Notes on GPT-4 Presentation

Introduction to GPT-4

• Arrival of GPT-4: Claims to surpass ChatGPT.
• Capabilities: Can convert drawings into functional websites and explain jokes from images.
• OpenAI's Live Stream: Introduced this multimodal AI with demonstrations.

Key Features of GPT-4

Multimodal Abilities

• Unlike previous versions (text-based), GPT-4 can process both images and text.
• Examples of Image Processing:
  • Explains jokes from a sequence of images (e.g., iPhone with VGA cable).
  • Can convert a napkin drawing of a website into functional code (HTML, CSS, JavaScript).

Performance and Integration

• Speed and Efficiency: Can produce a functional website in 10-20 seconds.
• Real-world Applications: Already integrated into companies like Khan Academy.
• Educational Assistance: Functions as a personal tutor with customization.

Statistical Performance

• GPT-4 outperforms previous models (GPT-3, GPT-3.5) in various evaluations:
  • Passed the LSAT and Bar exams in the top quarter percentile.
  • Previous models were in the lower quarter.
• Handles over 25,000 words of text, allowing for more complex tasks.

Creative and Reasoning Enhancements

• Creativity: Better at modifying and iterating technical and writing tasks.
• Example Task: Summarizing Cinderella with each sentence beginning with the next letter of the alphabet (A-Z).
• Advanced Reasoning: Better capable of scheduling appointments based on different calendars.

Safety and Accuracy Improvements

• Error Reduction: 82% less likely to generate disallowed content; 40% less likely to produce factually inaccurate responses.

Access to GPT-4

• Available through ChatGPT Plus (paid version).
• API access requires joining a waitlist.
• Differences in Model Performance:
  • GPT-3.5: Average reasoning, low conciseness, high speed.
  • GPT-4: High reasoning, high conciseness, slightly lower speed due to limitations.

Conclusion and Personal Experience

• User Feedback: GPT-4 showed improved comprehension, reasoning, and understanding over GPT-3.
• Attempted to trick GPT-4 with false math but it consistently gave correct answers.
• Awaiting API access for further business applications and comparisons with GPT-3.5.
• Anticipation about the future of GPT-3.5 with the introduction of GPT-4.