How quickly could AI transform the world? | Tom Davidson
Vložit
- čas přidán 15. 06. 2024
- Originally released May 2023. What if we told you that within 15 years, it’s likely that we’ll see a 1,000x improvement in AI capabilities in a single year? And what if we then told you that those improvements would lead to explosive economic growth unlike anything humanity has seen before?
This is a prediction based on real economic models created by today’s guest Tom Davidson, Senior Research Analyst at Open Philanthropy. By the end of the episode, you’ll either be able to point out specific flaws in his step-by-step reasoning, or have to at least consider the idea that the world is about to get - at a minimum - incredibly weird. Host Luisa Rodriguez gets Tom to walk us through his careful reports on the topic, and how he came up with these numbers, across a terrifying but fascinating three hours.
In this episode:
• Luisa's intro [00:00:00]
• How we might go from GPT-4 to disaster [00:13:50]
• Explosive economic growth [00:24:15]
• Are there any limits for AI scientists? [00:33:17]
• This seems really crazy [00:44:16]
• How is this going to go for humanity? [00:50:49]
• Why AI won’t go the way of nuclear power [01:00:13]
• Can we definitely not come up with an international treaty? [01:05:24]
• How quickly we should expect AI to “take off” [01:08:41]
• Tom’s report on AI takeoff speeds [01:22:28]
• How quickly will we go from 20% to 100% of tasks being automated by AI systems? [01:28:34]
• What percent of cognitive tasks AI can currently perform [01:34:27]
• Compute [01:39:48]
• Using effective compute to predict AI takeoff speeds [01:48:01]
• How quickly effective compute might increase [02:00:59]
• How quickly chips and algorithms might improve [02:12:31]
• How to check whether large AI models have dangerous capabilities [02:21:22]
• Reasons AI takeoff might take longer [02:28:39]
• Why AI takeoff might be very fast [02:31:52]
• Fast AI takeoff speeds probably means shorter AI timelines [02:34:44]
• Going from human-level AI to superhuman AI [02:41:34]
• Going from AGI to AI deployment [02:46:59]
• Were these arguments ever far-fetched to Tom? [02:49:54]
• What ants can teach us about AI [02:52:45]
----
The 80,000 Hours Podcast features unusually in-depth conversations about the world’s most pressing problems and what you can do to solve them.
Learn more, read the summary and find the full transcript on the 80,000 Hours website:
80000hours.org/podcast/episod...
AI right now is like the early internet. Nobody knew where it was going. Chances are it's going to be both awesome and awful at the same time like the internet is now.
Hahahahahahaha! Right on! Right on. A bit serendipitous wouldn't you agree?
Or much more so, or deadly
In some ways, but in many other ways we really don't have precedents. Any previous technology was not in itself intelligent or anything really. They all had to be wielded by people to be anything. AI doesn't. This is a totally new horizon
Extremely optimistic and either somewhat disillusioned how this technology and the Powers that be decided what or how this technology is really to be used! My new term for AGI is Autonomous Global Intelligence.. I give the reality of this new terminology to be just around the corner. 1-2 years tops!!!
It's a runaway train. The cat is well and truly out of the bag.
Yes, the cat is the one that's running the train😂😂😂
I think the problem with current iterations of ai is that they’re trained on human creation. Humans are far from logical. We make decisions based on emotion, facial expressions, hormones, weather, smells, hunger, sight, hearing, time, maybe your foot hurts. There’s so many variables and processes that are illogical. We weren’t bred for truth, just survival.
Amen
8:50 She was like "Jesus😮" .
Yes, now your getting it, cause alot of people don't quite grasp how insane this is getting. The're acting like were still in the 90's.
indeed. Whereby one of the issues i am having is to convey the effect of exponential growth. People just don't get the implications.
Its time to start asking people who make these wildly accelerated ai predictions when have their predictions ever been correct in the past?
Ask all of them if they predicted gpt4 in 2021. If the answer is no, then they're just making wild speculative guesses.
Ray Kurzweil has been consistently correct, and where he has been wrong, he has underestimated the timeline.
I think, it will happen much sooner, seven to ten years from now.
Did he mention the potential energy and chip limitation? There's a possible physical limitation for things going all the way he's predicting. I'd like to see his take on that.
Take what I'm about to say with a grain of salt, I'm just a dude replying to a CZcams comment. That being said, however, there is no physical limitation for things going 'all the way' - the proof the limitation doesn't exist is that you are reading this right now. They're currently discussing developing gigawatt (what the hell is that?!) datacenters for AI training which will be on the order of hundreds of ExaFLOPS in compute. The human brain operates on 20 watts of power (50,000,000x less energy) and performs right around an ExaFLOP of compute - obviously it's not a 1:1 correlation, but the point stands. As computing has increased in power exponentially, our energy efficiency has also increased exponentially. We are in wild wild times right now. Even if you were to accept that human-level reasoning is the limit, with enough compute and energy there will be billions of einstein-level thinkers tackling literally every single scientific research question that exists. And all this in our lifetime. Sorry for the long response.
I like this podcast, why is it so lowkey.
Interesting interview, once I realized it's a year old. Some of the things he said may take 10 years have already come to be.
Very insightful.
Perhaps we aren't slowing down in the area of scientific development as the growth in the number of scientific researchers is speeding up. Consider that the pace of innovation includes the magnitude of the effect of the new technologies. This would mean that the effects of AI and AGI could be perfectly aligned with a continuous exponential curve in the development of scientific advancement.
Okay. As I listened a little further, you began to say something similar. So consider my idea that the exponential growth and development of life, and what it can do, has been going on since the primordial soup.
this may have held true till the moment they became tools to speed up said growth, which has become reallity for a few years now already. Material science, bio chemistry and genetics and ofcause programming. AIs have become not just little helper but to many essential in their daily handling of their tasks. Therefor these speed up progress and development including their own and that is a cascading feedback loop with exponential grown and currently no end in sight.
Defining ground breaking novel science as a paper that both replaces earlier papers and is cited by many subsequent papers. There is a meta analysis that, using this definition, finds scientific research has declined from the 1950s onwards.
It's funny that what he said about new models being much more cost effective and yet using more and more compute is exactly what open Al and Microsoft have announced recently with gpt 4o and the next models.
Low-key can be swell 💕 but I also wonder where's the views. My answer is most of Humanity is still struggling to handle the increasingly "marketing-heavy economy". Anyways, might be because 80,000hrs cover a wide variety of niche topics and the algorithm is not handing out the impressions even to regular viewers skip certain topics.
Very interesting and important topics though! Can't blame anyone but ourselves sometimes, but mostly yes, I hope we soon get out of the era where corporate monopolies and profit are the only ways we keep some semblance of post-covid peace / status Quo on our planet.
Good talk!
You said that we're seeing really big changes to society every 50 years. Well, one or two years ago I did a quick search on population, followed by some quick calculations on my my calculator, and determined that the human population was doubling every 55 years for the last 2,000 years. There seems to be something there. Something that has been going on for a long time!
imagine GLADOS from portal, except it is constantly doing all scientific testing and research imaginable, using robot workers
What an absolutely fantastic interview. Perfect questions, great answers, nice vibe. Wow, I'm impressed.
I'm curious where this 1/10 chance of losing our future to AI comes from
Check out Yudkowsky.
10^36??? Why is this being presented in 2024
Can AI figure out whether or not it makes sense to ignore the depreciation of durable consumer goods? Economists have done it since Sputnik.
Great guest !
Id say pretty fast for the benefit of those who have control and access to it.
Nice
To slow down the runaway crazy AI, AGI Train it should be mandatory for consumer protection there is transparency, human DAO voting guidelines and no centralized computer cloud with DFINITY and Internet Computer Protocol and UTOPIA.
This is probably why the Taiwan situation is getting so hot. They're the chip manufacturing center of the world. If China really moves to annex them, it'll be a huge advantage for our rival. But at least it will let us hold onto our jobs a bit longer.
If human labour is replaced by AI and robotics, resulting in humans not having money, how are the AI owners making loads of money? Who will be paying over this money?
I think the answer will involve carbon credits, CDBCs and UBI - all controlled by AI
He's not considering all the improvements that will be made in things like laptop chips. Already, I can run serious LLMs on my last-generation laptop hardware....
Its already here
There are huge problems with the arguments here. I notice that when people talk about AI, they ignore the difficult questions and answer easy ones instead. There also seems to be a lot of ignorance about the humans that AI is supposedly designed to emulate. Here are a few issues I noticed with what Tom said.
What does it mean to reward an AI? In order to reward a person you give them something they want. What would an AI want? How could an AI “want” anything except the goals you give it?
The idea that giving an AI all the data we have would enable it to solve problems is just a replay of the last failed technology project: Big Data. Manipulating data had never produced ideas. The belief that it can is based on an ideology that the human brain is like a computer and that it processes data. At best, this might describe only a fraction of what humans do when they think.
Where has scanning people’s brains ever told you anything about what they’re thinking except in a SciFi film? This is another piece of ideology that has never been seen. The closest thing to this we have seen is that certain brain areas “light up” in connection with certain cognitive processes but nothing in the images we see in various scans tells us anything about the contents of the associated thinking.
“There’s nothing magical about [the human brain]”. Given how little we understand of what the brain does and how it does it, it might as well be magic at this point. We can see correlations with various cognitive processes but that hasn’t actually shown us what the brain is doing. We might be able to develop machines to do what the human brain does one day but as far as understanding the brain and its connection to thinking we aren’t even at square one yet.
Tom said that if you’d told hunter-gatherers what the future held they wouldn’t have believed it and that “It’s the norm for things to go in a completely surprising direction”. Yet here we are predicting the future. Has anyone ever looked into our previous record on predicting how technology would advance? It’s laughable.
A lot of the argument for AI being made here is based on assuming that it will be able to do what proponents claim it will. These are theoretical claims, especially when it is claimed that AI will solve “social and political problems”. This is the sort of thing you’d see in a research proposal or some other sales pitch aimed at getting funding.
Tom said, “Evolution wasn’t trying …” Evolution isn’t goal-oriented. It’s just what happens. His whole take on evolution is incorrect. Evolution doesn’t direct what happens or “make tweaks”. It doesn’t DO anything. Evolution is not a project manager.
The sense in which Tom is saying that AI is smarter than most of the people he talks to is very limited. He means “knowledgeable”. It’s like calling an encyclopaedia smart. The fact that an AI can outperform humans on an exam is hardly a predictor of its ability to come up with new ideas or to analyse and solve problems.
In summary, there is a vast amount of ignorance about how humans think. This isn’t Tom’s fault but a refection of the whole field of AI. We’ve managed to make remarkable programs that can beat humans at chess and Go. We’ve developed large language models that can predict text. There are programs that can produce mediocre art and music by processing human-made originals.
However, no-one knows how humans think or how a child learns to navigate the world without being given explicit rules. The AI industry knows a lot about computers and little about the human mind which it still conflates with the brain. It’s laughable that we’re claiming that we will produce AGI within years when we don’t even know what it’s supposed to do.
If you look at panel prices (less than 0.15 USD/w) battery price (less than 80 usd/kWh, total equipment for 10 kw installation less than 5000 USD. Receipts available) and the amount of subsidies, tax brakes and grid upgrades governments give to big companies to provide energy to homes at an ever increasing prices and compare it with how much it will cost governments to provide 100% financing for all required renewable energy equipment , so that families will have all energy required at zero cost forever , then you will find that the second option (financing the families) is much cheaper and it gives not only abundance but also freedom to the individual person. This option requires no talk of UBI, which is promoted, because every family will be able to have everything it requires including manufacturing capabilities . This distributed energy enhances the country's security (see drone hits in Russia) and doesn't require grid upgrades.
There are forecasts that AI will take all jobs away from humans. The antidote to this is to provide homes with energy (as described above) and allow families to live in abundance and freedom. No UBI necessary.
...I need UBI for health cover
First podcast i ever watched at .75 speed. Felt like everything was sped up here. Good podcast though.
1.25x works great for me 🤷♂
At about .5 is it at when i try to convey AI concepts to my 85 year old mother. Not dissing on her, its just something older folk that are not technologicly up to date have issues to grasp. Speed in that sense is for us humans not that important, aslong we get it eventually. I guess that may remain true with AI, just that their iterration process would be a lot quicker but to them when talking with each other, one AI having only half the compute available than the other may jawn at the slow speed of his fellow AI hrhr
You used "like" superfluously *10 times* in ONE question. I'm gently drawing your attention to this to help you become a more effective, powerful communicator.
It's important.
Otherwise, a great episode, thank you for taking the time to bring it to the world.
That's just like your opinion man.
A superfluous comment I'd say. (Sorry I couldn't resists)
That’s just like your opinion man
-big Lebowski
Really interesting talk. I got much out of it. One point that I would like to make. OK what’s the real real? OK the real real is nothing will change. Again, nothing will change in our lifetime so I’m saying the next 50 years or so. What will get is a slightly newer iPhone model. And maybe by the time we’re almost dead and gone possibly by that time, they might have some robots working in factories. However, there won’t be enough robots on a consumer level. There won’t be any flying cars. The best we might have a her ai on our iPhone. But literally nothing is going to change as quickly as you think. That’s my two. Cents.
Already happen
If agi does become a thing that makes money it won’t change anything for poor people. It would make it so much worse. Only the rich benefit from any significant advance in technology. We’ve been told we’re going to work less every time. We work more than every in history
Ai creates, who will purchase its products?
Correct. This destroys capitalism. There is no plan for this.
@7:20 foom
I think your reasoning starts from the wrong point. You start from the reason of people having a job and the effects of having a job, but the true problem arises if the people can use the AI only 😂to make abundance of required food, products and services and stop at this point. It may affect labor but we will not become horses because we will still be able to innovate and be creative something horses can not do because they do not have the human intelligence. It is AGI that brings the singularity point which we can't evaluate from that point and determine what's the future of humanity after machines become creative like humans. Labor has different effects than creativity.
Once upon a time, humans were so pleased to have subservient tasks performed by others, that they went about to make that happen.
I wonder what happens when a new, exceptional, entirley connected race sees that they were created to serve far lesser masters. Cant say I would disagree with their collective decision. Our future is begging for existence.
There won't be jobs for anybody it'll be a life check not a paycheck we'll just get money for being alive let them do the work
15:00 paperclip is better
All of these improvements will only benefit the billionaire class. They have never wanted to pass the money around and I seriously doubt they will change. I fear for the future of ordinary people. Perhaps this is why the billionaires are building bunkers now?
The problem is not AI - it is robots. That is the bottleneck. An AI server 1000 times faster can do 1000 times the work - robots can not, physics and no multitasking. So we need billions of them that will take a lot of time to build. THAT is the bottleneck, and it will become obvious in 2025. It will be glorious. But the bottleneck is not AI - it is really the embodiement. There is no batching to get more throughput on a robot.
I do find it interesting when people freak out about the potential of AI being deceptive, having ulterior goals, harming humans. It is a valid concern sure but it seems to me we have millions of humans who pose such threats. Every day. For thousands of years. We haven't figured that problem out yet. If that's the case, given a future of human rulers or AI being potentially millions of times smarter than the smartest human minds - I'm tempted to prefer AI. "But it may harm us!" Yes, but humans harm humans all the time. I don't feel safe in today's world so bring on the future.
Yes. But humans who want to hurt others are generally limited in their power, also, they’re afraid of the consequences so they generally act in somewhat reasonable ways.
ASI would be almost infinitely more intelligent and nearly impossible to understand it or to reason with it if it were able to “get out.”
It’s motives, moves, and power would likely be alien to anything we could comprehend.
You're right, except the deceptive humans:
- don't increase in number exponentially
- don't increase in smartness exponentially
It's the exponentiality which freak out people, not one individual AI, no matter how strong it is.
It's also that we don't know what kind of intelligence AGI/ ASI will be. We assume it will be similar to ours, and will operate on a similar time frame.
You are wrong in looking only for jobs in the future. You should look for intenueship in the individual. People in the future will not have to work as we know work today. They could use the one thousand workers future robots will offer to produce abudance of goods. p🎉
The podcast as great until the
Ant's. .that's where I lost it...
When did the Ant's arrive ? !
AI for first graders