WEBVTT - OpenAI Head of Policy Chris Lehane Talks AI Concerns

0:00:01.120 --> 0:00:02.960
<v Speaker 1>We're going to now head over to the AI Action

0:00:03.040 --> 0:00:06.160
<v Speaker 1>Summit in Paris, where Bloomberg's Tom mackenzie is sitting down

0:00:06.160 --> 0:00:09.560
<v Speaker 1>with open AI's chief Global affairs officer, Chris Lahane.

0:00:12.960 --> 0:00:14.960
<v Speaker 2>Guys, thank you very much doing yes at Le Grand

0:00:15.000 --> 0:00:17.400
<v Speaker 2>Ballet in Paris. I'll start with that question about deep

0:00:17.440 --> 0:00:20.040
<v Speaker 2>sea because it is a key question for people attending

0:00:20.200 --> 0:00:23.040
<v Speaker 2>this event. Chris Lhane, head of Global policy of course

0:00:23.120 --> 0:00:26.279
<v Speaker 2>at open AI deep Sea, how do you characterise the

0:00:26.400 --> 0:00:29.120
<v Speaker 2>impact of deep sea. We're just hearing from Demisisarvice, who

0:00:29.120 --> 0:00:32.000
<v Speaker 2>I sat down with earlier. He said their claims around

0:00:32.040 --> 0:00:34.880
<v Speaker 2>the way they built and trained this model, the cost

0:00:35.360 --> 0:00:36.920
<v Speaker 2>and the amount of chips they were using. He says,

0:00:36.920 --> 0:00:37.800
<v Speaker 2>that's exaggerated.

0:00:38.440 --> 0:00:40.159
<v Speaker 1>So I think there can be two things that they

0:00:40.200 --> 0:00:44.160
<v Speaker 1>are true here. One that they have built a really

0:00:44.600 --> 0:00:48.720
<v Speaker 1>impressive model. It basically competes with what open AI had

0:00:48.760 --> 0:00:50.879
<v Speaker 1>put out back in September. Now we've since put out

0:00:50.920 --> 0:00:55.680
<v Speaker 1>other more advanced models, but clearly a very capable model.

0:00:56.120 --> 0:00:58.880
<v Speaker 1>I think the second thing with Demis said can also

0:00:58.920 --> 0:01:02.240
<v Speaker 1>be true, which is perhaps and we've seen news reports

0:01:02.280 --> 0:01:05.919
<v Speaker 1>on this, that the costs, that how the technology was derived,

0:01:06.760 --> 0:01:09.600
<v Speaker 1>that you know, whether they access certain types of chips,

0:01:10.880 --> 0:01:14.000
<v Speaker 1>you know, whether what was initially said was not actually

0:01:14.360 --> 0:01:17.399
<v Speaker 1>the case. I do think the big takeaway though, even

0:01:17.400 --> 0:01:19.480
<v Speaker 1>if everything I just said turns out to be accurate

0:01:19.520 --> 0:01:23.800
<v Speaker 1>and crewe, it is still a very impressive and competitive model.

0:01:23.840 --> 0:01:26.080
<v Speaker 1>And so to me, the big, big, big takeaway in

0:01:26.120 --> 0:01:29.560
<v Speaker 1>all of this is that this really reaffirmed something that

0:01:30.080 --> 0:01:32.479
<v Speaker 1>open ai has been saying since the summer of twenty

0:01:32.600 --> 0:01:36.080
<v Speaker 1>twenty four that there are two countries in the world

0:01:36.360 --> 0:01:40.399
<v Speaker 1>that can build AI at scale, the US, the CCP

0:01:40.560 --> 0:01:44.160
<v Speaker 1>led China. And what that really means is that there

0:01:44.200 --> 0:01:47.520
<v Speaker 1>is a global competition right now between whether the world's

0:01:47.560 --> 0:01:51.400
<v Speaker 1>going to be built on small, d democratic AI rails

0:01:52.080 --> 0:01:56.520
<v Speaker 1>or authoritarian, autocratic AI rails. And that's the big takeaway

0:01:56.520 --> 0:01:56.800
<v Speaker 1>for this.

0:01:56.920 --> 0:01:58.080
<v Speaker 2>I want to get to that, but I want to

0:01:58.080 --> 0:01:59.680
<v Speaker 2>push you on what you'll see with discovery because I

0:01:59.720 --> 0:02:01.600
<v Speaker 2>know this is investigation that has been a problem within

0:02:01.680 --> 0:02:04.920
<v Speaker 2>open ai as to whether or not deepseek inappropriately use

0:02:04.960 --> 0:02:07.920
<v Speaker 2>some of the inferencing data distilling from your own models.

0:02:08.000 --> 0:02:09.600
<v Speaker 2>Have you come to a conclusional We're.

0:02:09.440 --> 0:02:11.840
<v Speaker 1>Still looking at it. Obviously, there's we and we've already

0:02:11.880 --> 0:02:14.799
<v Speaker 1>made public that we had seen some evidence of that

0:02:14.919 --> 0:02:18.320
<v Speaker 1>taking place, And just so folks understand, because distillation is

0:02:18.360 --> 0:02:20.560
<v Speaker 1>not a normal word, at least certainly that I don't

0:02:20.600 --> 0:02:23.680
<v Speaker 1>use at the breakfast table all the time. There is

0:02:23.800 --> 0:02:27.720
<v Speaker 1>different kinds of distillation. And I'll use an analogy. If

0:02:27.760 --> 0:02:29.519
<v Speaker 1>you go to the library and you take out a

0:02:29.560 --> 0:02:31.720
<v Speaker 1>book and you learn from that book, and that ultimately

0:02:31.720 --> 0:02:34.239
<v Speaker 1>informs some of your work. That's fine. That takes place

0:02:34.240 --> 0:02:36.360
<v Speaker 1>a lot of that. That's part of what happens in

0:02:36.400 --> 0:02:38.920
<v Speaker 1>the AI space. There's another version of it where you

0:02:38.919 --> 0:02:41.720
<v Speaker 1>go in to the library, take the book, keep the book,

0:02:41.960 --> 0:02:43.760
<v Speaker 1>put your name on the book, slap a cover on

0:02:43.800 --> 0:02:45.720
<v Speaker 1>the book, and hand it out as if it's your book.

0:02:46.040 --> 0:02:48.519
<v Speaker 1>And that's the replication. And I think that's what we're

0:02:48.520 --> 0:02:51.560
<v Speaker 1>concerned about, and again what we've seen some evidence of

0:02:51.600 --> 0:02:54.280
<v Speaker 1>and are continuing to review to have a better understand

0:02:54.320 --> 0:02:56.800
<v Speaker 1>We've talked with government officials about it, and we'll share

0:02:56.840 --> 0:02:57.720
<v Speaker 1>more as we learn more.

0:02:58.040 --> 0:03:00.960
<v Speaker 2>Well open AIS critics have said, then trend your models

0:03:01.000 --> 0:03:03.320
<v Speaker 2>on data, and you haven't been fully transparent in terms

0:03:03.320 --> 0:03:04.880
<v Speaker 2>of the use of that data. I don't know you

0:03:04.919 --> 0:03:05.560
<v Speaker 2>would push back.

0:03:05.560 --> 0:03:07.680
<v Speaker 1>That's when there's good calories and bad calories there's a

0:03:07.680 --> 0:03:10.240
<v Speaker 1>good distallation and problematic distillation.

0:03:10.400 --> 0:03:12.040
<v Speaker 2>Have any of your customers pushed back and said, look,

0:03:12.040 --> 0:03:14.640
<v Speaker 2>we're uncomfortable with the pricing of your models at this

0:03:14.800 --> 0:03:16.919
<v Speaker 2>at this junct shot.

0:03:17.160 --> 0:03:20.200
<v Speaker 1>Well, let me put it this way. You know, open

0:03:20.280 --> 0:03:23.839
<v Speaker 1>ai came out in November of two years ago, twenty

0:03:23.919 --> 0:03:28.200
<v Speaker 1>twenty two, right, and within two months was it a

0:03:28.320 --> 0:03:32.239
<v Speaker 1>one hundred million users? Right? Were well over three hundred

0:03:32.360 --> 0:03:37.760
<v Speaker 1>million today continuing to see you know, that really strong growth.

0:03:38.640 --> 0:03:41.720
<v Speaker 1>And what we try to do is have different models

0:03:41.880 --> 0:03:45.920
<v Speaker 1>priced at different levels depending on how you ultimately want

0:03:46.120 --> 0:03:48.840
<v Speaker 1>to use it. And that's something just given the nature

0:03:48.840 --> 0:03:51.280
<v Speaker 1>of how fast the technology is moving and the pace

0:03:51.680 --> 0:03:55.000
<v Speaker 1>this moving at that you're always constantly trying to evaluate.

0:03:54.600 --> 0:03:55.520
<v Speaker 2>Prices go down from it.

0:03:56.000 --> 0:03:58.720
<v Speaker 1>Oh yeah, So I think amongst the most interesting things

0:03:58.760 --> 0:04:01.400
<v Speaker 1>that we have seen this is a an interesting aspect,

0:04:01.440 --> 0:04:06.240
<v Speaker 1>which is the efficiencies of systems, particularly using the reasoning

0:04:06.320 --> 0:04:09.560
<v Speaker 1>technology is coming down. Sam, When our CEO put a

0:04:09.640 --> 0:04:12.560
<v Speaker 1>blog out last night, maybe technically this morning, I don't know,

0:04:12.600 --> 0:04:14.240
<v Speaker 1>but at some point in the last twenty four hours

0:04:14.720 --> 0:04:18.000
<v Speaker 1>that sort of looked at three observations of AI. The

0:04:18.120 --> 0:04:21.240
<v Speaker 1>first was that the more you spend on compute to

0:04:21.279 --> 0:04:24.720
<v Speaker 1>build frontier models. It's pretty logrhithmic. The more powerful of

0:04:24.720 --> 0:04:26.960
<v Speaker 1>those gets so you're going to require more and more compute,

0:04:27.160 --> 0:04:31.039
<v Speaker 1>more and more investment in infrastructure. Secondly, over the last

0:04:31.160 --> 0:04:34.000
<v Speaker 1>year or so, we've seen a big increase in efficiencies

0:04:34.320 --> 0:04:36.880
<v Speaker 1>which are bringing down the cost of a token. Of

0:04:37.000 --> 0:04:38.600
<v Speaker 1>that I think of a token as a price of

0:04:38.640 --> 0:04:40.880
<v Speaker 1>computer as a unit of computing about one hundred and

0:04:40.960 --> 0:04:44.760
<v Speaker 1>fifty percent. But even as the costs come down, the

0:04:44.800 --> 0:04:47.440
<v Speaker 1>amount of people using it go up, so that then

0:04:47.480 --> 0:04:50.040
<v Speaker 1>puts pressures back on computing. Like an analogy here is

0:04:50.400 --> 0:04:52.840
<v Speaker 1>car prices come down, more people drive cars, but you

0:04:52.960 --> 0:04:55.560
<v Speaker 1>then need more energy, more roads. Third, piecet that has

0:04:55.600 --> 0:04:57.440
<v Speaker 1>come out and his observation, sorry, just bear with me

0:04:57.520 --> 0:05:01.640
<v Speaker 1>for a second, is that is that the economic productivity

0:05:01.680 --> 0:05:03.280
<v Speaker 1>that you're getting in is super exponential.

0:05:03.680 --> 0:05:05.400
<v Speaker 2>Where are you getting the money to spend one hundred

0:05:05.440 --> 0:05:08.400
<v Speaker 2>billion initially on star gates, elon must says, you probably

0:05:08.400 --> 0:05:10.760
<v Speaker 2>don't have the money, even something Adela says, we're good

0:05:10.760 --> 0:05:12.880
<v Speaker 2>for our eight two billion. That's Onion.

0:05:13.040 --> 0:05:15.800
<v Speaker 1>So first of all, we have incredible partners. We have

0:05:16.000 --> 0:05:19.400
<v Speaker 1>soft Bank, which is a proven track record of raising

0:05:19.520 --> 0:05:22.200
<v Speaker 1>enormous syndicated money from sovereigns and pensions.

0:05:23.720 --> 0:05:24.839
<v Speaker 2>Thought is coming from nails.

0:05:24.920 --> 0:05:27.479
<v Speaker 1>And then and then we have Oracle, right, which actually

0:05:27.600 --> 0:05:30.599
<v Speaker 1>builds these. Then you have open aies. What are piece

0:05:30.680 --> 0:05:33.080
<v Speaker 1>of this? First of all, in terms of the media question,

0:05:33.400 --> 0:05:34.920
<v Speaker 1>there's one hundred billion that's going to be going out

0:05:34.960 --> 0:05:37.280
<v Speaker 1>the door in the immediate future. We already have a

0:05:37.360 --> 0:05:40.520
<v Speaker 1>facility in Abilene. You guys need to come down. You

0:05:40.600 --> 0:05:43.159
<v Speaker 1>can abline Texas. Sorry, we'd love to give you, guys.

0:05:43.200 --> 0:05:45.720
<v Speaker 1>You can hang out the Oracle. Guys have done an

0:05:45.720 --> 0:05:48.880
<v Speaker 1>awesome job down there. No, no, no, no, why don't

0:05:48.880 --> 0:05:53.320
<v Speaker 1>you come see it? Show? We like to show, not tell.

0:05:53.800 --> 0:05:56.280
<v Speaker 1>And then and then the open AI piece comes in

0:05:56.320 --> 0:05:58.160
<v Speaker 1>two different pieces here, and I think this is to

0:05:58.320 --> 0:06:01.080
<v Speaker 1>understand the economic model here. First of all, what we

0:06:01.440 --> 0:06:06.039
<v Speaker 1>bring to this is the IP and you can think

0:06:06.160 --> 0:06:08.960
<v Speaker 1>of compute the same way. Maybe you can think of

0:06:09.240 --> 0:06:13.560
<v Speaker 1>gas regular gas, medium gas, and then premium gas. The

0:06:13.640 --> 0:06:15.760
<v Speaker 1>premium gas is what people are gonna pay for. The

0:06:15.880 --> 0:06:19.240
<v Speaker 1>premium compute is going to be the most expensive compute

0:06:19.279 --> 0:06:22.360
<v Speaker 1>in the world because it's gonna be the highest level compute.

0:06:22.600 --> 0:06:26.360
<v Speaker 1>You only get that premium compute with our IP. IP

0:06:26.600 --> 0:06:29.680
<v Speaker 1>going too, the chip design IP going into the data centers,

0:06:29.960 --> 0:06:32.560
<v Speaker 1>IP going how the clusters are sort of structured. And

0:06:32.680 --> 0:06:36.600
<v Speaker 1>then we also are a buyer of the compute, right,

0:06:36.680 --> 0:06:38.960
<v Speaker 1>so we commit to buying a certain amount of the

0:06:39.000 --> 0:06:41.880
<v Speaker 1>compute that's coming out, which helps the whole economic model work.

0:06:42.040 --> 0:06:44.640
<v Speaker 2>Chris, you worked in the Clinton White House. Is the

0:06:44.680 --> 0:06:49.560
<v Speaker 2>Trump administration there go go to great headline is to

0:06:49.600 --> 0:06:52.239
<v Speaker 2>the Trump administration right to slash AI regulation.

0:06:52.920 --> 0:06:55.320
<v Speaker 1>I think what the Trump administration is focused on is

0:06:55.360 --> 0:06:57.920
<v Speaker 1>one thing, and one thing very clearly, who is going

0:06:58.000 --> 0:07:01.640
<v Speaker 1>to prevail in the competition between democratic That's what they

0:07:01.680 --> 0:07:03.600
<v Speaker 1>get up and think about every day, at least from

0:07:03.600 --> 0:07:06.279
<v Speaker 1>what I have seen and from what I have heard,

0:07:06.680 --> 0:07:08.560
<v Speaker 1>And so I think they understand that you do have

0:07:08.680 --> 0:07:11.680
<v Speaker 1>to really be leading in leaning into the innovation. If

0:07:11.680 --> 0:07:15.320
<v Speaker 1>you think about about where the sort of comparative advantages are.

0:07:15.720 --> 0:07:17.520
<v Speaker 1>Right at the end of the day, this is actually

0:07:17.600 --> 0:07:21.520
<v Speaker 1>pretty simple. Whoever has access to compute is going to

0:07:21.560 --> 0:07:26.760
<v Speaker 1>be in a strong position. What makes up compute, it's data,

0:07:27.200 --> 0:07:30.760
<v Speaker 1>it's energy, it's chips, and it's talent. Right, and if

0:07:30.800 --> 0:07:32.760
<v Speaker 1>you think of what the PRC has, they have an

0:07:32.840 --> 0:07:37.040
<v Speaker 1>enormous amount of data. Authoritarian State Energy ten nuclear facilities

0:07:37.120 --> 0:07:39.560
<v Speaker 1>last year, another ten coming on this year. Our chips

0:07:39.600 --> 0:07:43.080
<v Speaker 1>are better. They're throwing a ton of money at it talent,

0:07:43.440 --> 0:07:46.000
<v Speaker 1>and I think the talent piece is really interesting because

0:07:46.320 --> 0:07:51.440
<v Speaker 1>they do have great talent in China. But in capitalist systems,

0:07:52.080 --> 0:07:58.440
<v Speaker 1>it is capitalism that unleashes the developer, the builder, the entrepreneur,

0:07:58.840 --> 0:08:00.960
<v Speaker 1>the people who are actually making this stuff in the

0:08:01.040 --> 0:08:03.400
<v Speaker 1>tools that starve hauled problems. And that's where our advantage is.

0:08:03.640 --> 0:08:06.040
<v Speaker 2>Is that why you've joined up and partnered with Aurae'm

0:08:06.040 --> 0:08:08.200
<v Speaker 2>to thinking of the AI principles a Google. They've adjusted

0:08:08.240 --> 0:08:11.280
<v Speaker 2>them and Google is no longer ruling out building tech

0:08:11.640 --> 0:08:14.520
<v Speaker 2>for defense for the military. Yeah, you're signing up, you

0:08:14.600 --> 0:08:17.000
<v Speaker 2>partner with Angurill. How far does that relationship go? How

0:08:17.080 --> 0:08:19.640
<v Speaker 2>much are you prepared to embed tech into weapons system?

0:08:19.680 --> 0:08:21.840
<v Speaker 1>Yeah? And we also just announced a partnership with the

0:08:21.960 --> 0:08:25.400
<v Speaker 1>National Labs, which are the you know, which play an

0:08:25.400 --> 0:08:29.080
<v Speaker 1>incredibly important role in how the US government thinks about

0:08:29.320 --> 0:08:31.720
<v Speaker 1>national security. Is on the FILS, Yeah, yes, and the

0:08:31.840 --> 0:08:34.520
<v Speaker 1>labs you know obviously like Los Alamos and others, are

0:08:35.040 --> 0:08:39.319
<v Speaker 1>very big players in the broader national security ecosystem. So

0:08:39.720 --> 0:08:41.719
<v Speaker 1>you know, for US, right, we do want to be

0:08:41.760 --> 0:08:45.439
<v Speaker 1>a partner on innovation with the government. We do believe

0:08:45.480 --> 0:08:49.200
<v Speaker 1>it's incredibly important that democratic AI prevails, and that means

0:08:49.280 --> 0:08:51.280
<v Speaker 1>making sure that the government is getting access to the

0:08:51.360 --> 0:08:54.120
<v Speaker 1>highest capabilities. You know, we'll certainly do it consistent with

0:08:54.240 --> 0:08:56.640
<v Speaker 1>our values and our principles, but at the end of

0:08:56.679 --> 0:08:59.120
<v Speaker 1>the day, like we do believe it is very consistent

0:08:59.440 --> 0:09:02.040
<v Speaker 1>with their mission that is to make sure AI benefits everyone.

0:09:02.520 --> 0:09:04.280
<v Speaker 1>That you're ensuring that AI is going to be built

0:09:04.320 --> 0:09:06.439
<v Speaker 1>in a democratic way with democratic values.

0:09:06.760 --> 0:09:09.040
<v Speaker 2>Chris la Haye, thank you very much. Indeed, Global head

0:09:09.080 --> 0:09:10.360
<v Speaker 2>of Policy at Open AI.