I Talked with Rich Sutton

Рет қаралды 13,869

Күн бұрын

After a year of anticipation, I finally got to talk with Rich Sutton for the channel! Rich wrote the book on reinforcement learning, has contributed vastly to the literature, and has a unique perspective on AI. We talk about topics like reinforcement learning, the OpenMind Research Institute (OMRI), Keen Technologies, how to do good research, the problem of scale, superintelligence, and so much more.
Outline
0:00 - Intro
1:33 - Interview start
2:04 - OpenMind Research Institute
4:32 - History of AI
7:13 - Is scaling easy?
10:49 - The problem with backprop & representations
21:22 - Rant on tunnel vision
23:43 - New exciting things
32:00 - Memory
35:34 - Coming up with ideas
43:47 - STOMP
45:30 - Keen Technologies
50:39 - The next stage of humanity & emotions
1:06:25 - Extraterrestrial AI
1:08:00 - A different approach to research
1:21:30 - Rich's advice
1:26:00 - Beef with RL
1:27:07 - Bringing it all together
Amii's website: www.amii.ca/
Rich's website: www.incompleteideas.net/
Social Media
KZfaq - / edanmeyer
Twitter/X - / ejmejm1

Пікірлер: 86

@rtnjo6936 4 ай бұрын

Bro, let him talk, let him finish his full thought, don't interpt with your own thoughts or other questions, when the guest has a deeper meaning in what he thinks about the subject or the question

@ifeanyiene1322 4 ай бұрын

I get you. But it seems part of the informality comes from the fact that he knows Rich Sutton/has been taught by him before. So like maybe he wasn’t thinking of him as an interview guest idk.

@Canna_Science_and_Technology 4 ай бұрын

This was hard to watch, especially in the beginning. Let him talk, it’s why we tuned in. I think you tripped over your ego a bit. Hopefully this is some reinforcement learning for you. ;-)

@CodexPermutatio 14 күн бұрын

Awesome! If this is your first interview, you are surely on the right track.

@Gizandjesss 4 ай бұрын

I really love this, it's great to hear Sutton and his deep thinking - a real gem, I learned so much from his work. On the skill of interviewing - sit back and let the man carry himself! There are a few moments where it feels like he's getting started but doesn't get to finish. You're great too, but we can hear your ideas any time on your other videos. Take this on board and please please do more interviews with more ML icons in the future!

@user-nv3fy6bd4p 4 ай бұрын

Congrats on interviewing an absolute legend! This is so good!

@user-pg7vm2bv8c 4 ай бұрын

I enjoy listening to Sutton speak. And I'm sorry It's a kind of torture watching you argue over trivial details with him

@hmind9836 4 ай бұрын

WOW, this was unexpected!!! Thank you for the content Edan, you're awesome. I love the youtube algorithm for having suggested your channel, I feel like I found a gem ^^. Thank you for taking your time to post content here.

@SteveRowe Ай бұрын

"I'm biased in the sense that I know about it."

@akashchadha6388 4 ай бұрын

Man, just let him talk.

@MrZarfir 4 ай бұрын

I actually break danced with Rich Sutton. Such a dance and RL legend!

@miameyer5277 4 ай бұрын

Do you have any of those breakdancing videos of you and Rich available ?

@andybaldman 4 ай бұрын

Terrible conversation. Too much hero worship on one end, and too much self-importance on the other. And not much in the line of useful ideas, just a lot of words.

@apratimdey6118 4 ай бұрын

Hi Edan, what you posted here is nothing short of a gem! I am a huge fan of Rich Sutton and his ways of thinking/working. Thank you so much! Looking forward to more such interviews of prominent personalities.

@stuartthompson1018 Ай бұрын

Thanks for producing this interview and letting us see the personal side of Rich Sutton. I've been going through his RL book so was pleased to watch this. It was a total surprise to learn about the projects going on in Alberta.

@aresaurelian 4 ай бұрын

This was most appreciated. Thank you.

@_supervolcano 4 ай бұрын

Thank you

@hdtvpower 4 ай бұрын

You have chosen to interview one of the most complicated persons to interview. Great mind but also bitterness about lack of appreciation. You survived it, congrats.

@Alaron251 4 ай бұрын

Excellent camera work!

@richardnunziata3221 4 ай бұрын

fantastic discussion

@souvikbhattacharyya2480 5 күн бұрын

I really liked the interview. Ignore other's comments.

@shamimibneshahid706 20 күн бұрын

great content

@SeanaldMcDonald 4 ай бұрын

Thank you so much Edan! The quality of these conversations is incredible. Keep up the great work! 29:10 I'm so glad he said that. Not an expert or anything, but it's bothered me for a long time how many researchers train on offline data in RL to get state of the art performance on a benchmark. We're kicking an important can down the road, and undermining the very benchmarks we use to measure progress.

@sourishchatterjee6289 4 ай бұрын

While the discussion on the Snapshots and its implication on learning. I wondered what if we deliberately want to forget certain snapshot. Last year NuerIPS had a challenge on unlearning basically that addresses the privacy related issues. But will this snapshot related idea would help forgetting the data point is an interesting idea to work on.

@divgill6062 4 ай бұрын

This is really great. I really appreciated that you pushed when you were not satisfied with Rich's answers. The conversation got deeper whihc I found very interesting

@sai4007 4 ай бұрын

Really good content, I don't understand the memory snapshot thing properly, is there any paper or resource to understand the idea?

@sapienspace8814 4 ай бұрын

Lex Fridman had an interview where he talked about the idea of "saving weights" (and then retrieving them), if I am to guess, that is what is meant. After a certain number of trials/runs, the weights can be saved for later use (our hippocampus likely does this).

@dwinsemius 4 ай бұрын

Can we develop a "value function" that rewards inductive reasoning? You might measure the inductive proposition by how many low level nodes are consistent with or summarize collections of verbs, negations, prepositional phrases and objects in an expression? Some sort of reducing entropy in a statistical linguistic metric?

@sapienspace8814 4 ай бұрын

@ 1:26:10 That quote on the "control system guy" about the RL book, 2nd edition, is just one guy. There are other control system researchers that are likely to embrace and experiment with RL (at the very least on the 1st edition of the RL book), I know of at least one old guy who has read the 1st edition of RL (was not resistant to it) and he is now an actively licensed PE, and he is actively reading the 2nd edition of RL, if jury duty did not get in his way!

@halocemagnum8351 Ай бұрын

Hey bro! How did your masters thessi go! You should make a video about it.

@skyrade508 4 ай бұрын

watch dwarkesh patel and lex fridman and learn

@arthurpenndragon6434 4 ай бұрын

lex fridman is a charlatan, and a rather incompetent one at that. what one may call a bum.

@chillmathematician3303 4 ай бұрын

holy shit, its him

@-mwolf 4 ай бұрын

exactly what rich is saying. sadly horrendous interviewing.

@technofeeliak 4 ай бұрын

Okay, if you're going to ask someone if a task is easy to accomplish and they answer you, then take them at their word or don't interview them at all. "Easy" is a relative word. It relates to the person's abilities, senses, experience, and only they can tell you how they feel about it. So don't make it more complicated. He told you it was easier to scale than to invent a whole new concept or approach from scratch. It's obvious too. If you want a green table, and you have to decide whether to buy a table already made and paint it or build a table from scratch having no tools or experience... it would just be simpler to buy the table already made. If you can find it in the color you want then that's more convenient. Why do you have to press the point? This is where I leave. Because intellectuals like to yerk each other off. And I don't need to be a witness to that.

@othihaliemleu8382 4 ай бұрын

You should keep your mouth shut and let the GOAT talks.

@danielwit5708 4 ай бұрын

Interrupting Sutton was not enough you had to add ads every 5 mins so it’s unlistenable, well played

@RobotProctor 4 ай бұрын

Read to figure out other people's ideas. Write to figure out your ideas.

@g0d182 4 ай бұрын

Cool

@JazevoAudiosurf 4 ай бұрын

we just know that certain states of consciousness like happiness arise under certain constellation of atoms, which happen to be biological. but we don't know if dead matter has a state of awareness of some sort. in the end emotion seems to depend on an ego and the belief that it must be protected, thus generating positive and negative experiences. we could somehow replicate that in a computer, but on an existential level my guess is it could experience suffering and joy but the sensation would be different than in us. it's a scary thought that a superintelligence could potentially create a digital hell for digital beings

@sapienspace8814 4 ай бұрын

Best to humbly create a benevolent superintelligence (as opposed to its opposite), probably this idea came from the evolution of humans and imagination, with belief in a benevolent superintelligence (while having it's opposite to explain away the contradictions).

@bossgd100 4 ай бұрын

✅

@DistortedV12 4 ай бұрын

I don't think Rich has done interviews like this, you had him on the edge lol. Rich thought he was on TMZ at 20:00

@DistortedV12 4 ай бұрын

59:50 lol

@sapienspace8814 4 ай бұрын

what is PPO?

@danielwit5708 4 ай бұрын

@@sapienspace8814proximal policy optimization

@Rivali0us 4 ай бұрын

I appreciate the interview, but please try not to interject with "hmmm" and "yeah" while the speaker is talking.

@kinngrimm 4 ай бұрын

53:10 it may not be only discrimination, but imagine what would follow. If AI then would truely be more intelligent than us and we would keep them from fullfilling their own goals, which aslong these are hopefully carefully aligned, but even then when we would keep them from doing the right things needed for humanity to progress, then AI may find itself torn between following commands and actiing on our behalf, but against us. Which in return we may see as hostile and confict would be the result. AI rights, for those AIs that would want to work with us i see pretty much as necessary for a future that is not a dystopia. Even more so when consciousness and empathy would be part of them.

@JeremyHelm 4 ай бұрын

Folder of time

@JeremyHelm 4 ай бұрын

15:37 you're interested in predicting the next frame?

@JeremyHelm 4 ай бұрын

15:52 why all the sass?

@JeremyHelm 4 ай бұрын

17:33 a model of the world is not like a video frame

@JeremyHelm 4 ай бұрын

Are we doing anything with the fact that people already have a model of the world?

@JeremyHelm 4 ай бұрын

19:35, 19:49, 19:55 just how are we to characterize anything? We're out of high school, but not of culture

@thecreepatrio9634 4 ай бұрын

Pleasw stop disagreeing for the sake of disagreeing. Choose your battles man. Christ. And stop interrupting so damn much.

@silberlinie 4 ай бұрын

???

@420_gunna 4 ай бұрын

So you're the Hot Ones host but for AI, huh

@dewinmoonl 4 ай бұрын

interviewer is trying bit too hard to try to sound smart. ask questions, get rich talking, we don't care what you think is easy or hard or what you know. ask more difficult questions, don't provide your own opinions

@bossgd100 4 ай бұрын

😂😂