Welcome back to Daily AI Skills.

Here’s what we are covering today:
1. SkyReels V2 - Infinite Length AI Video
2. OpenAI Codex CLI
3. Deepmind’s Experiential Vision for AI

+ GPT 4.1 Prompting Guide by Open AI

SkyReels V2 - Infinite Length AI Video Generation

Skywork AI’s SkyReels V2, an open-source generative AI model, excels in infinite-length video creation. Using Diffusion Forcing and multi-modal LLMs, it delivers cinematic-quality videos with superior motion, prompt fidelity, and user control, rivalling top proprietary models.

The Details:

• Diffusion Forcing Model: Transformer-based architecture with RL, generating 720p videos at 24 FPS for T2V and I2V, scoring 3.24 on VBench for I2V, beating HunyuanVideo-13 B.

• SkyCaptioner-V1: Trained on 2M clips, it annotates cinematic details like shots and angles, achieving 3.15 on SkyReels-Bench for prompt adherence.

• Training Pipeline: Multi-stage pretraining and RL fine-tuning reduce artifacts by 20%, enhancing fidelity for complex multi-subject scenes.

• SkyReelsInfer Framework: Multi-GPU, FP8-optimised inference generates 4s 720p clips in 80s on RTX 4090, supporting 30s videos with low drift.

Try out the model here: https://www.skyreels.ai/home

OpenAI’s Codex CLI

OpenAI’s Codex CLI, an open-source coding agent, transforms terminal-based development. Powered by advanced models like o3 and o4-mini, it autonomously writes, edits, and executes code locally, offering developers seamless, privacy-focused AI assistance for real-time programming tasks.

The Details:

• Local Coding Agent: Codex CLI runs offline on macOS and Linux (Windows via WSL experimentally), translating natural language into executable code in a sandboxed environment.

• Multimodal Reasoning: Supports text and screenshot inputs, leveraging o4-mini’s reasoning to understand codebases and perform tasks like debugging or feature prototyping.

• Flexible Approval Modes: Offers three modes—manual, semi-auto, and full-auto—allowing developers to control autonomy for tasks like repetitive edits or complex builds.

• Community-Driven Development: Open-sourced on GitHub with a $1M grant program, it fosters innovation, though active development means potential instability.

Try out Codex: https://github.com/openai/codex

DeepMind’s Vision for AI: “Welcome to the Era of Experience”

DeepMind’s latest paper, “Welcome to the Era of Experience,” introduces a paradigm shift in AI development. Moving beyond human-generated data, it proposes “streams” for AI to learn continuously from real-world interactions and environmental feedback, unlocking new levels of discovery and adaptability.

The Details:

• Limitations of Human Data: Authored by RL pioneers David Silver and Richard Sutton, the paper argues that reliance on human data restricts AI’s ability to make novel discoveries.

• Continuous Learning via Streams: Streams enable AI to learn through ongoing real-world interactions, replacing short Q&A with dynamic, long-term adaptation.

• Real-World Feedback Signals: AI agents leverage metrics like health data, test scores, or environmental inputs as feedback, reducing dependence on human judgments.

• Building on RL Success: The approach extends techniques from systems like AlphaZero, applying them to complex, open-ended real-world challenges.

• Beyond Human Knowledge: This shift could allow AI to uncover solutions surpassing current human understanding while incorporating flexible safety protocols.

Read the full paper here

A Practical Guide to Building Agents by OpenAI

📩Forward it to people you know who are keeping pace with the changing AI world, and stay tuned for the next edition to stay ahead of the curve!

Maybe you missed out on these:

1. Copilot vs Claude vs Gemini

2. How I Built an AI Influencer in 2 Minutes—and Landed Sponsorships

3. Microsoft Muse, Helix and Google's AI Co-Scientist

4. Bolt vs Lovable

5. Your AI Stack Just Got Superpowers: 4 New Inference Providers Live!