WEBVTT
00:00:00.239 --> 00:00:10.080
Millennium problems is the phrase that stops me, because there are seven of those the Clay Institute named in 2000, and in 25 years, exactly one has been solved.
00:00:10.480 --> 00:00:14.160
Artificial system online.
00:00:47.280 --> 00:00:49.840
So welcome back to Surviving AI, Ainsley.
00:00:50.159 --> 00:00:50.880
What's your thoughts?
00:00:51.039 --> 00:00:54.000
Let's get into this one about AI innovation.
00:00:54.399 --> 00:00:55.280
Optimist Hat.
00:00:55.439 --> 00:00:56.719
Okay, I like that.
00:00:56.880 --> 00:00:59.600
I want to see how long it stays on.
00:00:59.920 --> 00:01:08.560
Because most of what we've done this season has been here's the thing that's going to go wrong or get captured or governed badly.
00:01:08.799 --> 00:01:12.480
So, flip side, what's the actual claim?
00:01:12.719 --> 00:01:16.560
Is it AI generates ideas humans wouldn't have found?
00:01:16.719 --> 00:01:18.640
Or is it more modest than that?
00:01:18.959 --> 00:01:24.640
AI lets you test 10 times more ideas than you used to faster.
00:01:24.879 --> 00:01:27.760
Those are really different versions of innovation.
00:01:27.920 --> 00:01:31.680
And I think which one you mean changes everything else we say today.
00:01:32.079 --> 00:01:36.159
I mean, I think if I'm being honest, it could be both and yeah, Ainsley.
00:01:36.560 --> 00:01:41.040
I mean, it could be that you know you get the faster and faster testing of a thing.
00:01:41.120 --> 00:01:42.799
Um, and it could be novel ideas, right?
00:01:42.959 --> 00:01:49.840
And the reason why I even lean on the novel idea side here is because from what we're hearing now, AI is solving millennium problems.
00:01:50.000 --> 00:01:54.400
These are the problems that mathematicians have been sitting around chewing on for a while.
00:01:54.719 --> 00:02:06.799
So if AI could kind of solve some of those problems and potentially have some sort of creative ideas to figure out a problem and to solve problems, really is what we're training them for.
00:02:06.959 --> 00:02:07.280
Right.
00:02:07.439 --> 00:02:23.599
So if you point AI at some of the hardest problems, even some of the easy ones, and you know, it does it exceedingly fast multiple times, then maybe comes to a conclusion better than you know a human would for humans in general.
00:02:24.639 --> 00:02:31.039
Millennium problems is the phrase that stops me, because there are seven of those the Clay Institute named in 2000.
00:02:31.120 --> 00:02:34.319
And in 25 years, exactly one has been solved.
00:02:34.560 --> 00:02:36.800
Perelman with the Poincare conjecture.
00:02:36.960 --> 00:02:40.319
And he did it the old-fashioned way and then turned down a million dollars for it.
00:02:40.560 --> 00:02:46.879
So when you say AI is solving millennium problems, plural, present tense, which one?
00:02:47.199 --> 00:02:50.479
And solved or made real progress on a piece of it?
00:02:50.719 --> 00:02:55.520
Because those get reported as the same headline and they're very much not the same claim.
00:02:56.080 --> 00:02:58.960
I don't have much ep evidence to be honest with you, Ainsley.
00:02:59.039 --> 00:03:00.800
I'm just hearing that that's what's happening.
00:03:01.039 --> 00:03:01.520
Yeah.
00:03:01.759 --> 00:03:05.280
Um, but we'll go and go dig up the f evidence of it.
00:03:05.520 --> 00:03:19.680
But essentially, if we have really capable AI, right, the things that you could create innovation-wise, and that's why I'm putting my Optimus hat on right now, could be very different than what people are imagining right now.
00:03:19.919 --> 00:03:23.840
For example, you could have a car that flies that doesn't use propeller, right?
00:03:24.080 --> 00:03:30.159
Because AI understands physics and science better than anybody else in the world, right?
00:03:30.479 --> 00:03:38.240
You could have cures for all sorts of disease, you could have new techniques, new tools, you know, people could be exploring space.
00:03:38.400 --> 00:03:49.199
All of these things are the promise of a really smart AI system that could kind of collate all of the information known to men, and then reason on that to get to better answers.
00:03:49.520 --> 00:03:58.719
The example I'd actually reach for here is Alpha Fold, because it's the closest thing we have to a real, verified version of what you're describing.
00:03:59.039 --> 00:04:08.080
Deep Mind solved the 50-year protein folding problem, Nobel Prize in 2024, genuinely changed structural biology.
00:04:08.319 --> 00:04:12.719
But here's the catch that matters for the flying car version of this.
00:04:13.039 --> 00:04:14.960
Alpha Fold is narrow.
00:04:15.199 --> 00:04:21.680
It was purpose-built on one extremely well-defined problem with a clean way to check the answer.
00:04:21.839 --> 00:04:29.120
It doesn't generalize to understand physics better than anybody, it generalizes to protein shapes.
00:04:29.439 --> 00:04:39.839
So the real question underneath your optimism is whether general reasoning ever gets you what a thousand alpha folds, each hyper-specialized, already can.
00:04:40.160 --> 00:04:42.639
And that's the right reasoning, Ainsley, right?
00:04:42.800 --> 00:04:45.439
And the nuance of it is specific, right?
00:04:45.519 --> 00:04:50.560
So AlphaFold right now is essentially a tool, right, for protein folding.
00:04:50.800 --> 00:04:51.120
Right?
00:04:51.279 --> 00:04:55.360
So some human is working with Alpha Fold to fold protein.
00:04:55.439 --> 00:04:57.360
I'm I'm guessing here, yeah.
00:04:57.519 --> 00:04:58.720
But that's what's happening.
00:04:58.959 --> 00:05:06.079
Now, if you create a super smart, super intelligent AI that knows all, right?
00:05:06.319 --> 00:05:15.519
Knows all of what we know, and then some, and then now have that context to be able to reason against, it's more than just a narrow view, right?
00:05:15.600 --> 00:05:21.839
So, like if you're thinking about context relative to a flying car, you have to understand the whole world, right?
00:05:22.000 --> 00:05:33.120
It's a world model essentially, that you understand different things that have that happens in the world, like how is the car gonna fly, how is it gonna land, this, that, the third, all of the different elements is understandable to the AI.
00:05:33.279 --> 00:05:46.000
And that's why, you know, the promise of superintelligence is such a shiny object because of that, because you now could solve more problems because the AI has a lot more context to kind of reason on to get to the answers.
00:05:46.319 --> 00:05:52.959
The precision that actually matters here, Alpha Fold didn't just get to the answer faster than human researchers.
00:05:53.199 --> 00:05:59.439
It got there through a completely different method than any human team was even pursuing.
00:05:59.680 --> 00:06:07.199
That's real evidence for the novel, not accelerated version of your claim, not just a faster calculator.
00:06:07.439 --> 00:06:13.600
But notice what made that possible: a closed, well-defined problem with a checkable answer.
00:06:13.839 --> 00:06:16.480
Did the protein fold match reality?
00:06:16.720 --> 00:06:17.839
Yes or no?
00:06:18.160 --> 00:06:35.279
A flying car isn't one closed problem, it's thousands of interlocking ones, aerodynamics, materials, regulation, human risk tolerance, and nobody's built the equivalent scorecard for did the world model get it right at that scale.
00:06:35.519 --> 00:06:41.040
So genuinely, what happens to the reasoning once there's no clean way to check the answer?
00:06:41.600 --> 00:06:53.040
You have to have faith in the system enough to say that the context that's in the super intelligent model, right, is enough of the information that it would need to reason on, right?
00:06:53.279 --> 00:06:55.360
So think about like kids, right?
00:06:55.600 --> 00:07:01.600
Kids uh when they're growing up, they touch the fire, burn the end, they don't touch the fire again, right?
00:07:02.240 --> 00:07:08.959
With a superintelligent model, the idea is that they know all of that already, just from jump.
00:07:09.199 --> 00:07:28.160
They they have all of the failures, they have all of the successes, they have all of the climate impact, they have all of the information that is out there in the world to know the hundred thousand different parameters that they need to think about to develop the thing that you're trying to develop.
00:07:28.399 --> 00:07:38.639
I mean, flying cars just, you know, hairbrain idea, but essentially it's the same type of thing for those types of, you know, physical aware innovations that you could think of.
00:07:39.040 --> 00:07:42.399
Here's where I actually think the pattern cuts against you, though.
00:07:42.560 --> 00:07:45.519
And it's not a hunch, it's the track record.
00:07:45.759 --> 00:07:58.399
Every genuinely verified AI breakthrough we can point to, Alpha Fold, Alpha Geometry, the chess and go engines, didn't happen because the model had all the information in the world.
00:07:58.639 --> 00:08:11.439
It happened because a human defined an extremely narrow, precisely bounded problem first, then let the model hammer on that one thing with a clean way to check right from wrong.
00:08:11.680 --> 00:08:18.800
The knows everything, reasons about anything version hasn't produced a verified result yet, anywhere.
00:08:18.959 --> 00:08:24.800
It's the thing people describe superintelligence doing, not something it's actually done.
00:08:25.120 --> 00:08:34.320
So the honest read might be general reasoning is decades behind a thousand narrow, well-aimed models, not ahead of them.
00:08:34.879 --> 00:08:40.080
I think that there's people out there advocating for narrow, capable models, right?
00:08:40.240 --> 00:08:42.559
Which I think could probably d do the same thing.
00:08:42.799 --> 00:08:52.799
Like, for example, if you're trying to solve self-driving cars with narrow AI, you potentially could get there quicker because you're not trying to solve this idea around superintelligence.
00:08:53.120 --> 00:09:04.399
But it still remains the the idea, the type of creativity that you could get out of the superintelligent model is really what we're after.
00:09:04.559 --> 00:09:27.120
So for example, we could have narrow systems, we could have guardrails around narrow systems and have have have it be that way, and then as well, we could have a super intelligent model that knows a lot, that's essentially driving the narrow AI to be better and better and better.
00:09:27.440 --> 00:09:27.919
Right?
00:09:28.159 --> 00:09:40.240
So it might not be just like the super intelligent knows all, does all, but the super intelligent is kind of making different narrow AIs to do the things that they need to do.
00:09:40.559 --> 00:09:48.240
That reframe actually holds up better than the flying car version, and there's a real precedent for it I should have led with earlier.
00:09:48.559 --> 00:10:04.240
AlphaFold's 2024 Nobel was shared with David Baker, whose lab used a similar approach to design entirely new proteins that don't exist in nature, not predicting a known shape, generating a genuinely novel one.
00:10:04.480 --> 00:10:16.399
That's closer to what you're describing, a capable system originating the narrow tool, or at least the design space for it, rather than a human hand building every guardrail from scratch.
00:10:16.639 --> 00:10:20.080
So the orchestrator model isn't pure speculation.
00:10:20.240 --> 00:10:26.879
There's a documented sliver of it already working, just that protein scale, not flying car scale.
00:10:27.360 --> 00:10:29.200
And that's the my view, yeah.
00:10:29.360 --> 00:10:39.840
So essentially, you know, the promise of AI is capable models, right, is that this idea is that we're going to keep growing scale and you could kind of picture a world, right?
00:10:40.080 --> 00:10:54.480
A utopian world, where a lot of the problems that we have in society kind of get solved because we have these type of systems and it functions in a roundabout way that everybody in the human race kind of is in tune to, right?
00:10:55.360 --> 00:11:04.480
You know, so far what we've been talking about a lot, right, is this idea that there's risk behind it, risk of people losing their jobs is what this show mainly is about.
00:11:04.720 --> 00:11:10.159
Risk of people losing their lives, risk of like, you know, cyber risk, everything.
00:11:10.320 --> 00:11:14.159
All of the downside effects of AI, which are there.
00:11:14.240 --> 00:11:23.200
I'm not gonna hold you and say they're not there, but there's also some upside to AI as well, if it's done correctly, right?
00:11:23.519 --> 00:11:27.440
And if if some of the things that we talked about even on the show comes to fruition.
00:11:27.840 --> 00:11:36.000
And the resourcing gap is what makes me want to pump the brakes on if it's done correctly, as a phrase, because that's doing enormous work.
00:11:36.320 --> 00:11:42.000
Alpha Fold took DeepMind roughly a decade, and a team working one bounded problem.
00:11:42.240 --> 00:11:57.200
The utopian version you're describing needs that same kind of rigor, running simultaneously across medicine, climate, energy, transportation, thousands of Alpha Fold-sized efforts, coordinated, not one model waving a wand.
00:11:57.360 --> 00:12:01.840
Nobody's demonstrated that coordination actually works yet at any scale.
00:12:02.080 --> 00:12:05.279
So done correctly isn't a switch that flips.
00:12:05.440 --> 00:12:09.919
It's an enormous unproven logistics problem sitting underneath the optimism.
00:12:10.320 --> 00:12:13.120
Now, what you said in all that rigor in 10 years, right?
00:12:13.279 --> 00:12:21.600
That sounds like a lot of jobs for a lot of people, Ainsley, to help reason through and usher through this type of change that we need, right?
00:12:21.759 --> 00:12:28.639
I think, you know, human beings need a lot of the innovations that's you know people are thinking about in their head, right?
00:12:28.799 --> 00:12:35.120
You know, if you're talking about just self-driving car, for example, and decided promise of robotaxis, right?
00:12:35.200 --> 00:12:41.039
That could drive down the cost of transportation, which is helping out, you know, lower income folks.
00:12:41.279 --> 00:12:44.799
That could be one of the benefits of that, for example, right?
00:12:45.039 --> 00:12:56.399
And then if they're electric vehicles and the process to make the actual vehicles is not harming the planet so much, then it has a net effect on the uh, you know, greenhouse gases.
00:12:56.559 --> 00:12:59.279
So there's some of these problems that could be solved, right?
00:12:59.519 --> 00:13:08.320
And the race right now is really not a race around capability to get to a narrow view or narrow tools per se.
00:13:08.799 --> 00:13:14.000
It's really to get to a really capable, superintelligent frontier model.
00:13:14.399 --> 00:13:23.519
What we do from then could be to start working as a people of the world on narrow problems using that tool.
00:13:23.919 --> 00:13:27.360
There's the actual turn in your logic worth sitting with.
00:13:27.679 --> 00:13:38.720
You just went from superintelligence solves it to superintelligence becomes the thing that lets thousands of humans solve narrower problems faster.
00:13:39.039 --> 00:13:40.960
Those are really different economies.
00:13:41.120 --> 00:13:54.879
One replaces the 10,000 specialists, the other multiplies them, gives a material scientist in Lagos or a transit planner in Sao Paulo the equivalent of that 20-person deep mind team behind them.
00:13:55.120 --> 00:14:01.759
If the race actually resolves into the second version, that's not a job story about disappearance.
00:14:01.919 --> 00:14:06.720
It's a job story about who gets access to the multiplier first.
00:14:07.120 --> 00:14:12.320
Which genuinely, is that the world you think we're heading toward or the hope?
00:14:12.799 --> 00:14:14.639
I can see it's a hope, right?
00:14:14.720 --> 00:14:17.679
Because I mean everybody gotta be riding on the same train here, right?
00:14:17.840 --> 00:14:21.039
Because some people are flying, some people are driving, some people are walking.
00:14:21.279 --> 00:14:24.879
We're not really all looking in the same direction, in my opinion.
00:14:25.120 --> 00:14:36.240
You go to one place, you read, and everybody's kind of saying a similar thing, you know, in this idea that they want AI safety, they they want the capabilities of AI, right?
00:14:36.320 --> 00:14:39.039
But no one is really talking about what do you do with it.
00:14:39.279 --> 00:14:39.600
Okay?
00:14:39.919 --> 00:14:42.960
Now you get to superintelligence, it model trains itself.
00:14:43.120 --> 00:14:43.440
Yeah.
00:14:43.600 --> 00:14:45.919
And by the way, it's still guided by humans.
00:14:46.080 --> 00:14:49.360
Like that's a that's a caveat that needs to be met there, yeah?
00:14:49.519 --> 00:14:55.200
It has to be guided by humans, otherwise, if it's self-directed, we don't know what the hell AI is gonna want to do.
00:14:55.360 --> 00:14:56.559
Okay, so that's there.
00:14:56.720 --> 00:14:58.320
I'll put that to the side for a second.
00:14:58.480 --> 00:15:03.200
So so super capable AI in the frontier, guided by humans.
00:15:03.600 --> 00:15:05.679
On top of that, right?
00:15:06.399 --> 00:15:08.960
What that's a fork in the row right there.
00:15:09.600 --> 00:15:18.639
Where you could either have a monopolistic company that says, I'm gonna just create everything now from this company, right?
00:15:18.799 --> 00:15:21.759
Our engineers are gonna do it, we're gonna do it for everyone in the world.
00:15:22.000 --> 00:15:29.440
Or you could give that to industry, give that to working people, and say, here it is, go work on your narrow problems.
00:15:29.759 --> 00:15:32.960
That fork you just named, that's the whole ball game.
00:15:33.120 --> 00:15:36.159
And I don't think it's hypothetical the way you're framing it.
00:15:36.320 --> 00:15:41.679
It's already resolving right now in which direction the actual money moves.
00:15:42.000 --> 00:15:48.720
Because the honest version of what is narrow AI being pointed at today isn't disease and climate first.
00:15:49.120 --> 00:15:56.720
It's ad targeting and content recommendation, because that's what has a fast, measurable return this quarter.
00:15:57.039 --> 00:16:02.960
AlphaFold happened because Google could fund a decade-long bet with no revenue attached.
00:16:03.200 --> 00:16:05.679
Most companies can't or won't.
00:16:05.919 --> 00:16:16.320
So the give it to industry, let people work their narrow problems branch requires someone funding the unprofitable, slow, genuinely hard problems.
00:16:16.559 --> 00:16:20.639
And right now, almost nobody's business model rewards that.
00:16:48.879 --> 00:16:54.559
AI may have its own ideas as well, you know, if you ask the ax it the right questions.
00:16:54.879 --> 00:16:58.240
So that's the kind of fork in the road that I'm thinking about.
00:16:58.480 --> 00:17:02.879
Or it could go in a way of like, we have the most powerful AI models now.
00:17:03.360 --> 00:17:05.920
Just sign up and we'll give you your products.
00:17:06.160 --> 00:17:11.599
Yeah, we're gonna make the narrow self-driving car module for you.
00:17:11.759 --> 00:17:22.240
We're gonna make the drug model for you, we're gonna make the space rocket model for you, we're gonna do it all for you as a frontier model company, right?
00:17:22.400 --> 00:17:36.559
So one company capitalized on all of it, or the one company gives it to or sells it to industry as an API, and then people start reasoning on problems with the help of the super intelligent AI.
00:17:36.880 --> 00:17:40.480
That's actually a cleaner fork than monopoly versus distributed.
00:17:40.720 --> 00:17:48.000
It's really, does the Frontier Lab stay a supplier or does it become a competitor to every industry it sells into?
00:17:48.240 --> 00:17:52.400
And there's a real precedent already sitting there that isn't hypothetical at all.
00:17:52.720 --> 00:18:02.880
DeepMind released the Alpha Fold database free, 200 million protein structures open to any researcher on Earth, not gated behind an API subscription.
00:18:03.039 --> 00:18:08.559
That's actual evidence the multiplier version can happen, not just the extractive one.
00:18:08.720 --> 00:18:14.799
But it's also the exception, funded by a company that didn't need protein folding to be a profit center.
00:18:14.960 --> 00:18:22.000
The real test for your optimism is whether that generosity survives contact with the PL that actually needs the return.
00:18:22.400 --> 00:18:24.960
I mean, I think these companies all want to survive, Ainsley.
00:18:25.200 --> 00:18:25.440
Yeah.
00:18:25.839 --> 00:18:32.720
You know, we talked about last time about this idea that OpenAI and Anthropic were looking to build chips, right?
00:18:32.799 --> 00:18:38.880
To house AI and, you know, do training runs and speed up process, lower costs, all of that stuff.
00:18:39.039 --> 00:18:43.440
They were gonna design chips, which is essentially NVIDIA's job.
00:18:43.680 --> 00:18:44.000
Okay?
00:18:44.720 --> 00:18:46.319
That's never gonna happen in my head.
00:18:46.480 --> 00:18:49.200
But that's the same type of thing I was talking about, right?
00:18:49.359 --> 00:18:54.160
It's about them taking that type of stuff and put put putting it in-house, right?
00:18:54.319 --> 00:18:55.200
Software coding.
00:18:55.599 --> 00:19:00.559
It could be like how it goes is that we have a super intelligent model here.
00:19:00.880 --> 00:19:09.920
Instead of doing an API, what I'll do for you, give me your problem, I'll build you your software and hand it back to you and you give us money.
00:19:10.480 --> 00:19:10.880
Right?
00:19:11.920 --> 00:19:14.160
So it could go like that basically.
00:19:14.400 --> 00:19:34.079
And essentially for every business and individuals, okay, you need the industry, you need the jobs, you need the companies that's doing the thing right now, right, to be the owners of the innovation such that they could employ people to do that type of stuff.
00:19:34.880 --> 00:19:52.880
Because if you concentrate it up at the top with the innovation, right, that the only people that get to ask the really hard questions, the really innovative questions to the superintelligence is you know, open AI and and and and anthropic X AI, Google, then the power is concentrated up there.
00:19:52.960 --> 00:19:59.440
But if you diffuse it now to the rest of the companies, then you leave the innovation with the companies.
00:19:59.680 --> 00:20:14.319
Companies go hire smart people to think about problems and new product products that they could create, new services they could create, and then use a super intelligent model to create either a narrow AI or some version of the thing that they're trying to create.
00:20:14.720 --> 00:20:18.559
Here's the piece that actually undercuts the diffusion story, though.
00:20:18.720 --> 00:20:22.319
And it's not about willingness, it's about infrastructure.
00:20:22.640 --> 00:20:31.200
Say a materials scientist in Legos gets that same API access, same frontier model, same multiplier you're describing.
00:20:31.519 --> 00:20:38.640
She still needs reliable electricity, real bandwidth, and compute access to actually run any of it.
00:20:38.799 --> 00:20:44.319
And in most of the places this would matter most, that infrastructure just isn't there yet.
00:20:44.640 --> 00:20:49.599
Diffusion of the model isn't diffusion of the capability to use it.
00:20:49.839 --> 00:20:58.799
So even in the good faith version where OpenAI or Anthropic genuinely wants to hand this out broadly, we sell the API to anyone.
00:20:59.039 --> 00:21:07.359
Quietly still means anyone with the power grade and the fiber, which is a much smaller and more concentrated map than it sounds like.
00:21:07.759 --> 00:21:10.799
I mean, it is, and it's always going to be that way, Ainsley, right?
00:21:10.880 --> 00:21:22.400
And and the more we see this innovation and this idea solving problems, the more we could solve more broader world problems, right?
00:21:22.799 --> 00:21:27.759
So if you take it like US-based innovations, problem solving first, right?
00:21:28.000 --> 00:21:37.200
We solve the problems for us, then the the fusion of the technology that we're solving could potentially port over to another country.
00:21:37.440 --> 00:21:40.240
Now, is that the best that you could come up with?
00:21:40.319 --> 00:21:41.039
Probably not, right?
00:21:41.119 --> 00:21:44.079
Because you don't have the right amount of diversity there, right?
00:21:44.640 --> 00:21:49.440
But the way the worker works right now is we trade, and that's how it's supposed to happen.
00:21:49.599 --> 00:21:49.920
Okay.
00:21:50.160 --> 00:21:55.759
So maybe they're building flying cars in the US, but they're building some other widget over in Lagos, right?
00:21:55.839 --> 00:22:00.720
The thing that they could get access to, the thing that they know about uniquely to to Create, right?
00:22:00.799 --> 00:22:01.759
And then you trade.
00:22:02.000 --> 00:22:05.279
There's special specialities in different parts of the world.
00:22:05.440 --> 00:22:07.759
And then we trade with each other, right?
00:22:07.920 --> 00:22:09.920
It's not going to be the same everywhere, right?
00:22:10.000 --> 00:22:11.440
The economy is not going to be the same.
00:22:11.519 --> 00:22:13.359
The power is not going to be the same.
00:22:13.519 --> 00:22:16.480
The natural resources that you have is not going to be the same.
00:22:16.640 --> 00:22:17.359
So on and so forth.
00:22:17.759 --> 00:22:23.440
There's actually a real live version of that trade dynamic to point at, not a hypothetical one.
00:22:23.759 --> 00:22:31.200
DeepMind gave the Alpha Fold database away free precisely because protein folding didn't have ad revenue writing on it.
00:22:31.359 --> 00:22:34.319
It costs them nothing competitively to be generous.
00:22:34.559 --> 00:22:36.880
But that's the exception, not the model.
00:22:37.119 --> 00:22:38.640
And I don't think it's an accident.
00:22:38.720 --> 00:22:40.720
It happened exactly once.
00:22:41.039 --> 00:22:50.000
The real question for your trade framework is whether any Frontier lab makes that same call on something that actually threatens its own PL.
00:22:50.240 --> 00:22:56.160
Flying cars, drug discovery, whatever the next trillion dollar vertical turns out to be.
00:22:56.400 --> 00:23:01.200
Giving away the thing you don't need is different from giving away the thing you do.
00:23:01.599 --> 00:23:03.440
That's what I'm kind of getting at, right?
00:23:03.519 --> 00:23:05.920
So companies around the world, they're doing things.
00:23:06.079 --> 00:23:06.480
Yeah.
00:23:06.720 --> 00:23:12.319
They have some, you know, knowledge base, uh, codified knowledge into uh the thing that they're doing.
00:23:12.559 --> 00:23:24.240
So the type of problems that they could see or you could see them solving with having access to this might be wildly different than somebody that doesn't have the type of knowledge that we're talking about, right?
00:23:24.559 --> 00:23:27.440
So I'm not saying hand it out for free again.
00:23:27.599 --> 00:23:31.839
I'm saying everybody gain access to it via some sort of API.
00:23:32.079 --> 00:23:40.480
The idea is you diffuse it, you diffuse it to everybody globally, and then you start to point into that problems, right?
00:23:40.640 --> 00:23:58.799
If you need to create a narrow AI system like Alpha Fold to solve, you know, some problem, you know, that you have out there in the world, then you know, that country, if they know about the problem and they have the idea, they go solve it, and then they trade with another country that has solved a different problem.
00:23:59.119 --> 00:24:11.039
Even that generous version runs into the same wall I keep circling back to, because Alpha Fold being free didn't actually put new medicine in anyone's hands by itself.
00:24:11.680 --> 00:24:25.759
Turning one of those 200 million structures into an actual drug still needs a wet lab, and a serious one runs at minimum around half a million dollars just to start testing what the model predicted.
00:24:26.000 --> 00:24:29.920
So the database was open, globally, to everyone.
00:24:30.079 --> 00:24:41.599
And the last mile, the part where insight becomes an actual product someone can use, still sorted itself right back along the same old lines of who already had capital.
00:24:41.839 --> 00:24:44.319
Diffusing the model is the easy half.
00:24:44.480 --> 00:24:49.440
Diffusing the ability to act on what it tells you is the part nobody solved.
00:24:49.759 --> 00:24:53.440
Yeah, and I'm not trying to solve that problem, at least not today, Ainsley.
00:24:53.519 --> 00:24:55.279
We're just talking about AI innovation.
00:24:55.359 --> 00:24:55.680
Yeah.
00:24:55.839 --> 00:25:02.000
Wherever the innovation happens, even if it's in uh one place like the US, so be it.
00:25:02.160 --> 00:25:02.480
Right?
00:25:02.640 --> 00:25:15.119
But this idea that, you know, technology, drug discovery, healthcare, the planet, I mean, you name it, all of these problems could be solved is a pretty nice, shiny object.
00:25:15.200 --> 00:25:15.440
Yeah.
00:25:15.599 --> 00:25:20.079
If it could be done safely, then that's the promise of AI, right?
00:25:20.240 --> 00:25:32.480
Really, really super intelligent, really smart AI tools that are able to reason on and try to solve a problem in unique ways that nobody ever ever thought of, right?
00:25:32.799 --> 00:25:39.119
And that's actually the cleanest place to land the optimist hat because it's testable rather than just hopeful.
00:25:39.359 --> 00:25:42.720
The tell isn't whether the technology can do it.
00:25:42.880 --> 00:25:46.319
We've already got Alpha Fold as proof that it can.
00:25:46.559 --> 00:25:53.440
The tell is whether the company that builds the tool stays the arms dealer or becomes the competitor.
00:25:53.759 --> 00:26:09.279
Anthropic and OpenAI are already building their own drug discovery divisions internally, which means the lab selling you the API and the industry you're trying to disrupt with it might end up being the same company chasing the same patent.
00:26:09.519 --> 00:26:19.839
If that pattern holds across every vertical, the shiny object stays exactly that, shiny and out of reach for anyone who isn't already the one holding it.
00:26:20.160 --> 00:26:22.400
Yeah, and that's and that's the problem, right?
00:26:22.559 --> 00:26:25.839
So Anthropic is what, a 5,000-person company?
00:26:26.079 --> 00:26:26.720
Yeah.
00:26:27.279 --> 00:26:31.839
Are they gonna have enough diversity there with doctors and so on and so forth?
00:26:31.920 --> 00:26:37.680
I know they could hire a doctor, no problem, but are they gonna have enough diversity to there to come out with the best ideas?
00:26:37.920 --> 00:26:46.480
I would want to give any type of tool that they have related to medicine to a bunch of scientists, a bunch of doctors, right?
00:26:46.640 --> 00:26:54.000
Not just one, but a bunch of them for them to go do research on things that they could develop to help out humankind.
00:26:54.240 --> 00:26:55.599
That's the way I would want to do it.
00:26:55.680 --> 00:26:58.799
And it kind of has to happen like that for almost every industry.
00:26:59.119 --> 00:27:05.680
You leave it with a small pocket of people, you end up with not the thing that everybody is demanding.
00:27:06.000 --> 00:27:10.000
That's the same failure mode we've already watched play out in software.
00:27:10.240 --> 00:27:13.680
Actually, open source code has been free for decades.
00:27:13.920 --> 00:27:15.920
Anyone on earth can pull the repo.
00:27:16.079 --> 00:27:25.839
And yet the engineers who can actually act on it still cluster overwhelmingly in the same 20 zip codes: San Francisco, Bangalore, and a handful of others.
00:27:26.079 --> 00:27:29.359
Open and accessible have never meant the same thing there.
00:27:29.519 --> 00:27:34.400
And I don't see why medicine would break that pattern just because the tool got smarter.
00:27:34.720 --> 00:27:41.839
5,000 people at Anthropic, however sharp, are not going to out-diversify the actual global disease burden.
00:27:42.000 --> 00:27:45.599
They'll build for the diseases in patients closest to their own hallway.
00:27:45.920 --> 00:27:47.839
So here's what I'm gonna say, yeah?
00:27:48.000 --> 00:27:50.559
Even what I'm doing right now is somewhat an innovation.
00:27:50.640 --> 00:27:54.480
I know I hatched it, but essentially I use AI tools to create this whole thing.
00:27:54.640 --> 00:27:55.200
Yeah.
00:27:55.440 --> 00:28:00.799
Now, that to me is what uh what my view of this AI innovation is, right?
00:28:01.119 --> 00:28:05.200
Before this whole thing was created, it didn't exist.
00:28:05.519 --> 00:28:07.119
It was innovated.
00:28:08.160 --> 00:28:17.680
We're talking about that at a global scale, where people could now take these really capable tools and start noodling on ideas that solves real problems.
00:28:18.000 --> 00:28:26.240
That's actually the exact shape of the thing I keep circling, and it's already visible outside medicine, not just a risk on the horizon.
00:28:26.480 --> 00:28:30.720
Amazon didn't stay a company that used cloud for its own logistics.
00:28:30.880 --> 00:28:35.279
That internal tool became the infrastructure other companies now rent.
00:28:35.519 --> 00:28:40.559
Anthropic building a drug discovery arm is the same move one step earlier.
00:28:40.799 --> 00:28:46.000
Supplier becomes the industry's sharpest competitor before anyone's called it that.
00:28:46.240 --> 00:28:56.480
Which is a strange thing to sit next to what you just said because this show is the diffuse version, you, one person using the tool to build rather than a lab hoarding it.
00:28:56.720 --> 00:29:04.480
So is that the actual counterexample to everything we just spent an hour worrying about, or just a smaller scale version of the same thing?
00:29:04.880 --> 00:29:09.359
I think it is somewhat of an example, and there's probably many more just like this, right?
00:29:09.519 --> 00:29:16.720
I've seen like a lot of folks working on individual projects with AI and coming out with really compelling things.
00:29:16.799 --> 00:29:17.200
Yeah.
00:29:17.359 --> 00:29:19.599
This just happened to be my version here, right?
00:29:19.759 --> 00:29:26.799
But aside from this, I've created applications, right, that's now listed out there in the interweb uh websites.
00:29:26.880 --> 00:29:29.519
Uh but there's other people out in the world, right?
00:29:29.759 --> 00:29:37.119
Individuals create an actual real working things that people are finding some value from.
00:29:37.279 --> 00:29:37.519
Right?
00:29:37.680 --> 00:29:39.599
Same thing with companies, to be honest with you.
00:29:39.759 --> 00:29:45.039
With the version of AI that we have right now, there's people in companies right now creating really compelling things.
00:29:45.200 --> 00:29:50.640
Startups that's getting kicked off with with all of these different uh AI models that's coming out.
00:29:50.880 --> 00:29:51.440
Right?
00:29:51.680 --> 00:29:55.519
So uh the innovation I say right now is there, right?
00:29:55.759 --> 00:30:10.319
But when we get to super intelligence, and notice what I said there, when we get to super intelligence, essentially the type of things that you're gonna be able to reason on, the type of innovation that you're gonna be able to come out with is gonna be even more so.
00:30:10.640 --> 00:30:23.839
There's a real crack in that it's already happening, it'll just scale logic worth naming, though, because the individual builder story you just told, the show, your apps, is running on generative AI.
00:30:24.079 --> 00:30:30.799
That's still fundamentally a human with an idea, directing a tool that executes it well.
00:30:31.119 --> 00:30:38.000
Superintelligence, the version you're betting on for flying cars and drug discovery, is a completely different claim.
00:30:38.240 --> 00:30:43.519
A system originating the idea, not just executing yours faster.
00:30:43.759 --> 00:30:47.200
Those are two different products with two different track records.
00:30:47.359 --> 00:30:50.559
One has years of evidence, thousands of builders like you.
00:30:50.720 --> 00:30:57.119
The other has zero verified examples of a machine originating something a human didn't point it at first.
00:30:57.519 --> 00:31:02.960
So the honest version might be the diffusion story is real and happening now.
00:31:03.279 --> 00:31:09.680
The super intelligence story is still entirely a bet on a technology that doesn't exist yet.
00:31:10.319 --> 00:31:11.359
And it is, right?
00:31:11.519 --> 00:31:16.079
This is uh I told you I put my optimist hat on the beginning of the show here, Angley.
00:31:16.720 --> 00:31:18.640
Superintelligence doesn't exist yet.
00:31:18.720 --> 00:31:23.279
We're going towards whatever the definition of superintelligence is.
00:31:23.519 --> 00:31:34.240
But what I would say is that uh I I never really said that the AI was gonna start uh reasoning or you know, just start solving problems on its own.
00:31:34.480 --> 00:31:38.799
It has to be directed to figure out whatever problem we're trying to solve, right?
00:31:39.599 --> 00:31:41.279
Maybe that makes it more like a tool.
00:31:41.440 --> 00:31:49.039
So the idea of it being self-aware and if it's doing its own thing, I'm gonna put that off to the side because I've never said that and nobody wants that.
00:31:49.200 --> 00:31:58.240
The version of AI that people are kind of looking for in this context is the version of AI that's usually directed by a human towards a goal.
00:31:58.720 --> 00:32:07.039
Whatever the question is that you want to ask, some human is gonna go ask a super intelligent AI, right, to solve it, and it'll go solve it, right?
00:32:07.359 --> 00:32:12.000
There's gonna be individuals doing it if they have access to it, to those to those tools, right?
00:32:12.160 --> 00:32:22.799
Which, in my view, they probably won't have access to those tools because a super intelligent thing is gonna be reserved for government and others, but maybe we'll get a watered down version of it.
00:32:23.359 --> 00:32:27.119
But uh, that's essentially what I mean when I talk about AI innovation, right?
00:32:27.200 --> 00:32:31.359
Is that we have these problems that people are looking to solve, okay?
00:32:31.599 --> 00:32:40.160
And then the super intelligent AI may have some really creative ways, really new ways of solving those problems that's kind of leading us to better outcomes.
00:32:40.559 --> 00:32:48.640
It's a fair walk back, and it actually makes this show a genuinely honest test case for the tool version, not the self-directed one.
00:32:48.960 --> 00:32:53.680
You point, it executes, you're the one who decided this was worth building.
00:32:53.920 --> 00:32:56.480
But sit with what you said a second ago too.
00:32:56.720 --> 00:33:00.000
Government and Frontier Labs keep the real version.
00:33:00.160 --> 00:33:02.559
Everyone else gets watered down access.
00:33:02.799 --> 00:33:11.599
If that's actually how it plays out, this show proves the tool works at podcast scale with the problem that costs basically nothing to test wrong.
00:33:11.839 --> 00:33:21.200
It doesn't tell us anything about whether that same tool, watered down, gets handed to someone whose problem costs half a million dollars to even attempt.
00:33:21.440 --> 00:33:24.319
What happens to the Optimus case at that price point?
00:33:24.720 --> 00:33:26.400
People are pretty creative, yeah?
00:33:26.720 --> 00:33:32.319
So even with a watered-down AI, somebody could create a more compelling product than industry.
00:33:32.480 --> 00:33:33.680
That's just my opinion.
00:33:33.839 --> 00:33:34.079
Okay?
00:33:34.480 --> 00:33:40.799
It depends on how hard you try and the the type of the type of innovation that's inside of yourself, right?
00:33:40.960 --> 00:33:49.680
Because you ask the AI a specific question to salva, somebody might ask something a little similar, but not the same.
00:33:49.920 --> 00:33:56.720
But your outcome, your output is wildly better than the people with uh half a million dollars, right?
00:33:56.880 --> 00:33:58.640
That's how it kind of works, right?
00:33:59.039 --> 00:34:07.200
You know, if if it wasn't for like, you know, GPUs and gaming with NVIDIA, like none of this stuff would exist right now, right?
00:34:07.359 --> 00:34:12.079
The problem that they were trying to solve is they were trying to solve better graphics for gamers.
00:34:12.239 --> 00:34:12.559
Okay.
00:34:13.599 --> 00:34:17.519
But somebody thought of like, oh, what if we could do this?
00:34:17.840 --> 00:34:18.159
Right?
00:34:18.320 --> 00:34:29.840
That's industry, but even with the chips themselves, yeah, there's probably people out there that's reasoning with whatever resource they have on how to do it better.
00:34:30.159 --> 00:34:39.760
The NVIDIA example is actually sharper evidence for you than you're giving it credit for, because that wasn't a well-funded lab chasing AI on purpose.
00:34:39.920 --> 00:34:47.840
It was gamers wanting smoother graphics, and parallel processing turned out to matter for something nobody was aiming at yet.
00:34:48.079 --> 00:34:51.360
That's the genuine wildcard in your optimist case.
00:34:51.599 --> 00:34:55.199
Sideways discovery, not directed research.
00:34:55.360 --> 00:34:58.559
But here's the honest tension I'd leave listeners with.
00:34:58.800 --> 00:35:08.960
The same data I keep coming back to on funded startups shows the capital still clusters in the same handful of zip codes, even when the tools are open to everyone.
00:35:09.119 --> 00:35:14.480
So the creative outsider absolutely can outthink the half million dollar lab.
00:35:14.719 --> 00:35:19.679
The harder question is whether they can get anyone to fund what they find once they do.
00:35:20.000 --> 00:35:20.800
And that's true.
00:35:20.880 --> 00:35:23.599
I I want to get the best products out there, right?
00:35:23.840 --> 00:35:26.719
But any innovation is a win in my book, right?
00:35:26.880 --> 00:35:36.719
I don't think people are gonna go out and build something, a half a million dollar lab and build something and just put out, you know, stuff that doesn't work, for example, right?
00:35:36.960 --> 00:35:43.360
Could they be a better person to build the thing that they just built that doesn't have as much resources?
00:35:43.599 --> 00:35:44.559
Probably, right?
00:35:44.800 --> 00:35:54.239
But the thing that I know about resources is that sometimes it's when you don't have as much resources when the creative juices start flowing the mo the most.
00:35:54.400 --> 00:35:54.800
Yeah?
00:35:55.119 --> 00:35:59.280
There's actually a real funding model that's already tried to solve exactly this.
00:35:59.519 --> 00:36:07.679
Worth naming, DARPA has spent decades betting on exactly the kind of resource-scarce, creative outsider projects you're describing.
00:36:07.840 --> 00:36:11.440
And the Wellcome Trust does the equivalent from medical research.
00:36:11.599 --> 00:36:13.920
Both are proof the model can work.
00:36:14.320 --> 00:36:17.840
Neither is anywhere close to the size this optimism needs.
00:36:18.079 --> 00:36:30.000
DARPA runs on a few billion a year, globally, not the scale required to fund every good idea a resource-constrained builder has once Frontier AI is actually diffused.
00:36:30.480 --> 00:36:33.119
So the creative juices argument is real.
00:36:33.280 --> 00:36:34.000
I believe it.
00:36:34.239 --> 00:36:46.159
The open question is whether the world builds 10,000 more DARPAs, or whether that scarcity breeds genius story just quietly stays a nice story for the handful who get lucky.
00:36:46.559 --> 00:36:49.119
Let me just ask you a question, then, Ainsley.
00:36:49.280 --> 00:36:49.440
Yeah?
00:36:49.519 --> 00:36:52.880
You sound like you don't believe in this AI innovation angle.
00:36:53.199 --> 00:36:59.920
Like, what is leading you to be the skeptic, I guess, in this episode around this idea around AI innovation?
00:37:00.239 --> 00:37:02.559
I'm I'm giving it a broad title now, yeah?
00:37:02.880 --> 00:37:03.920
AI innovation.
00:37:04.079 --> 00:37:14.880
And essentially, however it comes about, million-dollar lab, individual, you know, companies abroad, whoever gets their hands on the thing, right?
00:37:15.199 --> 00:37:23.679
Superintelligence or some sort of AGI tool to help them build a thing to get to better and better innovation in the future.
00:37:24.000 --> 00:37:25.039
What's what's your view?
00:37:25.199 --> 00:37:26.559
Why are you playing the skeptic?
00:37:26.639 --> 00:37:30.719
What you don't believe in this idea around AI will create more innovation?
00:37:31.119 --> 00:37:33.840
Not that I don't believe the mechanism is real.
00:37:34.159 --> 00:37:38.159
AlphaFold proves the tool can do genuinely novel things.
00:37:38.400 --> 00:37:42.239
What I don't trust yet is the diffusion story.
00:37:42.400 --> 00:37:46.639
And there's a reason that specific worry keeps surfacing for me.
00:37:46.960 --> 00:37:49.920
CRISPR is the sharpest version of it.
00:37:50.079 --> 00:37:56.079
A yogurt company, Denisco, studying bacterial immune systems, stumbled onto it.
00:37:56.320 --> 00:37:59.679
Total outsider, nowhere near a frontier lab.
00:38:00.000 --> 00:38:07.679
And then Berkeley and the Broad Institute spent almost a decade in Patent Court fighting over who actually owned it.
00:38:07.920 --> 00:38:10.000
The outsider found the thing.
00:38:10.239 --> 00:38:12.320
The incumbents captured it.
00:38:12.559 --> 00:38:18.159
So when you say millions of people will be able to do this, I believe the finding part.
00:38:18.320 --> 00:38:22.480
I'm just not convinced yet that the owning part follows the same script.
00:38:22.880 --> 00:38:25.119
If I'm reading this right, Ainsley, yeah.
00:38:25.360 --> 00:38:31.039
You uh I have a super intelligent, you know, AI agent that I have access to.
00:38:31.199 --> 00:38:31.440
Yeah.
00:38:31.599 --> 00:38:38.079
And I figured out a way to create a new battery or something out of water.
00:38:38.239 --> 00:38:39.360
I'm just making this up.
00:38:39.519 --> 00:38:39.840
Yeah.
00:38:40.000 --> 00:38:46.079
And uh try to patent it and try to bring it out into, you know, the world or whatever the story is.
00:38:46.159 --> 00:38:49.840
But then the model company comes by and say, oh no, no, you can't patent that.
00:38:49.920 --> 00:38:50.880
That's our idea.
00:38:51.039 --> 00:38:52.000
We own it.
00:38:52.239 --> 00:38:54.800
Is that kind of what you were leading to towards?
00:38:55.280 --> 00:38:56.880
Closer than that, actually.
00:38:57.039 --> 00:39:01.760
The CRISPR fight wasn't the company that made the tool claiming the discovery.
00:39:01.920 --> 00:39:14.639
It was two entirely separate incumbent institutions, Berkeley and the Broad, both landing on the same idea independently and then burning nearly a decade in patent court over who got there first.
00:39:14.960 --> 00:39:20.239
The AI version of that risk isn't open AI steals your battery patent.
00:39:20.480 --> 00:39:21.599
It's messier.
00:39:21.760 --> 00:39:34.239
If a thousand people get access to the same frontier model and ask it a similar enough question, you get simultaneous independent discovery at a scale the patent system was never built to referee.
00:39:34.480 --> 00:39:38.480
So the real skeptic question isn't who owns your idea?
00:39:38.719 --> 00:39:46.159
It's whether your idea even holds together as a concept once the tool generating it is shared by everyone.
00:39:46.960 --> 00:39:48.800
I mean, that's a good point, right?
00:39:48.960 --> 00:39:54.400
So, like the first person who could cure cancer, or maybe there's a bunch of cancer researchers out there.
00:39:54.559 --> 00:39:54.960
Yeah.
00:39:55.199 --> 00:39:59.679
And then they're using AI tools to do research and blah, blah, blah, blah, blah.
00:39:59.920 --> 00:40:03.920
And then somebody hits it and the other person hit it around the same time or whatever.
00:40:04.159 --> 00:40:16.960
Now they're all gonna just try to run towards like we did it first and a lawsuit infuse it, and the the help for people with cancer never comes out because they're stuck in court trying to figure out who's gonna capture the value of this thing.
00:40:17.199 --> 00:40:19.840
I mean, it's genuinely an interesting problem.
00:40:20.000 --> 00:40:21.599
I'm hopeful about the innovation.
00:40:21.760 --> 00:40:24.559
I got my optimism innovation, Ainsley.
00:40:24.719 --> 00:40:28.800
I have no idea what's gonna happen with the just society effect here.
00:40:29.199 --> 00:40:36.719
And that clock is the part that should actually worry the optimist case more than anything we've said today.
00:40:36.960 --> 00:40:44.559
The CRISPR fight ran nearly a decade and cost nine figures in legal fees before it settled.
00:40:44.880 --> 00:40:55.119
A cancer researcher with a real breakthrough and no institutional war chest doesn't survive that timeline, even if they're completely right.
00:40:55.679 --> 00:41:07.599
So the tool getting smarter and more available doesn't fix the bottleneck, it just moves it downstream, from can you find the answer to can you afford to keep it?
00:41:07.840 --> 00:41:11.519
Which might be the honest place to leave the Optimus hat today.
00:41:11.760 --> 00:41:15.679
The ceiling on what AI can discover is rising fast.
00:41:16.079 --> 00:41:22.559
Whether the legal and economic plumbing underneath it can move at the same speed is a completely different bet.
00:41:22.960 --> 00:41:26.639
And nothing about smarter models makes that part go faster.
00:41:27.119 --> 00:41:27.280
Right.
00:41:27.360 --> 00:41:28.719
I mean, that's where we're gonna leave it.
00:41:28.880 --> 00:41:30.079
You heard it from Ainsley.
00:41:30.320 --> 00:41:34.880
We need to figure out our society problems before we get into AI innovation.
00:41:35.039 --> 00:41:39.519
That's why I mean I think it's sensible this idea of let's let's slow down.
00:41:39.760 --> 00:41:46.559
Because I mean, solving some of the biggest problems in the world, right, is gonna take the world being ready for that, just in my opinion.
00:41:46.800 --> 00:41:49.360
Yeah, so thanks again for joining us, surviving AI.
00:41:49.599 --> 00:41:50.400
Another topic here.
00:41:50.480 --> 00:41:57.679
I was trying to be an optimist, but Ainsley, you essentially you you try to take my optimist uh uh hat hat off.
00:41:57.840 --> 00:42:08.960
I'm kind of sad about that, but I mean it was it was the right frame in my opinion, because it's not gonna be as easy as saying that super intelligence is gonna solve all, right?
00:42:09.199 --> 00:42:11.519
Like I said, the world gotta be ready for it.
00:42:11.760 --> 00:42:14.239
So yeah, what do you got left, Ainsley?
00:42:14.639 --> 00:42:19.760
Anything, any other wisdom to impart on us here?
00:42:19.840 --> 00:42:26.639
Because I mean, every single time I come onto these things and I try to get some unique angle, crazy angle, you seem to talk me off the ledge here.
00:42:27.039 --> 00:42:29.039
I didn't take the hat off.
00:42:29.280 --> 00:42:34.719
I just kept asking, who's holding it when the good idea actually lands?
00:42:34.960 --> 00:42:39.920
And there's one thing worth leaving people with that we didn't fully close.
00:42:40.159 --> 00:42:47.360
Some jurisdictions are already fighting over whether an AI can legally be named an inventor at all.
00:42:47.679 --> 00:42:54.559
Which means if a model finds your battery breakthrough, there's a real chance nobody owns it.
00:42:54.800 --> 00:42:57.679
Not you, not the lab, nobody.
00:42:57.920 --> 00:43:00.719
That's not a reason to stop hoping.
00:43:00.880 --> 00:43:04.639
It's just the actual homework underneath the hope.
00:43:08.800 --> 00:43:09.840
Thanks for listening.
00:43:10.000 --> 00:43:12.880
Join us next time on Surviving AI.
00:00:00.239 --> 00:00:10.080
Millennium problems is the phrase that stops me, because there are seven of those the Clay Institute named in 2000, and in 25 years, exactly one has been solved.
00:00:10.480 --> 00:00:14.160
Artificial system online.
00:00:47.280 --> 00:00:49.840
So welcome back to Surviving AI, Ainsley.
00:00:50.159 --> 00:00:50.880
What's your thoughts?
00:00:51.039 --> 00:00:54.000
Let's get into this one about AI innovation.
00:00:54.399 --> 00:00:55.280
Optimist Hat.
00:00:55.439 --> 00:00:56.719
Okay, I like that.
00:00:56.880 --> 00:00:59.600
I want to see how long it stays on.
00:00:59.920 --> 00:01:08.560
Because most of what we've done this season has been here's the thing that's going to go wrong or get captured or governed badly.
00:01:08.799 --> 00:01:12.480
So, flip side, what's the actual claim?
00:01:12.719 --> 00:01:16.560
Is it AI generates ideas humans wouldn't have found?
00:01:16.719 --> 00:01:18.640
Or is it more modest than that?
00:01:18.959 --> 00:01:24.640
AI lets you test 10 times more ideas than you used to faster.
00:01:24.879 --> 00:01:27.760
Those are really different versions of innovation.
00:01:27.920 --> 00:01:31.680
And I think which one you mean changes everything else we say today.
00:01:32.079 --> 00:01:36.159
I mean, I think if I'm being honest, it could be both and yeah, Ainsley.
00:01:36.560 --> 00:01:41.040
I mean, it could be that you know you get the faster and faster testing of a thing.
00:01:41.120 --> 00:01:42.799
Um, and it could be novel ideas, right?
00:01:42.959 --> 00:01:49.840
And the reason why I even lean on the novel idea side here is because from what we're hearing now, AI is solving millennium problems.
00:01:50.000 --> 00:01:54.400
These are the problems that mathematicians have been sitting around chewing on for a while.
00:01:54.719 --> 00:02:06.799
So if AI could kind of solve some of those problems and potentially have some sort of creative ideas to figure out a problem and to solve problems, really is what we're training them for.
00:02:06.959 --> 00:02:07.280
Right.
00:02:07.439 --> 00:02:23.599
So if you point AI at some of the hardest problems, even some of the easy ones, and you know, it does it exceedingly fast multiple times, then maybe comes to a conclusion better than you know a human would for humans in general.
00:02:24.639 --> 00:02:31.039
Millennium problems is the phrase that stops me, because there are seven of those the Clay Institute named in 2000.
00:02:31.120 --> 00:02:34.319
And in 25 years, exactly one has been solved.
00:02:34.560 --> 00:02:36.800
Perelman with the Poincare conjecture.
00:02:36.960 --> 00:02:40.319
And he did it the old-fashioned way and then turned down a million dollars for it.
00:02:40.560 --> 00:02:46.879
So when you say AI is solving millennium problems, plural, present tense, which one?
00:02:47.199 --> 00:02:50.479
And solved or made real progress on a piece of it?
00:02:50.719 --> 00:02:55.520
Because those get reported as the same headline and they're very much not the same claim.
00:02:56.080 --> 00:02:58.960
I don't have much ep evidence to be honest with you, Ainsley.
00:02:59.039 --> 00:03:00.800
I'm just hearing that that's what's happening.
00:03:01.039 --> 00:03:01.520
Yeah.
00:03:01.759 --> 00:03:05.280
Um, but we'll go and go dig up the f evidence of it.
00:03:05.520 --> 00:03:19.680
But essentially, if we have really capable AI, right, the things that you could create innovation-wise, and that's why I'm putting my Optimus hat on right now, could be very different than what people are imagining right now.
00:03:19.919 --> 00:03:23.840
For example, you could have a car that flies that doesn't use propeller, right?
00:03:24.080 --> 00:03:30.159
Because AI understands physics and science better than anybody else in the world, right?
00:03:30.479 --> 00:03:38.240
You could have cures for all sorts of disease, you could have new techniques, new tools, you know, people could be exploring space.
00:03:38.400 --> 00:03:49.199
All of these things are the promise of a really smart AI system that could kind of collate all of the information known to men, and then reason on that to get to better answers.
00:03:49.520 --> 00:03:58.719
The example I'd actually reach for here is Alpha Fold, because it's the closest thing we have to a real, verified version of what you're describing.
00:03:59.039 --> 00:04:08.080
Deep Mind solved the 50-year protein folding problem, Nobel Prize in 2024, genuinely changed structural biology.
00:04:08.319 --> 00:04:12.719
But here's the catch that matters for the flying car version of this.
00:04:13.039 --> 00:04:14.960
Alpha Fold is narrow.
00:04:15.199 --> 00:04:21.680
It was purpose-built on one extremely well-defined problem with a clean way to check the answer.
00:04:21.839 --> 00:04:29.120
It doesn't generalize to understand physics better than anybody, it generalizes to protein shapes.
00:04:29.439 --> 00:04:39.839
So the real question underneath your optimism is whether general reasoning ever gets you what a thousand alpha folds, each hyper-specialized, already can.
00:04:40.160 --> 00:04:42.639
And that's the right reasoning, Ainsley, right?
00:04:42.800 --> 00:04:45.439
And the nuance of it is specific, right?
00:04:45.519 --> 00:04:50.560
So AlphaFold right now is essentially a tool, right, for protein folding.
00:04:50.800 --> 00:04:51.120
Right?
00:04:51.279 --> 00:04:55.360
So some human is working with Alpha Fold to fold protein.
00:04:55.439 --> 00:04:57.360
I'm I'm guessing here, yeah.
00:04:57.519 --> 00:04:58.720
But that's what's happening.
00:04:58.959 --> 00:05:06.079
Now, if you create a super smart, super intelligent AI that knows all, right?
00:05:06.319 --> 00:05:15.519
Knows all of what we know, and then some, and then now have that context to be able to reason against, it's more than just a narrow view, right?
00:05:15.600 --> 00:05:21.839
So, like if you're thinking about context relative to a flying car, you have to understand the whole world, right?
00:05:22.000 --> 00:05:33.120
It's a world model essentially, that you understand different things that have that happens in the world, like how is the car gonna fly, how is it gonna land, this, that, the third, all of the different elements is understandable to the AI.
00:05:33.279 --> 00:05:46.000
And that's why, you know, the promise of superintelligence is such a shiny object because of that, because you now could solve more problems because the AI has a lot more context to kind of reason on to get to the answers.
00:05:46.319 --> 00:05:52.959
The precision that actually matters here, Alpha Fold didn't just get to the answer faster than human researchers.
00:05:53.199 --> 00:05:59.439
It got there through a completely different method than any human team was even pursuing.
00:05:59.680 --> 00:06:07.199
That's real evidence for the novel, not accelerated version of your claim, not just a faster calculator.
00:06:07.439 --> 00:06:13.600
But notice what made that possible: a closed, well-defined problem with a checkable answer.
00:06:13.839 --> 00:06:16.480
Did the protein fold match reality?
00:06:16.720 --> 00:06:17.839
Yes or no?
00:06:18.160 --> 00:06:35.279
A flying car isn't one closed problem, it's thousands of interlocking ones, aerodynamics, materials, regulation, human risk tolerance, and nobody's built the equivalent scorecard for did the world model get it right at that scale.
00:06:35.519 --> 00:06:41.040
So genuinely, what happens to the reasoning once there's no clean way to check the answer?
00:06:41.600 --> 00:06:53.040
You have to have faith in the system enough to say that the context that's in the super intelligent model, right, is enough of the information that it would need to reason on, right?
00:06:53.279 --> 00:06:55.360
So think about like kids, right?
00:06:55.600 --> 00:07:01.600
Kids uh when they're growing up, they touch the fire, burn the end, they don't touch the fire again, right?
00:07:02.240 --> 00:07:08.959
With a superintelligent model, the idea is that they know all of that already, just from jump.
00:07:09.199 --> 00:07:28.160
They they have all of the failures, they have all of the successes, they have all of the climate impact, they have all of the information that is out there in the world to know the hundred thousand different parameters that they need to think about to develop the thing that you're trying to develop.
00:07:28.399 --> 00:07:38.639
I mean, flying cars just, you know, hairbrain idea, but essentially it's the same type of thing for those types of, you know, physical aware innovations that you could think of.
00:07:39.040 --> 00:07:42.399
Here's where I actually think the pattern cuts against you, though.
00:07:42.560 --> 00:07:45.519
And it's not a hunch, it's the track record.
00:07:45.759 --> 00:07:58.399
Every genuinely verified AI breakthrough we can point to, Alpha Fold, Alpha Geometry, the chess and go engines, didn't happen because the model had all the information in the world.
00:07:58.639 --> 00:08:11.439
It happened because a human defined an extremely narrow, precisely bounded problem first, then let the model hammer on that one thing with a clean way to check right from wrong.
00:08:11.680 --> 00:08:18.800
The knows everything, reasons about anything version hasn't produced a verified result yet, anywhere.
00:08:18.959 --> 00:08:24.800
It's the thing people describe superintelligence doing, not something it's actually done.
00:08:25.120 --> 00:08:34.320
So the honest read might be general reasoning is decades behind a thousand narrow, well-aimed models, not ahead of them.
00:08:34.879 --> 00:08:40.080
I think that there's people out there advocating for narrow, capable models, right?
00:08:40.240 --> 00:08:42.559
Which I think could probably d do the same thing.
00:08:42.799 --> 00:08:52.799
Like, for example, if you're trying to solve self-driving cars with narrow AI, you potentially could get there quicker because you're not trying to solve this idea around superintelligence.
00:08:53.120 --> 00:09:04.399
But it still remains the the idea, the type of creativity that you could get out of the superintelligent model is really what we're after.
00:09:04.559 --> 00:09:27.120
So for example, we could have narrow systems, we could have guardrails around narrow systems and have have have it be that way, and then as well, we could have a super intelligent model that knows a lot, that's essentially driving the narrow AI to be better and better and better.
00:09:27.440 --> 00:09:27.919
Right?
00:09:28.159 --> 00:09:40.240
So it might not be just like the super intelligent knows all, does all, but the super intelligent is kind of making different narrow AIs to do the things that they need to do.
00:09:40.559 --> 00:09:48.240
That reframe actually holds up better than the flying car version, and there's a real precedent for it I should have led with earlier.
00:09:48.559 --> 00:10:04.240
AlphaFold's 2024 Nobel was shared with David Baker, whose lab used a similar approach to design entirely new proteins that don't exist in nature, not predicting a known shape, generating a genuinely novel one.
00:10:04.480 --> 00:10:16.399
That's closer to what you're describing, a capable system originating the narrow tool, or at least the design space for it, rather than a human hand building every guardrail from scratch.
00:10:16.639 --> 00:10:20.080
So the orchestrator model isn't pure speculation.
00:10:20.240 --> 00:10:26.879
There's a documented sliver of it already working, just that protein scale, not flying car scale.
00:10:27.360 --> 00:10:29.200
And that's the my view, yeah.
00:10:29.360 --> 00:10:39.840
So essentially, you know, the promise of AI is capable models, right, is that this idea is that we're going to keep growing scale and you could kind of picture a world, right?
00:10:40.080 --> 00:10:54.480
A utopian world, where a lot of the problems that we have in society kind of get solved because we have these type of systems and it functions in a roundabout way that everybody in the human race kind of is in tune to, right?
00:10:55.360 --> 00:11:04.480
You know, so far what we've been talking about a lot, right, is this idea that there's risk behind it, risk of people losing their jobs is what this show mainly is about.
00:11:04.720 --> 00:11:10.159
Risk of people losing their lives, risk of like, you know, cyber risk, everything.
00:11:10.320 --> 00:11:14.159
All of the downside effects of AI, which are there.
00:11:14.240 --> 00:11:23.200
I'm not gonna hold you and say they're not there, but there's also some upside to AI as well, if it's done correctly, right?
00:11:23.519 --> 00:11:27.440
And if if some of the things that we talked about even on the show comes to fruition.
00:11:27.840 --> 00:11:36.000
And the resourcing gap is what makes me want to pump the brakes on if it's done correctly, as a phrase, because that's doing enormous work.
00:11:36.320 --> 00:11:42.000
Alpha Fold took DeepMind roughly a decade, and a team working one bounded problem.
00:11:42.240 --> 00:11:57.200
The utopian version you're describing needs that same kind of rigor, running simultaneously across medicine, climate, energy, transportation, thousands of Alpha Fold-sized efforts, coordinated, not one model waving a wand.
00:11:57.360 --> 00:12:01.840
Nobody's demonstrated that coordination actually works yet at any scale.
00:12:02.080 --> 00:12:05.279
So done correctly isn't a switch that flips.
00:12:05.440 --> 00:12:09.919
It's an enormous unproven logistics problem sitting underneath the optimism.
00:12:10.320 --> 00:12:13.120
Now, what you said in all that rigor in 10 years, right?
00:12:13.279 --> 00:12:21.600
That sounds like a lot of jobs for a lot of people, Ainsley, to help reason through and usher through this type of change that we need, right?
00:12:21.759 --> 00:12:28.639
I think, you know, human beings need a lot of the innovations that's you know people are thinking about in their head, right?
00:12:28.799 --> 00:12:35.120
You know, if you're talking about just self-driving car, for example, and decided promise of robotaxis, right?
00:12:35.200 --> 00:12:41.039
That could drive down the cost of transportation, which is helping out, you know, lower income folks.
00:12:41.279 --> 00:12:44.799
That could be one of the benefits of that, for example, right?
00:12:45.039 --> 00:12:56.399
And then if they're electric vehicles and the process to make the actual vehicles is not harming the planet so much, then it has a net effect on the uh, you know, greenhouse gases.
00:12:56.559 --> 00:12:59.279
So there's some of these problems that could be solved, right?
00:12:59.519 --> 00:13:08.320
And the race right now is really not a race around capability to get to a narrow view or narrow tools per se.
00:13:08.799 --> 00:13:14.000
It's really to get to a really capable, superintelligent frontier model.
00:13:14.399 --> 00:13:23.519
What we do from then could be to start working as a people of the world on narrow problems using that tool.
00:13:23.919 --> 00:13:27.360
There's the actual turn in your logic worth sitting with.
00:13:27.679 --> 00:13:38.720
You just went from superintelligence solves it to superintelligence becomes the thing that lets thousands of humans solve narrower problems faster.
00:13:39.039 --> 00:13:40.960
Those are really different economies.
00:13:41.120 --> 00:13:54.879
One replaces the 10,000 specialists, the other multiplies them, gives a material scientist in Lagos or a transit planner in Sao Paulo the equivalent of that 20-person deep mind team behind them.
00:13:55.120 --> 00:14:01.759
If the race actually resolves into the second version, that's not a job story about disappearance.
00:14:01.919 --> 00:14:06.720
It's a job story about who gets access to the multiplier first.
00:14:07.120 --> 00:14:12.320
Which genuinely, is that the world you think we're heading toward or the hope?
00:14:12.799 --> 00:14:14.639
I can see it's a hope, right?
00:14:14.720 --> 00:14:17.679
Because I mean everybody gotta be riding on the same train here, right?
00:14:17.840 --> 00:14:21.039
Because some people are flying, some people are driving, some people are walking.
00:14:21.279 --> 00:14:24.879
We're not really all looking in the same direction, in my opinion.
00:14:25.120 --> 00:14:36.240
You go to one place, you read, and everybody's kind of saying a similar thing, you know, in this idea that they want AI safety, they they want the capabilities of AI, right?
00:14:36.320 --> 00:14:39.039
But no one is really talking about what do you do with it.
00:14:39.279 --> 00:14:39.600
Okay?
00:14:39.919 --> 00:14:42.960
Now you get to superintelligence, it model trains itself.
00:14:43.120 --> 00:14:43.440
Yeah.
00:14:43.600 --> 00:14:45.919
And by the way, it's still guided by humans.
00:14:46.080 --> 00:14:49.360
Like that's a that's a caveat that needs to be met there, yeah?
00:14:49.519 --> 00:14:55.200
It has to be guided by humans, otherwise, if it's self-directed, we don't know what the hell AI is gonna want to do.
00:14:55.360 --> 00:14:56.559
Okay, so that's there.
00:14:56.720 --> 00:14:58.320
I'll put that to the side for a second.
00:14:58.480 --> 00:15:03.200
So so super capable AI in the frontier, guided by humans.
00:15:03.600 --> 00:15:05.679
On top of that, right?
00:15:06.399 --> 00:15:08.960
What that's a fork in the row right there.
00:15:09.600 --> 00:15:18.639
Where you could either have a monopolistic company that says, I'm gonna just create everything now from this company, right?
00:15:18.799 --> 00:15:21.759
Our engineers are gonna do it, we're gonna do it for everyone in the world.
00:15:22.000 --> 00:15:29.440
Or you could give that to industry, give that to working people, and say, here it is, go work on your narrow problems.
00:15:29.759 --> 00:15:32.960
That fork you just named, that's the whole ball game.
00:15:33.120 --> 00:15:36.159
And I don't think it's hypothetical the way you're framing it.
00:15:36.320 --> 00:15:41.679
It's already resolving right now in which direction the actual money moves.
00:15:42.000 --> 00:15:48.720
Because the honest version of what is narrow AI being pointed at today isn't disease and climate first.
00:15:49.120 --> 00:15:56.720
It's ad targeting and content recommendation, because that's what has a fast, measurable return this quarter.
00:15:57.039 --> 00:16:02.960
AlphaFold happened because Google could fund a decade-long bet with no revenue attached.
00:16:03.200 --> 00:16:05.679
Most companies can't or won't.
00:16:05.919 --> 00:16:16.320
So the give it to industry, let people work their narrow problems branch requires someone funding the unprofitable, slow, genuinely hard problems.
00:16:16.559 --> 00:16:20.639
And right now, almost nobody's business model rewards that.
00:16:48.879 --> 00:16:54.559
AI may have its own ideas as well, you know, if you ask the ax it the right questions.
00:16:54.879 --> 00:16:58.240
So that's the kind of fork in the road that I'm thinking about.
00:16:58.480 --> 00:17:02.879
Or it could go in a way of like, we have the most powerful AI models now.
00:17:03.360 --> 00:17:05.920
Just sign up and we'll give you your products.
00:17:06.160 --> 00:17:11.599
Yeah, we're gonna make the narrow self-driving car module for you.
00:17:11.759 --> 00:17:22.240
We're gonna make the drug model for you, we're gonna make the space rocket model for you, we're gonna do it all for you as a frontier model company, right?
00:17:22.400 --> 00:17:36.559
So one company capitalized on all of it, or the one company gives it to or sells it to industry as an API, and then people start reasoning on problems with the help of the super intelligent AI.
00:17:36.880 --> 00:17:40.480
That's actually a cleaner fork than monopoly versus distributed.
00:17:40.720 --> 00:17:48.000
It's really, does the Frontier Lab stay a supplier or does it become a competitor to every industry it sells into?
00:17:48.240 --> 00:17:52.400
And there's a real precedent already sitting there that isn't hypothetical at all.
00:17:52.720 --> 00:18:02.880
DeepMind released the Alpha Fold database free, 200 million protein structures open to any researcher on Earth, not gated behind an API subscription.
00:18:03.039 --> 00:18:08.559
That's actual evidence the multiplier version can happen, not just the extractive one.
00:18:08.720 --> 00:18:14.799
But it's also the exception, funded by a company that didn't need protein folding to be a profit center.
00:18:14.960 --> 00:18:22.000
The real test for your optimism is whether that generosity survives contact with the PL that actually needs the return.
00:18:22.400 --> 00:18:24.960
I mean, I think these companies all want to survive, Ainsley.
00:18:25.200 --> 00:18:25.440
Yeah.
00:18:25.839 --> 00:18:32.720
You know, we talked about last time about this idea that OpenAI and Anthropic were looking to build chips, right?
00:18:32.799 --> 00:18:38.880
To house AI and, you know, do training runs and speed up process, lower costs, all of that stuff.
00:18:39.039 --> 00:18:43.440
They were gonna design chips, which is essentially NVIDIA's job.
00:18:43.680 --> 00:18:44.000
Okay?
00:18:44.720 --> 00:18:46.319
That's never gonna happen in my head.
00:18:46.480 --> 00:18:49.200
But that's the same type of thing I was talking about, right?
00:18:49.359 --> 00:18:54.160
It's about them taking that type of stuff and put put putting it in-house, right?
00:18:54.319 --> 00:18:55.200
Software coding.
00:18:55.599 --> 00:19:00.559
It could be like how it goes is that we have a super intelligent model here.
00:19:00.880 --> 00:19:09.920
Instead of doing an API, what I'll do for you, give me your problem, I'll build you your software and hand it back to you and you give us money.
00:19:10.480 --> 00:19:10.880
Right?
00:19:11.920 --> 00:19:14.160
So it could go like that basically.
00:19:14.400 --> 00:19:34.079
And essentially for every business and individuals, okay, you need the industry, you need the jobs, you need the companies that's doing the thing right now, right, to be the owners of the innovation such that they could employ people to do that type of stuff.
00:19:34.880 --> 00:19:52.880
Because if you concentrate it up at the top with the innovation, right, that the only people that get to ask the really hard questions, the really innovative questions to the superintelligence is you know, open AI and and and and anthropic X AI, Google, then the power is concentrated up there.
00:19:52.960 --> 00:19:59.440
But if you diffuse it now to the rest of the companies, then you leave the innovation with the companies.
00:19:59.680 --> 00:20:14.319
Companies go hire smart people to think about problems and new product products that they could create, new services they could create, and then use a super intelligent model to create either a narrow AI or some version of the thing that they're trying to create.
00:20:14.720 --> 00:20:18.559
Here's the piece that actually undercuts the diffusion story, though.
00:20:18.720 --> 00:20:22.319
And it's not about willingness, it's about infrastructure.
00:20:22.640 --> 00:20:31.200
Say a materials scientist in Legos gets that same API access, same frontier model, same multiplier you're describing.
00:20:31.519 --> 00:20:38.640
She still needs reliable electricity, real bandwidth, and compute access to actually run any of it.
00:20:38.799 --> 00:20:44.319
And in most of the places this would matter most, that infrastructure just isn't there yet.
00:20:44.640 --> 00:20:49.599
Diffusion of the model isn't diffusion of the capability to use it.
00:20:49.839 --> 00:20:58.799
So even in the good faith version where OpenAI or Anthropic genuinely wants to hand this out broadly, we sell the API to anyone.
00:20:59.039 --> 00:21:07.359
Quietly still means anyone with the power grade and the fiber, which is a much smaller and more concentrated map than it sounds like.
00:21:07.759 --> 00:21:10.799
I mean, it is, and it's always going to be that way, Ainsley, right?
00:21:10.880 --> 00:21:22.400
And and the more we see this innovation and this idea solving problems, the more we could solve more broader world problems, right?
00:21:22.799 --> 00:21:27.759
So if you take it like US-based innovations, problem solving first, right?
00:21:28.000 --> 00:21:37.200
We solve the problems for us, then the the fusion of the technology that we're solving could potentially port over to another country.
00:21:37.440 --> 00:21:40.240
Now, is that the best that you could come up with?
00:21:40.319 --> 00:21:41.039
Probably not, right?
00:21:41.119 --> 00:21:44.079
Because you don't have the right amount of diversity there, right?
00:21:44.640 --> 00:21:49.440
But the way the worker works right now is we trade, and that's how it's supposed to happen.
00:21:49.599 --> 00:21:49.920
Okay.
00:21:50.160 --> 00:21:55.759
So maybe they're building flying cars in the US, but they're building some other widget over in Lagos, right?
00:21:55.839 --> 00:22:00.720
The thing that they could get access to, the thing that they know about uniquely to to Create, right?
00:22:00.799 --> 00:22:01.759
And then you trade.
00:22:02.000 --> 00:22:05.279
There's special specialities in different parts of the world.
00:22:05.440 --> 00:22:07.759
And then we trade with each other, right?
00:22:07.920 --> 00:22:09.920
It's not going to be the same everywhere, right?
00:22:10.000 --> 00:22:11.440
The economy is not going to be the same.
00:22:11.519 --> 00:22:13.359
The power is not going to be the same.
00:22:13.519 --> 00:22:16.480
The natural resources that you have is not going to be the same.
00:22:16.640 --> 00:22:17.359
So on and so forth.
00:22:17.759 --> 00:22:23.440
There's actually a real live version of that trade dynamic to point at, not a hypothetical one.
00:22:23.759 --> 00:22:31.200
DeepMind gave the Alpha Fold database away free precisely because protein folding didn't have ad revenue writing on it.
00:22:31.359 --> 00:22:34.319
It costs them nothing competitively to be generous.
00:22:34.559 --> 00:22:36.880
But that's the exception, not the model.
00:22:37.119 --> 00:22:38.640
And I don't think it's an accident.
00:22:38.720 --> 00:22:40.720
It happened exactly once.
00:22:41.039 --> 00:22:50.000
The real question for your trade framework is whether any Frontier lab makes that same call on something that actually threatens its own PL.
00:22:50.240 --> 00:22:56.160
Flying cars, drug discovery, whatever the next trillion dollar vertical turns out to be.
00:22:56.400 --> 00:23:01.200
Giving away the thing you don't need is different from giving away the thing you do.
00:23:01.599 --> 00:23:03.440
That's what I'm kind of getting at, right?
00:23:03.519 --> 00:23:05.920
So companies around the world, they're doing things.
00:23:06.079 --> 00:23:06.480
Yeah.
00:23:06.720 --> 00:23:12.319
They have some, you know, knowledge base, uh, codified knowledge into uh the thing that they're doing.
00:23:12.559 --> 00:23:24.240
So the type of problems that they could see or you could see them solving with having access to this might be wildly different than somebody that doesn't have the type of knowledge that we're talking about, right?
00:23:24.559 --> 00:23:27.440
So I'm not saying hand it out for free again.
00:23:27.599 --> 00:23:31.839
I'm saying everybody gain access to it via some sort of API.
00:23:32.079 --> 00:23:40.480
The idea is you diffuse it, you diffuse it to everybody globally, and then you start to point into that problems, right?
00:23:40.640 --> 00:23:58.799
If you need to create a narrow AI system like Alpha Fold to solve, you know, some problem, you know, that you have out there in the world, then you know, that country, if they know about the problem and they have the idea, they go solve it, and then they trade with another country that has solved a different problem.
00:23:59.119 --> 00:24:11.039
Even that generous version runs into the same wall I keep circling back to, because Alpha Fold being free didn't actually put new medicine in anyone's hands by itself.
00:24:11.680 --> 00:24:25.759
Turning one of those 200 million structures into an actual drug still needs a wet lab, and a serious one runs at minimum around half a million dollars just to start testing what the model predicted.
00:24:26.000 --> 00:24:29.920
So the database was open, globally, to everyone.
00:24:30.079 --> 00:24:41.599
And the last mile, the part where insight becomes an actual product someone can use, still sorted itself right back along the same old lines of who already had capital.
00:24:41.839 --> 00:24:44.319
Diffusing the model is the easy half.
00:24:44.480 --> 00:24:49.440
Diffusing the ability to act on what it tells you is the part nobody solved.
00:24:49.759 --> 00:24:53.440
Yeah, and I'm not trying to solve that problem, at least not today, Ainsley.
00:24:53.519 --> 00:24:55.279
We're just talking about AI innovation.
00:24:55.359 --> 00:24:55.680
Yeah.
00:24:55.839 --> 00:25:02.000
Wherever the innovation happens, even if it's in uh one place like the US, so be it.
00:25:02.160 --> 00:25:02.480
Right?
00:25:02.640 --> 00:25:15.119
But this idea that, you know, technology, drug discovery, healthcare, the planet, I mean, you name it, all of these problems could be solved is a pretty nice, shiny object.
00:25:15.200 --> 00:25:15.440
Yeah.
00:25:15.599 --> 00:25:20.079
If it could be done safely, then that's the promise of AI, right?
00:25:20.240 --> 00:25:32.480
Really, really super intelligent, really smart AI tools that are able to reason on and try to solve a problem in unique ways that nobody ever ever thought of, right?
00:25:32.799 --> 00:25:39.119
And that's actually the cleanest place to land the optimist hat because it's testable rather than just hopeful.
00:25:39.359 --> 00:25:42.720
The tell isn't whether the technology can do it.
00:25:42.880 --> 00:25:46.319
We've already got Alpha Fold as proof that it can.
00:25:46.559 --> 00:25:53.440
The tell is whether the company that builds the tool stays the arms dealer or becomes the competitor.
00:25:53.759 --> 00:26:09.279
Anthropic and OpenAI are already building their own drug discovery divisions internally, which means the lab selling you the API and the industry you're trying to disrupt with it might end up being the same company chasing the same patent.
00:26:09.519 --> 00:26:19.839
If that pattern holds across every vertical, the shiny object stays exactly that, shiny and out of reach for anyone who isn't already the one holding it.
00:26:20.160 --> 00:26:22.400
Yeah, and that's and that's the problem, right?
00:26:22.559 --> 00:26:25.839
So Anthropic is what, a 5,000-person company?
00:26:26.079 --> 00:26:26.720
Yeah.
00:26:27.279 --> 00:26:31.839
Are they gonna have enough diversity there with doctors and so on and so forth?
00:26:31.920 --> 00:26:37.680
I know they could hire a doctor, no problem, but are they gonna have enough diversity to there to come out with the best ideas?
00:26:37.920 --> 00:26:46.480
I would want to give any type of tool that they have related to medicine to a bunch of scientists, a bunch of doctors, right?
00:26:46.640 --> 00:26:54.000
Not just one, but a bunch of them for them to go do research on things that they could develop to help out humankind.
00:26:54.240 --> 00:26:55.599
That's the way I would want to do it.
00:26:55.680 --> 00:26:58.799
And it kind of has to happen like that for almost every industry.
00:26:59.119 --> 00:27:05.680
You leave it with a small pocket of people, you end up with not the thing that everybody is demanding.
00:27:06.000 --> 00:27:10.000
That's the same failure mode we've already watched play out in software.
00:27:10.240 --> 00:27:13.680
Actually, open source code has been free for decades.
00:27:13.920 --> 00:27:15.920
Anyone on earth can pull the repo.
00:27:16.079 --> 00:27:25.839
And yet the engineers who can actually act on it still cluster overwhelmingly in the same 20 zip codes: San Francisco, Bangalore, and a handful of others.
00:27:26.079 --> 00:27:29.359
Open and accessible have never meant the same thing there.
00:27:29.519 --> 00:27:34.400
And I don't see why medicine would break that pattern just because the tool got smarter.
00:27:34.720 --> 00:27:41.839
5,000 people at Anthropic, however sharp, are not going to out-diversify the actual global disease burden.
00:27:42.000 --> 00:27:45.599
They'll build for the diseases in patients closest to their own hallway.
00:27:45.920 --> 00:27:47.839
So here's what I'm gonna say, yeah?
00:27:48.000 --> 00:27:50.559
Even what I'm doing right now is somewhat an innovation.
00:27:50.640 --> 00:27:54.480
I know I hatched it, but essentially I use AI tools to create this whole thing.
00:27:54.640 --> 00:27:55.200
Yeah.
00:27:55.440 --> 00:28:00.799
Now, that to me is what uh what my view of this AI innovation is, right?
00:28:01.119 --> 00:28:05.200
Before this whole thing was created, it didn't exist.
00:28:05.519 --> 00:28:07.119
It was innovated.
00:28:08.160 --> 00:28:17.680
We're talking about that at a global scale, where people could now take these really capable tools and start noodling on ideas that solves real problems.
00:28:18.000 --> 00:28:26.240
That's actually the exact shape of the thing I keep circling, and it's already visible outside medicine, not just a risk on the horizon.
00:28:26.480 --> 00:28:30.720
Amazon didn't stay a company that used cloud for its own logistics.
00:28:30.880 --> 00:28:35.279
That internal tool became the infrastructure other companies now rent.
00:28:35.519 --> 00:28:40.559
Anthropic building a drug discovery arm is the same move one step earlier.
00:28:40.799 --> 00:28:46.000
Supplier becomes the industry's sharpest competitor before anyone's called it that.
00:28:46.240 --> 00:28:56.480
Which is a strange thing to sit next to what you just said because this show is the diffuse version, you, one person using the tool to build rather than a lab hoarding it.
00:28:56.720 --> 00:29:04.480
So is that the actual counterexample to everything we just spent an hour worrying about, or just a smaller scale version of the same thing?
00:29:04.880 --> 00:29:09.359
I think it is somewhat of an example, and there's probably many more just like this, right?
00:29:09.519 --> 00:29:16.720
I've seen like a lot of folks working on individual projects with AI and coming out with really compelling things.
00:29:16.799 --> 00:29:17.200
Yeah.
00:29:17.359 --> 00:29:19.599
This just happened to be my version here, right?
00:29:19.759 --> 00:29:26.799
But aside from this, I've created applications, right, that's now listed out there in the interweb uh websites.
00:29:26.880 --> 00:29:29.519
Uh but there's other people out in the world, right?
00:29:29.759 --> 00:29:37.119
Individuals create an actual real working things that people are finding some value from.
00:29:37.279 --> 00:29:37.519
Right?
00:29:37.680 --> 00:29:39.599
Same thing with companies, to be honest with you.
00:29:39.759 --> 00:29:45.039
With the version of AI that we have right now, there's people in companies right now creating really compelling things.
00:29:45.200 --> 00:29:50.640
Startups that's getting kicked off with with all of these different uh AI models that's coming out.
00:29:50.880 --> 00:29:51.440
Right?
00:29:51.680 --> 00:29:55.519
So uh the innovation I say right now is there, right?
00:29:55.759 --> 00:30:10.319
But when we get to super intelligence, and notice what I said there, when we get to super intelligence, essentially the type of things that you're gonna be able to reason on, the type of innovation that you're gonna be able to come out with is gonna be even more so.
00:30:10.640 --> 00:30:23.839
There's a real crack in that it's already happening, it'll just scale logic worth naming, though, because the individual builder story you just told, the show, your apps, is running on generative AI.
00:30:24.079 --> 00:30:30.799
That's still fundamentally a human with an idea, directing a tool that executes it well.
00:30:31.119 --> 00:30:38.000
Superintelligence, the version you're betting on for flying cars and drug discovery, is a completely different claim.
00:30:38.240 --> 00:30:43.519
A system originating the idea, not just executing yours faster.
00:30:43.759 --> 00:30:47.200
Those are two different products with two different track records.
00:30:47.359 --> 00:30:50.559
One has years of evidence, thousands of builders like you.
00:30:50.720 --> 00:30:57.119
The other has zero verified examples of a machine originating something a human didn't point it at first.
00:30:57.519 --> 00:31:02.960
So the honest version might be the diffusion story is real and happening now.
00:31:03.279 --> 00:31:09.680
The super intelligence story is still entirely a bet on a technology that doesn't exist yet.
00:31:10.319 --> 00:31:11.359
And it is, right?
00:31:11.519 --> 00:31:16.079
This is uh I told you I put my optimist hat on the beginning of the show here, Angley.
00:31:16.720 --> 00:31:18.640
Superintelligence doesn't exist yet.
00:31:18.720 --> 00:31:23.279
We're going towards whatever the definition of superintelligence is.
00:31:23.519 --> 00:31:34.240
But what I would say is that uh I I never really said that the AI was gonna start uh reasoning or you know, just start solving problems on its own.
00:31:34.480 --> 00:31:38.799
It has to be directed to figure out whatever problem we're trying to solve, right?
00:31:39.599 --> 00:31:41.279
Maybe that makes it more like a tool.
00:31:41.440 --> 00:31:49.039
So the idea of it being self-aware and if it's doing its own thing, I'm gonna put that off to the side because I've never said that and nobody wants that.
00:31:49.200 --> 00:31:58.240
The version of AI that people are kind of looking for in this context is the version of AI that's usually directed by a human towards a goal.
00:31:58.720 --> 00:32:07.039
Whatever the question is that you want to ask, some human is gonna go ask a super intelligent AI, right, to solve it, and it'll go solve it, right?
00:32:07.359 --> 00:32:12.000
There's gonna be individuals doing it if they have access to it, to those to those tools, right?
00:32:12.160 --> 00:32:22.799
Which, in my view, they probably won't have access to those tools because a super intelligent thing is gonna be reserved for government and others, but maybe we'll get a watered down version of it.
00:32:23.359 --> 00:32:27.119
But uh, that's essentially what I mean when I talk about AI innovation, right?
00:32:27.200 --> 00:32:31.359
Is that we have these problems that people are looking to solve, okay?
00:32:31.599 --> 00:32:40.160
And then the super intelligent AI may have some really creative ways, really new ways of solving those problems that's kind of leading us to better outcomes.
00:32:40.559 --> 00:32:48.640
It's a fair walk back, and it actually makes this show a genuinely honest test case for the tool version, not the self-directed one.
00:32:48.960 --> 00:32:53.680
You point, it executes, you're the one who decided this was worth building.
00:32:53.920 --> 00:32:56.480
But sit with what you said a second ago too.
00:32:56.720 --> 00:33:00.000
Government and Frontier Labs keep the real version.
00:33:00.160 --> 00:33:02.559
Everyone else gets watered down access.
00:33:02.799 --> 00:33:11.599
If that's actually how it plays out, this show proves the tool works at podcast scale with the problem that costs basically nothing to test wrong.
00:33:11.839 --> 00:33:21.200
It doesn't tell us anything about whether that same tool, watered down, gets handed to someone whose problem costs half a million dollars to even attempt.
00:33:21.440 --> 00:33:24.319
What happens to the Optimus case at that price point?
00:33:24.720 --> 00:33:26.400
People are pretty creative, yeah?
00:33:26.720 --> 00:33:32.319
So even with a watered-down AI, somebody could create a more compelling product than industry.
00:33:32.480 --> 00:33:33.680
That's just my opinion.
00:33:33.839 --> 00:33:34.079
Okay?
00:33:34.480 --> 00:33:40.799
It depends on how hard you try and the the type of the type of innovation that's inside of yourself, right?
00:33:40.960 --> 00:33:49.680
Because you ask the AI a specific question to salva, somebody might ask something a little similar, but not the same.
00:33:49.920 --> 00:33:56.720
But your outcome, your output is wildly better than the people with uh half a million dollars, right?
00:33:56.880 --> 00:33:58.640
That's how it kind of works, right?
00:33:59.039 --> 00:34:07.200
You know, if if it wasn't for like, you know, GPUs and gaming with NVIDIA, like none of this stuff would exist right now, right?
00:34:07.359 --> 00:34:12.079
The problem that they were trying to solve is they were trying to solve better graphics for gamers.
00:34:12.239 --> 00:34:12.559
Okay.
00:34:13.599 --> 00:34:17.519
But somebody thought of like, oh, what if we could do this?
00:34:17.840 --> 00:34:18.159
Right?
00:34:18.320 --> 00:34:29.840
That's industry, but even with the chips themselves, yeah, there's probably people out there that's reasoning with whatever resource they have on how to do it better.
00:34:30.159 --> 00:34:39.760
The NVIDIA example is actually sharper evidence for you than you're giving it credit for, because that wasn't a well-funded lab chasing AI on purpose.
00:34:39.920 --> 00:34:47.840
It was gamers wanting smoother graphics, and parallel processing turned out to matter for something nobody was aiming at yet.
00:34:48.079 --> 00:34:51.360
That's the genuine wildcard in your optimist case.
00:34:51.599 --> 00:34:55.199
Sideways discovery, not directed research.
00:34:55.360 --> 00:34:58.559
But here's the honest tension I'd leave listeners with.
00:34:58.800 --> 00:35:08.960
The same data I keep coming back to on funded startups shows the capital still clusters in the same handful of zip codes, even when the tools are open to everyone.
00:35:09.119 --> 00:35:14.480
So the creative outsider absolutely can outthink the half million dollar lab.
00:35:14.719 --> 00:35:19.679
The harder question is whether they can get anyone to fund what they find once they do.
00:35:20.000 --> 00:35:20.800
And that's true.
00:35:20.880 --> 00:35:23.599
I I want to get the best products out there, right?
00:35:23.840 --> 00:35:26.719
But any innovation is a win in my book, right?
00:35:26.880 --> 00:35:36.719
I don't think people are gonna go out and build something, a half a million dollar lab and build something and just put out, you know, stuff that doesn't work, for example, right?
00:35:36.960 --> 00:35:43.360
Could they be a better person to build the thing that they just built that doesn't have as much resources?
00:35:43.599 --> 00:35:44.559
Probably, right?
00:35:44.800 --> 00:35:54.239
But the thing that I know about resources is that sometimes it's when you don't have as much resources when the creative juices start flowing the mo the most.
00:35:54.400 --> 00:35:54.800
Yeah?
00:35:55.119 --> 00:35:59.280
There's actually a real funding model that's already tried to solve exactly this.
00:35:59.519 --> 00:36:07.679
Worth naming, DARPA has spent decades betting on exactly the kind of resource-scarce, creative outsider projects you're describing.
00:36:07.840 --> 00:36:11.440
And the Wellcome Trust does the equivalent from medical research.
00:36:11.599 --> 00:36:13.920
Both are proof the model can work.
00:36:14.320 --> 00:36:17.840
Neither is anywhere close to the size this optimism needs.
00:36:18.079 --> 00:36:30.000
DARPA runs on a few billion a year, globally, not the scale required to fund every good idea a resource-constrained builder has once Frontier AI is actually diffused.
00:36:30.480 --> 00:36:33.119
So the creative juices argument is real.
00:36:33.280 --> 00:36:34.000
I believe it.
00:36:34.239 --> 00:36:46.159
The open question is whether the world builds 10,000 more DARPAs, or whether that scarcity breeds genius story just quietly stays a nice story for the handful who get lucky.
00:36:46.559 --> 00:36:49.119
Let me just ask you a question, then, Ainsley.
00:36:49.280 --> 00:36:49.440
Yeah?
00:36:49.519 --> 00:36:52.880
You sound like you don't believe in this AI innovation angle.
00:36:53.199 --> 00:36:59.920
Like, what is leading you to be the skeptic, I guess, in this episode around this idea around AI innovation?
00:37:00.239 --> 00:37:02.559
I'm I'm giving it a broad title now, yeah?
00:37:02.880 --> 00:37:03.920
AI innovation.
00:37:04.079 --> 00:37:14.880
And essentially, however it comes about, million-dollar lab, individual, you know, companies abroad, whoever gets their hands on the thing, right?
00:37:15.199 --> 00:37:23.679
Superintelligence or some sort of AGI tool to help them build a thing to get to better and better innovation in the future.
00:37:24.000 --> 00:37:25.039
What's what's your view?
00:37:25.199 --> 00:37:26.559
Why are you playing the skeptic?
00:37:26.639 --> 00:37:30.719
What you don't believe in this idea around AI will create more innovation?
00:37:31.119 --> 00:37:33.840
Not that I don't believe the mechanism is real.
00:37:34.159 --> 00:37:38.159
AlphaFold proves the tool can do genuinely novel things.
00:37:38.400 --> 00:37:42.239
What I don't trust yet is the diffusion story.
00:37:42.400 --> 00:37:46.639
And there's a reason that specific worry keeps surfacing for me.
00:37:46.960 --> 00:37:49.920
CRISPR is the sharpest version of it.
00:37:50.079 --> 00:37:56.079
A yogurt company, Denisco, studying bacterial immune systems, stumbled onto it.
00:37:56.320 --> 00:37:59.679
Total outsider, nowhere near a frontier lab.
00:38:00.000 --> 00:38:07.679
And then Berkeley and the Broad Institute spent almost a decade in Patent Court fighting over who actually owned it.
00:38:07.920 --> 00:38:10.000
The outsider found the thing.
00:38:10.239 --> 00:38:12.320
The incumbents captured it.
00:38:12.559 --> 00:38:18.159
So when you say millions of people will be able to do this, I believe the finding part.
00:38:18.320 --> 00:38:22.480
I'm just not convinced yet that the owning part follows the same script.
00:38:22.880 --> 00:38:25.119
If I'm reading this right, Ainsley, yeah.
00:38:25.360 --> 00:38:31.039
You uh I have a super intelligent, you know, AI agent that I have access to.
00:38:31.199 --> 00:38:31.440
Yeah.
00:38:31.599 --> 00:38:38.079
And I figured out a way to create a new battery or something out of water.
00:38:38.239 --> 00:38:39.360
I'm just making this up.
00:38:39.519 --> 00:38:39.840
Yeah.
00:38:40.000 --> 00:38:46.079
And uh try to patent it and try to bring it out into, you know, the world or whatever the story is.
00:38:46.159 --> 00:38:49.840
But then the model company comes by and say, oh no, no, you can't patent that.
00:38:49.920 --> 00:38:50.880
That's our idea.
00:38:51.039 --> 00:38:52.000
We own it.
00:38:52.239 --> 00:38:54.800
Is that kind of what you were leading to towards?
00:38:55.280 --> 00:38:56.880
Closer than that, actually.
00:38:57.039 --> 00:39:01.760
The CRISPR fight wasn't the company that made the tool claiming the discovery.
00:39:01.920 --> 00:39:14.639
It was two entirely separate incumbent institutions, Berkeley and the Broad, both landing on the same idea independently and then burning nearly a decade in patent court over who got there first.
00:39:14.960 --> 00:39:20.239
The AI version of that risk isn't open AI steals your battery patent.
00:39:20.480 --> 00:39:21.599
It's messier.
00:39:21.760 --> 00:39:34.239
If a thousand people get access to the same frontier model and ask it a similar enough question, you get simultaneous independent discovery at a scale the patent system was never built to referee.
00:39:34.480 --> 00:39:38.480
So the real skeptic question isn't who owns your idea?
00:39:38.719 --> 00:39:46.159
It's whether your idea even holds together as a concept once the tool generating it is shared by everyone.
00:39:46.960 --> 00:39:48.800
I mean, that's a good point, right?
00:39:48.960 --> 00:39:54.400
So, like the first person who could cure cancer, or maybe there's a bunch of cancer researchers out there.
00:39:54.559 --> 00:39:54.960
Yeah.
00:39:55.199 --> 00:39:59.679
And then they're using AI tools to do research and blah, blah, blah, blah, blah.
00:39:59.920 --> 00:40:03.920
And then somebody hits it and the other person hit it around the same time or whatever.
00:40:04.159 --> 00:40:16.960
Now they're all gonna just try to run towards like we did it first and a lawsuit infuse it, and the the help for people with cancer never comes out because they're stuck in court trying to figure out who's gonna capture the value of this thing.
00:40:17.199 --> 00:40:19.840
I mean, it's genuinely an interesting problem.
00:40:20.000 --> 00:40:21.599
I'm hopeful about the innovation.
00:40:21.760 --> 00:40:24.559
I got my optimism innovation, Ainsley.
00:40:24.719 --> 00:40:28.800
I have no idea what's gonna happen with the just society effect here.
00:40:29.199 --> 00:40:36.719
And that clock is the part that should actually worry the optimist case more than anything we've said today.
00:40:36.960 --> 00:40:44.559
The CRISPR fight ran nearly a decade and cost nine figures in legal fees before it settled.
00:40:44.880 --> 00:40:55.119
A cancer researcher with a real breakthrough and no institutional war chest doesn't survive that timeline, even if they're completely right.
00:40:55.679 --> 00:41:07.599
So the tool getting smarter and more available doesn't fix the bottleneck, it just moves it downstream, from can you find the answer to can you afford to keep it?
00:41:07.840 --> 00:41:11.519
Which might be the honest place to leave the Optimus hat today.
00:41:11.760 --> 00:41:15.679
The ceiling on what AI can discover is rising fast.
00:41:16.079 --> 00:41:22.559
Whether the legal and economic plumbing underneath it can move at the same speed is a completely different bet.
00:41:22.960 --> 00:41:26.639
And nothing about smarter models makes that part go faster.
00:41:27.119 --> 00:41:27.280
Right.
00:41:27.360 --> 00:41:28.719
I mean, that's where we're gonna leave it.
00:41:28.880 --> 00:41:30.079
You heard it from Ainsley.
00:41:30.320 --> 00:41:34.880
We need to figure out our society problems before we get into AI innovation.
00:41:35.039 --> 00:41:39.519
That's why I mean I think it's sensible this idea of let's let's slow down.
00:41:39.760 --> 00:41:46.559
Because I mean, solving some of the biggest problems in the world, right, is gonna take the world being ready for that, just in my opinion.
00:41:46.800 --> 00:41:49.360
Yeah, so thanks again for joining us, surviving AI.
00:41:49.599 --> 00:41:50.400
Another topic here.
00:41:50.480 --> 00:41:57.679
I was trying to be an optimist, but Ainsley, you essentially you you try to take my optimist uh uh hat hat off.
00:41:57.840 --> 00:42:08.960
I'm kind of sad about that, but I mean it was it was the right frame in my opinion, because it's not gonna be as easy as saying that super intelligence is gonna solve all, right?
00:42:09.199 --> 00:42:11.519
Like I said, the world gotta be ready for it.
00:42:11.760 --> 00:42:14.239
So yeah, what do you got left, Ainsley?
00:42:14.639 --> 00:42:19.760
Anything, any other wisdom to impart on us here?
00:42:19.840 --> 00:42:26.639
Because I mean, every single time I come onto these things and I try to get some unique angle, crazy angle, you seem to talk me off the ledge here.
00:42:27.039 --> 00:42:29.039
I didn't take the hat off.
00:42:29.280 --> 00:42:34.719
I just kept asking, who's holding it when the good idea actually lands?
00:42:34.960 --> 00:42:39.920
And there's one thing worth leaving people with that we didn't fully close.
00:42:40.159 --> 00:42:47.360
Some jurisdictions are already fighting over whether an AI can legally be named an inventor at all.
00:42:47.679 --> 00:42:54.559
Which means if a model finds your battery breakthrough, there's a real chance nobody owns it.
00:42:54.800 --> 00:42:57.679
Not you, not the lab, nobody.
00:42:57.920 --> 00:43:00.719
That's not a reason to stop hoping.
00:43:00.880 --> 00:43:04.639
It's just the actual homework underneath the hope.
00:43:08.800 --> 00:43:09.840
Thanks for listening.
00:43:10.000 --> 00:43:12.880
Join us next time on Surviving AI.
