Finding our place in the age of AI | Oliver Edholm | TEDxSSE

thank you this picture is called the scream and for me what this represents is The Dark Night of the soul the state in which you feel lost disconnected and for me really questioning the life in general and this this is a picture a very cute picture of me at the age of 13 and even though it doesn't necessarily look like I'm in an existential crisis I was definitely seeking a sense of a higher purpose something I could really live for and feel passionate about and yeah this makes life worth living and as you can see in the picture I I read this Hindu book The Autobiography of Yogi I was I was trying everything I was searching all over the place and throughout this search I felt like I landed at some personal conclusions some sort of viewpoints that uh I still have today and determines my sense of what makes life worth living one of those things I learned was secular Buddhist philosophy a way to relate to life that uh that feels very meaningful but almost as importantly I felt a lot of life purpose by reading this book super intelligence and what this book is about is the possibility to create uh humanlike intelligence inside of a computer and for me if we solve this we solve a lot of other problems humanity is facing we can solve climate change we can solve cancer we can really solve anything and as a result of reading this book I got really really determined this is my life purpose I want to be part of this and I want to steer it in a good direction this led me to drop out of high school to then I booked a flight ticket without my parents knowing to Singapore I I was doing AI research for a short while there and then starting my company dep pict where we raised over $2 million in funding and were appli in e-commerce and in this book there was a graph like this this graph represents more law and more law is the reason we have smartphones it's the reason we have such a modern society today because since the 70s we've seen a doubling every two years in the compute capability per dollar and you have to remember this is an exponential Trend so every two years we're doubling on top of all the previous progress we've made and what Nick Bostrom argued in this book is that at some point given enough computer power we could literally simulate Evolution because we know we could create humanlike intelligence through Evolution and it's quite a simple algorithm and if we have enough compute he argued that we will do the the same and if we move forward today 2024 we're seeing amazing AI breakthroughs like this Sora where we can create photo realistic videos just based on a prompt or we're having companies beating open AI at their own game taking overtaking Shaq GPT and we're also seeing really sci-fi examples such as being able to read human brains just based on an fmri scan we can recreate what someone is seeing visually and the reason we're seeing all of this progress comes back to what Nick bostron predicted and this what I'm going to show now is the main message I want you to take away so I want you to pay close attention the reason we're seeing all of this AI progress today is because we have figured out the equivalent of E equals mc² in physics but for AI and for me it's kind of insane that we're not talking about this all the time because as a result of this discovery that really got crystallized around four years ago only so we've had a long history of AI research but since 4 years ago I would say we've figured out the equivalent of E equals mc² and that's this thing called scaling loss which is this law where not exactly right the more computer power we can feed in to train huge language AI models and the more data we get the smarter they they become and this is a very predictable Trend that doesn't see any diminishing returns so this means that we can work backwards we know Mor law for instance we know that Mor law increases our compute capabilities every year exponentially and we can feed a lot of money into this because we know the financial implications of creating smarter and smarter a uh and then we have some Altman saying things such as it's almost like this law decided by God it just works and now what open Ai and other AI labs are doing is they're racing to figure out who can get the most computer power into these models hence we're seeing Sam Altman raising 7 trillion dollar for AI ships that's 5% of world GDP that's also why we're seeing Nvidia stock growing like crazy Nvidia is now the third most valuable company in the world and the reason is because a bunch of analysts have gone into the details of the scaling loss and they figured out that wow this actually works the more computer power the more data we put in the smarter these models get hence there's billions and billions of dollars invested in Nidia because they have Monopoly on computer ships so this is kind of unprecedented and I don't hear us talking about this this impacts us students because in a few years AI will be able to do certain things that you're studying for as companies you need to figure out what's the remote we have outside of just capabilities that future models will have and then also as citizens of society we need to take a stance on what we're going to do when AI gets more and more intelligent and I want to walk through some examples where we have GPD free is a very technical name for sha GPT gpt3 was the model that led to sha GPT which we all know about and the main difference between GPT between gpt2 which was the previous model and gpt3 was a 100x increase in compute so it's quite simple we have the same fundamental Computing building blocks but with 100x the amount to compute and you get one of the fastest growing applications in history moving forward to G gp4 which is the model we have today the main difference between gpt3 and gp4 was another 100x scale up in compute and a 100x scale up isn't as much as you think because we have more slw people are figuring out the financial returns of investing more and more money into this researchers are finding all these efficiency Improvement to utilize compute and to show an example of how much more intelligent gp4 got I want to show this this is the bar exam it's the main exam to get certified as a lawyer in the United States and before with gpt3 it had it answered the questions correctly 10% of the time and GPD 4 answered it correctly 90% of the time so it went from 10% to 90% And you as students you all know about um you know the the capabilities of using shat GPT for your schoolwork so you all know this very well and here it gets a bit more interesting because I don't see any reason for this scaling lot to not continue working so so by the end of 2024 by the end of this year we're very probably going to have a GPD 5 and the main difference between GPD 4 and GPD 5 is going to be another 100x increase in compute so it's very we don't need any fundamental scientific breakthrough here it's really a scaling operation that's happening here and we're probably going to see the equivalent intelligence increase as we had from gpt3 to GPD 4 happening again with GPD 5 and unless you as a worker a student or as a business think about what's my what's my Moe outside of the capabilities that GPD 5 will have let's say you will have a harder Ander hard time in the workforce or as a company competing against other companies uh I think that GPD 4 is probably a bit overhyped uh GPD 5 will probably lead to very very very impressive uh things uh that we can accomplish and to show where this is going the top AI labs are already preparing infrastructure for gpd6 and they're dealing with problems such as ensuring that local power grids don't get shut down as a result of how much com electricity they're using for the data centers those are the kind of bottlenecks we're dealing with and it's hard to predict how this will look but it will probably be very very intelligent and at this point I think we're going to see some professions getting kind of wiped out and also it's at this point we will probably really look at uh regulation people will wake up very intensively to the need for regulation on these kinds of things which we don't have today and just one example was that I was on a Swedish television a few weeks ago and I showed what happened when you take an open source models without any safety constraints I removed the safety constraints and I asked how can I kill thousands of people in Stockholm it's kind of a scary prompt and the answer I got was stepbystep instructions on how I can contaminate the water sources in Sweden I won't go into more detail but this is possible today some might argue you can get this information on the dark web but we know the capabilities are going to increase and increase and increase we don't have any protection against things like this so unless we get any regulation we're going to see another 100x scale up and by gpt7 we probably would have ai that's almost as good as human AI researchers at doing AI research so what you can do then is you can scale up billions of AI researchers that work 162 hours a week on making smarter versions of themselves and for every 1% Improvement they make they get 1% faster and they're working at superum speeds you know a fast computers are and here I would argue we will have AGI artificial general intelligence and how would I explain artificial general intelligence the most concrete way way to put it is that most of us will out be out of a job because AI will do it much better than us that's what AGI is and don't just listen to me listen to the CEOs of the top AI Labs these are the top three AI labs in the world and they all talk openly about the scaling loss and they there's of course a lot of unknowns in this we're in really Uncharted Territory but they all have been quoted saying we will probably given how things are panning out have AGI before 2030 and here we are assuming like everything is normal so to conclude we have figured out the equivalent of E equals mc² for AI it's these scaling laws where the more compute the more data we put in the more intelligent these models get and the top AI labs are in a race to get as much compute as possible Sam Alman raising 7 trillion dollars for this and what I would ask myself given this is as a student or as a worker what's my Moe what what how can I position myself in a world where we get more and more capabilities as a business what's my Moe and then most importantly as a citizen what do we want to do given this progress do we want all of the progress being in the hands of private compan where we have very little regulation today or do we want to take action to see something different so that's everything for me and I hope it was useful thank

Channel: TEDx Talks

Views: 465

Keywords: AI, Computers, English, Global issues, Research, Society, TEDxTalks, Technology, [TEDxEID:56934]

Id: h2CxZsW4kuE

Length: 17min 26sec (1046 seconds)

Published: Tue Jul 16 2024