인터뷰/예측
코히어 CEO 에이단 고메즈 "AI 기술이 경제 전반에 영향을 미치는 데에는 2~3년이 걸릴 것"
작성자
작성일
2024-10-10 22:41
조회
2793
https://open.spotify.com/episode/6rOLt2BtH4MO0U9FsDS8FC?si=lsvQeibLSU6EEiP5KP3K2A&nd=1&dlsi=40b18fdaf4824433
00:08:27s I think for real impact, we're probably about two to three years away. That's how long it's gonna take for this tech to reach all the systems and places that drive impact across the economy.
고메즈는 어린 시절 캐나다 시골에서 느린 인터넷 환경을 경험하며 기술에 대한 갈증을 느꼈고, 이것이 AI 분야로의 진로를 결정하는 계기가 되었다고 합니다. 그는 현재 AI 기술 발전 속도가 예상보다 훨씬 빠르다고 말하며, 자신이 예상했던 50년 뒤의 기술 수준에 7년 만에 도달했다고 놀라워했습니다. 특히 AI 모델의 능력과 신뢰성 향상에 큰 충격을 받았다고 언급했습니다.
고메즈는 AI 기술의 접근성 문제를 중요하게 생각합니다. 현재 AI 모델은 크기가 매우 커서 실행이 어렵지만, 코히어는 모델의 효율성을 높이고 크기를 줄이는 데 집중하고 있습니다. 그는 곧 개인용 기기에서도 AI 모델을 로컬로 실행할 수 있게 되어 인터넷 접근성이 낮은 사람들에게도 AI 기술이 보급될 것이라는 희망을 내비쳤습니다.
AGI 달성 시점에 대해서는 고정된 날짜는 없다고 봅니다. 이미 여러 분야에서 AI 모델이 인간과 동등한 수준의 성능을 보이고 있기 때문입니다. 다만 각 분야별로 AI가 인간 최고 전문가 수준에 도달하는 시점은 다를 것이며, 기술이 완성된 후에도 경제 시스템에 통합되어 실질적인 효과를 내기까지는 시간이 걸릴 것으로 예상합니다. 그는 기술 도입의 가장 큰 걸림돌은 자금이나 인간의 두려움이 아닌, 대규모 언어 모델(LLM)을 다룰 수 있는 전문 인력의 부족과 기술 통합에 필요한 시간 및 관심이라고 지적했습니다. 그에 따르면 AI 기술이 경제 전반에 영향을 미치는 데에는 2~3년이 걸릴 것이라고 예측했습니다.
고메즈는 현재 AI에 대한 논의가 부정적인 면에 치우쳐져 있으며, AI가 가져올 기회에 대한 논의가 부족하다고 지적했습니다. 그는 AI를 통해 인류가 노동 부담에서 벗어나 더 가치 있는 일에 집중할 수 있는 새로운 산업 혁명이 도래할 것이라고 전망합니다. 그는 향후 5~10년 안에 모든 사람이 AI 비서를 통해 업무 효율을 높이고 다양한 작업을 자동화하는 미래를 기대하고 있습니다. 또한, AI 기술 발전에 따라 기업들이 특정 공급자에 종속되는 것을 막기 위해 독립적인 AI 기업의 중요성을 강조하며, 정부가 경쟁을 보호하고 역동적인 시장을 조성하는 데 힘써야 한다고 주장했습니다.
스크립트
00:00:00s
Millions of people already count on Zelle to send and receive money, and our impact doesn't stop there.
00:00:05s
More than 40 million Americans saw our safety education content in 2023.
00:00:10s
So when you're sending money to people you know and trust, count on Zelle. Terms and conditions apply.
00:00:16s
Hey, welcome back to Politico Tech.
00:00:19s
Today's Wednesday, October 9th.
00:00:21s
I'm Stephen Overly.
00:00:23s
There's a lot of hand-wringing in tech these days about artificial general intelligence,
00:00:31s
the point at which computers will essentially be as smart as humans.
00:00:36s
Some worry it's just around the corner.
00:00:39s
Others say it's still a long way off.
00:00:42s
Aidan Gomez is the CEO of an AI company called Cohere,
00:00:46s
and he thinks there's actually not one fixed date when AGI will be achieved.
00:00:53s
But AI technology is progressing faster than he imagined.
00:00:58s
We've exceeded my wildest expectations.
00:01:01s
Like, the technology is where I thought it would be in maybe half a century.
00:01:07s
Cohear builds and sells AI models for other companies
00:01:10s
to automate their basic functions and make employees more efficient.
00:01:15s
And Aiden sees a lot of upside to the technology as it exists today,
00:01:20s
even if it hasn't totally surpassed humans.
00:01:24s
The trick, he says, is getting humans to embrace it.
00:01:28s
On the show today, Aiden tells me what that will take,
00:01:32s
how fast he thinks it will happen,
00:01:34s
oh, and how his own AI journey started with shoddy internet
00:01:39s
in the maple forests of Canada.
00:01:41s
Here's our conversation.
00:01:44s
Well, Aiden, welcome to Politico Tech.
00:01:47s
Thank you for having me.
00:01:48s
Excited to be here.
00:01:49s
Before we dive into business and policy, I actually want to ask a little bit about you, because I read that you grew up in rural Canada without high-speed internet access, and I was wondering how that experience propelled you into a career in AI.
00:02:04s
Yeah, I guess, you know, my dad might contest the high speed internet point, but we were definitely on dial up for way, way longer than anyone else in my little village. But yeah, I grew up in rural Canada in a forest. We had like maple trees that we would tap every March and make maple syrup. So it was a super Canadian upbringing and, you know, I'm very grateful for it.
00:02:30s
But definitely technology was hard to come by.
00:02:34s
And we basically only had the bare minimum.
00:02:37s
We had like a telephone cable that ran to our house in the middle of the woods.
00:02:41s
And we had to make do with that.
00:02:43s
And then eventually we flipped to satellite internet.
00:02:46s
But that was like very spotty.
00:02:48s
If it was a cloudy day, you just don't have internet.
00:02:51s
But I think that scarcity of access to technology and to the internet really drove my fascination.
00:02:57s
because it was something, it made the technology,
00:03:00s
it made the internet something that was just out of reach.
00:03:03s
Or like, I knew that I was getting access
00:03:06s
to a worse version than everyone else.
00:03:09s
And so I wanted to make my computer better,
00:03:13s
faster, more effective.
00:03:15s
And to do that, you have to understand how they work.
00:03:17s
And so I learned to code.
00:03:19s
Does that shape your outlook now as a creator of technology?
00:03:22s
Because there are people who won't have access to AI.
00:03:25s
Yeah, I think access is incredibly important. I think, you know, one of the unfortunate truths of the technology is that these models are huge. They're massive. By any standard of a computer program, it is a huge computer program. And so that makes them difficult to run.
00:03:43s
we've always prioritized efficiency.
00:03:46s
So trying to take those massive models,
00:03:48s
trying to make them smaller,
00:03:50s
compress them down, distill them.
00:03:52s
We've been pushing on that super hard
00:03:53s
and I think the field has made
00:03:55s
loads of progress.
00:03:57s
And so I'm hopeful
00:03:59s
we'll start to see compelling models
00:04:02s
that can fit on devices
00:04:03s
and can be local.
00:04:05s
You don't need to be able to hit
00:04:06s
some internet server
00:04:08s
with a big computer in the Midwest.
00:04:11s
I'm hopeful that that will happen soon.
00:04:13s
But frankly, we're not really there yet.
00:04:17s
We're not really at that state.
00:04:19s
But as the chips become faster in our devices
00:04:22s
and as we become better at compressing these models,
00:04:25s
I think we will get more access.
00:04:27s
And the kids like me who don't have great access to internet,
00:04:32s
they'll be able to install it locally
00:04:34s
and just run it right off their laptop.
00:04:37s
You know, it's been five years since you co-founded Cohere.
00:04:40s
has that been the biggest change you seen in AI technology over that time or what has been the biggest change In the past five years since starting Cohere like obviously I wouldn have started the company
00:04:52s
if I didn't think the technology was going to advance
00:04:55s
and we'd reach something extremely compelling
00:04:58s
and useful for the world.
00:05:00s
But we've exceeded my wildest expectations.
00:05:05s
Like the technology is where I thought it would be
00:05:08s
in maybe half a century.
00:05:10s
and it just showed up in like seven years.
00:05:14s
Interesting.
00:05:15s
So yeah, it's been really incredible to watch.
00:05:17s
I think the competence of these models,
00:05:21s
the level of intelligence and reliability
00:05:24s
that we've reached is really shocking
00:05:27s
and very, very promising.
00:05:30s
How do you feel about that?
00:05:31s
Because there are some people who hear that
00:05:33s
and I think would be afraid by the pace of development.
00:05:38s
I imagine you're maybe more excited
00:05:40s
What are your feelings on that?
00:05:42s
It still feels slow, to be honest, I guess because I'm in it every single day.
00:05:46s
And I know how long it takes to train one of these models and to collect the data.
00:05:50s
It's like a months-long process.
00:05:53s
And so when you're inside, it actually feels really slow.
00:05:55s
And you're frustrated, why isn't this coming faster?
00:05:58s
And I think I'm excited for a lot of the good this technology will do.
00:06:01s
And so I want to see it distributed and integrated into the enterprise world, into the consumer world.
00:06:09s
faster. From my perspective, when I look back over eight years, yes, stuff has happened quite quickly,
00:06:16s
but day to day, there's so much more we could be doing if we could move faster with the tech.
00:06:23s
What's your prediction on when we'll have artificial general intelligence? Do you have
00:06:27s
a date in mind for that? There's no hard version of AGI in my estimation. I think that there's no
00:06:37s
date where you can pick and say it's arrived. I think it's partially arrived in many different
00:06:42s
domains. The models are as good as humans across a bunch of different tasks. The question is,
00:06:50s
you have to be domain specific. Like when will a computer be as good in medicine as the best human
00:06:56s
worldwide or as good in physics or as good in mathematics or just in, you know, chit chat and
00:07:02s
entertaining people? So I think each of those questions has a different answer. In general,
00:07:08s
I would say we're heading towards something very, very compelling. And even once the technology
00:07:14s
arrives, once the technology is good across a large swath of domains and equal to humans,
00:07:22s
it will take a while for us to actually realize that benefit because then it needs to be
00:07:27s
integrated into the ecosystem. It needs to be integrated into the economy. And that's really
00:07:33s
what the project is becoming. Because even if you look at the technology as it stands today,
00:07:38s
it's incredibly capable, like extremely capable. But what's slowing us down is we need to implement
00:07:45s
it. And so that integration process really feels like the mission right now. Like that's where
00:07:52s
we're at. What's the biggest obstacle to that integration? Is it money? Is it human fear? Or
00:07:57s
is it something else? I think it's like time and attention. There's very few people who know how to
00:08:03s
work with large language models and do that integration job. And for that, we're trying to
00:08:09s
train as many as we can because there's like this huge technological shift that is actively underway.
00:08:16s
But at the moment, we don't have enough people. There's so much to do. There's so much change and
00:08:21s
implementation to be done. And we just don't have the time and attention to do it very quickly. And
00:08:27s
And so I think for real impact, we're probably about two to three years away.
00:08:34s
That's how long it's going to take for this tech to reach all the systems and places that
00:08:38s
drive impact across the economy.
00:08:41s
That doesn't seem that far away to me, actually.
00:08:43s
But I think that's the technologist's metabolism.
00:08:46s
Maybe you revs faster than most of us.
00:08:49s
You know, it's interesting, though, hearing you talk about the promise of the technology.
00:08:53s
When I started this podcast, I heard often from people about sort of fear of, you know, all-powerful AI. It feels like lately I hear more this argument that AI is overhyped and that, you know, companies and people don't feel like it has been as transformational as they expected. What do you make of that, like, disconnect?
00:09:14s
Well there two things The first is there whiplash So like you were saying ahead of the technology being ready people had very grand visions of I heard it described as like godlike AI or doomsday predictions that are inspired by science fiction And so that set extremely high expectations And we likely whiplashing or reacting from that
00:09:40s
with the reality of the technology,
00:09:42s
which is, okay, fine, not God-like AI,
00:09:45s
but extremely compelling, very useful, productive AI.
00:09:49s
And feeling, yeah, surprised at the reality,
00:09:55s
even if the reality is extraordinarily useful and good.
00:09:58s
And so there will be a settling into what this technology is actually good at and what we should actually deploy it to do. And so I think that's the first thing is that I'm quite empathetic to that whiplash. I don't think we should go too far the other way. We shouldn't call this technology a bubble or anything like that. It's demonstrably not. It's already providing value for hundreds of millions, if not billions of people. It's part of their day to day. They're using it at work. They're using it for fun, for learning.
00:10:28s
And so it's evidently valuable. It's evidently ready for production because it's literally in
00:10:34s
the hands of hundreds of millions of us. The second major thing I think is that I am super
00:10:42s
empathetic to those fears about doomsday scenarios because we've been, before AI even was really
00:10:49s
commonly contemplated as something that can be achieved, we've been telling stories about how
00:10:55s
it could go wrong. And so, so much media has been created to tell those stories. And it's a very
00:11:01s
salient and sort of eye-catching one. But I think they're a distraction from what we should actually
00:11:08s
be talking about, which is this technology is very capable, but we need to understand how to
00:11:14s
deploy it in specific domains and take a more pragmatic approach to thinking about risk. It's
00:11:20s
not terminators. It's, you know, how do we deploy this into healthcare in a way that doesn't put
00:11:25s
people at risk? How do we deploy this into finance in a way that doesn't put people at risk? Those
00:11:31s
are the more concrete and addressable problems that I think the conversation should be focusing
00:11:37s
on. And I hope it does in the future. And is that conversation something that should be happening
00:11:42s
largely within government, you know, or within industry? I know that Coher, for instance,
00:11:47s
has signed on to voluntary commitments from the White House, for instance, on AI safety.
00:11:52s
You've worked with governments in Canada and the UK on their own safety initiatives.
00:11:57s
I wonder how those have sort of changed, if at all, you think the way you're developing AI or
00:12:02s
the way that AI as an industry is progressing. Yeah, I mean, those voluntary commitments,
00:12:08s
like we're very supportive of good regulation and good regulation keeps people safe and it provides
00:12:14s
clarity. And so we're super supportive of that and general principles around building AI that's
00:12:20s
not biased and represents a broad spectrum of viewpoints and perspective, doesn't go one way
00:12:25s
or the other. I do think that guidance from government will be helpful and productive.
00:12:33s
I think that for our customers, we're an enterprise-focused AI platform. And our customers
00:12:39s
are the largest enterprises that have very high stakes deployments and they're very nervous.
00:12:45s
They're actually hesitant to adopt the technology because they're nervous about the consequences.
00:12:51s
And so they're the first ones to bring up, how do we make this safe? Are we being compliant?
00:12:55s
Et cetera, et cetera. And so it'll help our customers in adoption if they know the right
00:13:01s
ways to implement things, the safe way to implement things. And so we try to help them
00:13:05s
work through that. And I think government is trying to help them as well. What do you tell
00:13:11s
them now before those regulations are in place? How do you convince them that AI is safe, that it
00:13:17s
is capable, it is something they should be using today? So the key piece is benchmarking. If you
00:13:23s
can't measure the performance of the model, if you can't compare it to something, you're sort of
00:13:28s
operating with a black box. You don't understand how it's performing. The first thing that we do
00:13:32s
is we ask the customer to create an evaluation.
00:13:36s
Come up with some set of tests and metrics
00:13:38s
and let's measure the humans
00:13:40s
who are performing that task right now.
00:13:42s
How accurate are they?
00:13:43s
Let's measure language models.
00:13:45s
How accurate are they?
00:13:46s
If the language model is better,
00:13:48s
great, we're already there.
00:13:50s
If not, we're going to have to iterate on that technology
00:13:53s
and improve its accuracy and robustness
00:13:55s
to the point where it's useful to those humans.
00:14:02s
We see you next time fraud More than 40 million Americans saw our safety education content in 2023 and countermeasures like real fraud monitoring and detection help protect users
00:14:28s
So when you're sending money to people you know and trust, count on Zelle. Terms and conditions apply.
00:14:37s
Cohere, as we were saying earlier, has been around for five years. You're not a small company.
00:14:42s
You recently raised a half billion dollars from investors to keep growing, but you are competing in a space that dominant players are Google and Microsoft and Meta, OpenAI.
00:14:55s
I wonder how concerned you are about the really big tech players stifling competition as AI is getting off the ground.
00:15:04s
In the enterprise market, choice is extremely important.
00:15:07s
And so we've taken a strategy to be cloud agnostic, chip agnostic, so that our customers can come to us and say, we want you to deploy here on this platform. And I think that's something that needs to be protected. And I think that's borne out in previous generations of emergent technology.
00:15:28s
One good example is databases, which is obviously a key part of AI.
00:15:32s
But you can see that the market has chosen to buy from independents because they don't want to get locked in to one provider.
00:15:39s
And so to have a healthy, competitive, high-choice environment in AI, we need players like Cohere who are prioritizing that.
00:15:48s
Is that something you want to see the government engage more on?
00:15:51s
Because, you know, there are federal agencies like the Federal Trade Commission, for instance, or the Justice Department that are exploring, you know, whether this market is already anti-competitive or too small.
00:16:02s
I'm not sure what the government could do, but definitely we think it's very important that competition is preserved and that we have dynamic competitive markets.
00:16:12s
It's crucial to a healthy economy.
00:16:14s
We were talking earlier about kind of the, you know, Terminator threats and things and some of these.
00:16:19s
It sounds like you see sort of distractions from some of the real issues that should be talked about around AI.
00:16:25s
I'm curious to hear from you what you think is most missing in the debate we're all having right now about AI and how it's managed and, you know, for the future.
00:16:36s
For me, I think some discussion of the opportunity is necessary.
00:16:41s
We spend a lot of time on the fears and the concerns with AI.
00:16:45s
there's almost no discussion about what we want to do with it, what we want to achieve,
00:16:52s
what good could be accomplished, where a company like mine who's developing the technology
00:16:57s
should focus our efforts. It's virtually absent from the conversation at the moment because people
00:17:03s
are so preoccupied with the downsides. Can you paint that picture for me? You were saying earlier
00:17:09s
that some of the impact of this technology is only a few years away, potentially. What does
00:17:15s
that future look like where AI has been implemented and we are sort of seeing benefits from it?
00:17:23s
Humanity is very supply-side constraint. We want to do so much, but we don't have enough experts in
00:17:31s
that field to do it, or we don't have enough time, we can't do it fast enough. And what this
00:17:35s
technology promises is a new industrial revolution where you know if the steam engine took loads off
00:17:44s
of humans backs and animals backs and put it onto machinery instead so that the humans could go and
00:17:49s
do other things that were less burdensome this is doing that same thing for cognitive or intellectual
00:17:57s
labor. We can hand off burdensome tasks onto these models and free us up to do more. And so the
00:18:06s
reality that I want to see in the next half decade or decade is a world where as a worker, you do have
00:18:14s
an assistant that is extremely capable that you can ask to do a nearly arbitrary set of things for
00:18:20s
you and that you're sort of managing a little team of assistants that you can go ask to do tasks for
00:18:26s
you. You can guide them, direct them, tell them where to spend their time and effort instead of
00:18:31s
you having to do those low-level tasks yourself. And so it sounds boring, but the impacts are
00:18:37s
profound. It's boring, but profound. Well, listen, Aidan, appreciate you being here on Politico Tech.
00:18:44s
Yeah, my pleasure. Thank you for having me.
00:18:49s
That's all for today's Politico Tech. For more tech news, subscribe to our newsletters,
00:18:54s
Digital Future Daily and Morning Tech.
00:18:58s
Our managing producer is Annie Reis.
00:19:00s
Our producer is Afra Abdullah.
00:19:02s
I'm Stephen Overly.
00:19:03s
See you back here tomorrow.
END
00:08:27s I think for real impact, we're probably about two to three years away. That's how long it's gonna take for this tech to reach all the systems and places that drive impact across the economy.
고메즈는 어린 시절 캐나다 시골에서 느린 인터넷 환경을 경험하며 기술에 대한 갈증을 느꼈고, 이것이 AI 분야로의 진로를 결정하는 계기가 되었다고 합니다. 그는 현재 AI 기술 발전 속도가 예상보다 훨씬 빠르다고 말하며, 자신이 예상했던 50년 뒤의 기술 수준에 7년 만에 도달했다고 놀라워했습니다. 특히 AI 모델의 능력과 신뢰성 향상에 큰 충격을 받았다고 언급했습니다.
고메즈는 AI 기술의 접근성 문제를 중요하게 생각합니다. 현재 AI 모델은 크기가 매우 커서 실행이 어렵지만, 코히어는 모델의 효율성을 높이고 크기를 줄이는 데 집중하고 있습니다. 그는 곧 개인용 기기에서도 AI 모델을 로컬로 실행할 수 있게 되어 인터넷 접근성이 낮은 사람들에게도 AI 기술이 보급될 것이라는 희망을 내비쳤습니다.
AGI 달성 시점에 대해서는 고정된 날짜는 없다고 봅니다. 이미 여러 분야에서 AI 모델이 인간과 동등한 수준의 성능을 보이고 있기 때문입니다. 다만 각 분야별로 AI가 인간 최고 전문가 수준에 도달하는 시점은 다를 것이며, 기술이 완성된 후에도 경제 시스템에 통합되어 실질적인 효과를 내기까지는 시간이 걸릴 것으로 예상합니다. 그는 기술 도입의 가장 큰 걸림돌은 자금이나 인간의 두려움이 아닌, 대규모 언어 모델(LLM)을 다룰 수 있는 전문 인력의 부족과 기술 통합에 필요한 시간 및 관심이라고 지적했습니다. 그에 따르면 AI 기술이 경제 전반에 영향을 미치는 데에는 2~3년이 걸릴 것이라고 예측했습니다.
고메즈는 현재 AI에 대한 논의가 부정적인 면에 치우쳐져 있으며, AI가 가져올 기회에 대한 논의가 부족하다고 지적했습니다. 그는 AI를 통해 인류가 노동 부담에서 벗어나 더 가치 있는 일에 집중할 수 있는 새로운 산업 혁명이 도래할 것이라고 전망합니다. 그는 향후 5~10년 안에 모든 사람이 AI 비서를 통해 업무 효율을 높이고 다양한 작업을 자동화하는 미래를 기대하고 있습니다. 또한, AI 기술 발전에 따라 기업들이 특정 공급자에 종속되는 것을 막기 위해 독립적인 AI 기업의 중요성을 강조하며, 정부가 경쟁을 보호하고 역동적인 시장을 조성하는 데 힘써야 한다고 주장했습니다.
스크립트
00:00:00s
Millions of people already count on Zelle to send and receive money, and our impact doesn't stop there.
00:00:05s
More than 40 million Americans saw our safety education content in 2023.
00:00:10s
So when you're sending money to people you know and trust, count on Zelle. Terms and conditions apply.
00:00:16s
Hey, welcome back to Politico Tech.
00:00:19s
Today's Wednesday, October 9th.
00:00:21s
I'm Stephen Overly.
00:00:23s
There's a lot of hand-wringing in tech these days about artificial general intelligence,
00:00:31s
the point at which computers will essentially be as smart as humans.
00:00:36s
Some worry it's just around the corner.
00:00:39s
Others say it's still a long way off.
00:00:42s
Aidan Gomez is the CEO of an AI company called Cohere,
00:00:46s
and he thinks there's actually not one fixed date when AGI will be achieved.
00:00:53s
But AI technology is progressing faster than he imagined.
00:00:58s
We've exceeded my wildest expectations.
00:01:01s
Like, the technology is where I thought it would be in maybe half a century.
00:01:07s
Cohear builds and sells AI models for other companies
00:01:10s
to automate their basic functions and make employees more efficient.
00:01:15s
And Aiden sees a lot of upside to the technology as it exists today,
00:01:20s
even if it hasn't totally surpassed humans.
00:01:24s
The trick, he says, is getting humans to embrace it.
00:01:28s
On the show today, Aiden tells me what that will take,
00:01:32s
how fast he thinks it will happen,
00:01:34s
oh, and how his own AI journey started with shoddy internet
00:01:39s
in the maple forests of Canada.
00:01:41s
Here's our conversation.
00:01:44s
Well, Aiden, welcome to Politico Tech.
00:01:47s
Thank you for having me.
00:01:48s
Excited to be here.
00:01:49s
Before we dive into business and policy, I actually want to ask a little bit about you, because I read that you grew up in rural Canada without high-speed internet access, and I was wondering how that experience propelled you into a career in AI.
00:02:04s
Yeah, I guess, you know, my dad might contest the high speed internet point, but we were definitely on dial up for way, way longer than anyone else in my little village. But yeah, I grew up in rural Canada in a forest. We had like maple trees that we would tap every March and make maple syrup. So it was a super Canadian upbringing and, you know, I'm very grateful for it.
00:02:30s
But definitely technology was hard to come by.
00:02:34s
And we basically only had the bare minimum.
00:02:37s
We had like a telephone cable that ran to our house in the middle of the woods.
00:02:41s
And we had to make do with that.
00:02:43s
And then eventually we flipped to satellite internet.
00:02:46s
But that was like very spotty.
00:02:48s
If it was a cloudy day, you just don't have internet.
00:02:51s
But I think that scarcity of access to technology and to the internet really drove my fascination.
00:02:57s
because it was something, it made the technology,
00:03:00s
it made the internet something that was just out of reach.
00:03:03s
Or like, I knew that I was getting access
00:03:06s
to a worse version than everyone else.
00:03:09s
And so I wanted to make my computer better,
00:03:13s
faster, more effective.
00:03:15s
And to do that, you have to understand how they work.
00:03:17s
And so I learned to code.
00:03:19s
Does that shape your outlook now as a creator of technology?
00:03:22s
Because there are people who won't have access to AI.
00:03:25s
Yeah, I think access is incredibly important. I think, you know, one of the unfortunate truths of the technology is that these models are huge. They're massive. By any standard of a computer program, it is a huge computer program. And so that makes them difficult to run.
00:03:43s
we've always prioritized efficiency.
00:03:46s
So trying to take those massive models,
00:03:48s
trying to make them smaller,
00:03:50s
compress them down, distill them.
00:03:52s
We've been pushing on that super hard
00:03:53s
and I think the field has made
00:03:55s
loads of progress.
00:03:57s
And so I'm hopeful
00:03:59s
we'll start to see compelling models
00:04:02s
that can fit on devices
00:04:03s
and can be local.
00:04:05s
You don't need to be able to hit
00:04:06s
some internet server
00:04:08s
with a big computer in the Midwest.
00:04:11s
I'm hopeful that that will happen soon.
00:04:13s
But frankly, we're not really there yet.
00:04:17s
We're not really at that state.
00:04:19s
But as the chips become faster in our devices
00:04:22s
and as we become better at compressing these models,
00:04:25s
I think we will get more access.
00:04:27s
And the kids like me who don't have great access to internet,
00:04:32s
they'll be able to install it locally
00:04:34s
and just run it right off their laptop.
00:04:37s
You know, it's been five years since you co-founded Cohere.
00:04:40s
has that been the biggest change you seen in AI technology over that time or what has been the biggest change In the past five years since starting Cohere like obviously I wouldn have started the company
00:04:52s
if I didn't think the technology was going to advance
00:04:55s
and we'd reach something extremely compelling
00:04:58s
and useful for the world.
00:05:00s
But we've exceeded my wildest expectations.
00:05:05s
Like the technology is where I thought it would be
00:05:08s
in maybe half a century.
00:05:10s
and it just showed up in like seven years.
00:05:14s
Interesting.
00:05:15s
So yeah, it's been really incredible to watch.
00:05:17s
I think the competence of these models,
00:05:21s
the level of intelligence and reliability
00:05:24s
that we've reached is really shocking
00:05:27s
and very, very promising.
00:05:30s
How do you feel about that?
00:05:31s
Because there are some people who hear that
00:05:33s
and I think would be afraid by the pace of development.
00:05:38s
I imagine you're maybe more excited
00:05:40s
What are your feelings on that?
00:05:42s
It still feels slow, to be honest, I guess because I'm in it every single day.
00:05:46s
And I know how long it takes to train one of these models and to collect the data.
00:05:50s
It's like a months-long process.
00:05:53s
And so when you're inside, it actually feels really slow.
00:05:55s
And you're frustrated, why isn't this coming faster?
00:05:58s
And I think I'm excited for a lot of the good this technology will do.
00:06:01s
And so I want to see it distributed and integrated into the enterprise world, into the consumer world.
00:06:09s
faster. From my perspective, when I look back over eight years, yes, stuff has happened quite quickly,
00:06:16s
but day to day, there's so much more we could be doing if we could move faster with the tech.
00:06:23s
What's your prediction on when we'll have artificial general intelligence? Do you have
00:06:27s
a date in mind for that? There's no hard version of AGI in my estimation. I think that there's no
00:06:37s
date where you can pick and say it's arrived. I think it's partially arrived in many different
00:06:42s
domains. The models are as good as humans across a bunch of different tasks. The question is,
00:06:50s
you have to be domain specific. Like when will a computer be as good in medicine as the best human
00:06:56s
worldwide or as good in physics or as good in mathematics or just in, you know, chit chat and
00:07:02s
entertaining people? So I think each of those questions has a different answer. In general,
00:07:08s
I would say we're heading towards something very, very compelling. And even once the technology
00:07:14s
arrives, once the technology is good across a large swath of domains and equal to humans,
00:07:22s
it will take a while for us to actually realize that benefit because then it needs to be
00:07:27s
integrated into the ecosystem. It needs to be integrated into the economy. And that's really
00:07:33s
what the project is becoming. Because even if you look at the technology as it stands today,
00:07:38s
it's incredibly capable, like extremely capable. But what's slowing us down is we need to implement
00:07:45s
it. And so that integration process really feels like the mission right now. Like that's where
00:07:52s
we're at. What's the biggest obstacle to that integration? Is it money? Is it human fear? Or
00:07:57s
is it something else? I think it's like time and attention. There's very few people who know how to
00:08:03s
work with large language models and do that integration job. And for that, we're trying to
00:08:09s
train as many as we can because there's like this huge technological shift that is actively underway.
00:08:16s
But at the moment, we don't have enough people. There's so much to do. There's so much change and
00:08:21s
implementation to be done. And we just don't have the time and attention to do it very quickly. And
00:08:27s
And so I think for real impact, we're probably about two to three years away.
00:08:34s
That's how long it's going to take for this tech to reach all the systems and places that
00:08:38s
drive impact across the economy.
00:08:41s
That doesn't seem that far away to me, actually.
00:08:43s
But I think that's the technologist's metabolism.
00:08:46s
Maybe you revs faster than most of us.
00:08:49s
You know, it's interesting, though, hearing you talk about the promise of the technology.
00:08:53s
When I started this podcast, I heard often from people about sort of fear of, you know, all-powerful AI. It feels like lately I hear more this argument that AI is overhyped and that, you know, companies and people don't feel like it has been as transformational as they expected. What do you make of that, like, disconnect?
00:09:14s
Well there two things The first is there whiplash So like you were saying ahead of the technology being ready people had very grand visions of I heard it described as like godlike AI or doomsday predictions that are inspired by science fiction And so that set extremely high expectations And we likely whiplashing or reacting from that
00:09:40s
with the reality of the technology,
00:09:42s
which is, okay, fine, not God-like AI,
00:09:45s
but extremely compelling, very useful, productive AI.
00:09:49s
And feeling, yeah, surprised at the reality,
00:09:55s
even if the reality is extraordinarily useful and good.
00:09:58s
And so there will be a settling into what this technology is actually good at and what we should actually deploy it to do. And so I think that's the first thing is that I'm quite empathetic to that whiplash. I don't think we should go too far the other way. We shouldn't call this technology a bubble or anything like that. It's demonstrably not. It's already providing value for hundreds of millions, if not billions of people. It's part of their day to day. They're using it at work. They're using it for fun, for learning.
00:10:28s
And so it's evidently valuable. It's evidently ready for production because it's literally in
00:10:34s
the hands of hundreds of millions of us. The second major thing I think is that I am super
00:10:42s
empathetic to those fears about doomsday scenarios because we've been, before AI even was really
00:10:49s
commonly contemplated as something that can be achieved, we've been telling stories about how
00:10:55s
it could go wrong. And so, so much media has been created to tell those stories. And it's a very
00:11:01s
salient and sort of eye-catching one. But I think they're a distraction from what we should actually
00:11:08s
be talking about, which is this technology is very capable, but we need to understand how to
00:11:14s
deploy it in specific domains and take a more pragmatic approach to thinking about risk. It's
00:11:20s
not terminators. It's, you know, how do we deploy this into healthcare in a way that doesn't put
00:11:25s
people at risk? How do we deploy this into finance in a way that doesn't put people at risk? Those
00:11:31s
are the more concrete and addressable problems that I think the conversation should be focusing
00:11:37s
on. And I hope it does in the future. And is that conversation something that should be happening
00:11:42s
largely within government, you know, or within industry? I know that Coher, for instance,
00:11:47s
has signed on to voluntary commitments from the White House, for instance, on AI safety.
00:11:52s
You've worked with governments in Canada and the UK on their own safety initiatives.
00:11:57s
I wonder how those have sort of changed, if at all, you think the way you're developing AI or
00:12:02s
the way that AI as an industry is progressing. Yeah, I mean, those voluntary commitments,
00:12:08s
like we're very supportive of good regulation and good regulation keeps people safe and it provides
00:12:14s
clarity. And so we're super supportive of that and general principles around building AI that's
00:12:20s
not biased and represents a broad spectrum of viewpoints and perspective, doesn't go one way
00:12:25s
or the other. I do think that guidance from government will be helpful and productive.
00:12:33s
I think that for our customers, we're an enterprise-focused AI platform. And our customers
00:12:39s
are the largest enterprises that have very high stakes deployments and they're very nervous.
00:12:45s
They're actually hesitant to adopt the technology because they're nervous about the consequences.
00:12:51s
And so they're the first ones to bring up, how do we make this safe? Are we being compliant?
00:12:55s
Et cetera, et cetera. And so it'll help our customers in adoption if they know the right
00:13:01s
ways to implement things, the safe way to implement things. And so we try to help them
00:13:05s
work through that. And I think government is trying to help them as well. What do you tell
00:13:11s
them now before those regulations are in place? How do you convince them that AI is safe, that it
00:13:17s
is capable, it is something they should be using today? So the key piece is benchmarking. If you
00:13:23s
can't measure the performance of the model, if you can't compare it to something, you're sort of
00:13:28s
operating with a black box. You don't understand how it's performing. The first thing that we do
00:13:32s
is we ask the customer to create an evaluation.
00:13:36s
Come up with some set of tests and metrics
00:13:38s
and let's measure the humans
00:13:40s
who are performing that task right now.
00:13:42s
How accurate are they?
00:13:43s
Let's measure language models.
00:13:45s
How accurate are they?
00:13:46s
If the language model is better,
00:13:48s
great, we're already there.
00:13:50s
If not, we're going to have to iterate on that technology
00:13:53s
and improve its accuracy and robustness
00:13:55s
to the point where it's useful to those humans.
00:14:02s
We see you next time fraud More than 40 million Americans saw our safety education content in 2023 and countermeasures like real fraud monitoring and detection help protect users
00:14:28s
So when you're sending money to people you know and trust, count on Zelle. Terms and conditions apply.
00:14:37s
Cohere, as we were saying earlier, has been around for five years. You're not a small company.
00:14:42s
You recently raised a half billion dollars from investors to keep growing, but you are competing in a space that dominant players are Google and Microsoft and Meta, OpenAI.
00:14:55s
I wonder how concerned you are about the really big tech players stifling competition as AI is getting off the ground.
00:15:04s
In the enterprise market, choice is extremely important.
00:15:07s
And so we've taken a strategy to be cloud agnostic, chip agnostic, so that our customers can come to us and say, we want you to deploy here on this platform. And I think that's something that needs to be protected. And I think that's borne out in previous generations of emergent technology.
00:15:28s
One good example is databases, which is obviously a key part of AI.
00:15:32s
But you can see that the market has chosen to buy from independents because they don't want to get locked in to one provider.
00:15:39s
And so to have a healthy, competitive, high-choice environment in AI, we need players like Cohere who are prioritizing that.
00:15:48s
Is that something you want to see the government engage more on?
00:15:51s
Because, you know, there are federal agencies like the Federal Trade Commission, for instance, or the Justice Department that are exploring, you know, whether this market is already anti-competitive or too small.
00:16:02s
I'm not sure what the government could do, but definitely we think it's very important that competition is preserved and that we have dynamic competitive markets.
00:16:12s
It's crucial to a healthy economy.
00:16:14s
We were talking earlier about kind of the, you know, Terminator threats and things and some of these.
00:16:19s
It sounds like you see sort of distractions from some of the real issues that should be talked about around AI.
00:16:25s
I'm curious to hear from you what you think is most missing in the debate we're all having right now about AI and how it's managed and, you know, for the future.
00:16:36s
For me, I think some discussion of the opportunity is necessary.
00:16:41s
We spend a lot of time on the fears and the concerns with AI.
00:16:45s
there's almost no discussion about what we want to do with it, what we want to achieve,
00:16:52s
what good could be accomplished, where a company like mine who's developing the technology
00:16:57s
should focus our efforts. It's virtually absent from the conversation at the moment because people
00:17:03s
are so preoccupied with the downsides. Can you paint that picture for me? You were saying earlier
00:17:09s
that some of the impact of this technology is only a few years away, potentially. What does
00:17:15s
that future look like where AI has been implemented and we are sort of seeing benefits from it?
00:17:23s
Humanity is very supply-side constraint. We want to do so much, but we don't have enough experts in
00:17:31s
that field to do it, or we don't have enough time, we can't do it fast enough. And what this
00:17:35s
technology promises is a new industrial revolution where you know if the steam engine took loads off
00:17:44s
of humans backs and animals backs and put it onto machinery instead so that the humans could go and
00:17:49s
do other things that were less burdensome this is doing that same thing for cognitive or intellectual
00:17:57s
labor. We can hand off burdensome tasks onto these models and free us up to do more. And so the
00:18:06s
reality that I want to see in the next half decade or decade is a world where as a worker, you do have
00:18:14s
an assistant that is extremely capable that you can ask to do a nearly arbitrary set of things for
00:18:20s
you and that you're sort of managing a little team of assistants that you can go ask to do tasks for
00:18:26s
you. You can guide them, direct them, tell them where to spend their time and effort instead of
00:18:31s
you having to do those low-level tasks yourself. And so it sounds boring, but the impacts are
00:18:37s
profound. It's boring, but profound. Well, listen, Aidan, appreciate you being here on Politico Tech.
00:18:44s
Yeah, my pleasure. Thank you for having me.
00:18:49s
That's all for today's Politico Tech. For more tech news, subscribe to our newsletters,
00:18:54s
Digital Future Daily and Morning Tech.
00:18:58s
Our managing producer is Annie Reis.
00:19:00s
Our producer is Afra Abdullah.
00:19:02s
I'm Stephen Overly.
00:19:03s
See you back here tomorrow.
END
2027