Hobson LaneJul 25, 2024

The key takeaways from this week's Sam Charrington #TWIML #AI brain dump: "Building Real-World LLM Products ...with Hamel Husain" (episode #694,
starting at 00:37:24) https://chrt.fm/track/4D4ED/traffic.megaphone.fm/MLN5252789067.mp3?updated=1721769728#t=2244

1. Axolotl
2. instrument your #LLM UI
3. Curate your datasets

TL;DR do good data science
1/2

Hobson LaneJan 27, 2024

This week's #TWIML #AI podcast is a powerful brain dump on #DeepLearning and #LLMs -- some really promising trends are explained without hype. I'm looking forward to a second listen and review of the papers #ThomasDietterich (Oregon State) and Sam Harrington talk about. They're #brilliant together and I want to explore every hint of an idea that came out of their #brains in this convo.

Episode webpage: https://twimlai.com/podcast/twimlai/ai-trends-2024-machine-learning-deep-learning/

Media file: https://chrt.fm/track/4D4ED/traffic.megaphone.fm/MLN6485051406.mp3?updated=1704731315

AI Trends 2024: Machine Learning & Deep Learning with Thomas G. Dietterich | The TWIML AI Podcast

TWIML
Show threadBen WaberOct 14, 2023
First was a great conversation with Luke Zettlemoyer on scaling multi-modal generative AI on the #TWIML podcast. Zettlemoyer discussed some impressive advances on using transformer models for images, the importance of data and the limits of current scaling approaches, and why openness is essential for progress to continue in AI. Highly recommend https://www.youtube.com/watch?v=ZvfLYcG_8io&t=4s (2/8) #AI #GenerativeAI
Scaling Multi-Modal Generative AI with Luke Zettlemoyer - 650

YouTube
Show threadBen WaberOct 4, 2023
Next was a fabulous talk by @alex on #AI hype on the #TWIML podcast. Hanna reviews what drives AI hype, where AI can be useful, and DAIR's broader research agenda. Highly recommend https://www.youtube.com/watch?v=I4bdtG_SagE (4/9)
Pushing Back on AI Hype with Alex Hanna - 649

YouTube
Show threadBen WaberSep 5, 2023
Next was a fantastic conversation with James Zou looking at changes in ChatGPT's performance over the last few months on the #TWIML podcast. Zou convincingly demonstrates that #ChatGPT has gotten decisively worse on a number of metrics recently, and highlights some of the dangers of relying on #LLMs as stable arbiters of high-quality output. Highly recommend https://www.youtube.com/watch?v=2HXL89bqxx0 (3/7) #AI
Is ChatGPT Getting Worse? with James Zou - 645

YouTube
Show threadBen WaberAug 30, 2023
Next was an engaging conversation with Sophia Sanborn on neural networks, Fourier transform properties, and more on the #TWIML podcast https://www.youtube.com/watch?v=fK0kVYb32DA&t=4s (3/9) #AI
Why Deep Networks and Brains Learn Similar Features with Sophia Sanborn - 644

YouTube
Show threadBen WaberAug 23, 2023
First was a thought-provoking conversation with @gokul on inverse reinforcement learning without RL on the #TWIML podcast. I spent a lot of time after this thinking about how these techniques could be applied to LLM training as a step before RLHF to significantly reduce toxic content exposure, and I'm very excited to see where this work goes. Highly recommend https://www.youtube.com/watch?v=SPZLJwDNwpM (2/7) #AI
Inverse Reinforcement Learning Without RL with Gokul Swamy - 643

YouTube
Show threadBen WaberJul 12, 2023
Next was a fantastic conversation with Alice Xiang on #privacy vs. fairness in #ComputerVision on the #TWIML podcast. Xiang not only sets the table with a great summary of the historical arc of #AI #ethics, but explicitly examines the tension between individual privacy and algorithmic harms that is often ignored. Highly recommend https://www.youtube.com/watch?v=Ylu416ksuG8 (4/5) #AIEthics
Privacy vs Fairness in Computer Vision with Alice Xiang - 637

YouTube
Show threadBen WaberJul 5, 2023
Next was a great conversation with Mohit Bansal on unification in large #AI models on the #TWIML podcast. Bansal has a fabulous perspective on the state and trajectory of the field and discusses some impressive models - I was partial to applying multi-modal large models to spectrograms. Highly recommend https://www.youtube.com/watch?v=vVlRWjdwp4s (6/10) #MachineLearning #LLMs #GenerativeAI
Unifying Vision and Language Models with Mohit Bansal - 636

YouTube
Show threadBen WaberMay 31, 2023
Next was a rich conversation with Hugo Larochelle on transfer learning on the #TWIML podcast. As companies and researchers try to apply general models to applications that they weren't originally designed for, understanding how to retrofit these models so they can perform well without extensive retraining is essential, and examined in depth here https://www.youtube.com/watch?v=c--acLK_C9s (4/7)
Towards Improved Transfer Learning with Hugo Larochelle - 631

YouTube