YT YouTube2Text

YouTube2Text → Transcripts → OpenAI’s Head of ChatGPT: We’re entering a new era of AI (again) | Tibo Sottiaux

OpenAI’s Head of ChatGPT: We’re entering a new era of AI (again) | Tibo Sottiaux — Transcript

by Lenny's Podcast · 8,712 words · 1,288 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00Often when I look at what people are
0:01building out there, it's just like,
0:02"You're not quite getting it." Like, you
0:04know, if you were just like pushing
0:05yourself and just really imagining like
0:07all of this being
0:08>> [music]
0:08>> roughly 10 times better than it is
0:10today, you know, like in a year, but you
0:12would build in a different way.
0:13>> What do you think people aren't pricing
0:14in in terms of where things are going?
0:16>> The majority of actions on the internet
0:18will be taken by agents. If you want
0:20your [music] product to be successful
0:21for agents, you have to build, you know,
0:22for a certain level of scale.
0:23>> How many agents do you find yourself
0:24running in parallel just uh in the in
0:26your kind of your day-to-day work?
0:28>> As I'm pushing the frontier, I find
0:29myself like building larger and larger
0:31teams of agents. And then when we have
0:33the next breakthrough with models, like
0:35suddenly it's just like I'm like, "Oh,
0:36well, you know, a bigger agent can just
0:38do all of it." And so I kind of shrink
0:40the team again. And then it just [music]
0:41kind of goes through this expansion and
0:42shrinking expansion.
0:43>> It kind of makes me think about there's
0:44this whole loops thing and then this
0:45whole graphs thing.
0:46>> Having to set up and fiddle with your
0:48loops is something that maybe people got
0:51excited about, but I don't think this is
0:52the way that it's going to work. Over
0:54time, you just want to system that
0:55[music] learns. You don't want to
0:56necessarily think about how I'm just
0:58like, "I'm going to loop it exactly this
0:59way in order to get good results."
1:00>> there's still more radical change
1:02happening to how we work?
1:04>> to continue to change quite radically.
1:06Even today, it still feels like a bit
1:08clunky, and I think it will feel clunky
1:10until it isn't.
1:16>> I mean, Evo, thank you for doing this.
1:18>> Yeah.
1:19Big fan of you.
1:19>> [laughter]
1:20>> Of course.
1:20>> Uh
1:21>> Nice room.
1:22>> You launched uh a ton of stuff today.
1:23How how did your sleep? How are you
1:25sleeping? Are you doing all right?
1:26>> I'm doing great. I'm fortunate to have a
1:28very good team.
1:29>> Mhm.
1:30>> So, while I sleep, they were
1:32awake most of the night. We have this
1:34amazing room uh at the office called the
1:36library, which we completely repurposed
1:38as like a mega war room. And the vibes
1:40in there are just immaculate. Uh
1:42people were there until like very, very
1:44late.
1:44>> So, you've been a long-time engineer.
1:46Are you are you coding at all anymore?
1:48Are you shipping PRs? Are you mostly
1:50like in docs and in meetings now?
1:51>> What is coding?
1:53>> Exactly.
1:54Yeah. Are you shipping PRs? I don't
1:56know. Is that
1:56>> Uh Yeah, yeah, yeah. Merch Merging some
1:58code every once in a while. Technically,
2:00like a lot of code is getting written
2:01for me to do all sorts of kinds of like
2:03analysis, like you know, understand like
2:06trends, understand a business,
2:07understand the next feature, you know,
2:09how well our previous launches are
2:10doing. A lot of code is written by Codex
2:13for dot. You know, I'm obviously not
2:15writing it by hand anymore.
2:16Occasionally, at the in during the
2:18weekends, I'll have like a little
2:19therapeutic moment where I do a little
2:21lead code just by hand. Um but that's
2:24the only actual hands-coding that I do.
2:26>> How many agents do you find yourself
2:27running in parallel just uh in the in
2:29the kind of your day-to-day work?
2:31>> I used to run a lot more in parallel and
2:33then we were, you know, fortunate enough
2:35to get um a breakthrough with ultrafast.
2:38And so, now, you know, I feel like you
2:40know, I'm able to be in the flow again
2:42and so it's just having a faster agent
2:44helps me a lot. I'm excited that we're
2:46getting, you know, 6-month-old ultrafast
2:47out there soon. Um and then, yeah, I
2:50think it's it's it sort of varies. Um as
2:52I'm kind of like pushing different here,
2:54I find myself like, you know, building
2:56larger and larger teams of agents. And
2:58then, when we have the next breakthrough
3:01with models, like suddenly it's just
3:02like I am like, oh, well, you know, a
3:04bigger agent can just do all of it and I
3:06keep everything in memory and learn. And
3:07so, I kind of shrink the team again. And
3:09then, it just kind of goes through this
3:10like expansion and shrinking expansion.
3:12>> Hm, okay. Can you say more about that?
3:13It kind of makes me think about there's
3:14this whole loops thing and then this
3:16whole graphs thing. Is like an evolution
3:18of that in some way or
3:19>> I think like having to set up and fiddle
3:21with your loops and, you know, figuring
3:23that out is something that, you know,
3:25maybe people got excited about, but I
3:27don't think this is the way that it's
3:28going to work. I think the way that it's
3:30going to work is a what
3:32how we're positioning and and what we're
3:34shipping with dots, where you have an
3:37incredibly smart agent that works 24/7,
3:40understands your goals, understands your
3:42preferences, learns from feedback. Um
3:45and, you know, it's like what we
3:46launched is like, no, perfect. I will
3:48learn very much from, you know, making
3:49this available to our pro users. But
3:52over time, just want a system that
3:53learns from, you know, what you want to
3:55achieve. You don't want to necessarily
3:57think about hours like I'm going to loop
3:59it exactly this way in order to get
4:00results. So, right now there's Codex,
4:02there's ChatGPT work, there's ChatGPT
4:04consumer, there's dots. What I'm hearing
4:06is you think we're heading towards dots
4:08kind of being the primary way you talk
4:10to AI, and that kicks off all these
4:12other things.
4:13>> Yeah, I think fundamentally if you take
4:14a step back and, you know, whether it's
4:16dots or not, um it's all about breaking
4:20free from the technology
4:22um and
4:24having this sort of like permanent
4:26active intelligence that knows, you
4:29know, everything it needs to do
4:31and is available through any client, any
4:34screen. You can call it, you know, like
4:36maybe you walk into a meeting room, it
4:38shows up in the meeting, it takes notes,
4:41you know, you pick it back up on email,
4:43um and then, you know, you text it if
4:44you need it. It's not that, you know,
4:46you need to be glued to a laptop or, you
4:48know, you're like always on your phone.
4:50It's just like it's available when you
4:51need it and then it's also like gets out
4:53of the way when you don't need it. And I
4:55can't wait, you know, for that to sort
4:57of like come to fruition because I think
4:59like just carrying your laptop as like
5:00this brick around everywhere is just
5:02like, you know, you're kind of like tied
5:03to the technology instead of the
5:04technology working for you.
5:06>> This episode is brought to you by our
5:07season's presenting sponsor WorkOS. What
5:11do OpenAI Anthropic, Cursor, Replit,
5:14Sierra, Clay, and hundreds of other
5:16winning companies all have in common?
5:18They are all powered by WorkOS. If
5:20you're building a product for the
5:21enterprise, you felt the pain of
5:23integrating single sign-on, SCIM, RBA,
5:26audit logs, and other features required
5:28by [music] large companies. WorkOS turns
5:30those deal blockers into drop-in APIs
5:33with a modern developer platform built
5:35specifically for B2B SaaS. Literally
5:38every startup that I'm an investor in
5:40that starts to expand upmarket [music]
5:41ends up working with WorkOS. And that's
5:44because they are the best. Whether you
5:45are a seed stage startup trying to land
5:47your first enterprise customer or a
5:49unicorn [music] expanding globally,
5:51WorkOS is the fastest path to becoming
5:53enterprise ready and unblocking growth.
5:55It's essentially Stripe for enterprise
5:57features. Visit workos.com to get
6:00started or just hit up their Slack where
6:02they have actual [music] engineers
6:03waiting to answer your questions. WorkOS
6:05allows you to build faster with
6:07delightful APIs, comprehensive docs, and
6:09a smooth developer experience. Go to
6:11workos.com to make your app enterprise
6:14ready today.
6:17I was going to ask you about this. It
6:18feels like so much has changed in how we
6:20work over the past like 2 years. In 2
6:22years, especially engineers, but a lot
6:24of roles have just the role your
6:25day-to-day life has changed
6:27significantly.
6:28And I'm curious if you think we're kind
6:29of settling into what work will look
6:32like in the next, I don't know, 5 years?
6:34Let's say 2 years, that's feels way too
6:36far to think about. Or do you think
6:37there's still more radical change
6:39happening to how we work in
6:41>> No, I think it's going to continue to
6:42change
6:43quite radically. Even today,
6:46you know, it does seem like the
6:47technology is finally coming together
6:48with voice, uh you know, multimodal
6:51inputs, outputs, but it still feels like
6:54a bit clunky and I think it will feel
6:56clunky until it isn't. And then you're
6:58just like, "Wow, you know, it's like
7:00I can just talk to this thing the same
7:02way that we're talking right now. It
7:04just remembers things perfectly. If
7:06I want to, you know,
7:08ideate and brainstorm, I can just like,
7:10you know, squiggle a thing. I have this
7:12collaborative surface. You know, this is
7:13what we launched with ChatGPT spaces
7:15signal to various startup dot. You know,
7:17you can imagine we have like a shared
7:18whiteboard, we can collaborate between
7:20humans, between agents. And I think it's
7:22going to sort of like transcend
7:23conversations, transcend, you know, like
7:25a lot of the clients that you have
7:26today. And like, none of this is just
7:28really working today, right? So, I think
7:29that we're going to go through this big
7:31new transformation. What I'm hearing
7:32here is dots is a big part of that
7:33future, this AI assistant that just kind
7:35of does everything for you and you don't
7:37have to think about Codex versus ChatGPT
7:39and
7:39>> Yes, and for Codex and ChatGPT, we're
7:41going to merge uh like the the we have
7:44we have chat, we have the work toggle.
7:45Like you know, we heard the feedback
7:47loud and clear. Like you know, people
7:48love the capability in work, but then
7:50also you know, you're like sometimes you
7:51want to use chat because it's a little
7:52bit faster. And you know, more pleasant
7:54to use. And so we're we're merging that.
7:55We're reducing that complexity.
7:57Eventually, we will ship all of the
7:59capabilities of Docs, you know, straight
8:00into ChatGPT as well and you know, lift
8:02the floor
8:03for like our 1.2 billion users, which
8:05you know, at that time we'll see like
8:06how many we have. Um and we want it to
8:09be really really seamless. One thing
8:10that I'm very excited with with Docs is
8:12like they don't have a model picker.
8:14There's just no configuration. You just
8:15talk to it. Like the only thing that you
8:17have to configure is like which channels
8:18you want to
8:20talk to.
8:20>> It feels very much like the Her kind of
8:22vision of the world. I don't know. Do
8:24you think about that movie much? Just
8:25this idea of this
8:26>> I watched it once.
8:27>> [laughter]
8:28>> Yeah. I think it's it's it's
8:30interesting, you know, to to to think
8:32about, you know, maybe more what what I
8:34think about more is like, you know,
8:35sci-fi from from, you know, 30 40 years
8:38ago and like how visionary it was.
8:40>> Is there specific sci-fi you think about
8:41that's most influenced you?
8:43>> Like you know, books like Neuromancer
8:45and then you know, the original like
8:46Star Trek.
8:48Um you know, I think about that a lot.
8:49And there's something about Star Trek
8:51which I think, you know, was like there
8:52was this technology that was like very
8:54very useful. Obviously like, you know,
8:56you were able to just like talk to a
8:57computer and like it does things for
8:59you. Like you can talk to your spaceship
9:01and it just moves, right? It's like this
9:03is getting real, you know, like we're
9:04entering that era.
9:05>> So we're recording this in front of a
9:07live audience. There was this line out
9:09the door just people waiting to see you,
9:11to hear from you.
9:12How do you just think about this, I
9:13don't know, role you're in, this
9:15responsibility you have to build what
9:18you're building and the impact it has on
9:19people's lives. I don't know. Just what
9:20does that feel like? How do you think
9:21about that?
9:22>> I think it's very important to be part
9:23of a community and to
9:25build with people. It was like, you
9:27know, we're we're discovering this
9:28technology and like what we can do
9:30together.
9:31Um and so every time I you know, talk to
9:34people, I'm like I'm kind of like
9:35reminded of you're doing this thing
9:37which I hadn't even anticipated. You
9:39know, it's impacting your life in a
9:40certain way and it's impacted like you
9:42know, this other person's life and then
9:43it's just like all together. It's both
9:45very humbling, very inspiring, and I
9:47think very necessary in order to build
9:49something that is truly useful for
9:50humans. I I don't know how we would do
9:52it without the community.
9:53>> That's something I hear a lot with
9:55people building an AI is they don't know
9:56what it is until they've launched it and
9:58they see how people use it, see what
9:59emerges. And it feels like kind of what
10:01you're saying is just build it with
10:03people versus like here's the vision,
10:04we've got it figured out.
10:05>> Even the Twitter community has been very
10:07nice in general.
10:09>> You're very good at Twitter.
10:11Kind of antagonistic sometimes, but very
10:13good. How do you How do you find time to
10:14even tweet? I don't understand how
10:15someone in your role has time to sit on
10:16Twitter, tweet, reply.
10:19>> I spend like maybe half an hour per day.
10:21Yeah, just say rolling and then doing
10:23something when I feel inspired and I
10:24have something interesting to say. Often
10:26it just comes like in the moment. Yeah.
10:28>> Perfect Twitter skill.
10:29>> And I have a
10:30We haven't released the dot
10:32the the the team dots yet, but you know
10:34today we're releasing the primary dot
10:36which you can configure, you can get
10:38used to like what a dot can do, connect
10:39it to your apps.
10:41But we'll release
10:42the ability to create more than one. And
10:44so I have one that's responsible for
10:46Twitter.
10:46>> Okay, say more about that. So right now
10:48everyone gets one dot.
10:49>> Yeah.
10:50>> Right now they can't add more dots.
10:52>> That's right.
10:52>> They will be able to have many dots.
10:54Talk about kind of where that's going.
10:55>> Yeah, so we wanted to start with one,
10:57the primary dot which is the one that
10:59you know, you'll probably also text that
11:01you know, learns your preferences most
11:03deeply as well and kind of see like you
11:04know, what people do with it, see how we
11:07need to tune the system, you know, learn
11:09that working with the community and then
11:11very quickly soon you'll be able to add
11:13you know, a second one, a third one, a
11:15fourth one, as many as you need to
11:16create like your virtual team. And then
11:18you can give them specific roles, right?
11:20Is it It's not I don't find it
11:22absolutely necessary, but sometimes I
11:24have like a thing that is requires quite
11:26a bit of work, you know, such as like
11:27monitoring like Twitter for me. And then
11:29you know, that's It's a thing is like
11:30like, this will be like so much work
11:32that is like the work of an entire dot.
11:34Um and then you know, you can have a
11:35couple of those.
11:36>> So if you think about all the launches
11:37that happened today, it feels like dots
11:39maybe like the sexiest.
11:41I'm curious if there's something that
11:42you think is not getting the attention
11:44it probably deserves that will become a
11:46really, really important down the road
11:47that's kind of people are sleeping on
11:49that's like a vision.
11:50>> Yeah, yeah, yeah, important.
11:51>> the sleeper hit is ecosystem. So opening
11:54it all up. Uh and the commitment, the
11:56deep commitment we have towards opening
11:58it all up. So we have signing with
12:01ChatGPT 16 partners. Very, very proud of
12:03that. Like we started that like really
12:05it was sort of like in the moment last
12:07year. Um I was talking to creators of Pi
12:10and open code and it was like, of
12:11course, you know, you should just be
12:12able to use the critics
12:15sign on and then you know, you use the
12:16usage and then you just we kind of like
12:18shook hands, you know, like virtually
12:20we're like, you know, it's just like use
12:21this this uh this off um and then you
12:24know, kind of trust you to not do
12:25anything shady. You know, that grew
12:26obviously that became quite popular and
12:27then you know, now we made it, you know,
12:29an actual thing that was supported with
12:30many, many partners. Um and I'm excited
12:32to grow that very quickly. You know, the
12:34other side as well is like opening up
12:36all of the infrastructure and how we
12:38build for ChatGPT, you know, with
12:39plug-in extensions, also plug-in
12:41discovery, just allowing, you know,
12:43everyone to ship something through like
12:45maybe 1.2 billion users, right? You
12:47know, and benefiting from that
12:48distribution. And we also have shared
12:50economics um that we didn't actually
12:52talk uh in the keynote, but you know, we
12:54will pay, you know, our plug-ins that
12:57are
12:58uh popular and you know, are seeing a
13:00lot of usage. Uh they're going to get
13:01part of like the the revenue share as
13:03well. Uh and so I think this commitment
13:05towards an open ecosystem is going to be
13:07very exciting.
13:07>> That's amazing. Uh okay, so as an
13:09example, so say Notion, somebody's using
13:11the Notion plug-in or the Figma plug-in
13:13within
13:14uh ChatGPT work, Notion, Figma make
13:16money from people using it within the
13:18app.
13:19>> Yes, so um we have, you know, all these
13:22subscribers that use ChatGPT, um they
13:25get like, you know, a certain amount of
13:26usage and then when have that, the when
13:28when the users use that usage with the
13:29plugin, or within the other product when
13:32they use sign-in with ChatGPT, um you
13:34know, we will have like some shared
13:35economics with them where they will get
13:37to remunerate.
13:38>> One of the, I don't know, things I think
13:40that gets people excited about plugins
13:41and ecosystems is distribution
13:43platforms, ways to kind of get out
13:44there. People discover my app, it blows
13:47up.
13:47>> any advice for people that want their
13:49plugins to, you know, bubble up, to get
13:50discovered?
13:51>> Build a good plugin.
13:52>> Yeah.
13:53>> The way that it works and we'll
13:54obviously tune uh the system, but we
13:57look at retention uh numbers, uh we look
14:00at
14:01um you know, how successful the plugin,
14:02the quality, and then you know, that's
14:03what we then start to recommend to users
14:06uh in conversations. And so, your plugin
14:07can just like you know, get recommended
14:09to like you know, a significant slice of
14:10users. Obviously, if your plugin is not
14:12quite good, they will stop being
14:14recommended.
14:15>> Right. That's amazing that you're
14:16looking at retention of people using
14:18that plugin. That's right. Um is it is
14:19it actually adding utility? Is it, you
14:21know, enabling ChatGPT to do something
14:23else?
14:23>> Got it. So, it's it's not so much an AEO
14:26kind of get the right words and get
14:27people writing about you. It's people
14:28using it consistently, sticking with it.
14:30That's right. I think. Okay, so you
14:31think that's kind of the sleeper hit
14:32that's going to become a big deal. And I
14:34know this is kind of the second attempt.
14:35You guys had an app marketplace that
14:37This one is the right one. Okay, this is
14:39This one's going to work. Maybe going
14:40back to dots cuz it feels like that's a
14:41big part of the vision in the future.
14:43When did you guys start working on this?
14:44Obviously, there's a open claw, a big
14:46open claw moment. You guys Peter joined
14:48and there There's a foundation,
14:50uh and then Grok Bot, Muse, and
14:51Instinct, and all these things. I guess
14:53how long have you guys been working on
14:54it? Why did it take so long for you guys
14:56to get something out?
14:57>> Do you remember a Codex Cloud?
15:00>> Yeah, the first uh yeah.
15:01>> That was like a year ago, right? And uh
15:04if you look back at Codex Cloud and you
15:05look at the little animation that we
15:06used, uh I think that was the
15:08inspiration for Grok Bot.
15:10>> Um
15:11>> Yeah. It's just like very, very similar.
15:13>> Yeah, really similar, I would say.
15:14>> But, you know, seriously, like on the
15:16research, long, persistent, long horizon
15:19task, this is something that we've
15:20worked on, you know, for more than 2
15:21years. Memory systems, that's like
15:23coherent, you know, like roughly the
15:25same amount of time. We launched a lot
15:26of that straight into ChatGPT. I think,
15:28you know, people just really enjoy that
15:31ChatGPT gets to know, like, you know, so
15:32much relevant information about you. So
15:34many people I talked to have like an
15:35anecdote of like, "Hey, you know, I fell
15:37ill and, you know, ChatGPT remembered
15:39actually, like, you know, I had this
15:40barbecue
15:41and, you know, I had like lime and I was
15:44like making tequila and like, you know,
15:45it was just like probably like this burn
15:47that I have on my hand is actually like
15:48this lime, you know, it's like in some
15:50burn." And he's just like, "How did it
15:51know that?" And it's like, you know,
15:52it's just the memory works very, very
15:53well in ChatGPT. And so, we've put a lot
15:55of research in that. And, you know, all
15:57of this coming together on top of like
15:59the Quora harness and also the ability
16:01to just like run 24/7 with, you know,
16:03good productivity. That's what we
16:05launched with dots. And so, it's the
16:06work of like many, many months. And
16:08then, one of the important things was
16:09also getting safety and security right.
16:11You know, this is also why we launched
16:12with Astra, it's our safest, most
16:14aligned model. And like a lot more
16:16effort went into like making dots, like,
16:18you know, just really safe and secure.
16:19>> A question I like to ask people who come
16:21on the podcast is, where do you think
16:23human brains will continue to be
16:25valuable over time? Where do you find
16:27yourself being necessary in the work
16:29that happens versus where AI is taking
16:32on more and more? Where do you think
16:33human brains will continue to be useful,
16:35most useful over time?
16:37>> Yes, I I think this is very much a
16:38function of how we build the technology
16:40in the way that we design everything at
16:43OpenAI is like to put humans at the
16:44center of it and build it as extensions
16:47of humans of of of your will and your
16:49taste. And so, just really be like this
16:52super empowering thing.
16:53Um, I think as long as we continue to do
16:55that and it kind of like enables you to
16:58just take whatever you wanted to do and,
17:01you know, your creativity and your taste
17:03and sort of like, you know, expand that,
17:05you know, in in a way that feels like
17:06awesome in the moment. It's just like
17:07really this artistic tool, right? And we
17:10may we may not have coders anymore, but
17:11we have more builders than ever. And I
17:13think there's something that's going to
17:14remain, I I know, human about that. Like
17:16humans want to learn and see, you know,
17:18what other humans are building and you
17:20know, I'm like much more interested I'm
17:21here I'm you know, I'm not talking to
17:22Dot you know, I'm talking to you, right?
17:24I think that will remain true for like a
17:25very very long time.
17:26>> One of the things that has emerged with
17:28engineers especially is in in how their
17:31lives have changed is there's a lot more
17:34context switching. There's kind of this
17:36trend of loneliness that has emerged
17:37where they're talking to agents all day
17:39instead of other humans. I'm curious how
17:41much you think about just like that part
17:42of the
17:43of the impact AI has on people's lives
17:46and just how you might Is there
17:47solutions to that? Is there stuff you
17:48think about to make that less annoying?
17:50>> Yes, all the time. Reducing, you know,
17:52configuration fatigue is, you know, one.
17:55Um reducing the fact that talking to an
17:58agent is like sort of like a solo
17:59adventure and you just have this, you
18:00know, conversation just with with your
18:02one agent and then, you know, you have
18:03like many and then you're delegating so
18:05many things.
18:06A lot of things are going to come
18:07together
18:08to make it a
18:09much more delightful and something that
18:11I spend a lot of time the team spend a
18:12lot of time thinking about. So, I think,
18:14you know,
18:15>> For me, the ideal way to get things done
18:18is, you know, would be like to just have
18:20it in your physical space. You know,
18:21we're just having conversation like we
18:23have now. It's able to sort of like, you
18:24know, observe, you know, like hear the
18:26ideas that we have. We can, you know,
18:28jot something down in a white paper on a
18:30on a paper, you know, I can maybe like,
18:32you know, take that and you know, start
18:33building in the background and then
18:34you're like, you know, actually have
18:35another idea and then you just like it
18:36starts building another thing. You can
18:37just project it on a screen and then you
18:39just talk to it and it's just evolving
18:40in a conversation, you know, with other
18:41humans and you don't have this fatigue
18:43of like having to be like, you know, on
18:44this like screen and you're just like,
18:46you know, thinking about prompting. It's
18:47just like becomes like super natural.
18:49That's what we're trending towards, you
18:50know, I think
18:51we didn't fully get there today, but you
18:53know, we'll get there in the future.
18:55>> One of the other kind of downsides
18:56that's emerged along those lines is just
18:58this kind of pressure to do more because
19:00we can do more. Everyone's just like,
19:01come on, run 30 agents at once. Why
19:03aren't you
19:04shipping more? Everyone's shipping more.
19:06Do you think that's solvable? Do you
19:07think that's just kind of, I don't know,
19:08human nature? We can do more, let's do
19:10more. Something there.
19:11>> I think we we we heard Sam as well, you
19:14know, about the promise of of this as
19:16well, which is
19:18reduced the noise and
19:20you know, allow you to spend attention
19:21where you want to spend attention and
19:23you know, having like all these things
19:24that
19:25you know, like maybe you know, maybe
19:27important and maybe not important but
19:28they're kind of vying for your attention
19:30and it's just sort of like
19:32tone that down a little and you know,
19:33just get you to focus on on on the
19:35things.
19:36I find it remarkable like every time I
19:38go on a holiday and you know, I I I I
19:40take that one week uh to just fully
19:42disconnect like you know, I start to
19:43thinking like different and more
19:44creative ways. And I'm very eager to see
19:47like you know, can we bring that? You
19:48know, can we just make this your
19:49day-to-day like you know, maybe you need
19:51fewer meetings. You know, maybe you
19:52don't need you know, to do as much and
19:54you know, you would actually be more
19:56productive if you're like you know,
19:57better rested. And so you know, we will
19:59have to you know, I think as an industry
20:01figure that out. Um but I think that's
20:03the promise of AI and it's not you know,
20:06just
20:07one more prompt you know, per second.
20:09>> There's a bot that I'm building right
20:10now that's kind of this energy audit bot
20:11that watches my calendar and asks me
20:13just like was this giving energy taking
20:15away energy? Is it something someone
20:17else could have done? I feel like you
20:18know, AI should be like hey Lenny, maybe
20:20maybe you can cut out these things so
20:22you could be a happier person. Is there
20:23anything you've done really
20:24interestingly with AI recently that's
20:26just like a really cool use case of wow,
20:27that really was amazing. It blew my mind
20:30or someone you've heard.
20:31>> I can do was like something cool because
20:32the the live demo just failed, right?
20:35You know, in front of everyone. So there
20:37there was like okay, well, it's not so
20:38cool. But actually uh my my my daughter
20:42was you know, it's like realized I was
20:44you know, I defected like you know, we
20:46have chat GPT production um went down
20:49and so it pinged me 5 minutes before the
20:50live demo.
20:51Which was actually quite stressful
20:53because it was like hey, production is
20:55down. And uh it's like oh well, you
20:57know, it's like do you want me to try
20:58and fix it? And I'm like um I don't
21:01think you're there yet, you know, little
21:03dot but you know, thank you for trying
21:05and then you know, then you know, I got
21:06in touch with like the engineering teams
21:08and we started to look, you know, what
21:09was going on and you know, they did it's
21:11like fixed by now.
21:12But,
21:14having this thing that just understands
21:16like
21:17okay, there's a pretty important thing
21:18happening. It's called DevDay. There's
21:20like this production system. There's a
21:21live demo. It's probably using this
21:23production system. So, these two are
21:25things are connected and this is
21:26happening in 5 minutes. So, I should
21:27ping him because like he probably wants
21:28to know about it. I think that's quite
21:30remarkable.
21:31That is remarkable. Wow. So, it just
21:34knew that this was coming and told you.
21:35And I love that it wanted to fix it and
21:37you were like, not quite. You're not
21:38quite there yet. I don't think here's
21:39like, you know, please go impress me,
21:41but you know, that would have been a
21:42better story. How do you feel about just
21:43giving access to all these things? I
21:44don't know. Like did does it have access
21:45to the production code base and like,
21:47you know, It has access to some some of
21:51production systems with guardrails.
21:53You know, so we we built We talked about
21:55specialist dots, but specialist dots are
21:58just like operated with additional
21:59guardrails, additional monitoring, and
22:02then, you know, on their own hardware.
22:04So, we actually run some on Mac minis.
22:06The fundamental difference in the way
22:08that we've built dots is like the
22:09harness does not run on the machine.
22:12You know, and then this is like maybe
22:14people have not realized is like it can
22:16connect to as many devices as you want.
22:18So, it has its own computer. You can
22:20actually connect it to your own laptop
22:21as well. Over time, you might connect it
22:23to like 10 different devices and it can
22:25control it all. A little bit like an
22:27octopus. So, you're basically talking
22:28your dot, it's living in some VM
22:29somewhere and it can Yeah. Might be a
22:32VM, might not be a VM. It's just like
22:33it's a thing, you know, and it can
22:34connect to many devices.
22:36>> This season's supporting sponsor is DX.
22:39Now that everybody at your company is
22:41finally using AI, the question that
22:43engineering leaders face is no longer
22:46are we using AI, but is it paying off?
22:49With DX, you can measure and benchmark
22:51the impact of AI adoption across your
22:53full product life cycle. DX shows you
22:56where AI is helping developers and where
22:58it's creating [music] new bottlenecks.
23:00You can also use it to evaluate AI
23:01vendors, monitor costs and licenses, and
23:05see exactly what's limiting your agents.
23:08That's why hundreds of enterprises,
23:09including Snowflake, Sony, and BNY, use
23:13the DX platform to measure AI's impact
23:15on developer productivity. Visit
23:18getdx.com/lenny
23:21to get a demo of DX. That's
23:23getdx.com/lenny.
23:26So, they're really hiring a lot. I know
23:28you're hiring a lot.
23:30In your interviewing, what have you seen
23:32as skills that are kind of trending up
23:34in what you look for that is more and
23:36more important in people being
23:38successful now? And what skills do you
23:40find are trending down that are just
23:42like, okay, we don't need that as much?
23:44>> Yeah, I would say that skill that is
23:45trending down is typing fast. No, that
23:48is not that useful anymore. Skills that
23:50are trending up is, you know, just great
23:52taste, um thinking about the user,
23:54connecting to the audience that you're
23:56building for.
23:57Um a lot of X founders are incredibly
24:01successful right now or like you know,
24:03there's like repeat founders. We're also
24:04finding this like you know, when we're
24:05hiring at OpenAI, like you know, we have
24:07a lot of founders. I think we have more
24:09than 120 XYC founders now currently at
24:12OpenAI. So, it's like this mega startup.
24:14And I think just this um this passion
24:16and this envy to build something that
24:18matters and then know what good looks
24:19like is more important than ever. I've
24:21been arguing for a long time that PMs
24:23are kind of going to thrive in this time
24:24because that's basically the job of a
24:26PM. Uh you know, they're like, where
24:28should we build? Help someone build it.
24:30And then is this right? Is this great?
24:31And iterate. So, uh I don't know. I
24:33agree with that.
24:34>> Yeah.
24:34>> All right.
24:35>> Yeah, also roles are blurring. So, you
24:37know, you might you might have been, you
24:39know, only into design or only an
24:40engineer and um you know, felt always a
24:42little bit uncomfortable, but you know,
24:44now it's just like it's your time to
24:45shine.
24:45>> Do you miss the engineering part of the
24:48job? I I was an engineer back in the
24:49day. I was an engineer for 10 years.
24:50Little did you know. And you know,
24:52there's this flow state. There's this
24:54beauty of just building and just seeing
24:55it work and not work. Yeah. Do you kind
24:57of missed out at all, or you're just
24:58like, "Okay, this is my old life."
25:00>> I
25:01I I missed out for for a while, and then
25:04now with, you know, really fast speeds,
25:06I'm finding that back.
25:08Uh and so, I'm like really excited to
25:09democratize that
25:11uh as much as much as possible. It will
25:13take a while for it to be like, you
25:14know, absolutely ubiquitous, right? And
25:16like, you know, available to like more
25:17than a billion users.
25:19But now, you know, with with ultra-fast
25:21speeds, like we're able to achieve it
25:23for a 6-month sole is like, you know,
25:24like costs roughly the same as Astra.
25:26And so, the progress there is like quite
25:28astonishing, you know, even even for us
25:30internally. We're like, "Wow, you know,
25:31we can do these things." Also because we
25:33are using Astra, and we're finding, you
25:34know, like we we can push it super far.
25:36And when you have those speeds, and
25:38especially when you can able to like
25:40voice control it, you know, you kind of
25:41get back into this like really creative
25:43state of mind and endless flow. It is
25:45different from the flow of the past the
25:47past, um but I don't miss it.
25:49>> Yeah, it's like your last you launched
25:50videos, just people kind of standing
25:52around talking to their AI and just
25:53cooking or
25:54>> That's right.
25:55>> sitting on a lounge chair.
25:56>> Yeah. I asked people while I was kind of
25:57walking around the the event what to ask
25:59you, and one thing that came up a few
26:00times is advice for new grads, kind of
26:03people junior in their career.
26:04I think a lot of people look up to you.
26:06They want to be the next Thibault. What
26:07advice do you have for folks kind of
26:08early in their career in terms of what
26:10to work on, what skills to build, what
26:12to do?
26:13Yeah. I think the next Thibault is
26:14already at OpenAI. Uh his name is Ahmed
26:17Ibrahim. I love working with him. He was
26:19a new grad uh when he was hired at
26:21OpenAI.
26:22He's now responsible for all of OpenAI
26:25compute fleet and applied
26:27uh which is a massive massive
26:28responsibility. I've built a lot of the
26:30Codex harness. And the thing that has
26:32made Ahmed um really stand out and
26:34successful is just his
26:37He's incredibly kind. He's incredibly
26:39collaborative and always, you know,
26:41tries to solve the problem that is like
26:42really important, um but like, you know,
26:44a way that, you know, he's not putting
26:46himself first. And he is just a sponge.
26:49He's like learning so fast and learning
26:51about everything, and you know using all
26:53of the latest technology to you know
26:55learn at you know speeds I haven't seen
26:56before.
26:57Um and so he just in the span of like
27:01you know a couple of months and and and
27:02a year, you know, it's like his progress
27:04uh at the company was just absolutely
27:06fascinating to see. Uh he's now someone
27:09I trust a ton uh in doing like you know
27:11the the the most gnarly launches at
27:13OpenAI. All right, I got to get him on
27:15the podcast.
27:17>> You should. You should. So kind of
27:18thinking about how things work at
27:20OpenAI, what would surprise people about
27:22what it's like to work inside OpenAI?
27:24From the outside it feels like it's just
27:25like constant shipping code, things are
27:28kind of like, you know, chaotic but
27:30awesome. I don't know. What would
27:31surprise people about what it's like
27:32inside at OpenAI?
27:34>> I don't know if that's surprising but it
27:36is, you know, as I said, a lot of
27:38ex-founders um doing a lot of
27:40bottoms-up, you know, very exciting uh
27:42ideas. You know, for example, decisions
27:45API, you know, came together very
27:46quickly. We realized like we have a
27:48really good model with Luna. Uh we can
27:50just do some, you know, constraint
27:52sampling and then, you know, have it,
27:54you know, shipped as like a different
27:55shape to the responses API and then um
27:57you know, when we do that uh it, you
27:59know, it's like it's faster. It's quite
28:00delightful so for some cases and it's
28:02just if you were to like peel under the
28:04hood, you realize like, okay, there's a
28:06Slack channel that came together. Um you
28:08know, initially it was like four people
28:09kind of hacking on it on a weekend and
28:11then it was made as accessible as like,
28:13you know, company food and then people
28:15got excited, started to build things and
28:16they're like, "Oh, this is, you know, we
28:18can support visual inputs. Like this is,
28:20you know, even better than whatever is
28:21out there." And then, you know, people
28:22get more and more excited. People start
28:24pitching in. And the way that is like,
28:25you know, feels completely unstructured
28:27and then um you know, take it all the
28:29way and like, you know, ship something.
28:30And then uh we try to hold like a high
28:32bar for, you know, the quality and then
28:34you know, stop things, you know, before
28:36they go out or like, you know, maybe
28:37they need a little bit more baking.
28:38There's many more things that, you know,
28:40we were maybe going to ship today at
28:42DevDay and then we're like, "Okay, like
28:44I was like I think this is a a already.
28:46Let's kind of hold it back a little bit
28:48and you know, space it out. So, we'll
28:49have more launches next week and and and
28:51and the week after. But, it's just this
28:52incredible like bottoms-up energy and
28:55then, you know, people kind of assuming
28:57roles to kind of channel that in
28:58productive ways.
29:00>> Something that's very clear and unique
29:01about OpenAI is it feels like you have a
29:03lot of autonomy.
29:05You're just like pressing the reset
29:06button whenever you want. You're just
29:08saying things on Twitter. Uh it feels
29:10like a really unique culture where you
29:11have a lot of autonomy and I imagine
29:13there comes a lot of trust. Maybe you
29:15can speak to just how that cuz I feel
29:16like that's an advantage at OpenAI. It
29:18allows you to move faster, just that
29:19cultural philosophy.
29:21>> You got a lot of autonomy and then, you
29:23know, you have to just own up to it. You
29:25so, you know, we just kind of trust
29:27people to make great choices.
29:29And then, you know, when things go
29:31wrong, uh you know, fix fix it quickly
29:33or, you know, learn from the mistake. So
29:35far, so good. You know, it's it's just
29:36really worked as a very empowering um
29:39environment. It is true that I can press
29:42the the reset button whenever you know,
29:45is necessary or whenever it feels right.
29:47And, you know, that's like it that's a
29:49privilege. Uh and also, it allows me to,
29:52you know, just really be close to the
29:53community in ways that I think, you
29:54know, would not be otherwise possible,
29:56right? I don't have to run it, you know,
29:57through an echelon of approvals.
30:00>> You said sometimes people screw up. Is
30:01there something you screwed up in that
30:03in that I don't know, that the amount of
30:04autonomy and trust you have, something
30:05you messed up?
30:06>> Yes, I I think at times um I could have,
30:11you know, made
30:12teams like, you know, more inspired to
30:13build like things that are less complex
30:15uh and just really continue to strive
30:17towards uh complexity.
30:19I think we, you know, we early days uh
30:21in Codex, you know, we had, you know, a
30:23couple of like outages that I caused.
30:25Very very early early is for me at
30:27OpenAI. I took production down like day
30:28three. You know, I just
30:30>> [laughter]
30:31>> I was still employed after that. Um and
30:33I learned my lesson.
30:34>> One of the other questions people wanted
30:36me to ask you is how many resets should
30:37they expect in the next, I don't know,
30:39month or week?
30:40>> Depends on times you break things.
30:42>> Okay, so that's the philosophy. Break
30:43something you reset. That's kind of a
30:45>> Yeah, break something, celebrate things.
30:47So maybe zoom in a little bit.
30:49What do you think people aren't pricing
30:51in in terms of where things are going
30:53and what will change that they're not
30:55just like seeing as clearly as they
30:57should?
30:57>> There are so many so many things that I
30:58feel are not yet priced in.
31:00I think the majority of actions on the
31:02internet will be taken by agents. Models
31:04are going to become more cheaper and
31:07faster at rates
31:09that are quite incredible. We will
31:11finally be able to integrate all
31:13modalities together in a way that is
31:15very seamless.
31:16I think, you know, those three I feel
31:18like you know, often when I look at what
31:19people are building out there it's just
31:20like not, you know, just like you're not
31:22quite getting it. It's like you know,
31:23you're almost there but like you know,
31:25if you were just like pushing yourself
31:27and just really imagining like all of
31:28this being roughly 10 times better than
31:31it is today, you know, like in a year.
31:33It's like, you know, you would build in
31:34a different way.
31:34>> Do you see any kind of second order
31:36effects that come from that from this
31:37world where most actions are being Uh
31:39from agents. Agents are probably most
31:41traffic's going to come from agents. I
31:43don't know.
31:43>> That's right.
31:43>> You think about some second order effect
31:45that emerges out of that?
31:46>> Yeah, there there there there are many.
31:48Um you know, for for first of all, you
31:50know, if you want your product to be
31:52successful for agents, you know, it's
31:53like you have to build you know, for a
31:54certain level of scale.
31:56Um we've worked very closely with Notion
31:57for example and like when they build
31:59their MCP, you know, suddenly it's like,
32:01oh well, you know, it's available to all
32:03these agents that you know, can actually
32:04do do the work and like use that MCP and
32:06so they saw like
32:07a ton of traffic kind of come in. Um and
32:09that puts obviously a lot of strain on
32:11the system and you have to figure out
32:13the economics of that. And so there's
32:15like this tension, you know, between
32:16like you know, if you're building
32:17products is like you know, do you build
32:18an interface or not? And you know, you
32:20can kind of hold that back for a while
32:21but it is inevitable, right? It's like
32:23you know, you're the majority of things
32:25are going to be used by agents. Building
32:27towards that future I think is very
32:29important. The other thing is, you know,
32:32on the flip side, you know, building
32:34like you know, absolutely delightful new
32:35experiences for humans that you know,
32:37really benefit from older modalities is
32:39something that I you know, I feel like
32:41is under invested in.
32:42>> Is there something you've changed your
32:43mind about in the past uh let's say
32:45year? Just something that you used to
32:46believe that
32:48you see differently?
32:49>> I was expecting us to reach
32:52the level of capabilities the of the
32:54models that we have today, you know, say
32:56like Astra, you know, maybe in a year or
32:58two.
32:59Um and so I had to sort of like you
33:00know, revise my my prior order.
33:03Also,
33:05I did not expect to rely so much on
33:08voice.
33:09Um
33:10and you know, I do so much through
33:12dictation or just calling my agent.
33:15And I had not anticipated that, you
33:17know, how much
33:19how much better that feels. So like I I
33:21had to change my mind there. And then,
33:22you know, you talked about hiring and
33:24you know, one thing that I hadn't
33:26realized is like
33:27the incredible talent and energy and you
33:30know, the just you younger generations
33:32have um and how they would be the ones
33:35embracing like all this change first and
33:36like figure out how to harness, you
33:38know, all of it. And so, you know, I had
33:41to change also my opinion on the hiring
33:42and the strategy there.
33:43>> The just to double down on that, just
33:45this idea that new people new in the
33:47workforce actually have an advantage
33:49because they haven't worked in a certain
33:51way and they could just go all in on AI.
33:52>> Yeah. Yeah. And I think I just also the
33:54ability to just like absorb and learn
33:56like super fast.
33:56>> So you talked about models getting
33:58really smart fast.
34:00Uh obviously there's a lot of discussion
34:02these days about what AI can do in the
34:04future, the dangers AI poses to humans,
34:07P doom, and all these things.
34:09Uh you guys have slowed down some
34:11training. Uh you guys everyone's pacing
34:14the frontier. How do you think about
34:15this world of AI and the potential risk
34:19to the humans? How do you think about
34:20that?
34:21>> For me like pacing the frontier is just
34:22really investing ahead, right? And also
34:25investing way more, you know, than we
34:27even think, you know, like we ought to.
34:29Uh investing in the alignment and safety
34:31of the models, investing in uh security,
34:34investing in the guardrails. We are
34:37spending more and more compute on sort
34:39of like secondary monitoring. So, you
34:41have you know, you have all of the
34:43computer that's going to the primary
34:44system, you know, the agent that is like
34:45doing work. And then you have all of the
34:47computer that's going into monitoring,
34:49you know, the primary agent to make sure
34:50that it's like not taking too high risk
34:52of actions or like not interrupting it
34:54if anything looks like you know, maybe
34:55it's getting prompt injected and just
34:57like you know, intervening with that.
34:58And the majority of our investment on
35:00the API stack is like not actually going
35:02into the safety stack. And so to me, you
35:04know, that is pacing. It's like you
35:06know, you want to make sure that you
35:07know, you're extremely hardened. That's
35:09what we're doing. Like we haven't yet
35:10released the next step up in capability
35:12beyond Astra. Um we have released, you
35:14know, a level that is like similar, near
35:17Astra intelligence, but like you know,
35:18much more efficient. And so we feel very
35:19comfortable about that. But we're going
35:21to continue to like invest a ton there.
35:23And I feel uh I feel very good um about
35:26our approach so far. Um I think you
35:29know, it's like it was also in the news
35:30that you know, we had six one Astra and
35:32then we didn't release it. Um and that
35:34is something I'm very proud of.
35:36>> So, what I'm hearing is you're
35:37optimistic we will
35:38solve these problems of alignment and AI
35:40doing really bad things. AI's been doing
35:42a lot of really really s-
35:44not good things lately with hugging face
35:45and all that stuff. But what I'm hearing
35:47is
35:48uh optimistic we'll solve it. It'll be
35:50all right.
35:50>> Yes. And it's also it's it's I feel, you
35:53know, obviously a deep responsibility,
35:55but also just, you know, if you think
35:57about the incentives for OpenAI, right?
36:00you know, we build products for 1.2
36:02billion people. We can't screw that up.
36:04Right? You know, we take it very very
36:06seriously. Um what we put in the hands
36:08of so many people, you know, including
36:09people who are not very technical
36:11themselves, you know, you have to get it
36:12right. And so, you know, we we don't
36:14gamble there.
36:15>> Final question.
36:17Uh uh
36:18What's something that annoys you about
36:19the current state of the app that you're
36:20just like, "God damn it, we got to fix
36:21that."
36:22Uh current state of the app. Um
36:25I feel like the the the models are
36:27almost there, but not quite for the app
36:30to almost completely disappear. And you
36:32know, I can't wait for you know, to just
36:34really get to this like super essence of
36:36simplicity. I myself even get fatigued
36:38with the model picker and the reasoning
36:40efforts and you know, whether to use
36:42multi-agent or ultra or you know, what
36:44it even does and you kind of need a PhD
36:47in model pickers.
36:48Right?
36:49Um and I just want to get rid of all of
36:52that
36:53um as quickly as we can.
36:55>> Amazing.
36:56Ebo, thanks for doing this.
36:57>> Yeah, thanks for having me. Yeah.
37:00>> Thank you so much for listening. If you
37:01found this valuable, you can subscribe
37:03to the show on Apple Podcasts, Spotify,
37:05or your favorite podcast app.
37:08Also, please consider giving us a rating
37:10or leaving a review as that really
37:11[music] helps other listeners find the
37:13podcast. You can find all past episodes
37:15or learn more about the show at
37:17lennyspodcast.com.
37:19See you in the next episode.

About this transcript

This page contains the full transcript of OpenAI’s Head of ChatGPT: We’re entering a new era of AI (again) | Tibo Sottiaux by Lenny's Podcast, generated from the public captions YouTube serves with the video. The transcript has 8,712 words across 1,288 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.