WEBVTT - Greylock Partners Partner Reid Hoffman Talks Our AI Future

0:00:02.520 --> 0:00:07.040
<v Speaker 1>Bloomberg Audio Studios, podcasts, radio news.

0:00:08.039 --> 0:00:10.639
<v Speaker 2>Let's stick with tech. The impact of generative AI under

0:00:10.680 --> 0:00:14.400
<v Speaker 2>scrutiny as some investors question massive capex spending in the space.

0:00:14.680 --> 0:00:17.840
<v Speaker 2>Read Halfman, the co founder of Malice AI and author

0:00:17.840 --> 0:00:21.040
<v Speaker 2>of Superagency What could possibly go Right with our AI Future?

0:00:21.320 --> 0:00:23.400
<v Speaker 2>Joined us now read It's good to see you. I

0:00:23.440 --> 0:00:25.160
<v Speaker 2>said as you walked into the room. There's a book

0:00:25.200 --> 0:00:26.880
<v Speaker 2>I need to read What Could Go Right? Because I'm

0:00:26.960 --> 0:00:29.680
<v Speaker 2>very worried about what things could go wrong. Let's put

0:00:29.680 --> 0:00:32.160
<v Speaker 2>it that way. You in the book approach something really

0:00:32.200 --> 0:00:36.040
<v Speaker 2>important as we make these technological revolutions, these advancements over

0:00:36.240 --> 0:00:40.440
<v Speaker 2>human history. Actually it benefits human agency, and I'm worried

0:00:40.440 --> 0:00:42.839
<v Speaker 2>that maybe with AI it guys in a different direction.

0:00:43.120 --> 0:00:44.839
<v Speaker 2>Help me out and tell me why it could go right.

0:00:45.120 --> 0:00:47.479
<v Speaker 3>Well, the shorting answer is it gives you superpowers, right,

0:00:47.520 --> 0:00:49.560
<v Speaker 3>and it gives you superpowers across a large number of things.

0:00:49.600 --> 0:00:54.120
<v Speaker 3>Everyone should go try Chat, GPT, Inflections, Pie and Trapicsquad.

0:00:54.720 --> 0:00:56.800
<v Speaker 3>Give it a try. You can find what should I

0:00:56.840 --> 0:00:59.120
<v Speaker 3>make for dinner? How do I get medical advice that's

0:00:59.160 --> 0:01:01.720
<v Speaker 3>available to me? How I am I to have a

0:01:01.760 --> 0:01:04.520
<v Speaker 3>difficult conversation with a colleague, you know, how might I

0:01:05.120 --> 0:01:07.840
<v Speaker 3>actually do things in parenting that might be different? You know,

0:01:07.920 --> 0:01:11.640
<v Speaker 3>all of those things are are our superpowers that we get.

0:01:12.120 --> 0:01:14.760
<v Speaker 2>And each run. Lisa under the bus, I am, do

0:01:14.760 --> 0:01:16.760
<v Speaker 2>you want to share with the audience what you do?

0:01:16.319 --> 0:01:18.840
<v Speaker 4>You can manipulate what your kids look at if you

0:01:18.880 --> 0:01:22.080
<v Speaker 4>send them links to things that then change their algorithms

0:01:22.080 --> 0:01:25.840
<v Speaker 4>that then give them more news feeds in their in

0:01:25.880 --> 0:01:27.280
<v Speaker 4>their feed. I mean there are ways that you can

0:01:27.360 --> 0:01:28.920
<v Speaker 4>kind of play the system a little bit.

0:01:28.840 --> 0:01:29.840
<v Speaker 1>And I think that that's proper.

0:01:29.920 --> 0:01:32.000
<v Speaker 3>Yeah, exactly. And by the way it helps them learn,

0:01:32.160 --> 0:01:34.640
<v Speaker 3>it gets some news about the environment around them. I mean,

0:01:34.840 --> 0:01:36.039
<v Speaker 3>this is one of the things that is if you're

0:01:36.080 --> 0:01:40.000
<v Speaker 3>not actually in fact using these AI agents to learn

0:01:40.080 --> 0:01:42.000
<v Speaker 3>right now, you should. And like, for example, one of

0:01:42.040 --> 0:01:44.360
<v Speaker 3>the ways that I do when I approach a difficult subject,

0:01:44.400 --> 0:01:47.520
<v Speaker 3>say for example, quantum computing, I stick a technical paper

0:01:47.520 --> 0:01:49.440
<v Speaker 3>in and I say, explain this to me like I'm

0:01:49.440 --> 0:01:52.360
<v Speaker 3>twelve superpowers superpowers.

0:01:52.480 --> 0:01:54.480
<v Speaker 2>Yes, okay, So my next question is going to be

0:01:54.680 --> 0:01:57.800
<v Speaker 2>how should we participate and how we shake this intelligence?

0:01:57.800 --> 0:01:59.520
<v Speaker 2>And Lisa has given us one idea, could you give

0:01:59.560 --> 0:02:00.000
<v Speaker 2>us another time?

0:02:00.640 --> 0:02:03.960
<v Speaker 3>Well, I think part of the key thing is my

0:02:04.080 --> 0:02:07.360
<v Speaker 3>very first chapter is about chat GBT, where humanity enters

0:02:07.360 --> 0:02:09.639
<v Speaker 3>the chat. You got hundreds of millions of people who

0:02:09.639 --> 0:02:13.200
<v Speaker 3>are interacting with it, and the company watches what works

0:02:13.200 --> 0:02:15.280
<v Speaker 3>and what doesn't work and then improves it. So by

0:02:15.480 --> 0:02:18.720
<v Speaker 3>simply going and experimenting and getting which kind of things

0:02:18.760 --> 0:02:22.600
<v Speaker 3>are the things that help you elevate your agency, then

0:02:23.040 --> 0:02:27.119
<v Speaker 3>actually in fact helps the company know, oh, these things

0:02:27.160 --> 0:02:29.160
<v Speaker 3>really work, Oh these things need to be improved.

0:02:29.520 --> 0:02:32.480
<v Speaker 1>At the same time, you talk about technical papers, and

0:02:32.560 --> 0:02:35.280
<v Speaker 1>one thing that AI does really well is work with

0:02:35.320 --> 0:02:38.919
<v Speaker 1>technical details, not so much, the social not so much.

0:02:38.919 --> 0:02:41.560
<v Speaker 1>When you have to have certain biases that are baked

0:02:41.639 --> 0:02:45.000
<v Speaker 1>into some of these algorithms, what are the limitations to

0:02:45.080 --> 0:02:47.520
<v Speaker 1>this in sort of broad raised application at a time

0:02:47.560 --> 0:02:51.040
<v Speaker 1>where people are saying this could do qualitative jobs and

0:02:51.080 --> 0:02:54.240
<v Speaker 1>there could potentially be a lot of biases baked in.

0:02:54.800 --> 0:02:59.920
<v Speaker 3>So every major group is working intensely on biases. And actually,

0:03:00.000 --> 0:03:02.600
<v Speaker 3>one of the things that we did with inflection and

0:03:02.800 --> 0:03:06.320
<v Speaker 3>PIE personal intelligence PUN intended was to teach it to

0:03:06.360 --> 0:03:10.600
<v Speaker 3>be kind and empathetic, to have EQ as important as IQ.

0:03:11.200 --> 0:03:13.440
<v Speaker 3>And actually you see that now in anthropics Claude and

0:03:13.440 --> 0:03:15.760
<v Speaker 3>other things is it's kind of spreading, and so you

0:03:15.800 --> 0:03:19.120
<v Speaker 3>can actually have these kinds of interactions be trained in them,

0:03:19.200 --> 0:03:21.760
<v Speaker 3>and this is kind of reinforcement learning by human feedback

0:03:22.040 --> 0:03:24.600
<v Speaker 3>to actually be much better than your average human.

0:03:24.760 --> 0:03:27.040
<v Speaker 1>The reason why I find this so fascinating is because

0:03:27.600 --> 0:03:29.960
<v Speaker 1>I think about my children and what careers they potentially

0:03:30.000 --> 0:03:32.480
<v Speaker 1>could have, and I'm like, no, not programmer, because that's

0:03:32.480 --> 0:03:34.200
<v Speaker 1>going to be dead. You know, all of these other ones.

0:03:34.200 --> 0:03:37.320
<v Speaker 1>Probably if you can be empathetic human, those human skills

0:03:37.360 --> 0:03:39.360
<v Speaker 1>are probably going to be more needed. That's what people say.

0:03:39.480 --> 0:03:40.720
<v Speaker 1>Are you saying those two are going to go out

0:03:40.720 --> 0:03:41.000
<v Speaker 1>the waydo?

0:03:41.000 --> 0:03:41.200
<v Speaker 2>I mean?

0:03:41.200 --> 0:03:42.280
<v Speaker 1>What is this due to employment?

0:03:42.520 --> 0:03:46.720
<v Speaker 3>So first, I actually think programming skill jobs aren't going

0:03:46.720 --> 0:03:48.840
<v Speaker 3>out the door. I think there's human application, but I

0:03:48.840 --> 0:03:52.040
<v Speaker 3>think it's also human application of like EQ as well.

0:03:52.080 --> 0:03:55.440
<v Speaker 3>So the fact that you have a for example, say

0:03:55.440 --> 0:03:57.960
<v Speaker 3>a medical assistant that can help you with something, or

0:03:58.000 --> 0:03:59.720
<v Speaker 3>a coach that might be able to talk with something,

0:04:00.000 --> 0:04:02.320
<v Speaker 3>doesn't mean you don't want a human coach. It's more

0:04:02.360 --> 0:04:05.920
<v Speaker 3>coaches the better, right, and being able to make it work.

0:04:06.160 --> 0:04:08.480
<v Speaker 5>You're talking about AI for the good, but what about

0:04:08.480 --> 0:04:10.080
<v Speaker 5>AI in the hands of adversaries?

0:04:11.400 --> 0:04:16.160
<v Speaker 3>Well, so superpowers superpowers for terrorists, superpowers for criminals, superpowers

0:04:16.200 --> 0:04:18.000
<v Speaker 3>for rogues states. That's one of the things that we

0:04:18.080 --> 0:04:20.760
<v Speaker 3>need to be careful about how we navigate. Once again,

0:04:21.120 --> 0:04:23.720
<v Speaker 3>all of the major groups actually have security groups that

0:04:23.800 --> 0:04:25.520
<v Speaker 3>are working on this to try to make sure and

0:04:25.839 --> 0:04:28.400
<v Speaker 3>that's I think part of the having call so called

0:04:28.480 --> 0:04:30.960
<v Speaker 3>red teaming plans and kind of trying to make sure

0:04:30.960 --> 0:04:33.080
<v Speaker 3>that that's as minimized as much as possible.

0:04:33.160 --> 0:04:35.040
<v Speaker 5>How do think the US government should be approaching this?

0:04:35.320 --> 0:04:38.160
<v Speaker 5>The market's freaked out because of deep seek and how

0:04:38.279 --> 0:04:41.640
<v Speaker 5>much potentially the CCP is putting money into their version

0:04:41.760 --> 0:04:44.360
<v Speaker 5>of AI. But when you're in China and you want

0:04:44.360 --> 0:04:46.560
<v Speaker 5>to look at things on deep seek that are particularly

0:04:46.600 --> 0:04:49.240
<v Speaker 5>critical to the CCP, it will not show you. So

0:04:49.279 --> 0:04:51.280
<v Speaker 5>how does the US government go about this?

0:04:51.880 --> 0:04:54.440
<v Speaker 3>Well, So I think it's important that the US government

0:04:55.440 --> 0:04:57.880
<v Speaker 3>kind of continues some of the threads that the previous

0:04:57.920 --> 0:05:02.039
<v Speaker 3>executive order did, which is you have red teaming things

0:05:02.080 --> 0:05:04.919
<v Speaker 3>that you actually do world leadership about what kind of

0:05:04.920 --> 0:05:07.480
<v Speaker 3>things are happening. I mean, there's connections in the UK

0:05:08.480 --> 0:05:10.359
<v Speaker 3>AI Safety and Suit and the US one and the

0:05:10.360 --> 0:05:12.479
<v Speaker 3>French one, and I think these are important to do.

0:05:13.560 --> 0:05:15.120
<v Speaker 3>But I think that this is one of the things

0:05:15.120 --> 0:05:17.080
<v Speaker 3>where there will be multiple AIS and part of the

0:05:17.120 --> 0:05:20.080
<v Speaker 3>reason why I want AI not to just be artificial

0:05:20.080 --> 0:05:24.159
<v Speaker 3>intelligence or amplification intelligence, but also American intelligence as part

0:05:24.160 --> 0:05:24.760
<v Speaker 3>of what we're doing.

0:05:24.960 --> 0:05:28.000
<v Speaker 5>What do you make then, of the tech industry being

0:05:28.160 --> 0:05:30.800
<v Speaker 5>very close to this administration. I'm not just talking about

0:05:30.800 --> 0:05:32.320
<v Speaker 5>your former front Elon Musk. I was at the White

0:05:32.360 --> 0:05:34.640
<v Speaker 5>House yesterday and Bill Gates was walking into the West

0:05:34.640 --> 0:05:37.600
<v Speaker 5>wing for meetings. Do you think they can help the

0:05:37.640 --> 0:05:42.400
<v Speaker 5>administration basically get to there faster when it comes to

0:05:42.440 --> 0:05:43.080
<v Speaker 5>safe AI?

0:05:44.080 --> 0:05:47.039
<v Speaker 3>In the short absolutely, I think it's if you said,

0:05:47.200 --> 0:05:50.760
<v Speaker 3>of any kind of Western democracy government the world, most especially,

0:05:50.760 --> 0:05:54.040
<v Speaker 3>of course, you know, American presidency, American government, should they

0:05:54.040 --> 0:05:56.200
<v Speaker 3>be connected with the tech industry talking about how tech

0:05:56.320 --> 0:05:57.960
<v Speaker 3>creating the future? What are the things that we can

0:05:58.000 --> 0:06:00.920
<v Speaker 3>do to make better American industry, better American society, better

0:06:00.960 --> 0:06:03.839
<v Speaker 3>lives for American citizens? The short answer is absolutely.

0:06:04.160 --> 0:06:06.080
<v Speaker 2>What you make in the back lash against Dale Musk

0:06:06.160 --> 0:06:09.320
<v Speaker 2>more recently about he's in film from the government, well.

0:06:09.680 --> 0:06:13.159
<v Speaker 3>In government, the move fast and break things maybe should

0:06:13.160 --> 0:06:16.000
<v Speaker 3>be you know, a little bit more compassionate and judicious.

0:06:16.360 --> 0:06:17.600
<v Speaker 2>Is that what you think the apprice should be.

0:06:18.360 --> 0:06:21.480
<v Speaker 3>I think that the question should be is trying to say, hey, look,

0:06:21.800 --> 0:06:24.440
<v Speaker 3>there's a reason government doesn't work like companies, even though

0:06:24.480 --> 0:06:26.880
<v Speaker 3>I think there's efficiency that's really good and we need

0:06:26.920 --> 0:06:28.200
<v Speaker 3>to respect those reasons too.

0:06:28.600 --> 0:06:32.360
<v Speaker 5>Do you feel that you're going to basically get some

0:06:32.360 --> 0:06:35.840
<v Speaker 5>sort of repudiation because of where your politics were ahead

0:06:35.880 --> 0:06:39.120
<v Speaker 5>of this election and given how Trump has basically cosied

0:06:39.240 --> 0:06:40.960
<v Speaker 5>up to a lot of individuals in the tech world.

0:06:41.000 --> 0:06:42.800
<v Speaker 5>I have not seen you yet at mar Lago in

0:06:42.880 --> 0:06:44.880
<v Speaker 5>inauguration or walking into the West wag.

0:06:46.480 --> 0:06:49.560
<v Speaker 3>Look, I think the thing that we most want as

0:06:49.560 --> 0:06:52.839
<v Speaker 3>American citizens is for our government to succeed, to succeed

0:06:52.880 --> 0:06:56.440
<v Speaker 3>for our industry, succeed for our citizens, And so that's

0:06:56.480 --> 0:06:59.039
<v Speaker 3>what I most want, and I think we all do.

0:06:59.520 --> 0:07:02.279
<v Speaker 1>You have been an incredible investor. You are an early

0:07:02.320 --> 0:07:06.320
<v Speaker 1>investor in open Ai, You invested in LinkedIn, just so

0:07:06.440 --> 0:07:09.479
<v Speaker 1>many of the behemoths currently. Do you see a lot

0:07:09.480 --> 0:07:12.240
<v Speaker 1>of upstarts right now that look like they could become

0:07:12.800 --> 0:07:16.680
<v Speaker 1>the next big thing in some of this machine learning technology?

0:07:17.800 --> 0:07:21.400
<v Speaker 3>So the short answer is yes, although picking them is

0:07:21.400 --> 0:07:23.440
<v Speaker 3>one of the difficult things about being a venture capitalist,

0:07:23.480 --> 0:07:27.240
<v Speaker 3>and there's a lot of different options. Back a few

0:07:27.280 --> 0:07:29.800
<v Speaker 3>years ago, I was arguing that we were five large

0:07:29.800 --> 0:07:32.720
<v Speaker 3>tech companies heading to ten. But I wouldn't necessarily have

0:07:32.720 --> 0:07:35.840
<v Speaker 3>predicted that Nvidio would have so quickly become one of those,

0:07:36.040 --> 0:07:38.360
<v Speaker 3>you know, into the seven and then going to ten.

0:07:38.400 --> 0:07:40.520
<v Speaker 3>And so I think that's we are in that progress.

0:07:40.560 --> 0:07:42.480
<v Speaker 3>And I think there are a number of startups that

0:07:42.560 --> 0:07:45.480
<v Speaker 3>are creating amazing technology. And then the question is product

0:07:45.480 --> 0:07:46.440
<v Speaker 3>market fit at scale?

0:07:46.600 --> 0:07:49.200
<v Speaker 1>What do you think the next technological evolution is going

0:07:49.240 --> 0:07:52.160
<v Speaker 1>to be? They could potentially be a sort of chat

0:07:52.200 --> 0:07:55.240
<v Speaker 1>GPT moment, or a deep seek moment, or one of

0:07:55.280 --> 0:07:57.640
<v Speaker 1>these sort of breakthrough ahaz.

0:07:58.600 --> 0:08:00.160
<v Speaker 3>One of the things I think we're going to to

0:08:00.200 --> 0:08:04.560
<v Speaker 3>see this year, all of the major AI companies are

0:08:04.600 --> 0:08:07.760
<v Speaker 3>working on amplifying coding. And one of the things like,

0:08:07.800 --> 0:08:10.320
<v Speaker 3>for example, you know for your kids, other things will

0:08:10.360 --> 0:08:13.920
<v Speaker 3>be yes, you have to bring up the UH will

0:08:13.960 --> 0:08:17.400
<v Speaker 3>be actually all of us will have a software engineering

0:08:17.440 --> 0:08:19.640
<v Speaker 3>assistant and that will make all of our work a

0:08:19.680 --> 0:08:21.680
<v Speaker 3>lot better. UH. And that's one of the things I

0:08:21.680 --> 0:08:23.800
<v Speaker 3>think hasn't yet entered the consciousness. So I think that's

0:08:23.840 --> 0:08:26.200
<v Speaker 3>one of the things that's coming. UH. And then of course,

0:08:26.320 --> 0:08:28.520
<v Speaker 3>you know, with what I'm doing with manus I'm hopeful

0:08:28.560 --> 0:08:31.840
<v Speaker 3>that we will have some really great early results in

0:08:32.480 --> 0:08:33.520
<v Speaker 3>hearing cancer.

0:08:33.320 --> 0:08:36.000
<v Speaker 2>And how we can accelerate the scientific process. I say,

0:08:36.040 --> 0:08:38.640
<v Speaker 2>we're all up with you on that page. You're very successful.

0:08:38.679 --> 0:08:40.720
<v Speaker 2>Thank you, sir, read to appreciate your time. Thank you

0:08:40.760 --> 0:08:42.839
<v Speaker 2>so good to see you. Ride half and there manus

0:08:42.920 --> 0:08:44.319
<v Speaker 2>Ai and a whole lot more