클로드 끝났어요. GPT아스트라 200% 활용하는 법

Summary

The video presents GPT’s newly released Astra as a major step up for complex creative work, especially compared with Claude’s Fable 5.1. The host argues that the difference is most apparent on demanding tasks—not ordinary text chat—and demonstrates two workflows: motion-graphic video creation and building a 3D scene with Blender.

Key points and demonstrations

• Motion-graphic video workflow: The host asks GPT to create an Apple-style promotional video. After connecting an external tool to process video files and extract frames, GPT analyzes a reference and produces an opening sequence. The host then connects an image- and video-generation tool to add rocket and space footage and assemble the final edit.

  • The first version is described as simple, while the later result includes animated text, transitions, and edited footage.
  • The host estimates the work took roughly 10–15 minutes.

• 3D scene and camera-work workflow: Using Blender, GPT first creates a 2D floor plan or bird’s-eye view, then builds a basic 3D space and plans a camera path. Astra is used to produce a Blender file and coordinate a camera move through the scene. An external tool is then used to add imagery and textures to the rendered footage.

  • The example follows a camera through an office, including a desk and a view through a floor-to-ceiling window.
  • The host recommends starting with a simple layout and camera path rather than asking the AI to create a fully detailed 3D scene immediately. This is faster and uses fewer tokens and fewer computer resources.

Practical guidance

• Use Astra for complex deliverables, not routine chat. The host says it performs well but consumes tokens quickly. Use a less costly model for planning or discussion, then switch to Astra for demanding production work.
• Use the appropriate GPT workspace or app. The video says Codex is accessed through the GPT app and describes it as a place to create outputs such as presentations, spreadsheets, and videos, rather than just chat responses.
• Connect external tools when needed. The host’s workflow relies on external integrations to analyze video frames and generate or apply visual assets; the video does not present these capabilities as purely built-in GPT features.
• Keep prompts brief and iterate. The host argues that newer agent-style models can interpret short requests, build on them, and handle more of the workflow themselves, reducing the need for extremely long prompts.
• Treat comparisons as demonstrations. The host describes Astra’s output as substantially better than Fable 5.1 in a 3D villa comparison and cites NVIDIA CEO Jensen Huang’s comments about Astra and the AI era. These are presented as commentary and examples, not as a controlled benchmark.

Main speakers and sources

• Main presenter: Dasol Kim (“Ro-Sajang”), who demonstrates the workflows and shares her assessments.
• Other speaker: An interviewer or co-presenter who asks questions about the demonstrations and workflow.
• External reference cited: NVIDIA CEO Jensen Huang. The video also compares GPT Astra with Claude Fable 5.1.