Stepfunction Podcast
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 [https://www.theverge.com/2023/9/25/23886699/chatgpt-pictures-voice-commands-ai-chatbot-openai] for ChatGPT. * DALL-E 3 update [https://techcrunch.com/2023/09/20/openai-unveils-dall-e-3-allows-artists-to-opt-out-of-training/]. * Google Converse [https://www.gadgetsnow.com/featured/how-googles-converse-generative-ai-is-different-from-google-search/articleshow/103252562.cms] aka Google SGE [https://blog.google/products/search/generative-ai-search/] is still better than DALL-E. * Midjourney [https://zapier.com/blog/how-to-use-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 [https://blogs.nvidia.com/blog/2019/03/18/gaugan-photorealistic-landscapes-nvidia-research/] by NVIDIA. * Associated paper at arXiv [https://arxiv.org/abs/1903.07291] and code at GitHub [https://github.com/NVlabs/SPADE]. * From Jeff's workshop: * Definitions for the G,P, and T in "ChatGPT" * Generative (as in generative AI--see this entire podcast 😉). * Pre-trained [https://stats.stackexchange.com/questions/193082/what-is-pre-training-a-neural-network]. * Transformer [https://en.wikipedia.org/wiki/Transformer_(machine_learning_model)]. * Meta/FB's Llama2 [https://ai.meta.com/llama/] (7 Billion parameters). * Fine-Tuning–one of part of many methods to optimize a base model. See charts in this NVIDIA article [https://developer.nvidia.com/blog/selecting-large-language-model-customization-techniques/]. * Low-Rank Adaptation: * Conceptual article about LoRA [https://huggingface.co/docs/peft/conceptual_guides/lora] at HuggingFace. * Original LoRA 2021 paper [https://arxiv.org/abs/2106.09685]. * May 2023 QLoRA paper [https://arxiv.org/abs/2305.14314]. * August 2023 LoRA-FA paper [https://arxiv.org/abs/2308.03303]. * Short Wikipedia description of LoRA [https://Wiki%20article%20--%20https//en.wikipedia.org/wiki/Fine-tuning_(deep_learning)#Low-rank_adaption]. * 2019 programmer joke [https://www.reddit.com/r/ProgrammerHumor/comments/bgrlu4/stackoverflow_in_a_nutshell/] about using Google and StackOverflow [https://en.wikipedia.org/wiki/Stack_Overflow]. Send questions/comments to stepfunctionpod@gmail.com and find us on the web at www.stepfunction.org
19 afleveringen
Reacties
0Wees de eerste die een reactie plaatst
Meld je nu aan en word lid van de Stepfunction Podcast community!