NemoClaw Knowledge Wiki

GPT 5 - Prompt Engineer channel

5 min read

https://www.youtube.com/watch?v=RyI17GMrMDs This video provides a first look at outputs from OpenAI’s alleged GPT-5 model, accessed via API, showcasing its capabilities in code generation, visual rendering, and complex problem-solving. Here’s a detailed summary of the demonstrations and observations: 1. General Impressions & Key Capabilities:

  • The speaker claims to be showcasing the first confirmed GPT-5 API outputs, distinct from public chatbot arena versions.
  • The model is described as “really good,” especially for coding, capable of generating visuals “that I haven’t seen from any other model.”
  • It’s a multimodal model, accepting images as input (video capability is unknown) and generating images.
  • It features a “Reasoning Effort” setting (Low, Medium, High), similar to Claude 3 models.

2. Visual & Code Generation Examples:

  • Procedural Planet (0:00, 3:57): Generates a realistic, procedurally generated 3D planet in Three.js with detailed terrain, biomes, atmosphere, lighting, and interactive controls (rotation speed, pause, reset, zoom). The speaker notes the quality is highly dependent on the detailed prompt provided.
  • Sci-Fi UI Dashboard (0:12, 5:09): Generates a complex, futuristic UI dashboard (Holographic Nav Core, Reactor Core, System Console) with dynamic elements like energy analyzers, mesh network topology, and system status logs. The speaker highlights its creativity and visual quality compared to other models.
  • Professional Websites (0:19, 2:34): Ledgerline Landing Page (0:19): Generates a visually appealing and functional landing page for a financial service, including an interactive code snippet for TypeScript and cURL, and claims about audit-grade features and uptime. Legendary Pokédex (1:36): Creates a functional Pokédex website featuring 25 legendary Pokémon with search, filtering by type (Dragon, Electric, Fire, etc.), and a theme toggle for dark/light mode. The speaker notes the visual appeal and functionality, but points out “AI tells” like a misplaced theme button. Ledgerline Pricing Page (2:34): Generates a pricing page with hover-to-flip card animations and FAQs. While visually impressive, some text on the flipped cards appears upside down, indicating minor rendering issues that could likely be fixed with further iteration on the prompt.

3. Complex Problem Solving:

  • Mathematics Olympiad Problem (0:27, 8:04): Successfully solves a complex Math Olympiad-level geometry problem (IMO 2025 Solution Notes, Problem 1) in about 10 minutes. The model provides a reasoning process and the correct final answer (possible values for k are 0, 1, 3), matching the official solution. The speaker notes this is one of the few models capable of solving such problems.
  • Rubik’s Cube Simulation (0:21, 5:49): 5x5 Rubik’s Cube: Generates a 5x5 Rubik’s Cube simulation that can be scrambled and solved using Kociemba’s Algorithm. The speaker states that only a few models (including Gemini 2.5 Pro) can solve complex Rubik’s Cubes. Performance: In three attempts to solve a scrambled 5x5 cube in the video, it succeeded twice but failed once, noting that initial conditions might matter and there were occasional rendering glitches (missing blocks) which appear to be visual bugs rather than logical errors.
  • Wolf-Goat-Cabbage Riddle (8:58): Tests the model’s reasoning capabilities with a slightly modified version of the classic “wolf-goat-cabbage” riddle, where the question specifically asks only about transporting the goat safely, disregarding the wolf and cabbage for the core problem. Initial Response: The model initially provides the full, classic solution, implying it “misguidedly recognized” the unmodified problem (due to frequent occurrence in its training data, a concept from the “Misguided Attention” dataset). Self-Correction with Nudge: When prompted, “do you see any issues with your solution?”, the model correctly identifies that the user’s specific question only asked about the goat, not the full puzzle, and offers the simpler, correct answer of just taking the goat over. This demonstrates strong instruction following and reasoning even when facing common patterns in its training data.

4. Comparisons & Limitations:

  • GPT-5 vs. Other Models (Planet): The procedural planet generated by GPT-5 is deemed “much better” in visual quality compared to outputs from other models tested on LM Arena (e.g., “Anonymous Chatbot 0717” / o3-Alpha, and “Nectarine”).
  • “AI Tells”: Despite impressive outputs, some generated visuals show imperfections or odd placements (e.g., the misplaced “Theme” button on the Pokédex, upside-down text on flipping cards, occasional rendering issues in Rubik’s Cube).
  • Prompt Quality Matters: The speaker emphasizes that GPT-5 requires “highly detailed prompts” to achieve optimal results, as it closely follows instructions rather than improvising.
  • Not AGI (Yet): The speaker cautions against hyperbole, stating that GPT-5 is “nowhere close” to AGI and that there are often “hype cycles” around new model releases.

5. Overall Conclusion: GPT-5, based on these early API outputs, appears to be a significant leap forward for OpenAI, particularly in its ability to generate complex and functional code for interactive visuals and web applications, demonstrating impressive instruction following and reasoning capabilities even in tricky scenarios. While not flawless, its adherence to detailed prompts and capacity for self-correction indicate a powerful and versatile model.

Graph View