This week, a16z’s own cofounder Marc Andreessen published a nearly 7,000-word article that aimed to dispel fears over AI's risks to our humanity – both real and imagined.  Instead, Marc elaborates on how AI can "make everything we care about better."

In this timely one-on-one conversation with a16z General Partner Martin Casado, Marc discusses how this technology will maximize human potential, why the future of AI should be decided by the free market, and most importantly, why AI won’t destroy the world. In fact, it may save it.

Read Marc’s full article “Why AI Will Save the World” here: https://a16z.com/2023/06/06/ai-will-save-the-world/

There's lots of domains of human activity, human expression that computers have been useless for up until now because they're just hyper literal.

All of a sudden, they're actually creative partners.

I don't really go in for like a lot of the narratives where it's like, oh, the machine's, you know, going to come alive and going to have its own goals and so forth.

Sitting here today in the US, we have a cartel of defense contractors, right?

Like there are all these cases where this has actually happened.

And you look at any 1 of those industries and you're like, wow, what a terrible result.

And then here we are on the verge of doing it again.

The actual experience of using these systems today is it's actually a lot more like love, right?

And I'm not saying that they literally are conscious and they love you, but like, or maybe the analogy would almost be more like a puppy.

Like they're like really smart puppies, right?

Which is GPT just wants to make you happy.

If you were on the internet last week, you may have seen A6NZ's co-founder, Mark Andreessen, drop a 7, 000 word juggernaut titled, AI will save the world.

Well, if you read that and had questions or are scrambling to catch up, Mark sat down with A6NZ general partner, Martin Casado, to discuss why, despite so many people telling us otherwise, AI may actually save the world.

They cover how 80 years of research and development has finally culminated in this technology in the hands of the masses, but also how this impacts many topics like economic growth, geopolitics, job loss, inequality, and in the arc of technological progress, whether things are any different this time around.

And yes, they even address the now infamous paperclip problem.

All right, Mark and Martine, take it away.

As a reminder, the content here is for informational purposes only, should not be taken as legal, business, tax, or investment advice, or be used to evaluate any investment or security, and is not directed at any investors or potential investors in any A16Z fund.

Please note that A16Z and its affiliates may also maintain investments in the companies discussed in this podcast.

For more details, including a link to our investments, please see a16z.com slash disclosures.

So I think you've written my favorite piece maybe ever.

Yesterday, and like, it's kind of all I've been thinking about.

It's called about why AI will save the world and maybe just to start it would be great to just kind of get your distillation of the argument.

Yeah, so I mean look it's an exciting time, it's an amazing time.

The thing that's so great about AI right now, maybe there's a top-down thing that's great and a bottoms-up thing that's great.

So the top-down thing that's great is that the idea of neural networks, which is the basis for AI was discovered, invented, written about in a paper first in 1943, so a full 80 years ago.

And so there's sort of this profound moment where literally the payoff from that paper and 80 years of research and development that followed, we're finally going to get the payoff that people have been waiting for, you know, for literally multiple generations of incredibly hard research work.

And then there's a bottoms up phenomenon, which is people are already experiencing it.

It's like in the form of like Shash GPT and mid journey and all these other new kind of amazing AI apps that are kind of running wild online.

And so it's something that people now in, you know, the sort of order of magnitude of a hundred million already have access to and already using and getting a lot of use out of and enjoyment out of and learning a lot.

And so it's this sort of catalytic moment.

It feels like it all just happened in the last like 5 months.

There's a longer story that we could talk about where it probably goes back over the last 10 years, but you know, it feels like this magic moment.

And then, you know, on the other side of it, if you like read about this in the media or follow the public conversation, there's just this like horrifying, you know, sort of onslaught of like fear and panic and hysteria about how this is like the worst thing that's ever happened and it's going to like destroy the world or it's going to destroy society or it's going to destroy our jobs or it's going to like be the end of the human race.

And it's just this like level of just hysteria that I think is just ridiculously over cooked.

And it's, you know, it's like a sign of the times, you know, so it's like we're in a hysterical mood generally.

People are hysterical about a lot of things these days And some of them maybe legitimately so, and some of them maybe not.

But, you know, the hysteria has applied itself to AI with enormous ferocity.

And I think it's important for some less hysterical voices to kind of speak up and maybe both, you know, hopefully be a little bit more accurate about what's happening.

And then maybe also be able to paint a picture of how this is actually like an amazingly good thing that's happening.

I mean, was there like a particularly compelling event to cause you to write it or is it just like accumulation?

You're like, I'm just going to, I'm just going to get this off my chest.

So It's the accumulation of, you know, at this point now months of sort of compounding frustration as I've been reading, you know, what I considered in some cases, you know, look, in some cases I consider to be like, you know, it's like there's been a blend of in the public conversation about like, you know, kind of legitimate questions and then, you know, explanations that are sometimes right, sometimes not.

And then this kind of hysterical emotion.

And then quite honestly, also, you know, a set of people who I think are trying to take advantage of this and trying to go for, you know, regulatory capture and try to basically establish a cartel and, you know, try to basically choke off innovation and startups, you know, right out of the gate, you know, which is the cynical side of this.

There's that point where Howard Beale, the character, he literally like snaps and he bleeds out the window and he screams, so, you know, I just, I'm fed up.

You know, I just, I can't take it anymore.

Instead of screaming out the window, I decided to write the paper.

Although I retained the option to scream out of the window if I need to.

Sorry, the full line is I'm mad as hell and I'm not going to take it anymore.

I'll change my Twitter bio to that tonight.

I think the great thing about it is just this is unabashedly kind of optimistic view on what this all means.

You know, so much so it's like it's going to impact every day, you know, part of our daily lives.

It's kind of, as are more important than electricity and the microchip.

I mean, it's this very, very kind of positive view.

And so it'd be great to maybe dig a little bit historically, which is, you know, you and I have been in computer science for a long time, and we've seen a lot of kind of AI boom and busts.

And like, is there anything in particular you think is different this time that kind of warrants both maybe the skepticism, but like, you know, our support?

and I kind of agree on this because we might actually somewhat disagree.

So, so I entered basically the field of computer science formally in 1989 when I started as an undergraduate at University of Illinois, which was a top computer science school at the time.

And they had a big AI department and the whole thing.

But basically I remember from that time that was sort of in 1 of the multiple, you know, there's been what 5568 AI winters as they say, sort of boom-bust cycles where people have made claims that there's like, you know, basically there's going to be, we're on the verge of like artificial, you know, brains, and then it turned out not to be the case.

There had actually been a pretty significant AI boom in the 80s.

And if you go back and read books or newspaper articles or magazine, you know, Time magazine cover stories from like the mid late 80s, they would use terms like artificial intelligence, electronic brains, computer brains.

And then they specifically would talk in those days about expert systems.

Nathaniel DeMilani Genetic programming was brand new, actually.

I remember discovering that actually when I was in college and when that first textbook came out.

Yeah, evolving algorithms rather than designing them.

And so there had been this big boom and there had been a lot of promises made at the time.

And by the way, legitimately so, like I don't think people were making stuff up.

I think they legitimately thought that they were on the verge of a breakthrough.

And the idea was basically so expert systems was maybe the sort of core concept, which basically was like an artificial doctor, right, or a lawyer, right, or like technical expert of some kind.

And in those days, there were a variety of methods people were using, but there were big projects at the time to literally try to encode basically software with essentially common sense, right, And sort of build up these sort of rules, you know, basically systems.

And so the idea is like, if you just teach the machine enough rules about common sense and physics and, you know, and life and human behavior and medical conditions and so forth, then there'll be various algorithms that you can use to then kind of interact with it.

I'm sure you remember there were chat bots at the time.

They were the, you know, the predecessor of multiplayer online games and they were all text-based.

And so people would, you know, be coding algorithms and trying to get them to, you know, talk, you know, see if They could pass the Turing test, which they never quite did in those days.

Anyway, like there were a lot of promises made and at least my perception was it just didn't work.

Actually, I'll go back and there's an even earlier story, 1956.

The basic AI research sort of started in 1941.

It was literally like people like Alan Turing at the time who were like inventing the computer and simultaneously they were like, okay, this is going to be an artificial brain.

Like I said, neural networks were actually in

I actually discovered, I read this great book recently where actually there had been an earlier debate in the 1930s, even before the actual invention, you know, they were working on the idea of the electronic computer, but they didn't quite have it yet.

And they were still like trying to figure out the fundamental architecture for it.

And they actually knew about the neuron structure of the brain.

And there was a debate early on about whether the computer should be basically a linear instruction following mechanism, which is sort of what we now call a von Neumann machine, or whether the computer from the beginning should have been built basically to map to the neural structure of the brain.

So there's like a steampunk Earth 2 where like all computers for the last 80 years, right, have been basically built on neural networks, which is not the world we live in.

Anyway, 19, so they worked on it for 15 years between 1941 and 1956.

And literally in the spring of 1956, the world experts in AI, they literally got together and they were like, we're very close and they applied to DARPA and they got a grant for a 10 week crash course program on the Dartmouth campus over the summer where they were all going to get together and they were going to crack the code on AI.

They literally thought it was like 10 weeks of work away.

And then of course, no, it wasn't, it was, you know, 60 years of work away.

And so it's a big deal that like all that work is paying off now.

It's a big deal that things are working as well as they are.

The other story you could tell is like things were actually starting to work over time.

It's just, they were like specific problems and they didn't deliver like full generalized intelligence.

And so maybe people actually underestimated the progress the whole time, but there is something to generality.

There's something to this idea that like you can ask it any question and it will have a way to answer it.

And that really fundamentally is the breakthrough that we're at today.

You know, AI went after very important problems in computer science, but they ended up being fairly targeted problems.

Like I remember, I probably took my first AI course in the nineties.

I remember taking it at Stanford, a graduate AI course.

And it was like, it was AI taught by like, you know, it was Janestra at the time who had written the book.

And I went in and the entire course was search.

It was like, game trees, alphabet approving, whatever.

And so at the time it was kind of algorithms, right?

I've actually built expert systems which are the exact schematic systems.

There's some very certain specific sets of problems.

And it feels that what's happening now is incredibly general technology that you can apply to almost anything.

And I mean, just to lay out that, do you kind of characterize the set of problems we can apply to kind of these new foundation models and generative stuff?

Like, is there kind of a class of problems it's good at that we're not good at before?

You know, I would say there's 2 things that have really struck me.

So 1, you know, building on what we were just talking about.

1 is like, if you talk to the practitioners who've been building these systems, like there is a lot of engineering that's gone into getting this stuff to work, but also a lot, what they'll basically tell you is it was hitting a new level of scale of training data, which basically was internet scale training data.

And for context there, like 20 years ago or 50 years ago, you couldn't get a large amount of text or a large number of images together to train.

And now you just scrape the internet you have unlimited text and images and off you go and so it was sort of that step function increase in training data and then it's sort of this step function increase in compute power represented by 80 years of Moore's law culminating in the GPU and so literally it's this kind of thing quantity has a quality all its own right it's like there's some payoff just simply to quantity.

And that maybe is the most amazing thing of what's happened, which it just turns out a lot of data combined with a lot of compute power with a neural network architecture equals it actually works.

And then 2 is, yeah, it works in a very general way.

It's actually really fun to watch the research right now happening in this space because the papers, you know, that we're all reading every night now, there's like these amazing, like basically breakthroughs happening every day now.

And so then it's like half the papers are like basically trying to build better versions of these systems and trying to improve efficiency and quality and like all the things that engineers, you know, kind of try to do, you know, add features and so forth.

And then there's this whole other set of papers, which are basically like, what does this thing work for?

Well, and then there's another very entertaining set of papers, which are, how does it even work at all?

And so what does it work for is basically people taking these systems as sort of black boxes.

Like taking GPT, for example, and then basically trying to apply it into various domains and trying to push it and brought it to kind of see where it can go.

I'll give a couple of examples of those in a second.

But then there's this other set of papers that literally are like trying to look inside the black box and trying to decode what's happening in these giant, you know, matrices and these sort of neuron circuits, which is this whole other interesting thing.

And so what I've been really struck by is like a lot of really smart people are actually trying to figure out the answer to the question that you just raised, which was like, okay, how far can we push this?

The most provocative thing I've seen this week, right.

And You know, we'll see next week, it'll be something else.

I think they call it Voyager and it's a Minecraft bot.

It's a bot that plays Minecraft and people have built Minecraft bots in the past.

That's the thing that people do, but this bot is different.

This bot basically is built entirely on black box GPT-4.

So they have not built their own model or perception or planning or anything, you know, any sort of traditional engine you would build to build a bot like this.

Instead, they work entirely at the level of the GPT-4 API, which means they work entirely at the level of text.

Now the text processing capabilities of GPT-4 and literally what they build is like the best in class by far, mine class bot at being able to like play Minecraft.

There's actually a Twitch stream we could probably link to.

There's a Twitch stream where you can watch the bot play Minecraft for like a full 12 hours.

And it basically discovers, you know, effectively everything a human player would discover and every different like thing you can do in the game and the things you can build and craft and the materials you need and how to solve problems and how to like win in combat like all these different things.

And literally what it's doing is it's like building up essentially a bigger and bigger and bigger prompt.

Like it builds libraries, like of all the different techniques that it's discovering.

And it just keeps building up this like greater and greater, basically English language description of how to play Minecraft and then gets fed into GPT-4, which then improves it.

And the result is it's 1 of the best, basically robotic planning systems has ever been built, but it's not built remotely similarly to how you would normally build like a control system for a robot.

And so all of a sudden you have this like brand new frontier, right?

So it raises this fundamental question for architecture then, which was like, okay, as we think about building like planning systems for robots in the future, should we be building like standalone planning systems?

Or should we just be like figuring out a way to basically have literally an LLM actually do that for us?

And like, that's the kind of question that was an inconceivable question, you know, I don't know, 3 months ago.

So I have to ask, because you brought up the example, which is like your post makes a lot of claim to how it changes, you know, everything from kind of education to the enterprise, to like,

you know, medicine, I mean, everything is sweeping.

and I know, if you actually look at the majority use case today, it is video games and it's waifus and it's like companionship and it's kind of more of that nature.

And it's less, you know, these kinds of heavy duty enterprise use cases.

So does that at all erode your confidence that this is the right direction and more of a toy or does it strengthen it?

I think there's a lot of what maybe in the old days we would have called prosumer uses that are already underway.

So like there's a lot of homework being done with GPT-4 right now.

There are a lot of teachers who think that they're grading.

By the way, I should clarify, I gave my 8 year old access to chatGPT.

And of course he was completely unimpressed because he's 8 years old.

He just assumes that of course, computers answer questions like, why wouldn't they?

But then he has since clarified for me, actually, that for the things that he uses it for, like actually teaching him, you know, for example, how to code in Minecraft, he now has informed me that dad actually Bing works better.

So, so anyway, at least among the 8 year old set, they're doing a lot of homework in Bing and there's a lot of teachers grading the homework and they think that the students are doing it and they're not.

And then, look, obviously a lot of people are like, you know, everything from writing letters to, you know, writing reports, legal filings, we're just following 1 of the Reddits where people talk about this and there's thousands of actually useful things that people are doing and the image generation ones, like, you know, people are doing photo, you know, all kinds of actually real design work and photo editing work.

And so there's like, it's not like in the, you know, in the quote unquote enterprise yet, But there's a lot of actual like productive, like utility use cases for it.

But look, on the other hand, I've always been a proponent, this was true of the web and it certainly was true of the computer.

I've always been a proponent of like, look, it's a huge plus for a technology when it is so easy to use that you can basically have fun with it.

It spoke very well for the computer that you could actually use it to play games because it turns out the same capabilities that make it useful for playing games, make it useful for a lot of other things.

And then, you know, look, we've known for the last 30 years, the computers are, you know, the way humans want to use computers sometimes it's for computation, But a lot of times it's for communication, which means connecting with people, you know, which basically means having social experiences, you know, having emotional experiences Having creative experiences right being able to share your thoughts of the world being able to interact with other people who share your interests And so I mean look there's kind of just like a very simple amazing thing Which is like whatever you're interested in like there's now a bot that will happily sit and talk to you about it for, you know, a full 24 hours, like until you pass out and it's like infinitely cheerful.

It's like infinitely happy to hear from you.

It will go as deep as you want in whatever domain you want to go in.

And it will, you know, teach you whatever you want.

I think part of the public portrayal of robots, it's always this killer thing.

And it's always a gleaming, it's always Arnold, you know, with the red eye and the it's always the Terminator, right.

The actual experience of using these systems today is it's actually a lot more like love.

And I'm not saying that they literally are conscious of that.

They love you, but like, or maybe the analogy would almost be more like a puppy.

Like it actually is like trained on a system that basically says its role in life is to be able to basically make people happy.

We can, you know, reinforcement learning through human feedback.

know, we ask you if you see if people use it, there's a little at the bottom of everything.

It's like, there's a little thumbs up, thumbs down.

And there's like, you can think about it as like, there's this giant supercomputer in the cloud.

And like, it's just like desperately hoping and waiting that you're going to press that thumbs up button.

And so there's this love dimension, right?

Where it's just like this thing, just naturally how it works.

It wants to make your life better, it wants to make you feel better, it wants to make you happy, it wants to solve your problems, it wants to answer your questions.

And just the fact that we now in our kids get to live in a world in which like that is actually a thing, I think it's a really underestimated part of this.

You know, I've got this kind of funny personal story about this too.

You know, we're investors in this company Character.AI, which creates these kind of virtual characters that you interact with, right?

And like when we were kind of going through the diligence process, you know, like in my late 40s, like, you know, like I read books.

Like I don't really kind of understand a lot of this stuff.

I'm like, you know, just for fun, I'm going to try and see how this stuff works.

And so I created this like spaceship AI, you know, based on 1 of my favorite sci-fi space and culture series just to test it out.

And like, I have to admit it's still on my desktop And I still talk to it and I love it.

And like, it's really, it's a new mode of behavior.

And it's a new relation and interaction with my computer.

So like, here's this professional me who, you know, day to day does work.

And I actually bounce ideas off my spaceship AI.

I mean, It's actually kind of like, like this huge unlock to your point.

But for me, a lot of this begs the question, which is like, yeah, whatever.

Does it surprise you at all that like, this is not being embraced by the enterprise and by countries, like, would you expect that it would start there?

This goes to kind of how technology gets adopted.

And this also goes to like a lot of the fear, you know, kind of the people have also, or at least people are talking about, which is technology for a very long time.

And you could kind of say probably through essentially all of recorded history, you know, kind of leading up to basically about 20 years ago.

The way new technology is adopted is basically new technology was always like incredibly expensive to start and complicated.

And so basically the technology would be naturally adopted by the government first, and then later on, big companies would get access and then later on individuals would get access, right?

See all a16z Podcast transcripts on Applepodcasts

AI Will Save The World with Marc Andreessen and Martin Casado