Last Week in AI #225 - GPT 5.1, Kimi K2 Thinking, Remote Labor Index
473 snips
Nov 21, 2025 
Michelle Lee, founder and CEO of a company creating physical AI scientists, joins to dive into the latest in AI tech. They discuss the arrival of GPT-5.1, which boasts a warmer personality, and Baidu's ambitious Ernie 5.0 and Apollo Go developments. The duo explores the boom in AI-generated music that is topping charts and the launch of voice marketplaces for iconic voices. They also highlight advancements in remote labor and implications for automation in the workforce, plus a peek at Kimi K2 Thinking's innovative tool-use model.
AI Snips
Chapters
Transcript
Episode notes
Real Freelance Work Still Hard To Automate
- The Remote Labor Index tests AI on real freelance tasks and finds only ~2.5% automation rate today.
- Andrey and Michelle stress long-horizon context and teamwork limit current agents' ability to automate remote work.
Sparse Training For Mechanistic Interpretability
- OpenAI paper explores training sparse transformers to make circuits more interpretable than post-hoc methods.
- Michelle notes sparse training is inefficient and currently limited to smaller models.
Linear Attention Challenges Full Attention
- Kimi Linear proposes a linear-attention architecture that claims better efficiency and competitive quality.
- Andrey says this could be a step beyond standard transformers if results scale to very large models.