S3 E2 Stanford Prof Chelsea Finn: How to build AI that can keep up with an always changing world
Ғылым және технология
S3 E2 Stanford Prof Chelsea Finn: How to build AI that can keep up with an always changing world (Host: Pieter Abbeel)
What's in this episode:
00:00:00 - Chelsea
00:01:40 - sponsors: Index Ventures, Weights and Biases
00:02:51 - what can AI not (yet) do today? (Spoiler Alert: Distribution Shift)
00:05:55 - handling spurious features
00:07:55 - targeted LLM editing
00:12:30 - LLMs that read the news
00:13:46 - tackling distribution shift in robotics
00:20:35 - when is generalization beyond the training distribution possible
00:23:50 - single life reinforcement learning (RL)
00:27:10 - rewards in RL
00:34:37 - should we just be training our RL agents longer
00:37:45 - robotics data-sets, data collection at home
00:40:55 - meta-learning, MAML
00:50:12 - what will happen in the next 5-10 years in AI
00:52:20 - Chelsea’s growing up and career trajectory
00:58:01 - productivity, working exciting things, swimming, sleep
01:01:27 - what looking for in students
01:04:01 - advice for high school students
Guest Links:
ai.stanford.edu/~cbfinn/
Sponsor Links:
www.indexventures.com/
wandb.ai/site
Subscribe and listen to our podcast today:
Apple: bit.ly/42yvIsw
Spotify: bit.ly/407FXCK
Amazon: bit.ly/3lvPaFJ
Google: bit.ly/3JEbHIm
Acast: bit.ly/408HFUe
Host: Pieter Abbeel
Production: Bo Obradovic
Пікірлер: 9
Great Pieter and Chelsea; superb interview. BUT, from two ace roboticists I expected to hear more about limitations of current actuator and sensor suites (more real-time stuff) and the pace of rollout of new devices. (And, in re: meta ... the difficulties of combining robot DBs across different device suites.) But... great show! Bob Blum, MD, PhD Stanford (doing ML there in 1976!!)
Great interview!
Amazing! I really appreciate when you try to get into more details and not just talk abstractly. Thank you for this. I'm looking forward for more 😊
A good interview.
Great interview.
Could you make the youtube automatically-generated subtitles work please?
Simple ML has this stuff all worked out... It need to be generalized to DNN
"When a robot is trying to learn something, they'll be waving their arms around in the air randomly, or be doing things that are just very strange, that are very different from what humans would ever do." - 33:42. Sounds like a baby who is still ... untrained? Humans at this stage (as I understand it) do best with a mixture of supervised and unsupervised learning.
@BillBakerB
Жыл бұрын
Oh lol then Pieter immediately brings up his 3-month-old, and they discuss that, and how to share data among ML projects, and how hard that is. "Take robots home and collect data in something close to 1000 different environments."