- 6 settimane fa
Categoria
📺
TVTrascrizione
00:05Grazie a tutti per avercivito per avercivito a Developer Day.
00:07Grazie a tutti per avercivito in San Francisco per once.
00:10Io amo San Francisco.
00:11No, è un evento di più, abbiamo avuto persone lineate per entrare in,
00:14e c'è un po' di parlare.
00:16Voglio parlare con l'AI safety debate.
00:18C'è un po' di parlare, ma la notizia che abbiamo confermato è che OpenAI
00:22ha esentemente abbandonato a lanciare un nuovo modello
00:25perché di risultati di risultati.
00:27C'è un po' di risultati di risultati di risultati di risultati di risultati.
00:57ma lo abbiamo detto, che questo si è benissimo,
01:00monitor e sicuramente hanno bisogno di stare in fronte.
01:05Abbiamo avuto dei modelli in cui stanno dicendo che non ci sono quasi al bar
01:08che vogliono fare un altro e fare un nuovo modello e lanciare.
01:11Ma c'è più attenzione di questo.
01:13Ok.
01:14What about this particular model worried you or freaked you out to the point where you
01:18wouldn't launch it?
01:19I wouldn't over rotate on this one thing.
01:21It was like, you know, this model was a little bit worse on a few of the evals we look
01:25at.
01:25This was, think of this in the abundance of caution category.
01:29There's not a big scary thing in this case.
01:30Does this count as a pause?
01:32And I wonder what the president would say in a moment like this.
01:35If he is concerned about slowing down too much, what have your conversations been like?
01:39And do you worry that he looks at a headline like this and says, this counts as a pause
01:43here?
01:45Again, I would put this in like the normal course category of, you know, often we build
01:52a model, we test it, it doesn't meet our standards, we change it, we launch it later.
01:56But as we have talked about, we are pacing our progress, which includes sometimes not
02:02training the model.
02:07Everybody wants great models.
02:08Well, actually not everybody.
02:09Many people want great models.
02:10We want great models.
02:12But we have to do these in a way where we can make very confident safety cases and claims
02:18so that people don't have a bunch of anxiety and can rely on us for the models we're putting
02:22out into the world.
02:24We are now in a period where the models are getting so good so fast that we think it is
02:29prudent and wise to make more progress on, since we were talking about safety, alignment,
02:34monitoring, so that we can continue with progress.
02:36But I think this idea of pacing is really important.
02:38It's not just like stop it because, you know, we don't want smarter models, of course we
02:41want better models.
02:42It is make sure that we improve the areas we need to confidently and safely continue with
02:47progress.
02:48And I think the president...
02:49So pacing doesn't always mean slowing.
02:50And I think that's something that has confused people.
02:52Yeah.
02:52It means doing it in a responsible way with guarantees that the world can rely on.
02:57Hugging face.
02:58We've got to talk about that for a second.
02:59And that was just this watershed moment.
03:01And it feels like we've had examples every other day since then.
03:05Did you underestimate at the time the depth of these issues?
03:09I mean, there's been reports that there are thousands of these.
03:12Has it surprised you?
03:13We're trying to...
03:14I think part of the way you want to do this is to really over communicate and be extremely
03:17transparent.
03:18So there's nothing else that I'm aware of or that we're aware of that is as serious as
03:22the Hugging Face incident.
03:23But we are trying to put out every small or otherwise case we can find.
03:28And we really, as part of our mission, want the world to understand what these models are
03:35capable of.
03:36One of the things that's special about OpenAI is we have a lot of breadth.
03:40We do consumer.
03:40We do enterprise.
03:41We do developers.
03:42And people have very different ways of thinking about where we are with AI.
03:47And we want to make sure that all of these people that are touching our products and
03:50living in the world get a sense of where the technology is and what the kind of risks we
03:56have to mitigate are.
03:57Do you feel like there's just more attention on OpenAI?
03:59It feels like others have not had the same level of incidents?
04:03That's kind of what I was trying to get there.
04:06We just touch so many people and we do so much different stuff.
04:09Okay.
04:09So Jensen Wang, NVIDIA's CEO, put out this basically piece of software that he said would
04:14have prevented Hugging Face and he said it's an engineering problem.
04:17OpenAI didn't sign on to that project.
04:21Why not?
04:21Anthropik and others did.
04:22Do you agree with Jensen that it could have been solved?
04:25From what I know about it, and I'm not super into the details, I think it's a good thing.
04:29I think we are doing similar things.
04:31I don't think it's a full solution.
04:33And I worry that if we treat AI safety as only an engineering problem, we will miss the very
04:40important point that we have a science problem in front of us.
04:45We still have discovery about how to align these models and we have to solve that scientific
04:49problem too.
04:50Okay.
04:50At the same time, as we mentioned, we are here at Developer Day.
04:52Yeah.
04:52You have, I mean, it looks like hundreds of developers, companies, I would think investors
04:56here too.
04:57How are you towing the line between trying to get these people to use your technology,
05:02show them how great it is, but at the same time tell them that it's safe?
05:07I mean, I think it's a similar thrust, like people want to be able to use this technology
05:12and know that it's going to be safe.
05:14So I think part of it being great technology is to have it be the safest, most aligned,
05:17most dependable AI in the industry.
05:20One question on liability.
05:21So if something does go wrong, so say one of these startups here builds a great app using
05:25OpenAI and something goes terribly wrong.
05:27Whose fault is that?
05:30The liability framework is being debated a lot in the world right now.
05:34I suspect, like other industries, there will be a sort of multi-level thing.
05:38Like if you think about the auto industry, for example, if there's faulty parts, you
05:42know, you can be liable in some ways there.
05:44If the car manufacturer puts them together in a way that's outside of the spec of what they're
05:48supposed to do, then the car manufacturer could be, and if someone's driving junk, it's
05:52their fault.
05:52So, you know, depending on whether the fault is in the model and how it's used or someone
05:57misusing it intentionally, I assume we'll have like a, you know, a new liability framework.
06:01What is your pitch today to developers?
06:04Cost seems like it's been a big theme for you guys, but what are you going to tell people
06:07on stage in terms of why OpenAI, why ChatGPT, in your mind, is the best option for these guys?
06:13We have more than 20 announcements today.
06:15So we'll be telling them a lot of things.
06:18Give us a flavor, by the way.
06:19new sort of like new ability to have AI systems working for you and helping you all the time.
06:27New models.
06:29We'll talk about the cost issue.
06:30Yeah.
06:31We really, you know, a lot of new tools for developers.
06:33It's really a lot of, I think, exciting stuff.
06:36It is cool to see all of this come together and the degree to which I think we just have
06:40the best offering for developers across the board.
06:42Cost is such a central topic here and we've also talked before about the issue of distillation.
06:46I mean, some of the open source models that are cheaper, you and Anthropic have accused of basically copying your
06:52technology.
06:53How are you feeling about distillation and those models?
06:56I think we have the best models to offer at every price point.
06:59Yeah.
07:00And I don't think, like, you know, people do talk about open source, but I don't think there's any open
07:04source model out there that you can run that beats us on a cost price comparison.
07:08And speaking of cost generally, come watch.
07:10We've got some cool stuff.
07:11Okay.
07:11Well, agents as well.
07:12So Mark Zuckerberg and Muse have made a lot of progress, it seems like, in agents, in hardware.
07:17What do you make of Muse?
07:18Have you tried it at all?
07:20I've tried it.
07:21It seems like a nice product.
07:22Are you threatened by Zuckerberg at all?
07:23No.
07:25Not at all?
07:26Well, I mean, he's like a very impressive person and that company has, like, great resources, but I feel very
07:30good about it.
07:31He's also showed what's possible with hardware.
07:33Where's Johnny Ive?
07:36I don't want to spoil your surprise, but I think we will have something that's worth waiting for.
07:40Today we may have some hardware news.
07:41Is that what you said?
07:42Okay.
07:42Well, what is the plan in hardware and does it put more pressure to have an even better product if
07:47Mark Zuckerberg seems to have kind of gotten out of the gates first, even though maybe some of their stuff
07:52is still in the works?
07:53I am not worried about being first with new hardware.
07:56New hardware is very hard to do.
07:58Like, new computing form factors come along very rarely.
08:00And the quality bar that we want to hit for something that, like, you will love and really be a
08:05key part of the way you interface with AI, that's going to take a while.
08:09And we're going to do that at a high level of quality.
08:11I also want to ask about the perception.
08:13We were just talking about the SNL weekend update.
08:15There is obviously a movie about you.
08:17I don't know if you've seen the trailer.
08:19There's some issues in terms of public perception and just this thought that maybe you, maybe Dario Amade as leaders,
08:27don't have the best reputation right now.
08:29Um, I mean, I think everyone doing something this important to society deserves great scrutiny.
08:37I did see the movie trailer.
08:39I mean, it looks like an awful movie, but I appreciate people's creativity and I really do support that.
08:42So I hope, I hope they have a fun time with it.
08:45Um, the, but this is like a big thing and I get why people have anxiety and also if they
08:52need to like project that anxiety on someone or have like a little bit of relief and poke fun at
08:56me, it doesn't bother me.
08:57I know we got to go in a second, but Greg Brockman, one of your co-founders is in Washington,
09:01D.C. today.
09:02You were there at the state dinner last night.
09:04What is your expectation of what's possible in D.C.
09:07And what are the conversations looked like behind the scenes, especially on international coordination?
09:14Um, I don't want to be too optimistic because I have been wrong before when I thought something was going
09:19to happen in D.C.
09:20and, you know, it didn't. Uh, but I do feel like people are taking it seriously this time and I'm
09:26optimistic.
09:26I'm optimistic in spite of previous experience. I'm optimistic.
09:29Okay. IPO timeline. I know you don't like to talk about the IPO, but you said it's ill-advised to
09:33go this year.
09:34Is anything going to change next year? I mean, the safety debate, will that be in the rearview mirror when
09:38you go out potentially in 27?
09:39I don't have a particular timeline in mind. Um, once we have run a few safety cases, once we feel
09:47like we understand how to contend with this next level of AI and do it safely, I think we could.
09:53But I, you know, our mission first is about benefiting people with AGI and helping and making sure that we
10:01can ensure safety and barreling forwards with an IPO while I think things feel so in flux.
10:08And I really want to be able to just focus on the mission and not with all of the maybe
10:12challenges of being a newly public company.
10:14I just, I don't want to do that. I really think this is the time to put safety and mission
10:17first.
10:18Fair enough, Sam. Well, we'll let you get to your keynote. We will be tuning in as well. We really
10:21appreciate your time.
10:22I hope you enjoy it. Thanks for coming. Will do. All right, Scott, back over to you. Thanks so much.