# 📆 Oct 9, 2025 — Dev Day’s Agent Era, Samsung’s 7M TRM Shock, Ling‑1T at 1T, Grok Video goes NSFW,...

Source: https://www.youtube.com/watch?v=Ffu7rgT99r4
Recap page: https://rapidrecap.app/video/Ffu7rgT99r4
Generated: 2025-10-10T02:32:33.016+00:00

---
## Quick Overview

The Thursday, October 9th episode of the AI-focused show covered major developments including Microsoft's release of the user-simulating User-LM-AB model, AI21 Labs' Jamba Reasoning 3B model featuring a hybrid SSM-transformer architecture, Inclusion AI's trillion-scale Ling-1T reasoner claiming performance matching Gemini 2.5, and Samsung's groundbreaking 7-million-parameter Tiny Recursive Model (TRM) achieving high accuracy through recursive self-critique.

**Key Points:**
- Microsoft released User-LM-AB, an 8B Llama 3 fine-tune trained to simulate the user role in conversations to develop more robust assistants, useful for synthetic data generation.
- AI21 Labs released Jamba Reasoning 3B, a tiny model using a hybrid SSM-transformer architecture, which leads tiny reasoning models on benchmarks like IFBench and Instruction Following Bench.
- Inclusion AI debuted Ling-1T, a one-trillion-parameter efficient reasoner using dynamic routing that reportedly matches Gemini 2.5 performance while using half the compute footprint.
- Samsung released the TRM (Tiny Recursive Model) with only 7 million parameters, which achieves 45% test accuracy on RKGI-1 by iteratively critiquing and refining its logic up to 16 times.
- OpenAI Dev Day announcements included CodeX going General Availability (GA) with an SDK, the launch of ChatGPT Apps built on the MCP standard using iframes for UI presentation, and the Agent Kit.
- The hosts expressed excitement and simultaneous fear about the rapid pace of AI development, noting that builders feel like they are 'running with the bulls' due to potential disruption by large labs.
- OpenAI launched the Realtime-Mini API, which is significantly faster and approximately 80% cheaper than the existing Realtime voice API.

**Context:** The hosts, Alex Falov, Wolram, Yam Pelleg, LDJ, and Ryan Carson, along with guest Eric Provener from RepoPrompt, provided their weekly update on AI news for Thursday, October 9th, emphasizing the blurring lines between physical and digital reality due to advancements in generative video models like Sora. The discussion focused heavily on major open-source model releases, key announcements from OpenAI's Dev Day, and benchmarking tools like RepoBench.

## Detailed Analysis

The episode opened with a reflection on the increasingly post-reality world driven by video generation, before diving into open-source breakthroughs. Key releases included Microsoft's User-LM-AB for simulating user dialogue, AI21's Jamba Reasoning 3B utilizing a hybrid SSM-transformer design for efficiency, and Inclusion AI's Ling-1T, a one-trillion-parameter model claiming parity with Gemini 2.5 on several benchmarks while using less compute. The most astonishing reveal was Samsung's TRM, a 7-million-parameter model that achieves high accuracy by employing recursive self-critique, sparking excitement about hobbyist research potential due to its small scale and open-source code. Transitioning to OpenAI Dev Day, the panel discussed the launch of CodeX GA and its SDK, highlighting the paradigm shift of self-modifying apps via code execution. Furthermore, OpenAI officially launched ChatGPT Apps, leveraging the MCP standard to embed iframed experiences directly into the chat interface, marking their third attempt at creating an app ecosystem, prompting discussion on whether OpenAI is becoming the 'Windows of AI' and acting as a gatekeeper for app discoverability. Other notable mentions included Google DeepMind's Gemini 2.5 Flash, uncensored video generation from Grok, and the introduction of Serverless RL via CoreWeave's partnership with OpenPipe.

### Open Source LLM Innovations

- Microsoft released User-LM-AB to simulate user roles for better assistant training
- AI21 released Jamba Reasoning 3B, a tiny hybrid SSM-transformer model
- Inclusion AI launched Ling-1T, a 1T parameter model matching Gemini 2.5 performance.

### Samsung TRM

- The Tiny Recursive Model (TRM) uses only 7 million parameters and obtains 45% accuracy on RKGI-1 by refining logic recursively up to 16 times
- The researcher behind it, Alexia, is noted for releasing groundbreaking, simple, and insightful work
- The code and data are fully open-sourced on GitHub.

### OpenAI Dev Day Key Announcements

- CodeX reached General Availability (GA) with an SDK, enabling programmatic integration of agents inside applications
- ChatGPT Apps launched, utilizing iframes and the MCP standard to integrate third-party tools directly into the chat interface
- Agent Kit was also announced.

### Ecosystem Growth and Platform Control

- OpenAI reported 800+ million weekly active users and 6 billion tokens per minute on their API, signaling a massive ecosystem push
- The move to in-chat apps via iframes positions OpenAI as a potential platform gatekeeper, controlling which apps are presented to users.

### API Updates and Benchmarking

- OpenAI released GPT-5 Pro in the API, though noted as very expensive
- The Realtime-Mini API launched, offering an 80% price reduction over the existing Realtime voice mode
- Guest Eric Provener introduced RepoBench, focused on evaluating models' ability to perform code editing grunt work.