S3 Ep 10 MacArthur Genius Prof Yejin Choi: Teaching AI Common Sense and Morality

Подписаться 14 тыс.

Просмотров 3,2 тыс.

50% 1

S3 E10 MacArthur Genius Prof Yejin Choi: Teaching AI Common Sense and Morality (Host: Pieter Abbeel)
What's in this episode:
00:00:00 Yejin Choi
00:01:13 sponsors: Index Ventures and Weights and Biases
00:02:21 limitations of ChatGPT / Large Language Models
00:05:32 ChatGPT's training and abilities
00:10:04 truth seeking pre-training
00:12:05 discovering mistakes of LLMs
00:14:15 is scaling up all that’s needed
00:17:05 the low compute regime
00:18:48 academia vs industry
00:21:42 inference time algorithms
00:27:15 power of high quality data
00:29:08 common sense
00:32:19 language as the best medium for reasoning
00:35:40 can machines learn morality
00:40:09 any benefits to smaller models
00:41:33 syntax, semantics, pragmatics
00:45:51 real-world applications
00:50:55 growing up and road to the AI profession
00:58:39 ways to relax
Guest links:
/ yejinchoinka
homes.cs.washington.edu/~yejin/
www.ted.com/talks/yejin_choi_...
Sponsor Links:
www.indexventures.com/
wandb.com
Our podcast on other platforms:
Apple: bit.ly/3Wdry6M
Spotify: bit.ly/437ifY7
Amazon: bit.ly/459zyty
Google: bit.ly/439InSx
Acast: bit.ly/3Ok6euf
Host: Pieter Abbeel -- www.cs.berkeley.edu/~pabbeel
Production: Bo Obradovic

Наука

Опубликовано:

20 июл 2024

Ссылка:

Скачать:

Готовим ссылку...

Добавить в:

Мой плейлист

Посмотреть позже

Комментарии : 15

@duudleDreamz Год назад

GPT4 now correctly answers the "5 pieces of clothing to dry in the sun" question. It even explains that drying clothes in the sun is not a sequential task but a parallel task. Great video, but already outdated regarding this example from Yejin.

@PieterAbbeel Год назад

Thanks for checking and reporting, everything is moving so fast, it's amazing

@tylerparks5656 Год назад

Thank you for these wonderful discussions.

@One00042 29 дней назад

👍🏿

@ChrisStewart2 Год назад

It looks to me like mll's build a language model based on general data (the web in general) and then require specific training to learn facts.

@tostane Год назад

I have been trying to teach "The Bard" not to say I when it refers to itself. Also trying to get it to say The Bard.

@MarcoMugnatto 11 месяцев назад

Human intelligence is never about annnn... hmm... eehhh... predicting the next word...

@netscrooge Год назад

Cognitive psychologist Daniel Kahneman calls "System 1" thinking fast, automatic, and often based on heuristics or 'rules of thumb.' This is in contrast with "System 2" thinking, which is slower, more deliberate, and more logically rigorous. GPT-4 was designed to mimic aspects of human "System 1" thinking: generating responses quickly in a single pass, without a System 2 error correction stage. When I asked GPT-4 the clothesline question, it got it wrong. When I simply asked it to double check its answer, it immediately found its mistake. So go ahead and laugh at GPT-4 for being less intelligent than a young child, but you're laughing at it for being exactly what we designed it to be. Yes, we can create systems with even better System 1 thinking. But no matter how good those get, we will probably be able to improve them greatly by adding a System 2 layer on top.

@netscrooge Год назад

Update: I just found the paper "Tree of Thoughts: Deliberate Problem Solving with Large Language Models" that came out a few days ago. Very relevant.

@ChrisStewart2 Год назад

The RL training is happening very fast now that millions of people are voluntarily training it. This has actually been happening for several years where OpenAI has been using large groups of beta testers and is a big reason why GPT4 can answer a much wider range of questions.