# AI News: This Video Model Has Everyone Freaked Out!

Source: https://www.youtube.com/watch?v=msrbd-d6lWk
Recap page: https://rapidrecap.app/video/msrbd-d6lWk
Generated: 2026-02-13T18:08:56.307+00:00

---
## Quick Overview

The video reviews recent developments in AI, focusing on ByteDance's Seodance 2.0 video generation model, Google's Gemini 3 Deep Think, the release of Minimax M2.5, a new open-source coding model from Ziai, Meta's AI-powered animation for Facebook profile pictures, and OpenAI's deprecation of older GPT models in favor of newer ones, alongside a demonstration of the speed and capability of the new models.

**Key Points:**
- ByteDance launched Seodance 2.0, a next-generation video creation model supporting four input modalities (text, image, audio, video), which demonstrated highly realistic video generation, including consistent character and motion control (0:19).
- Google released Gemini 3 Deep Think, a specialized reasoning mode that achieved state-of-the-art results on several benchmarks, outperforming GPT-4 and Claude Opus 4.6 in areas like reasoning, coding, and physics (10:21).
- Minimax released M2.5, an open-source model that is significantly faster (10x faster inference) and cheaper (1/10th the cost) than frontier models while achieving competitive performance benchmarks, including beating GPT-4.5 on some tasks (5:59, 10:00).
- OpenAI announced the deprecation of legacy models (GPT-4, GPT-3.5, etc.) in favor of newer models, with GPT-5.3-Codex-Spark showing massive speed improvements over GPT-5.3-Codex in coding tasks (3:38, 12:31).
- Meta is rolling out AI-powered features to animate Facebook profile pictures and allow users to control their Threads feed content via 'Dear Algo' prompts (1:34, 17:12).
- Krea AI introduced a prompt-to-workflow feature allowing users to generate complex node-based workflows from text instructions, demonstrated by creating a Game Boy Advance emulator (2:57).
- Waymo began fully autonomous operations with its 6th-generation Waymo Driver, featuring enhanced sensor resolution and AI processing capabilities, which is notably different from previous Jaguar-based models (29:30).

![Screenshot at 0:19: demonstration of Seodance 2.0 generating high-fidelity video of a figure ice skating, showcasing character and motion consistency.](https://ss.rapidrecap.app/screens/msrbd-d6lWk/00-00-19.jpg)

**Context:** This video serves as a news roundup summarizing several significant recent announcements and product releases across the competitive landscape of generative AI and related technology, including advancements in video generation (Seodance 2.0), large language models (Gemini 3 Deep Think, Minimax M2.5), coding assistants (GPT-5.3-Codex-Spark), social media AI features (Meta AI on Threads/Facebook), and autonomous driving hardware (Waymo's 6th Gen Driver).

## Detailed Analysis

The video covers a flurry of recent AI and technology news. First, ByteDance launched Seodance 2.0, an advanced video creation model supporting multi-modal inputs (text, image, audio, video) that excels at maintaining character and motion consistency, demonstrated by generating a realistic video of Will Smith eating spaghetti (0:19) and a figure skating sequence (1:23). Next, Google announced Gemini 3 Deep Think, which sets new state-of-the-art scores on benchmarks like ARC-AGI-2 (84.6%) and various academic reasoning tests, outperforming competitors like Claude Opus 4.5 and GPT-4.5, particularly in complex reasoning and coding (10:17). Minimax also released M2.5, an open-source model that is significantly faster and cheaper than frontier models, as demonstrated by creating a functional, playable Vampire Survivors clone in just 50 seconds (13:34). OpenAI is beginning to deprecate legacy models like GPT-4o and GPT-4-Turbo in favor of newer models, including the very fast GPT-5.3-Codex-Spark, which was shown to generate code significantly faster than the standard GPT-5.3-Codex (12:31). Furthermore, Meta is rolling out AI features to animate Facebook profile pictures and allow users to control their Threads feed content using 'Dear Algo' prompts (17:12). Krea AI introduced a prompt-to-workflow feature, demonstrated by generating a complex node-based workflow to build a Game Boy Advance emulator (2:47). Finally, Waymo began deploying its 6th-generation Waymo Driver, which features significantly improved sensing capabilities over previous models, highlighted by a dramatic nighttime driving simulation (29:30). The presenter concludes by emphasizing the rapid pace of development across all these domains.

### Video Generation

- ByteDance released Seodance 2.0 supporting four modalities
- Demonstrated by generating realistic, consistent video clips, including Will Smith eating pasta (0:19) and a figure skating routine (1:23)

### LLM Benchmarks

- Google's Gemini 3 Deep Think leads on many academic/reasoning tests, outperforming GPT-4.5 and Claude Opus 4.5 (10:17)
- Minimax M2.5 shows high performance while being 10x faster and 1/10th the cost of competitors (10:00)

### Coding & Speed

- OpenAI released GPT-5.3-Codex-Spark, demonstrating significantly lower latency compared to the standard Codex model when generating code (12:31), exemplified by building a functional Snake game in 50 seconds (13:56)

### AI Agent Capabilities

- GLM-5 demonstrates long-term task capability by building a GBA emulator autonomously over 24 hours (20:22)
- Krea AI introduced node-based prompt-to-workflow for complex agent construction (27:45)

### Social Media AI

- Meta is rolling out AI features to animate Facebook profile pictures and allow users to customize their Threads feed via 'Dear Algo' prompts (17:12)

### Hardware News

- Waymo launched its 6th-generation Waymo Driver with enhanced sensor resolution and AI capabilities for fully autonomous operations (29:30)

![Screenshot at 0:22: Highlighting the four input modalities supported by Seodance 2.0 \(text, image, audio, video\) in the official announcement.](https://ss.rapidrecap.app/screens/msrbd-d6lWk/00-00-22.jpg)
![Screenshot at 10:31: Bar chart comparing Gemini 3 Deep Think's performance against Claude and GPT models across four major benchmarks, showing clear outperformance in reasoning and coding.](https://ss.rapidrecap.app/screens/msrbd-d6lWk/00-10-31.jpg)
![Screenshot at 13:50: Side-by-side comparison of GPT-5.3-Codex vs. GPT-5.3-Codex-Spark generating a simple HTML Snake game, illustrating the massive speed difference.](https://ss.rapidrecap.app/screens/msrbd-d6lWk/00-13-50.jpg)
![Screenshot at 17:17: Graph showing GLM-5 performance across various benchmarks, demonstrating its lead over other open-source models and closeness to closed models like Claude Opus 4.5.](https://ss.rapidrecap.app/screens/msrbd-d6lWk/00-17-17.jpg)
![Screenshot at 29:49: Image of the 6th-generation Waymo Driver vehicle, highlighting its new sensor suite and exterior design.](https://ss.rapidrecap.app/screens/msrbd-d6lWk/00-29-49.jpg)