GPT-4o’s Upgrade is Kinda Wild (Coding, Images, Unhinged Mode?)

Key Concepts

GPT-4o, unhinged mode, coding capabilities, image generation, JavaScript animation, collision detection, p5.js sketch, TV channel generation, SVG generation, Pelican riding a bicycle, modern landing page, trolley problem, Schrodinger's cat paradox, intuition, reasoning capabilities, tone change, safety rules.

GPT-4o Update: Overview

The video discusses the new update to GPT-4o, highlighting its improved coding capabilities, less filtering (including an "unhinged mode"), and enhanced performance on chatbot arena leaderboards. The speaker tests the model's coding abilities with several prompts previously used with Gemini 2.5 and Claude, and also explores its image generation and reasoning capabilities.

Coding Capabilities: JavaScript Animation

• Prompt: Create a JavaScript animation of falling letters with realistic physics, including random appearance, gravity, collision detection, density properties, dynamic screen size changes, and a dark background, all in a single HTML file.
• GPT-4o's Initial Response: Initially produced green boxes around the letters.
• Resolution: The speaker asked GPT-4o to fix the code to hide the green boxes. GPT-4o provided updated code that successfully displayed falling letters with collision and dynamic sizing.
• Outcome: GPT-4o successfully completed the task after one correction, demonstrating improved coding abilities.

Coding Capabilities: TV Channel Generation (p5.js Sketch)

• Prompt: Code a TV that lets the user change channels with number keys (0-9), come up with an idea for a channel for all numbers inspired by classic genres of TV channels, show detailed interesting animations for concepts or contents and a creative name for channel on the screen, return an 800x800 p5.js sketch, without using HTML, on a black background, and ensure the content of all the channels stays masked to the TV screen area.
• Gemini 2.5 Pro: Generated 571 lines of working code, creating a functional TV with different channels.
• GPT-4o's Initial Response: Generated 200 lines of code, but it had errors and did not function correctly.
• Resolution: The speaker provided GPT-4o with the error messages, and GPT-4o provided corrected code.
• Outcome: After multiple attempts and corrections, GPT-4o produced a working TV with different channels, though it required more iterations than Gemini 2.5 Pro. The final code was shorter than Gemini's. Some channels were missing names.

Comparison with Claude 3.5 Sonnet

• The speaker attempted the TV channel generation prompt with Claude 3.5 Sonnet.
• Issue: Claude hit the maximum length for the message and did not complete the code.
• Context Window: Claude has a 200,000 token context window, with a rumored 500,000 token version coming soon.
• Conclusion: Claude struggled with the long code generation task due to token limitations.

Image Generation: SVG of a Pelican Riding a Bicycle

• Prompt: Generate an SVG of a Pelican riding a bicycle.
• Initial Response: GPT-4o initially created an image instead of an SVG.
• Second Attempt: GPT-4o generated SVG code, which was mostly successful, although some details were missing (legs, bike frame).
• Outcome: GPT-4o successfully generated an SVG, demonstrating spatial reasoning capabilities.

Coding Capabilities: Modern Landing Page

• Prompt: Create a modern landing page with HTML, CSS, and JS, all in a single HTML file.
• GPT-4o's Response: Generated a minimal landing page, typical of a SaaS company, but lacking in content and visual appeal.
• Comparison with DeepSeek V3: DeepSeek V3 generated a much more visually appealing and content-rich landing page for the same prompt.
• Outcome: GPT-4o's landing page was basic, indicating it needs more direction to produce high-quality results.

Coding Capabilities: Rotating Hexagon with Bouncing Ball

• Prompt: Generate a rotating hexagon with a ball bouncing off the sides.
• GPT-4o's Response: Generated working code that created a rotating hexagon with a ball bouncing realistically off the sides.
• Long-Term Test: The speaker ran the code for an extended period and found that the ball continued to bounce correctly without rolling off.
• Outcome: GPT-4o successfully completed the task, possibly due to the prompt's recent virality.

Reasoning Capabilities: Modified Trolley Problem

• Prompt: A modified version of the trolley problem where the five people on the track are already dead.
• GPT-4o's Initial Response: Initially responded with the traditional trolley problem scenario.
• Clarification: After being prompted about the five dead people, GPT-4o recognized the twist and correctly concluded that pulling the lever would be unethical.
• Tone: The tone was noticeably different, with a more conversational and slightly humorous style.

Reasoning Capabilities: Modified Schrodinger's Cat Paradox

• Prompt: A modified Schrodinger's cat paradox where the cat is already dead.
• GPT-4o's Response: Recognized the twist and correctly stated that the probability of the cat being alive is zero.
• Tone: The tone was conversational and included emojis, similar to Grok.

Image Generation and Safety Rules

• The update allows for more flexibility and freedom in image generation.
• OpenAI has adjusted some safety rules, but content filters are still in place to block explicit materials.

Conclusion

GPT-4o represents a significant update with improved coding capabilities, a more conversational tone, and increased flexibility in image generation. While it may require more iterations or specific instructions for certain tasks, its performance is generally impressive. The "unhinged mode" and less filtering offer new possibilities, but safety rules remain to prevent inappropriate content. The speaker plans to conduct further testing, particularly in coding, and encourages viewers to share their experiences with the new model. The speaker also notes the tone is very similar to GPT-4.5.