Seymour and Jeff discuss the recently announced updates from OpenAI, especially regarding image generation in GPT-4 and DALL-E 3. Our ranking of image generation AI's from best to worst: (1) Midjourney, (2) Google Search Generative Experience (SGE), and finally (3) DALL-E. 

Jeff closes by talking about the recent LLM workshop he conducted for junior high and middle school students.

Links:

• OpenAI announces new voice chat and image features for ChatGPT. 
• DALL-E 3 update.
• Google Converse aka Google SGE is still better than DALL-E.
• Midjourney is still the best.
• Regarding earlier deep learning methods of translating sketches into finished drawings, Jeff was thinking of NVIDIA's GauGAN, based on SPatially-Adaptive DEnormalization (SPADE).
  • 2019 blog post by NVIDIA.
  • Associated paper at arXiv and code at GitHub.
• From Jeff's workshop:
  • Definitions for the G,P, and T in "ChatGPT"
    • Generative (as in generative AI--see this entire podcast 😉).
    • Pre-trained.
    • Transformer.
  • Meta/FB's Llama2 (7 Billion parameters).
  • Fine-Tuning–one of part of many methods to optimize a base model. See charts in this NVIDIA article.
  • Low-Rank Adaptation:
    • Conceptual article about LoRA at HuggingFace.
    • Original LoRA 2021 paper.
    • May 2023 QLoRA paper.
    • August 2023 LoRA-FA paper.
    • Short Wikipedia description of LoRA.
• 2019 programmer joke about using Google and StackOverflow.

Send questions/comments to [email protected] and find us on the web at www.stepfunction.org