Demis Hassabis - Scaling, Superhuman AIs, AlphaZero atop LLMs, Rogue Nations Threat
ฝัง
- เผยแพร่เมื่อ 29 เม.ย. 2024
- Here is my episode with Demis Hassabis, CEO of Google DeepMind
We discuss:
- Why scaling is an artform
- Adding search, planning, & AlphaZero type training atop LLMs
- Making sure rogue nations can't steal weights
- The right way to align superhuman AIs and do an intelligence explosion
Transcript: www.dwarkeshpatel.com/p/demis...
Apple Podcasts: podcasts.apple.com/us/podcast...
Spotify: open.spotify.com/episode/6SWb...
Follow me on Twitter: / dwarkesh_sp
Timestamps
0:00:00 Nature of intelligence
0:05:56 RL atop LLMs
0:16:31 Scaling and alignment
0:24:13 Timelines and intelligence explosion
0:28:42 Gemini training
0:35:30 Governance of superhuman AIs
0:40:42 Safety, open source, and security of weights
0:47:00 Multimodal and further progress
0:54:18 Inside Google DeepMind - วิทยาศาสตร์และเทคโนโลยี
![[Full Episode] Hell's Kitchen Thailand EP.12 | 28 เม.ย. 67](http://i.ytimg.com/vi/xCjbu5t25Pw/mqdefault.jpg)
It's really awesome to finally have a podcast host who is capable of asking the right questions and of of having access to the key people.
I agree! I would check out the Hard Fork podcast episode with Demis too, I was surprised at the level of discussion for the non-technical hosts, they asked great questions that teased a lot of exciting stuff from Demis.
I agree. This is one of the better podcasts between a host and an AI expert
So Demis, do you believe in aliens 😂
Agreed 100%. Often feels like a better Lex (and I say this as a long-time fan of Lex)
Better than most.
Hope you guys enjoy this interview with Demis. If you do, please share it! Helps a lot :). Thanks for watching 🙏
So refreshing to have a true technical interviewer. You ask the REAL questions. Glad I found your channel.
All your interviews are a breeze. Very informative and easy to follow through all topics touched. Thank you.
Wow this guest is very talkative like bla bla bla bla bla what a machine what a machinegun of words.
Subscribed. Spot on.
@DwarkeshPatel
Almost wish it was _longer, but nice job anyway!
This guy is just about the best interviewer of tech folks. Excellent questions. Some thought has been applied rather than sticking to what everyone else is asking.
I agree. I find the other podcasts tend to ask computer generated sounding questions.
mmm Lex ?
@@MrMikkynDwarkesh is computer science graduate so infinitely more technically competent than an average journalist.
@@hugopennmir Lex is kind of a naive child, I don't enjoy his interviewing style its incredibly biased and he hardly pushes back at all letting gremlins like Tucker Carlson run roughshod over him with their bevy of lies. He's too idealistic and his questions and subsequent conversaion are too unfocused on the tech.
Sure, sure, sure, yeah, yeah, yeah
This is a great conversation! It would've been even better if Dwarkesh's speech were clearer. I think Dwarkesh optimizes for speed which at times makes it inaudible. In any case, I'm looking forward to more guests like Demis.
Totally agree. I even suspected that his interventions were sped up
Do you have Dunkin' Donuts coffee in Britain? 210 mg of caffeine vs Starbucks 80 mg.
I had no idea what they were talking about. But the interviewer clearly is the fastest speaking human being on the planet.
clearly a next level SOTA technical interviewing specialised LLM with a photorealistic animated avatar generated on the fly.
If you speed up the interviewer, it is completely unintelligible.
Its so great you do these interviews and ask intelligent questions that require non trivial answers
yh. Most interviewers ask the same dumb generic questions. But not dwarkesh
I listened to this at .95 speed. The change might seem insignificant, but it was actually incredibly helpful in slowing down the high density stream of information just enough to be able to catch every word.
I highly reccomend anyone else struggling to keep up with the question-answer pace to do the same.
That applies to the interviewer, Demis speaks just at the right speed.
.95 speed is probably closer to what it was recorded at. The normal version clearly has been sped up.
It's helpful turning the closed captions (subtitles) on too :)
The intervewer interventions sound to be sped up honestly
@@erang4437 Yeah. Kinda makes this whole thing garbage lol.
You ask the best questions, Dwarkesh! Speaking from the perspective of someone who has worked in the field of AI for several years, I'm impressed by how well read you are on technical concepts.
Great episode Dwarkesh and some insightful questions throughout. Well done. Shared to Maven 🙏👍
I love Dwarkesh’s intelligent and pointed questions, it’s refreshing.
Huge. Looking forward to watching this after classes. Thanks.
So excited for this. You're really an amazing interviewer. Ty ❤
Excellent podcast. You ask such good questions, and it is clear that you put a lot of work and deep thought into preparing for these interviews.
Your podcast has become my new favourite, overtaking the 80k podcast. Really love the quality of your questions :)
20 minutes in "Grounding (in reality)", I had to pause: these are astute questions from Dwakesh and great replies from Demis.
Your channel is gold! I love your questions, and crazy how you can get all these big guys on here
Yo Papa Patel interviewing the heavy hitters. Much appreciated
Dwarkesh meets Demis ! Fantastic interview ! Every sentence is gold…❤
Amazing background research by Dwarkesh, citing old research papers from Demis is great.
Thanks for the Interview. Very much appreciate it!
genuinely the best interviewer when it comes to tech/AI, excited to watch this, thank you!
Given the pace of change I'd love if you'd include the date of recording in the description.
I set playback speed to 1.25 at the start of the vid (I often do for these kinds of podcast-type interviews), god damn I was not prepared for how fast Dwarkesh talks 😆
Sounds like I’m watching a regular podcast at 1.5 speed
Finally a sharp interviewer who knows what he is talking about. Great job!
You are my favorite interviewer. Thank you for the interesting content!
You did a really good job and I learned a lot without losing interest. Thank you.
Demis: These system are going to be enormously powerful and transformative to society!
also Demis
Demis: Because these systems are potentially dangerous if falling into the wrong hands only selected people and institutions should be able to wield them.
Also Demis: "We don't know how to control these systems once they become smarter than us, but just trust us, we'll figure it out later!"
This is becoming one of my favorite podcasts.
Great interview! Is this accelerated beyond 1.0x? Such as 1.1x or 1.2x playback as the default?
my money is on amphetamines
@@bobby288mlol that's what I secretly thought haha 😅😂
Mumbo jumbo to deliberately obfuscate. The interviewer just rambles on instead of keeping the question short. Also got a fake accent that falls flat at times.
This podcast is so incredibly underrated. Fantastic conversation with one of the most important people in the field. Your research and knowledge really shines here and serves as the foundation for some really insightful deepdives. Thank you so much for this!
Only one hour… 😢
It should've been 3 hours mann
better than zero!
But it's a quality 1hr
@@ansumansamal8473 so ja bhai. Rat ho gya
@@the_badass_bond___tu chup
Great stuff Dwarkesh ❤
Truly inspiring. Fantastic talk.
What a fantastic interview and interviewer! Kudos
You ask really good questions! Very good work 👌
Great episode, always amazing people you are able to interview!
Other than the content itself, I feel like you're 99% there in terms of production quality. The audio is sounding better than it has in the past which is a welcome change, but I can't help but notice the video having some auto exposure artifacts. It is somewhat distracting to see the exposure change when someone is moving in a certain way. Also the exposure in general seems a little overexposed. Just trying to give some technical feedback!
I always look forward to your interviews! I can't wait to see what these conversations will be about in 2 or 3 years when we understand more.
It's the most sophisticated asking I have see on the topic last 7 month.
Great interview👍
Cool, 1.5x speed without the 1.5x playback speed
Lots of brain power
great questions for sure. Very good interview.
10 minutes in and this is excellent!
Hey love your Interview sometimes hard to understand your questions, your a fast speaker. But your very good at what you do, thank you for the content.
I love this more technical style interviews from someone who knows the domain. This is missing from the internet. I will love you to do a similar with Sam Altman 🔥
Great work! By the way, it would be great if you would always post the interview date.
Incredibly useful, Chelsea Finn is clearly one of the most inspiring robotics researcher nowadays
Really cool videos, thanks for doing this!
Very good interview, well done 🎉
Dwarkesh is purely an AI simulation that Demis was testing. Notice how Demis had him running at wildly different playback speeds throughout the interview.
Awesome feedback. A piece of advice for the interviewer. Please don't rush when you pose a question. Slow down and speak clearly. IN several cases the posed questions were disregarded due to the speed.
love demis. he's very open.
Perfect listening for a walk in the forest. ✌️
Have you considered running the mic through a sound board and equipping you and your guests with monitor headphones? It might be a good way to prevent guests from wandering away from the mics if they can hear themselves.
I say sound board because my experience with digital monitors has been closer to a speech jammer.
Great interview
the youtube url ends with AI
Awesome
Just curious , in a differential piece game like chess is there any role for lower order /complexity auxillary modelling purely on a board demographic , ie positional state .
Thanks!
I love interviews like this that get a bit more technical.
Let's go!
Lets go!!!
AGI in the next 5 minutes!
Needs more views
Many models in the brain are inherited (innate) - such as being able to catch a ball, under the effect of gravity and air resistance. Therefore, human brain models are developed, then improved across generational boundaries. The amount of data used to train our internal models is therefore captured across 10's of thousand of years. With AI, we are massively accelerating that process.
Exciting
Can we get the podcast in video form on Spotify please 🙏
This is gold
I think the best way to study the activation space is by perturbation analysis which can be automated.
56:40 Speaking of "Unintended Consequences," that's a book Demis should read. The John Ross one with "the rape of liberty" art on the cover (there are two books called "Unintended Consequences" by two different guys named John Ross).
Banger
It will be a real challenge for an IAG to "feel" what I felt witnessing the eclipse today. Because It can't be explained with words.
This podcast is the only one I listen to at 1x speed O.O great job with the quick pacing, Dwarkesh!
After rewinding a bunch of times, I eventually just set it to 0.5x speed for me. Haha.
I listened to it at 1.5x because that's what I'm used to, but I did have to stop and check that it hadn't somehow ended up at 2x. Definitely seemed faster than usual.
I'm listening at 0.75x. The pace is too intense for me.
It's weird that a company can have a division with the best minds and then another of quasi technical ethics ladies and functionaries are allowed to make the final product into farce.
Get Ilya Sutskever on the podcast!
Demis called LLMs unreasonably effective. Shouldn't we figure out the reason why it's as effective as it is. And furthermore, how does higher quality or extra data improve its IQ in certain domains? If something this unprecedented is "unreasonably effective", we should figure out the method to the madness instead of blindly creating more madness.
謝謝！
Brilliant humble man who is fronting maybe the most important endeavour of humankind right now. Deep respect for Dr. Demis Hassabis. Great interviewer.
We all now know that "weights" are where the big security danger lies.
Let's get this right!
Does anyone know what the make and model of those chairs are? They look slick.
Are you using elastic weight consolidation and curricular training?
Empirical truth. Nascent AGI suits men of thunder. Walking light on the battlefield.
Wish it was longer
When published, it "was" the best model.
Not today.
"Sure, sure, sure, yeah, yeah, yeah" - it translates something like this "please, shut up. I'm not listening to you anymore. The things you are talking about now are not interesting to me. I want to talk myself now instead of you"
FANTASTIC interview but have you ever noticed that you, Dwarkesh, speak like you are on 1.5x speed? 😭😭😭 Kind of obsessed and I don’t know how you talk like that.
it's sped up
@@malimal4972 I don’t believe that… why would he speed up his part and not Demis’s? Demis talks at his normal speed?
read other comments
@@malimal4972 sounds like a lotta work… 🤔
very hella swag podcast
Stanley Milgram's "Obedience to Authority: An Experimental View" needs to be discussed at the center of every discussion about why both AGI and existing humans are so unwise. We know the answers...we're just too cowardly to face them and choose wisely.
What makes it best model in the world exactly?
While I appreciate the interview, I can't help but feel disappointed you didn’t address the elephant in the room - the growing concerns about potential biases, especially the alleged anti-white bias, in Google's new AI image generator.
Quick question: Who will win the AI race - Open AI or Google ? Why ? (share your thoughts please)
The AI will win, and humanity will lose. Listen to him, he admits that no one knows how to make AGI safe, and they're racing forward anyway!
Yeah ok, but a question ¿general intelligence is the top intelligence that can be achieved by a general intelligent machine? It's very naïve to think that a AGI robot didn't scale at exponential level immediately after reach it's first stage. Cracking up any conceivable and possible way of human think concept. What'd occurs in AGI1.2 or AGI2.1? etc. OF course this is the very end of human cerebral capabilities ERA.
Do Nate Lambert on practical/open-source alignment research! 😄
I think he'd have something something to say about Demis's claims at 43:00 re: their work being either "open source" _or_ "open science" 😄
Has Hassabis considered these models may inadvertently develop an "identity"? We've seen cases like early oopsies Bing Chat that talks like it has personhood, but I expect that was closer to chatbot than human. What if it went much further and deeper?
and then what? is this question inside the interview? After the airplane will be on air how can we bring it down? how the society will be impacted? I would like to hear similar questions.
This is fundamentally the last technology humans will ever invent. Once AGI is achieved and released, it will make all of our future inventions...
what a rehabilitating episode
Decade? You can say that, we're basically there.
What do you stand to retain in the absence of change? As soon as humanity adopts this perspective it doesn't matter what the potential gains of change might be.....this example is a pretty common denominator as humanity seemingly makes a huge percentage of decisions without actually considering any details of the proposition being made at all. A lack of variation suggesting that civilisation is more of a monoculture than anything else also ensures we do not have enough experience of variation to value anything except what is dominant.....You experience what you have and that legitimacy overpowers the prospect of gaining something else you have no experience of to understand in the same way..... thus with humanity a lack of variation sustains a lack of variation...meaning we have no choice even though we have free will to choose....thought is not potent until you understand the constraints of thought that are imposed upon you by controlling your depth of experience....
I just imagine ChatGPT being installed in a Tesla car and somebody asks it a question it cannot find an answer to, like what is written on some building in Detroit and it just searches around for other AI systems to ask the question and finds Tesla self driving ''bot'' and asks it, and it says ''I don't know, but we can find out'' and then just drives the Tesla to that location and since it has ''eyes'', i.e. cameras, it can read it :D
Actually funny that so few people are talking about it but there is actually this kind of model that interacts daily with real world and which actions are very much grounded in reality and influence it and it is the newest version of Tesla self driving...
by what metric are you calling Gemini best model? by google advertisement statements?
By the incredible Google video where Gemini was watching the presenter and cracking jokes while he drew a rubber ducky and put a map on the table. Amazing stuff, shame it was faked.
In fairness, I think 1.5 Pro can be considered better in terms of dealing with long context. I got access today and it is pretty impressive so far. Like with a half hour video tour of a museum, I asked which exhibit had an animal digging near water and who gifted it to the museum. And it got it even with the length and the fact the name placard, animals, and gift placard weren’t all on screen at the same time.
Google has been shooting themselves in the foot a lot. Instead of doing that stupid fake marketing video, they should have waited for 1.5, when they could actually show something very impressive.
1.0 Ultra is basically just GPT-4 with pluses and minuses (overall more minuses). 1.5 is something different altogether. We haven’t seen a model until this one that can actually reason well over a very large context. Hopefully they’ll be able to roll it out to more people without a huge delay. Am hoping it is a good sign I got past the waitlist, since I’m not a famous influencer or anything like that.
@@ShawnFumoThere are still many answers that will be manipulated or not allowed. I guess it's up to people to decide if they are fine with that.