- 20 minutes ago
Category
🤖
TechTranscript
00:28They love you already.
00:30It's for you.
00:31It's for you.
00:34We met, not in person, but virtually a couple years ago.
00:40I wrote the story where I talked to all eight co-authors of that wonderful paper,
00:47Attention is All You Need.
00:50And when I did the story, none of the eight were at Google anymore.
00:58One returned, but then just this morning we learned that Noam Shazir has left Google
01:05and joined OpenAI.
01:09But it's beautiful how the co-authors all went and did their own thing.
01:15I don't think any have founded a company that has been done as well as Cohere, the company
01:23you started in Toronto.
01:25So we're going to, and you build models for enterprise that actually work.
01:31That's the idea, isn't it?
01:32So we'll talk about that, but I want to ask you first about where you were yesterday.
01:40You want to tell me about what it was like to be an AI CEO of the G7 conference?
01:46Yeah, it was super special.
01:48Obviously, you don't get the leaders of some of the largest democracies in the world all
01:53in the room together very often, but you certainly don't get them with the AI CEOs pretty much ever.
02:00So that was an exciting opportunity.
02:03And I think the conversation was surprisingly, for me, collaborative and productive.
02:10There was a very clear sense that they have consensus they need to collaborate much more.
02:18And that's both on seizing the opportunity, deploying the technology into their economies
02:22as quickly as possible.
02:24We want the benefits to be felt by democracies as fast as possible.
02:29We need to lead.
02:30We need to see the upside.
02:32But of course, there's downsides that need to be mitigated, things like child safety
02:37and, of course, a cyber threat that is very real and prominent right now.
02:43And so there was a consensus around figuring out a way to align standards between the democracies.
02:51So when you say the conversation, how is that conducted?
02:54You know, I sometimes wonder whether they're just bringing this on.
02:58You know, we got some CEOs here and, you know, we're going to make speeches at them
03:05and maybe they'll make some speeches back at us.
03:08But do you actually have room for productive conversations in a session like that?
03:12Well, it was a lunch.
03:14And so we weren't going to solve all the problems in the room.
03:17And there were certainly a lot of speeches.
03:19I can say that.
03:21But no, I think the leaders, both in private sector and public sector,
03:29came into the room with a lot of alignment.
03:32The differences were sort of set aside and we agreed we need to lead,
03:36democracies need to lead on this technology.
03:39And so it was about trying to solve that problem.
03:45The Macron gave a talk and he, what I understand,
03:50he was talking about there should be an alternative to U.S. dominance in the field.
03:57Did you feel that the, you know, leaders, besides the leader of the U.S.,
04:04were united in that feeling?
04:09I think the U.S. is leading in the AI race.
04:13And that's very true.
04:14What I said at the meeting was that we need to ensure that a democracy occupies the number two position.
04:21And that's not true today.
04:22And so I think all of the G7 understands that we need to create more champions.
04:28We need a diverse supply chain of AI providers.
04:33Cohere, I think, is the most promising effort towards that.
04:37As a good example of that, there is the Canadian-German digital alliance
04:42and Cohere's merger with Aleph Alpha.
04:44That's the most prominent and, I think, promising example of two G7 nations,
04:51two democracies coming together to pool resources and really build and scale a champion.
04:58Instead of doing what Canada has done for so long,
05:01and I'm sure Europe has done as well,
05:03which is spreading resources across hundreds of bets that won't work on this file.
05:09AI requires concentrated resources, and so we need to direct them behind champions.
05:15You know, folks may not be familiar with that deal you made with Aleph Alpha.
05:19Can you talk about that?
05:21And, you know, Aleph Alpha is a German AI company.
05:25If you go to their website, like virtually every page that says,
05:30oh, and we are operating on German infrastructure,
05:33and we're, you know, local,
05:35and we're not, we're alternative to what else you use.
05:39So I think it's fascinating that you folks made this alliance with them.
05:43But tell us a little bit more about what's involved in that.
05:46Yeah, I mean, their whole focus was very similar to ours
05:49in the sense it was on critical industries, high security settings,
05:52the things that power our economy.
05:54Think like the grid, water treatment, telecommunications, healthcare, finance.
06:00If any of these switch off, if any of these systems go down,
06:03the economy stops.
06:05And that is the biggest risk to national security,
06:08especially in the new world of cyber threats
06:11and the fact humans are writing less and less of the code
06:16that powers these critical systems.
06:18So they were extremely aligned to us,
06:22and we saw an opportunity to bring two major economies.
06:26Obviously, Germany is the third largest economy in the world.
06:29Canada is in a very unique position
06:30in the ability to bring online a lot of energy
06:34to support the build-out of computing infrastructure.
06:37And we have a very capable model development function at Cohere,
06:42bringing together the strengths of these two companies
06:45and two countries behind a champion for democracies.
06:51So is your strategy then to look for alliances like this in other countries?
06:56Absolutely.
06:57Yeah, it wasn't supposed to just be Canada and Germany.
07:00Canada and Germany are the first two to come together,
07:02but the door is wide open to bring others in.
07:05A few weeks ago, it was with the king of Spain
07:08to sign an MOU with Indra,
07:10which is the largest tech company in Spain.
07:12So we want to bring, and of course I should say that my mom is British,
07:16I live in England, my dad is Spanish,
07:18so I have a very deep commitment.
07:21I love the way you just dropped,
07:22oh, I launched the king of Spain, you know.
07:24Yeah, that was the coolest thing ever.
07:26You know, as a son of a Spanish guy,
07:29I think it was, he was so nice.
07:32He's such an incredible person.
07:33It's good to be king.
07:34It's good to be king.
07:36But yeah, he was so lovely.
07:38Like, he actually studied in Toronto when he was a child.
07:42So he came back to visit his, I guess, primary school
07:46that he had spent a year or two at.
07:48And he was taking photos with everyone,
07:50and he was just so accessible, welcoming.
07:54Yeah, it was an incredible experience.
07:56So, how about France?
07:58France, you know, it's been...
08:00There is a company here, which I understand does a pretty good AI business.
08:04I'm familiar, yeah.
08:05And I'm a big fan.
08:06I'm a big fan of Mistral.
08:08I was with Arthur yesterday.
08:10Tons of respect.
08:11And I think we need more democratic champions.
08:15I mean, would you want to find a partner
08:18other than Mistral to work with, or maybe them?
08:22So I think in France, we've had a presence here for a long time.
08:27A big chunk of our team is on the modeling side,
08:31on the model development side, based here in France.
08:34And we want to continue to invest in that and scale that up.
08:37We're going to be announcing an expansion to our office here very, very soon.
08:42So, and the other thing to say is that Canada has a very deep history with France.
08:48Our mother nation, our parent nations are the UK and France.
08:53And so there's a very close affinity.
08:56We want to contribute to this ecosystem.
08:58We want to contribute to European sovereignty.
08:59That absolutely includes France.
09:01And the talent here is essentially second to none.
09:05Like the level of quality of ML researchers that the schools here produce,
09:11it's world class.
09:13And so the ability to create and scale champions,
09:16I think France is in a phenomenal position.
09:19So, Cohere focuses on the enterprise.
09:25You build powerful models, large language models.
09:30What's the secret you have of competing with the giants,
09:36who probably want to sell the same customers that you do?
09:40Well, the strategy is not try to spend the most money.
09:46That's not going to be a winning strategy.
09:48Instead, where we focused in these critical industries,
09:51oftentimes it's highly compute constrained.
09:54And so even if you had a 10 trillion parameter model, et cetera,
09:58they can't adopt it.
09:59They can't deploy it.
10:00And so the market can't even purchase that scale of model.
10:05Over time, as compute gets more efficient, cost effective, et cetera,
10:09those models will be unlocked.
10:11And at that point, Cohere will have that scale of model
10:13to deploy into this critical infrastructure.
10:17But the reality is today we can operate 10, 20, 30 times more efficiently than them
10:22because our customers couldn't consume something 10, 20, 30 times more expensive.
10:29You know, a lot of companies, you know, we hear are, you know,
10:34put a big investment in AI and then they're finding it's not really giving them a commensurate return.
10:42You know, for all the disruption and expense they have, you know,
10:46they're not being more productive.
10:48Are you avoiding that?
10:49How?
10:51I would say that's not what I'm seeing.
10:54I think the demand is continuing to go through the roof.
10:59You can see that in our own revenue numbers and the revenue numbers of our peers.
11:04And that is because it's incredibly useful.
11:07Employees are demanding access to it.
11:09And if you don't give them secure access to it,
11:11they'll take insecure access to the technology.
11:14So the demand is there and growing.
11:19And the ROI, I think there was a question a year ago or 18 months ago,
11:24we were in this POC phase.
11:26Lots of little tests and nothing really at scale.
11:30And many of those POCs failed because they were poorly run
11:33or because the technology wasn't ready, the use cases were poorly chosen.
11:37At this stage, the maturity of market
11:39and the maturity of organizations like Cohere to deliver for their needs
11:43has increased so much that I think the success rate of those POCs
11:47has gone up massively.
11:49And we're starting to see those earlier POCs roll out into production.
11:54It's why we see such an immense compute constraint on the world stage.
11:59You know, looking at the history of Cohere,
12:03at a certain point, you kind of did a reset.
12:05You know, you were following one path
12:08and then, you know, you rebuilt from scratch.
12:11Is that right?
12:14I don't know if I would use those terms,
12:16but certainly, like, Cohere has...
12:17How do you know the word reset came in there?
12:19Cohere has, like, reinvented itself many times and found our...
12:24Okay, reinvent.
12:24But what was the lesson that you had to take
12:28in order to, you know, reinvent yourself
12:31into what currently seems to be something that's working?
12:35I think very early on, we made a strategic decision
12:39not to get locked into any ecosystem.
12:42We didn't do one of these massive deals with a hyperscaler
12:45because it would have meant we'd have to be exclusive to them.
12:48And we didn't want to get locked in.
12:51And that decision led us towards being interoperable
12:55with a lot of different compute providers, compute platforms.
12:58And that played directly into critical infrastructure,
13:02which refuses to get locked in to a single provider,
13:05which refuses to make sacrifices on sovereignty.
13:09So that early strategic decision led to a product
13:14that was uniquely well-positioned
13:16for this sovereign critical infrastructure play.
13:20But I think the market had to mature and be ready for that.
13:24Like, in many ways, I actually think
13:25Cohere has stood in the same place
13:28for more than half a decade.
13:29And the market has come to us.
13:31You know, first consumer took off,
13:33then it was, like, general enterprise use cases
13:37that were not so secure.
13:39And now it's this critical government workloads,
13:43telecoms, that type of thing,
13:44which is doing an extraordinary ascent.
13:48And so the market has come to us in many ways.
13:53And there's a lot of concern about labor and AI.
13:58Have you been able to see the impact on workforce
14:03of the people who use Cohere
14:05and integrate it into their operations?
14:08Yeah, to the best of my knowledge,
14:10we don't see any sort of, like, dismissals happening
14:13as a product of AI.
14:14It's being deployed to drive growth.
14:17Certainly productivity is a part of that,
14:20but that's to drive growth at the organization.
14:23So it's not to get rid of people, displace people.
14:25It's to empower them to do more,
14:27make more business for the company.
14:29I think that that's not to say there isn't a risk.
14:34And we're still very early in the adoption of this technology.
14:38We're just scratching the surface.
14:41So we do need to start thinking about
14:43what potential scenarios could look like.
14:46If there is displacement,
14:47I think governments need to have that conversation actively,
14:49whether it's skills, retraining programs.
14:51There's more extreme ideas like UBI.
14:55There's ideas like right to work or job guarantees.
14:58I do think these need to be well-studied
15:01and understood in advance of a potential crisis.
15:04You know, when you folks did Transformers
15:10and in the years immediately after that,
15:14when OpenAI applied and did ChatGPT,
15:17there was a lot of excitement and optimism about AI.
15:22And what we're seeing now is a lot of pushback.
15:27A lot of people actually despise AI.
15:34Maybe the labor thing is one aspect of it,
15:37but there's a number.
15:38How do you regard that?
15:40You know, how much do you pay attention to that?
15:43You know, some leaders seem to feel,
15:46oh, they'll use it anyway, things like that.
15:49But, you know, I'm just curious about your reaction
15:52to what is, you know, a major pushback against AI.
15:57I think it's understandable.
16:02Like, it feels like a technology
16:05that is completely transforming
16:08the way that our economy and our society
16:12is going to operate.
16:14And if you...
16:15And it's something that no one can predict
16:17that far out into the future.
16:18They can't predict what's coming next year.
16:20They can barely predict what's coming in the next six months.
16:22And so when you have such high uncertainty
16:25about the future
16:26with so many different possible outcomes,
16:29many of which don't look so great,
16:32that causes a lot of fear and resistance.
16:35You layer on top of that
16:36a broader economic boom
16:40or concentration into the field
16:42and the prioritization of data centers
16:46over consumer grid consumption,
16:49that type of thing,
16:50raising prices for people.
16:53I think it's a very understandable pushback.
16:56And I think it's on us
16:57as the ones developing this technology
16:59and deploying it
17:01to ensure people are brought into that process,
17:05that they're protected,
17:05that we aren't building data centers
17:08in highly congested, under-pressure regions
17:11where the grid just can't sustain it.
17:13We need to be strategic
17:14about where we place data centers.
17:15There are places with excess.
17:17There's places with completely clean energy sources.
17:20That's where we should be investing in the infrastructure.
17:23And we need to bring people on board
17:25with the creation and deployment of these tools
17:28as tools that empower
17:30rather than tools that displace.
17:34I think today, overwhelmingly,
17:36these tools are used to empower.
17:38I think there has been.
17:39You can see it in the numbers.
17:41Incredible job growth.
17:43Unemployment rates aren't skyrocketing.
17:46And it might be that Jevons paradox,
17:48this idea that when a massively efficiency-enhancing technology comes out,
17:54you would think that it displaces people
17:56and that a bunch of jobs will go away.
17:57But instead, it actually creates more jobs.
18:00The cost of a service goes down.
18:02The demand shoots up.
18:03And so you actually need more people
18:05to support the management of that demand.
18:07It might be the case that Jevons paradox is correct.
18:11And as we start to see how this rollout goes,
18:14I hope the consumer confidence,
18:17the public's confidence will return.
18:19So the people who boo the graduation speakers,
18:23you know, they mention AI.
18:25AI, they just don't understand the value of it.
18:29You know, it's a communications problem.
18:32No, I think it's...
18:35I think they're uncertain about the future
18:37and they're afraid for what it means for them.
18:40And I think there has been examples of places
18:42where energy prices have gone up
18:44because a data center was built there
18:45and that was a really poor choice
18:47of where to place a data center
18:48and we need to protect against that.
18:51They're not wrong.
18:54And it's good pressure
18:55to steer things in the right direction.
18:58But I do think this technology
19:00is going to do extraordinary good
19:03across healthcare, across education,
19:06across making us more efficient, more effective.
19:08One of the things, you know, here in France,
19:10in Canada, in Germany, in England,
19:15many advanced economies over the past 15 years,
19:19growth has stagnated.
19:21Like, GDP per capita has actually flatlined
19:24or started to decline.
19:26That's extremely problematic
19:28because what happens is when the pie isn't growing,
19:31if the pie is growing,
19:32your slice of it is just growing.
19:34And so you don't need to compete with others.
19:36It's not a zero-sum game,
19:37but if the pie is staying the same
19:40or contracting,
19:41you enter into a psychology of zero-sum.
19:44It's a competition.
19:45If someone tries to take something,
19:48they have to take it from me.
19:49And so I think that mentality,
19:53the promise of AI is restoring growth,
19:57restoring a growing pie for everyone
19:59so that these psychological dark sides
20:04of humanity feel less pressure and prominence.
20:09So you're one of the co-authors
20:12of Transformers paper.
20:13A lot of people, you know,
20:15I had Jan LeCun on the stage yesterday.
20:17He's not bullish on large language models.
20:21Do you feel, you know,
20:23and quickly,
20:24it's our time running out,
20:26are you bullish?
20:27Do you feel this is something
20:29that could sustain, you know,
20:30the way they're going with, you know,
20:34you mentioned data centers,
20:35bigger data centers,
20:37more compute to get the more advanced models?
20:40Yeah.
20:41Well, I mean, on Jan,
20:42not supporting large language models,
20:45what does he support?
20:47I would love to see the alternative.
20:48Physical AI.
20:49He said he was up here yesterday.
20:51Okay.
20:52What's powering that?
20:53What is the model behind the physical AI?
20:56So I think he's got to answer that question first.
20:59And we're all waiting and very excited.
21:03But, you know,
21:04like any reasonable scientist,
21:07I don't want the Transformer to be dominant forever.
21:10It's been, I guess,
21:11next year it'll be like the 10-year anniversary
21:13of the Transformer.
21:14It's absurd that we're still using that technology.
21:17You invented it 10 years ago.
21:18Exactly.
21:19Exactly.
21:20So I hope something new comes.
21:22Great.
21:23We can go on, but we can't.
21:25Thank you so much.
21:26Thank you, Stephen.
Comments