#71: Alex O'Connor — Transformers, Generative AI, and the Deep Learning Revolution
Episode description
Alex O’Connor—researcher and ML manager—on the latest trends of generative AI. Language and image models, prompt engineering, the latent space, fine-tuning, tokenization, textual inversion, adversarial attacks, and more.
Alex O’Connor got his PhD in Computer Science from Trinity College, Dublin. He was a postdoctoral researcher and funded investigator for the ADAPT Centre for digital content, at both TCD and later DCU. In 2017, he joined Pivotus, a Fintech startup, as Director of Research. Alex has been Sr Manager for Data Science & Machine Learning at Autodesk for the past few years, leading a team that delivers machine learning for e-commerce, including personalization and natural language processing.
Favorite quotes
- “None of these models can read.”
- “Art in the future may not be good, but it will be prompt.” Mastodon
- Machine Learning Systems Design by Chip Huyen
- Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow by Aurélien Géron
- The Illustrated Transformer by Jay Alammar
- Attention Is All You Need by Google Brain
- Transformers: a Primer by Justin Seonyong Lee
- Alex in Mastodon ★
- Training Dream Booth Multimodal Art on HuggingFace by @akhaliq
- NeurIPS
- arxiv.org: Where most papers get published
- Nono’s Discord
- Suggestive Drawing: Nono’s master’s thesis
- Crungus is a fictional character from Stable Diffusion’s latent space
- Stable Diffusion
- Arcane Style Stable Diffusion fine-tuned model ★
- Imagen
- DALL-E
- CLIP
- GPT and ChatGPT
- BERT, ALBERT & RoBERTa
- Bloom
- word2vec
- Mupert.ai and Google’s MusicLM
- t-SNE and UMAP: Dimensionality reduction techniques
- char-rnn
- High-performance computing (HPC)
- Transformers and Attention
- Sequence transformers
- Quadratic growth
- Super resolution
- Recurrent neural networks (RNNs)
- Long short-term memory networks (LSTMs)
- Gated recurrent units (GRUs)
- Bayesian classifiers
- Machine translation
- Encoder-decoder
- Gradio
- Tokenization ★
- Embeddings ★
- Latent space
- The distributional hypothesis
- Textual inversion ★
- Pretrained models
- Zero-shot learning
- Mercator projection
- 00:00 · Introduction
- 00:40 · Machine learning
- 02:36 · Spam and scams
- 15:57 · Adversarial attacks
- 20:50 · Deep learning revolution
- 23:06 · Transformers
- 31:23 · Language models
- 37:09 · Zero-shot learning
- 42:16 · Prompt engineering
- 43:45 · Training costs and hardware
- 47:56 · Open contributions
- 51:26 · BERT and Stable Diffusion
- 54:42 · Tokenization
- 59:36 · Latent space
- 01:05:33 · Ethics
- 01:10:39 · Fine-tuning and pretrained models
- 01:18:43 · Textual inversion
- 01:22:46 · Dimensionality reduction
- 01:25:21 · Mission
- 01:27:34 · Advice for beginners
- 01:30:15 · Books and papers
- 01:34:17 · The lab notebook
- 01:44:57 · Thanks
I'd love to hear from you.
Submit a question about this or any previous episodes.
Join the Discord community. Meet other curious minds.
If you enjoy the show, would you please consider leaving a short review on Apple Podcasts/iTunes? It takes less than 60 seconds and really helps.
Show notes, transcripts, and past episodes at gettingsimple.com/podcast.
Thanks to Andrea Villalón Paredes for editing this interview.
Sleep and A Loop to Kill For songs by Steve Combs under CC BY 4.0.
Follow Nono