shown in the video progression, focusing on key visual shifts in the presentation slides on the left panel as well as the main content area. The slide sequence seems to jump around slightly in the visual panel vs the spoken topic, but I will stick to the main content visibility for accurate context. I will replace the previous screenshot with the correct one for the context being discussed at that time \(which was CLI/Skills\). Let's use the correct slide index based on the visible list on the left side or the content displayed on the main slide area. Based on the current view, the speaker is discussing CLI/Skills, which is slide 11 on the left panel, corresponding to the content around 2:26-2:50. I will use a better representation of that section instead of the one from 1:47. Re-selecting a screenshot for the CLI/Skills section: 2:26:Comparison of Claude Code for the Web vs OpenAI Atlas Web Browser, showing UI differences and feature sets for web interaction tools. Re-selecting a screenshot for the Gemini URL Grounding section \(slide 10\): 10:39:Diagram showing the RAG and URL Context Grounding workflows for Gemini's API feature. Re-selecting a screenshot for the LLM on Android section \(slide 12\): 12:47:Slide detailing methods to run LLMs locally on Android, showing the comparison chart for Term2, Kitty, etc. Re-selecting a screenshot for the CLI/Run Multiple Instances section \(slide 11\): 14:25:Overview of the 'Cline CLI - run multiple instances' capabilities, including setting multiple agents and handling agent communication patterns like MCP/Callback. Re-selecting a screenshot for the Jobs slide \(slide 26\): 34:25:Chart showing Tech Layoffs in 2024-2025, highlighting significant drop-off in 2025 projections and the Jevons Paradox mention. Re-selecting a screenshot for the Andrej Karpathy Interview slide \(slide 20\): 25:30:Key takeaways from the Andrej Karpathy interview, contrasting LLMs vs. different kinds of intelligence and discussing reinforcement learning challenges. Re-selecting a screenshot for the Google VISTA Agent slide \(slide 21\): 26:11:Workflow diagram of the Google VISTA agent, detailing the 4-stage process: Planning, Tournament Selection, Multi-Agent Critique, and Deep Thinking Prompting. Re-selecting a screenshot for the Jobs slide \(slide 26\): 34:25:Chart showing Tech Layoffs in 2024-2025, highlighting significant drop-off in 2025 projections and the Jevons Paradox mention. Re-selecting a screenshot for the 'All Updates - 2' slide \(slide 20\): 21:25:Updates section covering Google Vibe 3.1, Nvidia Blackwell chip production in the US, and the RTX 4000 series GPU memory increase, alongside ChatGPT Memory management settings.](https://ss.rapidrecap.app/screens/l1TIERzS-Zw/00-02-50.png)