Dario Amodei — “We are near the end of the exponential”

Dario Amodei · 2h 22m · youtube.com ↗

29 korrents from this recording

Dario Amodei did not write this page.

Every claim below is a statement made in this recording, quoted word for word and linked to the second it was said, so you can hear it rather than take our word for it. The wording comes from the transcript published alongside the recording; the sentence above each quote is our reading of the claim, not their wording.

0:00:48 · watch on youtube.com↗

Our reading
The most surprising thing about the last three years is not the technology but how few people recognise we are near the end of the exponential.

Their wordsI will tell you though what the most surprising thing has been. The most surprising thing has been the lack of public recognition of how close we are to the end of the exponential.

Interesting
9 min later
0:09:23 · watch on youtube.com↗

Our reading
Pre-training is not the process by which humans learn; it sits somewhere between human learning and human evolution.

Their wordsI think there's something going on that pre-training it's it's not like the process of humans learning. It's somewhere between the process of humans learning and the process of human evolution.

Interesting
2 min later
0:11:20 · watch on youtube.com↗

Our reading
RL environments exist to make a model generalise, not to teach it each skill one at a time — exactly as pre-training does.

Their wordsthe goal is not to teach the model every possible skill within RL just as we don't do that within pre-training, right? Within pre-training, we're not trying to expose the model to, you know, every every possible you know, way that words could be put together, right? You know, we're it's it's rather that the model trains on a lot of things and then and then it reaches generalization across pre-training, right?

Interesting
3 min later
0:13:51 · watch on youtube.com↗

Our reading
There is a 90% chance of a country of geniuses in a datacentre within ten years, and no honest forecast can go much higher than that.

Their wordson the basic hypothesis of you know, as you put it, within 10 years we'll get to, you know, you know, what I call kind of country of geniuses in a data center. I'm at like 90% on that. Um and it's hard to go much higher than 90% cuz the world is so unpredictable.

Interesting
0:13:51 · watch on youtube.com↗

Our reading
We will have a country of geniuses in a datacentre within ten years.

Their wordson the basic hypothesis of you know, as you put it, within 10 years we'll get to, you know, you know, what I call kind of country of geniuses in a data center. I'm at like 90% on that. Um and it's hard to go much higher than 90% cuz the world is so unpredictable.

Interesting
1 min later
0:15:00 · watch on youtube.com↗

Our reading
The one genuine uncertainty left about AI capability is tasks nobody can verify — planning a Mars mission, discovering CRISPR, writing a novel.

Their wordsMy one little bit, the one little bit of of fundamental uncertainty even on long time scales is this thing about tasks that aren't verifiable. Like, planning a mission to Mars, like, uh you know, doing some fundamental scientific discovery like like CRISPR, like, you know, writing a writing a novel. Hard to hard to verify those tasks.

Interesting
1 min later
0:16:07 · watch on youtube.com↗

Our reading
Models already generalise substantially from tasks that can be verified to tasks that cannot.

Their wordsWe already see substantial generalization from things that that verify to things that don't verify. We're already seeing that.

Interesting
2 min later
0:18:22 · watch on youtube.com↗

Our reading
Predicting that AI would write 90% of the code was never a prediction that we would need 90% fewer software engineers.

Their wordsBut but that's actually a very weak criterion, right? People thought I was saying like we won't need 90% of the software engineers. Those things are worlds apart, right?

Interesting
9 min later
0:27:11 · watch on youtube.com↗

Our reading
Even a country of geniuses in a datacentre will not be an infinitely compelling product; it buys three to ten times growth a year, not instant adoption.

Their wordsbut it's not an infinitely compelling product, and I don't think even AGI or powerful AI or country of geniuses in a data center will be an infinitely compelling product. It will be a compelling product enough maybe to get three or five or 10x a year growth even when you're in the hundreds of billions of dollars, which is extremely hard to do and has never been done in history before, but not infinitely fast.

Interesting
10 min later
0:36:45 · watch on youtube.com↗

Our reading
A lab under real commercial pressure cannot fool itself about whether AI makes it productive, because the output is a model launch every few months.

Their wordsLike we see the end productivity every few months in the form of model launches. Like there's no kidding yourself about this. Like the models make you more productive.

Interesting
1 min later
0:37:43 · watch on youtube.com↗

Our reading
Coding models give AI research a 15 to 20% total-factor speedup today, up from about 5% six months ago.

Their wordsLike I would say right now the coding models give maybe I don't know a a like 15 maybe 20% total factor speedup. Like that's my view. Um and 6 months ago it was maybe 5% and so and so it didn't matter.

Interesting
5 min later
0:42:40 · watch on youtube.com↗

Our reading
Continual learning will probably be solved within a year or two, but the trillions of dollars do not depend on solving it.

Their wordsI think the trillions of dollars a year market, maybe all of the national security implications and the safety implications that I wrote about in adolescence of technology can happen without it, but I I I also think we, and I imagine others, are working on it. And I think there's a good chance that that, you know, that we get there within the next year or two.

Interesting
0:43:07 · watch on youtube.com↗

Our reading
Longer context is an engineering and inference problem, not a research problem — nothing prevents it from working.

Their wordsThere's There's nothing preventing longer context from working. You just have to train at longer context and then learn to to serve them at inference. And both of those are engineering problems that we are working on and that I would assume others are working on as well.

Interesting
9 min later
0:51:45 · watch on youtube.com↗

Our reading
If a lab is off by a single year in its revenue forecast, no hedge on Earth stops the compute it already bought from bankrupting it.

Their wordsAnd if my if my revenue is not a trillion dollars, if it's even 800 billion, there's no force on Earth. There's there's no hedge on Earth that could stop me from going bankrupt if I if I buy that much compute.

Interesting
1 min later
0:53:03 · watch on youtube.com↗

Our reading
Some of the other AI companies have not written down the spreadsheet — they are committing hundreds of billions without understanding the risk.

Their wordsI kind of get the impression that, you know, some of the other companies have not written down the spreadsheet, that they don't really understand the risks they're taking. They're just kind of doing stuff cuz it sounds cool.

Interesting
7 min later
0:59:35 · watch on youtube.com↗

Our reading
In this industry, profit means you underestimated demand and loss means you overestimated it, because the datacentres are bought years ahead.

Their wordsI actually think profitability happens when you underestimated the amount of demand you were going to get and loss happens when you overestimated the amount of demand you were going to get because you're buying the data centers ahead of time.

Interesting
11 min later
1:10:19 · watch on youtube.com↗

Our reading
Every frontier model is profitable on its own; the company loses money only because it is already training one ten times more expensive.

Their wordsbut at the same time, we're spending $10 billion to train the next model because there's an exponential scale up. And so the company loses money. Each model makes money, but the company loses money.

Interesting
3 min later
1:13:35 · watch on youtube.com↗

Our reading
AI will settle the way cloud did — three or four players, kept few by the cost of entry rather than by network effects.

Their wordsThe way you get industries in which there are small number of players are very high costs of entry, right? Um so, you know, uh cloud is like this. I think cloud is a good example of this. You have three, maybe four players within cloud. I think I think that's the same for AI, three maybe four.

Interesting
2 min later
1:15:05 · watch on youtube.com↗

Our reading
Models are far more differentiated from one another than clouds are, so AI will not commoditise the way cloud did.

Their wordsLike models are good at different types of coding. Models have different styles. Like I think I think these things are actually, you know, quite different from each other. And so expect more differentiation than you see in in um cloud.

Interesting
3 min later
1:17:35 · watch on youtube.com↗

Our reading
The real danger in AI-driven growth is geographic: 50% growth in and around Silicon Valley and business as usual everywhere else.

Their wordsa worry I have is that the growth rate could be like 50% in Silicon Valley and, you know, parts of the world that are kind of socially connected to Silicon Valley and, you know, not that much faster than its current pace elsewhere. And I think that'd be a pretty messed up world.

Interesting
2 min later
1:19:32 · watch on youtube.com↗

Our reading
Robotics will be revolutionised a year or two behind the country of geniuses, and getting there does not require human-like learning.

Their wordsdoes that mean the robotics industry will also be generating trillions of dollars of revenue? My answer there is yes, but there will be the same extremely fast but not infinitely fast diffusion. So, will robotics be be revolutionized? Yeah, maybe tack on another year or two.

Interesting
1 min later
1:20:48 · watch on youtube.com↗

Our reading
The history of machine learning is of proclaimed barriers dissolving inside the big blob of compute.

Their wordsIn fact, I would point to the history in in ML of people coming up with things that are barriers that end up kind of dissolving within the big blob of compute, right?

Interesting
4 min later
1:24:32 · watch on youtube.com↗

Our reading
The chatbot has hit its ceiling — a smarter model no longer helps the average consumer — and that is a limit of the product, not of AI.

Their wordsThe the chatbot is already running into limitations of, you know, making it smarter doesn't really help the average consumer that much. But I don't think that's a limitation of AI models. I don't think that's evidence that, you know, the models are are the models are good enough and they're they're, you know, them getting better doesn't matter to the economy.

Interesting
2 min later
1:26:51 · watch on youtube.com↗

Our reading
Not every token is worth the same amount, so AI will be sold by results and by the hour as well as by the token.

Their wordsso so I think we're definitely going to see business models that that recognize that, you know, at some point we're going to see, you know, pay for results or you you know, in some in some form or we may see forms of compensation that are like labor. Um, uh you know, that that kind of work by the hour.

Interesting
11 min later
1:38:17 · watch on youtube.com↗

Our reading
Banning state AI regulation for ten years with no federal rule to replace it is a crazy thing to do, because ten years is an eternity.

Their wordsgiven the serious dangers that I lay out in adolescence of technology around things like the you know, kind of biological weapons and bioterrorism, autonomy risk and the timelines we've been talking about, like 10 years is an eternity. Like that's that's a that's a I I think that's a crazy thing to do.

Interesting
8 min later
1:45:52 · watch on youtube.com↗

Our reading
Chip export controls to China keep failing not because the case is weak but because there is too much money riding on the sale.

Their wordsThe counterarguments against it are I'll politely call them fishy. Um, uh and yet it doesn't happen and we sell the chips because there's there's so much money. There's so much money riding on it.

Interesting
17 min later
2:03:18 · watch on youtube.com↗

Our reading
In an AI world growth and economic value come easily; distribution, wealth and political freedom are the hard parts, and that is where policy belongs.

Their wordsWhat I would say is that you know we are we are about to be in a world where growth and economic value will come very easily. If right if we're able to build these powerful AI models, growth and economic value will come very easily. What will not come easily is distribution of benefits, distribution of wealth, political freedom, um you know, these are the things that are going to be hard to achieve.

Interesting
3 min later
2:06:41 · watch on youtube.com↗

Our reading
Training a model on principles rather than a list of rules is not an ideology but an empirical finding: its behaviour is more consistent and covers edge cases.

Their wordsit's it's kind of purely a practical and empirical thing that we've observed that by teaching the model principles, getting it to learn from principles, its behavior is more consistent, it's easier to cover edge cases, and the model is more likely to do what people want it to do.

Interesting
7 min later
2:14:04 · watch on youtube.com↗

Our reading
The hardest thing for a future historian to recover will be how little the world outside understood, because whatever happened looks inevitable in retrospect.

Their wordsOne is at every moment of this exponential, the extent to which the world outside it didn't understand it. This is This is a bias that's often present in history where anything that actually happened looks inevitable in retrospect.

Interesting