# Have you heard these exciting AI news? - September 26, 2025 AI Updates Weekly

Source: https://www.youtube.com/watch?v=ZXRKJ9yKhMY
Recap page: https://rapidrecap.app/video/ZXRKJ9yKhMY
Generated: 2025-09-26T19:04:49.016+00:00

---
## Quick Overview

The AI updates for September 26, 2025, highlight exponential growth across the sector, exemplified by Grok 4 Fast being 47x cheaper than Grok 4, significant LLM data compression via LLM-Deflate, the release of new Qwen3 models by Alibaba, the launch of AI-powered tools like Google Stitch and Convex Chef, and the impending release of Python 3.14 with major performance improvements.

**Key Points:**
- AI is growing 2-3 times faster than Moore's Law, with exponential growth demonstrated across infrastructure, investment, usage, and quality.
- Grok 4 Fast offers cost-efficient reasoning, being 47x cheaper than Grok 4 and supporting a 2M token context window for multimodal (text & images) tasks.
- LLM-Deflate compresses training data by extracting text into billions of parameters, reversing the process to extract structured datasets from trained models.
- Alibaba's FP8 builds on Qwen3 models (Next-80B-A38 and Qwen3-235B-A22B) featuring new MoE reasoning models.
- New developer tools include Google Stitch (text-to-design), Crush CLI (open-source AI coding agent for terminal), Lumi.new (AI no-code app builder), and Klink 2.5 Turbo (video generation model).
- The threat of AI displacing jobs is highlighted, with translators and data entry clerks facing 100% automation, while roles requiring empathy and creativity are least affected.
- Python 3.14 introduces major changes, including official support for free-threaded Python (removing GIL parallelism) and providing a 30% faster execution overall.

![Screenshot at 00:06: AI Updates for September 26, 2025, shown on a slide detailing advancements in leaderboards, exponential growth, and various company releases.](https://ss.rapidrecap.app/screens/ZXRKJ9yKhMY/00-00-06.png)

**Context:** This video presents the weekly AI Updates for September 26, 2025, delivered by Ray Kurzweil, summarizing significant developments across LLMs, hardware, software tools, job market impacts, and programming language updates. The overarching theme is the accelerating pace of AI progress, which is now outpacing Moore's Law, leading to massive productivity gains and societal shifts.

## Detailed Analysis

The presentation begins by emphasizing that AI is growing 2-3 times faster than Moore's Law, showcasing exponential growth in infrastructure, investment, usage, and quality. The Crowd-sourced 'LM Arena' Leaderboard shows models like Mistral Magisterial and Grok 4 Fast achieving high scores. The exponential growth section details how AI investment is expanding at a 24.4% CAGR, with corporate AI investment reaching $252.3 Billion in 2024, and how job displacement is accelerating, with translators and data entry clerks facing 100% automation. The segment on outsourcing jobs lists several roles heavily impacted (e.g., telemarketing, accounting) versus roles least affected (e.g., those involving empathy and creativity). Updates on major companies include Google Chrome browser having AI now, Groq securing $750M, and OpenAI sharing revenue with Microsoft decreasing to 8% by 2030. Software tool updates feature Google Stitch for text-to-design, Crush CLI for open-source AI coding, Lumi.new for no-code app building, and Kling 2.5 Turbo for video generation. Claude Opus 4.1 is noted for fixing bugs, while DeepSeek V3.1 is highlighted for its powerful, fast, and cost-efficient open-source agentic LLM capabilities. The presentation also covers Tencent's Hunyuan 3D 3.0 for creative professionals and Leann Vector DB offering 30x less space via pruned/recomputed embeddings. AI for Creative Media is exemplified by Fable Studio, a Showrunner platform that generates entire worlds. Finally, Python 3.14 is announced for October 7th, featuring official support for free-threaded Python (removing the GIL lock) and providing a 30% faster execution. The presenter concludes with a section on the speculative nature of the current AI boom, noting inflated valuations and the potential for a 'bubble burst' focusing on durable value.

### Crowd-sourced LM Arena Leaderboard

- Groq 4 Fast is competitive with huge models like Mistral Magisterial
- Claude Opus 4.1 is fixed and can run Docker and backend/frontend processes robustly
- Alibaba's FP8 builds on Qwen3 models for faster inference

### Exponential Growth

- AI infrastructure investment is expanding at a 24.4% CAGR
- Moore's Law is now growing 2-3 times faster than its historical rate
- Job displacement is accelerating, with 100% automation projected for translators and data entry clerks.

### Outsourcing Jobs to AI

- Translators (spoken/written) and Data entry clerks are 100% automated; roles involving empathy, judgment, and creativity are least affected
- Emerging hybrid roles include AI trainers and workflow designers.

### AI Monthly Payment Plans

- Perplexity Free/Pro Individual $20/Enterprise $40/Max Individual $200/Max Enterprise $325
- Claude Pro Individual $20/Max $100/Enterprise custom
- Gemini Individual $20/Workspace $20/Enterprise $30/Ultra $200-$250
- OpenAI Plus $20/Pro $200/Enterprise $30+

### AI Updates - 1

- Google AI updates Chrome browser and Search with AI mode
- Groq secured $750M at $6.9B valuation; Cohere secured another $100M at $7B valuation
- Tencent's Hunyuan 3D 3.0 offers precision and ultra-HD modeling
- Convex Chef platform opens source with Apache 2.0 license.

### AI Updates - 2

- OpenAI revenue sharing with Microsoft decreasing from 20% to 8% by 2030
- Alibaba FP8 builds on Qwen3 models for faster inference
- Fei-Fei Li's World Labs demo shows interactive 3D worlds from a single image
- Major LLMs (OpenAI, Google, Anthropic) pass all three levels of the CFA financial analyst exam.

### AI Coding & Vector DB

- GitHub Spec Kit framework for structured three-phase workflow
- LEANN Vector DB uses 3% of storage space compared to traditional vector databases by storing only pruned index and recomputing embeddings in real-time.
- Spiking Brain Models use 'event-driven processing' to mimic biological neurons, offering 69% computation sparsity and massive energy savings.

### Python 3.14 & Jobs

- Python 3.14 introduces GIL-free parallel execution on multiple CPU cores and provides 30% faster execution
- Tech layoffs in 2025 forecast 90K (US only), with Amazon redirecting $5B into AI systems, and Meta cutting $40B.

![Screenshot at 00:06: AI Updates - September 26, 2025 slide overview.](https://ss.rapidrecap.app/screens/ZXRKJ9yKhMY/00-00-06.png)
![Screenshot at 00:22: Crowd-sourced 'LM Arena' Leaderboard showing English and Coding model rankings.](https://ss.rapidrecap.app/screens/ZXRKJ9yKhMY/00-00-22.png)
![Screenshot at 01:10: Slide detailing 'Exponential growth' statistics, including pond doubling analogy.](https://ss.rapidrecap.app/screens/ZXRKJ9yKhMY/00-01-10.png)
![Screenshot at 03:31: Slide detailing job outsourcing statistics and the three categories of job impact.](https://ss.rapidrecap.app/screens/ZXRKJ9yKhMY/00-03-31.png)
![Screenshot at 04:49: Slide detailing Compute Owners vs Underclass polarization.](https://ss.rapidrecap.app/screens/ZXRKJ9yKhMY/00-04-49.png)
![Screenshot at 06:08: Slide detailing AI monthly subscription plans for Perplexity, Claude, Gemini, and OpenAI.](https://ss.rapidrecap.app/screens/ZXRKJ9yKhMY/00-06-08.png)
![Screenshot at 07:15: Slide titled 'AI Updates - 1' showing updates from Google, Groq, Cohere, Tencent Hunyuan, and Convex Chef.](https://ss.rapidrecap.app/screens/ZXRKJ9yKhMY/00-07-15.png)
![Screenshot at 09:18: Slide titled 'AI Updates - 2' detailing revenue sharing, Qwen3 models, Fei-Fei Li's World Labs demo, and programmer AI usage.](https://ss.rapidrecap.app/screens/ZXRKJ9yKhMY/00-09-18.png)
![Screenshot at 11:00: Slide titled 'AI Updates - 3' detailing new developer tools like Stitch, Crush CLI, Lumi, and Kling 2.5 Turbo.](https://ss.rapidrecap.app/screens/ZXRKJ9yKhMY/00-11-00.png)
![Screenshot at 12:06: Slide titled 'AI Updates - 4' detailing Skywork for professional reports and Grok Supernova as a new free AI coder, including its capabilities like analyzing broken website layouts and fixing CSS issues. \| 17:37:Slide titled 'AI Updates - 6' discussing Grok 4 Fast, LLM-Deflate, Tencent Parallel R1, and Ollama Web Search API availability in the cloud. \| 18:21:Slide titled 'AI Coding - 1' covering AI coding updates, including GitHub Spec Kit framework and OP Code desktop GUI for Claude code sessions. \| 22:01:Slide titled 'Preventing Fake Alignment' detailing Apollo Research's work on anti-scheming training and safety specs. \| 23:10:Slide titled 'DeepSeek-R1 Develops Advanced Reasoning' detailing RL-incentivized reasoning in LLMs, achieving SOTA performance across various tasks. \| 26:17:Slide titled 'LEANN Vector DB' highlighting its 30x space saving via pruned indices and real-time computation. \| 27:37:Slide titled 'Google Gemini TV' detailing Gemini AI built directly into TV software with presence-sensing technology. \| 31:17:Slide titled 'Jobs' showing tech layoff data for 2024-2025 and investments by major companies into AI systems. \| 31:24:Slide titled 'About the Speaker' featuring Lev Selector, Ph.D. \| 31:25:Final 'Thank You!' slide.](https://ss.rapidrecap.app/screens/ZXRKJ9yKhMY/00-12-06.png)