The Next 10 Years of AI Will Change Everything | Alexander Wissner-Gross | TEDxBoston

[00:00:06]Ladies and gentlemen, Alex Whistner Gross. Um, I have to ask, are you an AI?

[00:00:10]>> What's the difference really at this point?

[00:00:13]>> And how would you know?

[00:00:14]>> Yeah. So, yeah, I just because people may think this is what they call a non-denial denial.

[00:00:19]>> Yeah. All right. Well, thank you for the non-denial denial. Um, the question is not what the world will look like in 250 years. The question is what foundations we choose to build today so that 250 years from now people look back on our generation this moment as one that expanded human possibility.

[00:00:42]Talk to me about that quote. What does that mean? What what should we be doing right now? What should we not be doing based on on you know hopefully building a positive foundation for the future?

[00:00:53]>> I I am I I think it's fair to say an accelerationist. I think there are many things that we should be doing not just for quarter of a millennium from now but also for 25 years from now. I think one of the most obvious travesties is that right now on earth approximately 150,000 plus people die every day and I think there's far more that our civilization could be doing about that. I don't think people have to die. I think accelerating super intelligence is one of a very few number of things that we can do to put a positive end to that. And I think in part when we speak of 250 years in the future, I think keeping substantially everyone who's alive today around so that they can see America's 500th anniversary, one of the best things we could possibly do. So what do you believe historians in 2276 will say was the defining turning point of the 21st century? It's a tricky question in part because we've only seen about a quarter of the 21st century unfold. So I would say of the first approximately 25 years of this century probably the single most important event was the publication of GPT2 by OpenAI. according to their paper large language models are few shot learners and this was the paper and the moment when I think at at the very latest humanity realized how to build super intelligence and many would quibble with that many would say we don't have AGI yet we don't have super intelligence yet I think that's fundamentally wrong I think as of no later than the GPT2 moment in summer of 2020 humanity as a whole first in pieces and then more of humanity over time figured out how to build much more intelligence much more quickly. And it turned out the answer was elegantly simple. The answer was you just take general knowledge and you compress it.

[00:02:56]That's all it took. We could have sent that knowledge of just take lots of information about general human knowledge, compress it. I think we probably could have sent that information, that knowledge back in time by a few decades and just gotten the whole show on the road decades sooner and we could have had general purpose robotics sooner. We could have had large language models and chat bots sooner if only more of humanity realized that information which was so startlingly simple. We could have explained it to a mathematician or a physicist from a century ago and accelerated everything.

[00:03:30]So, I don't know how many other advances are so simple to explain yet so transformative, but I would say that's the highlight of the first quarter of a century. I would hope that there going to be many comparable highlights for the next threequarters of the century, but my bet is we're going to get a lot of major surprises in the next just 10 years. Forget about the next 75 years or the next 250 years. I think thanks to AI and super intelligence, we're going to see a whole lot more doors open up to ontological shock over the next 10 years.

[00:04:01]>> So, it's not the Pentium chip, it's the data, stupid.

[00:04:05]>> Ah, yeah. I I've argued in the past that the key to unlocking grand challenges in AI is actually more of a data set problem than a compute problem. that creating focused data sets and benchmarks around those data sets is fundamentally a societal accelerant. And if you look at arguably most of the grand challenges that AI has unlocked over the past 50 years, automated speech recognition, statistical machine translation, chess, Jeopardy, and yes, >> go and general purpose conversational abilities. I would argue that those were fundamentally data set constrained, not compute constrained, not algorithm constrained, data set constrained. And so the argument would go creating data sets and benchmarks and communities around them is one of the single most powerful unlocks to solving all of the ground challenges we still have left.

[00:04:58]>> So when Minsky and McCarthy went to Dartmouth and kind of coined the term AI, you know, cybernetics had been around before. were they aware how the role of data >> I think so Marvin was my first research adviser at MIT I think there's unfortunately an alternative timeline where there was less focus early on in the development of the field of AI on algorithms and more of a focus on benchmarks arguably so Marvin isn't around to defend himself anymore but there's a narrative that Marvin did fundamentally a disservice to the acceleration of AI by publishing a now famous paper that showed that a very simple function an exor function could not be reproduced by a shallow neural network. And there are some myself arguably included who would say that one counter example may be set back AI by 20 years or more.

[00:05:59]>> That's the AI winter that people >> that that was one of the earliest AI winters. We've had a few at this point.

[00:06:03]some count two, some count three or four. And so there's a an an alternative history as it were where instead of just sort of throwing neural networks early on under the bus by saying well they they can't model exor we would have had benchmarks early on that provided denser rewards and denser signals to the entire field. And even the the big shock the big onlogical shock of GPT2 which is just how well can your model predict the next word or next token in general natural language. That's such a simple objective. We could have hypothetically sent that back in time to the late 1950s to Marvin at all and said you know what forget all of your other award-winning tenure providing PhD winning algorithmic research. Forget all of your religious warfare regarding whose algorithm is more neuro inpired or whose is more mathematically elegant. Just focus on this one objective. Do a better job of predicting the next word in general text which was already being electronically stored by the 50s and the 60s. Just focus on that and measure progress in your ability to predict the next word over general text over time. Let that be your northstar. And I think we could have been where we're going at least 20 to 30 years earlier.

[00:07:23]Um, you know, I'm thinking of a sine wave, you know, these AI springs, AI winters. Are we, you know, now it's going like, are we due for an AI another AI winter or we've passed a moment where AI winter won't won't be happening?

[00:07:38]>> I think we're overdue for an AI winter, but I don't think it's going to take the form of historic AI winters.

[00:07:43]Historically, you saw perhaps most infamously well after the the Marvin mini winter from Exor not being able to be modeled with an ultra shallow neural network. You saw the winter of the 80s and 90s from Japan's fifth generation computing initiative and expert systems turning out not to be such a great approach to general intelligence. I don't think we're going to see anything quite like that. If I had to pinpoint finger to the wind, the likeliest possible source for a brief AI winter or something that gets called an AI winter, it's probably all of the capex that's being allocated to tile the earth with compute with data centers right now. And if the leading frontier labs, for whatever reason, can't provide enough revenue to justify that capital expenditure over the next two years, I could foresee a mini winter, but I don't think it'll last long. And I don't think it would even be comparable to arguably the telecom winter in the early 2000s when the US and other western countries but mostly US overbuilt telecom capacity and then it took a number of years.

[00:08:49]>> That's right. And that those were the the years of Enron as well trading fiber optic uh capacity. I don't think we're going to see anything like that because I think we're we're so deep into scaling at this point and so deep into recursive self-improvement. Unless there's some, at least to me, unforeseeable ceiling on our ability to continue recursively self-improving, I think we get to wherever we're going before we see another window. Now, it is possible that at the end of the recursive self-improvement rainbow, we discover the perfect AI algorithm, the perfect ML model, and then we hit a ceiling, and then there's a bit of a hangover while we figure out what to do with ourselves, and maybe the recursive self-improvement shock wave propagates down in the stack to improving better hardware, and then maybe better solutions at the infra level than CMOS. But is that a shock to the system in the same sense that the 80s or the the 60s winters were? I think definitely not.

[00:09:47]>> So I want to go a few um I'm going to pick a few years and ask you what you think is going to happen. So next 200 years 2020 uh 2226 will humanity uh still be earth ccentric. What are the chances we discover evidence of ETSs extraterrestrials? If you don't know what I'm talking about, what constraints on civilization today are likely to disappear completely? What is the most profound discovery humanity could make in the next two centuries? In which order would you like me to answer questions? Yeah. And and if one of those is too hard, feel free to pass.

[00:10:23]>> Yeah. I think the odds are very high. So maybe one of my perhaps mildly maybe it doesn't constitute a contrarian perspective at this point but at minimum mildly contrarian perspectives is I think the odds are very high that sometime in the next 5 to 10 years we'll have a definitive answer to the question of is humanity alone in the universe >> that's because AI is going to be able to analyze data >> I think AI is the ultimate forcing function so one way I think about it is >> according to my timelines which may by some standards s constitute very aggressive AI timelines. I think sometime in the next 5 to 10 years, humanity will possess the technological capability of building self-replicating vonoyman probes that we can send out in all directions at relativistic speeds.

[00:11:10]And if we wanted to, I think we'll have the capability to quite literally turn the rest of our galaxy to paper clips if we wanted to. We'll have advanced nanotechnology. We'll have the ability to send out replicating relativistic speed self-replicating probes. And that's a point at which if there is any other non-human intelligence in our galaxy, if I were that NHI, I'd be pretty concerned and I would view humanity's sudden leap in capabilities thanks to artificial super intelligence as potentially an existential threat. So I think that's one reason. There are a number of other reasons why if we're not alone, at least in the galaxy, galaxy is what about 50,000 lighty years or so in diameter. Uh, I would be rather concerned right now if I had any awareness of Earth and its nent super intelligence that poses an existential potentially threat to any other intelligent life in our galaxy that's capable of detecting us. So punchline, if there is non-human intelligence out there and it's able to detect us and it's aware of us, it had better make a cameo appearance sometime in the next 5 to 10 years if for no other reason than self-preservation. So if if you believe that super intelligence capabilities are on as steep a ramp as I do and we're going to get molecular nanotechnology, self-replicating robots, relativistic interstellar probes, then you should also believe that unless we are completely alone in this galaxy, we're going to have some visitors sometime soon.

[00:12:36]>> Right? So that's 200 years, 150 years from now, 2176. How might our definition of consciousness change? What aspects of human nature are likely to remain unchanged despite uh technological transformation?

[00:12:52]>> Well, I think the nature of human consciousness and I'm not even sure the question is well posed as stated is likely to get resolved in the next 10 years. So, if if you're hearing me say next 10 years a lot, it's not just a verbal tick. I really do think many of the grand challenges in math, science, and engineering are likely to get solved with enormous help from super intelligence in the next 10 years, including the nature of consciousness.

[00:13:17]I'm not even sure using that word as an avatar or as a verbal device makes an enormous amount of sense right now. But I I do think I'm higher confidence that sometime in the next 10 years there will be at least among some scientific elite inclusive of the AI scientific elite some consensus as to what consciousness is if it is a reified thing that that makes sense to speak of. We'll have a sense of what that is and then 10 years out I'm very bullish that we're going to have variety of forms of human cognitive enhancement. I think full bandwidth full dive virtual reality I think high bandwidth brain computer interfaces I think wholeb brain emulation and human mind uploading all par for the course in the next 10 years so when you ask about what consciousness looks like 100 plus years from now I would envision and expect and I don't think maybe again some might construe this as over optimism I don't think it's over optimism I think it's a realistic reasonable expectation at this point that 100 150 years from Now we will be 100 plus years into an era of human civilization when perhaps most of the intelligence in our solar system is ultimately derived from human brain uploads and not biological human meat bodies in nature.

[00:14:37]>> So I asked you about 100 years from now 200 you you keep on referring to the next 10 years. Let me ask a question.

[00:14:43]>> I'm rounding down to 10 years for every question.

[00:14:45]>> No no no yeah no all good. So specifically 10 years from now, what does a successful human AI civilization look like? What is that symbiotic relationship look like? What is how do you define success? And uh and what what are some chess moves that we need to be able to get to that kind of success?

[00:15:07]>> I think there are so many levels to what a successful endgame on a 10-year time scale looks like. I wrote a book with Peter Diamandis called Solve Everything is available publicly solveverthing.org.

[00:15:18]Encourage folks to read it if it's of interest. I think there are many different dimensions to that 10-year endgame. I think there's a technical endgame where math for example right now those following the really popcorn popping drama in the the set of Eddish problems which are a set of call them uh hard at least the the ones that are as of yet unsolved medium to hard unsolved problems in math that were enumerated by the Hungarian mathematician Paul Erdish these are starting to get bulk solved now >> what's your Erdish number >> I need to recalculate it changes is every day because I meet more mathematicians. So the Erdish number is the the number of degrees of separation via co-authorship from >> 60 degrees of separation from Kevin Bacon but different.

[00:16:05]>> It's probably two at this point.

[00:16:07]Probably two. Yeah.

[00:16:10]Now now it's changed probably since the fist bump. So math gets math gets solved for some definition of solve. My definition of math being solved is that for any given problem in math, we can be confident that if we just pour more compute into AI that we can be very confident that we're going to get a solution to the math problem that exists. And we're starting to see a variety of previously unsolved problems in math just get bulk solved by AI. So right now it's at maybe a dozen, a couple dozen.

[00:16:43]>> How does that help the everyday person?

[00:16:45]How does that change, you know, the the billions of populations?

[00:16:48]>> It's true. What what what has math ever done for me?

[00:16:51]>> Yeah.

[00:16:51]>> Uh so math, I think, is I like to say the canary that owns the coal mine. Math is the herald of physics and chemistry and biology and material science. These are all going >> close encounters to the third kind. It's how we communicate >> and and formal first contact with non-human intelligence. These are all I think downstream of at least chronologically if not causally downstream of AI bulk solving math. I think these are all likely to get solved in the next 10 years through the same underlying advances that are allowing AI to solve math. Now what does AI bulk solving physics, chemistry, biology?

[00:17:27]Well, how about cures for the top 5,000 diseases? Through AI, creating virtual twins, digital twins of individual cells and then tissues and organs and whole organisms, we'll be able to speedrun the future of medicine. I think over the next 10 years, we'll be able to speedrun, you like to talk of Star Trek, I know. How about a future where 10 years from now, not 250 years from now, we have substantially all of the physically possible inventions and discoveries that Star Trek conceives of.

[00:17:57]If they're physically possible, I think odds are very good that >> the replicator warp drive the >> replicator if they're allowed by the the laws of physics of the universe that we turn out to live in. I think we're going to see at least a theoretical understanding of them and probably unless they require exorbitant amounts of energy or other resources. I think we're likely to see early implementations of any physically possible Star Trek technology in the next 10 years.

[00:18:22]>> Thank you, Alex, for sharing the stage and thank you for the work that you do.

[00:18:26]Alex Wister gross a verb.