S3 E18 John Schulman of OpenAI on ChatGPT: invention, capabilities and limitations
Ғылым және технология
S3 E18 John Schulman (co-founder OpenAI) on ChatGPT: invention, capabilities and limitations
What's in this episode:
00:00:00 John Schulman
00:01:14 sponsors: Index Ventures and Weights and Biases
00:02:17 what is ChatGPT
00:03:17 how is ChatGPT built
00:05:29 generating text better than any human
00:06:42 tasks which bots don't perform well
00:07:29 maybe this is just a beginning
00:08:36 bringing in Reinforcement Learning (RL)
00:10:48 uses of ChatGPT that are surprising
00:12:32 ChatGPT is good at programming
00:13:30 why hallucinations happen and how to avoid them
00:16:06 training ChatGPT to be aware of hallucinations
00:18:42 retrieval vs. training
00:21:34 single pass training
00:22:50 future in scaling up
00:23:49 closed vs open source
00:27:03 need for another approach to push the development in future
00:30:06 does fine-tuning decrease generalization ability
00:32:42 LLM domination and maybe something more on the horizon
00:37:01 John's curious career path
00:40:47 transition from imitation learning
00:43:10 potential in Academia with a limited budget
00:45:40 interesting problems for Ph.D. students
00:47:54 is there any future in tiny data sets
00:50:24 academia or industry as a Ph.D. student today
00:51:57 being John Schulman
00:55:49 things to do to relax
Links:
ChatGPT: chat.openai.com
John Schulman: joschu.net/
Sponsor Links:
www.indexventures.com/
wandb.com
This episode on other platforms:
Apple: bit.ly/3YhMvP5
Spotify: bit.ly/3Ks0bBn
Amazon: bit.ly/3OkR4Dx
Google: bit.ly/3q7igOg
Acast: bit.ly/3Qox9pO
Host: Pieter Abbeel
Production: Bo Obradovic
Пікірлер: 11
That's been one of the most inspiring interviews about this topic (ChatGPT and RL) I've seen for a long time. Thanks to you both!
Incredible inside view into the model details!!!
Wonderful. Thank you!
Great talk, thank you for sharing! That is interesting how he did some studying on cart-pole problems (Barto, Anderson, and Sutton have an interesting, though old, IEEE paper on that). I hope to learn more about Proximal Policy Optimization and scaling. A connection between language and mathematics is Fuzzy Logic, and when combined with RL can discover it's own inference rules about it's interaction with the world. It is sort of like a "Lego" building block of intelligence. It is interesting to me that the "Transformer: Attention is all you need" paper was not discussed. I recall Ilya Sutskever saying that RL might replace the transformer...
✨ thanks to you both. Enjoyed listening to you John. Keep up the rock climbing 🧗🏻♀️ and like that you are raising chickens in your backyard.
Hey Pieter, Do you believe that Deep RL is going to get us to AGI or at least something revolutionary in robotics?
interacting with raw pixels.. feeding raw pixels sounds like tesla vision for gpt video.. Andrej Karpathy sounds like a good fit for open ai future ;p
the pace of the answers are sometimes too slow...at times it is too boring and tiring to watch...a good idea is a good idea, when it goes on forever it loses a lot of its value..
Even at 2.75x he is too tiresome and annying to listen to :( So many pauses, uhm, erm, hm... Also why do so many ppl who clearly could afford a good mic give no fecks and use those horrible apple pod? ...
Erm Uh Erm Uh ... Really tiresome to listen to him ://