Latest / Agentic Conversations (formally mlops.community) / Agents & the $40M Bet on Multiplayer AI

Transcript

Agents & the $40M Bet on Multiplayer AI · Agentic Conversations (formally mlops.community)

• 0:00
  Even if the technology again plateaus, even with current
• 0:04
  model shape, the way we're going to walk in 10 years is probably
• 0:07
  nothing compared to the way we work today.
• 0:09
  And so there's many, many, many opportunities to build for for
• 0:12
  everyone we. Should probably start with this
• 0:21
  multiplayer versus single player.
• 0:23
  AII want to mention to folks that you've got a decorated
• 0:28
  background, you're at Stripe in the early days, you then you're
• 0:32
  also at Open AI and now you're doing your own thing with dust.
• 0:36
  So it's super cool to see this progression.
• 0:40
  Are there things that in those early days of Stripe scaling up
• 0:43
  or in the early days of Open AI scaling up that you have taken
• 0:49
  away from that and you want to infuse into dust as you're
• 0:54
  scaling up? Oh yeah, a ton.
• 0:56
  Mostly Stripe over open eye first.
• 0:59
  Yeah, I've, I've been very lucky.
• 1:01
  I've been very lucky. So we joined Stripe through an
• 1:04
  acquisition actually with my current Co founders.
• 1:07
  It's not the first time we've started a company.
• 1:10
  The big difference that it was 15 years ago and we, we were
• 1:13
  young, we had absolutely no clue about what we were doing, but it
• 1:16
  was fun. But if we can chat about that
• 1:19
  later, it's a whole subject of its own.
• 1:21
  Stripe, we saw the 150 people to 3000.
• 1:24
  It was absolutely incredible. And I think I owe very dear to
• 1:30
  my heart the 150 to 500 people. I think it was pretty magical at
• 1:34
  Stripe because there was really very minimal management and it
• 1:40
  still was working really well. And I'm trying, I've, I've been
• 1:43
  trying since then to understand what was happening at the time
• 1:46
  for that to work so well, do you think?
• 1:48
  That's the writing culture. So, yes, so I've got a, I've got
• 1:53
  a bit of a theory about that which we we actually turn into
• 1:57
  operating principle at dust. But my my big theory about the
• 1:59
  time it's slightly bullshit, but I think it's interesting
• 2:02
  nonetheless. My theory is that we were
• 2:05
  actually implementing A flocking algorithm.
• 2:07
  So you know what the flocking algorithm is.
• 2:09
  It's like when you so when you want to program a flock of birds
• 2:13
  or flock of whatnots in computers, there are kind of a
• 2:17
  autonomous groups of things that auto organize around obstacles.
• 2:22
  There's actually only three things to there's only three
• 2:24
  rules to make it work. It's local inspiration, so small
• 2:29
  files that push the the people away, distant attraction and
• 2:35
  alignment and that gives you kind of some form of autonomous
• 2:40
  behavior around obstacles. You can, you can look at
• 2:44
  flocking algorithm online and you'll see the small videos.
• 2:47
  And I think at Scribe, the writing culture was the distant
• 2:50
  attraction. So Scribe was very early into
• 2:54
  implementing that kind of a open writing culture.
• 2:57
  It was mostly through e-mail at the time, interestingly.
• 3:01
  And what it was great. It was great.
• 3:04
  Oh yeah. I don't like slag that much
• 3:07
  personally, and I love e-mail, but that's another take that we
• 3:10
  can spend a lot of time on. It's surprising me here.
• 3:14
  All right. But anyway, so distant traction.
• 3:16
  Why? It's because we had those
• 3:18
  mailing lists. And so there were people within
• 3:21
  the org that were spending a lot of time kind of sifting through
• 3:23
  those main lists. And so that means that somebody
• 3:25
  would see someone else talk about subject A over there and
• 3:29
  see another person talk about the same subject over there in
• 3:32
  the org and was able to tell them, hey, you're talking about
• 3:34
  the same thing, you should chat. And so that's kind of a distant
• 3:36
  traction. False.
• 3:39
  I think at Scribe, one thing that really struck me is the
• 3:41
  amount of trusts between people at the time, it was just
• 3:46
  ingrained in the culture in a very nice way.
• 3:48
  And that's kind of a local separation force because you
• 3:51
  don't look over the shoulder. If somebody says I'm I've got
• 3:53
  this, they've got this and you stop caring about that.
• 3:57
  And then finally, the last one was which is, which is almost
• 4:00
  the, the the most bullshit one is alignment.
• 4:02
  But I think at Scribe were we were and which is very
• 4:04
  interesting because it's very hard to reproduce that in modern
• 4:07
  times with AI, but we had a very strong alignment on the vision.
• 4:11
  It was very easy to understand. We are a dev API for payments.
• 4:17
  Basically there's a tensor of countries, payment methods,
• 4:20
  payout methods, and you just just want to to fill it in.
• 4:24
  It was very easy to understand. And so that means that you are
• 4:26
  every is kind of aligned very naturally.
• 4:28
  And so that gives you the three properties of looking, which
• 4:31
  gives you kind of a autonomous behaviour in a sense, emergent
• 4:36
  autonomous behaviour, which explains why.
• 4:38
  But then surely we didn't need that that much management and it
• 4:40
  was so nice to work in that environment.
• 4:44
  It feels like we could draw a parallel here and heads up
• 4:48
  Segway into multiplayer AI, single player versus multiplayer
• 4:53
  with the Flock algorithm. And how have you thought about
• 4:56
  humans working with agents in that Flock style?
• 5:01
  Do you think that there's anything that you could look at
• 5:03
  there? Yeah, it's probably a ton of
• 5:05
  stuff, but it's very hard to anticipate because today we are
• 5:08
  still very much stuck in the single player mode and.
• 5:11
  What do you mean by that single player mode?
• 5:14
  It's mostly that most of the agents interactions we have, we
• 5:17
  just have them alone for two reasons.
• 5:21
  I think the main 2 reasons I see is first, the timer reason of
• 5:26
  the tasks an agent can take is still pretty small according to
• 5:33
  the meter benchmark we are topping at right now at a couple
• 5:38
  hours, let's say 1/2 a day and mostly in coding tasks by the
• 5:42
  way. And you know half a day is still
• 5:44
  very much a single player thing. You I can walk half a day on the
• 5:49
  task. You can walk half today on this
• 5:50
  task and you walk alone and it's fine because it's a kind of well
• 5:54
  shaped tasks. The other thing is that the the
• 5:57
  they're still very imperfect in many ways.
• 5:59
  And so you interact a lot with them.
• 6:02
  It's messy really. When you look at the an agent
• 6:05
  trace within a human and an agent, you look at the all the
• 6:09
  sausages made and you're like, oh man, you did something funky
• 6:12
  there. Or it's like, oh, you were so
• 6:16
  close dude. A little bit over to the right
• 6:20
  or left you would have been good.
• 6:22
  And so on those time horizon and, and in terms of
• 6:24
  capabilities, the models are super jagged, which means that
• 6:28
  you on that you're looking in that direction.
• 6:31
  The thing is clearly superhuman. You move a few degrees off on a
• 6:37
  slightly different task, you fall into a valley and the, the,
• 6:40
  the thing, that thing is the dumbest, dumbest thing on Earth.
• 6:43
  And so that's why it's messy. And so all of that makes it very
• 6:47
  kind of unnatural to collaborate in an agent loop with multiple
• 6:51
  humans to share your sessions with agents because they are
• 6:55
  messy. You share the outputs, you don't
• 6:56
  share the session. And so that's that's why I think
• 7:00
  we're still very much single player AI in the way we we walk.
• 7:03
  We all spend a lot of day in, in codecode or codecs.
• 7:06
  We all spend a lot of day in our favorite agent harness to walk
• 7:10
  on presentation and stuff. And it's we pretty much do it
• 7:15
  alone and we share in the in the other platforms the I put and we
• 7:19
  collaborate there be it on GitHub, be it on Notion on
• 7:21
  Google Drive, what not. I think the thing that the the
• 7:25
  the thing that is supposed to is, is, is about to switch is
• 7:28
  probably as the models are getting better and they seems to
• 7:31
  be getting better consistently. It's always an open question,
• 7:33
  but when you look at the past few months, well, the the time
• 7:38
  horizon of those tasks was increased.
• 7:40
  And if you start considering a task like that represents a week
• 7:43
  of man walk for most job descriptions, those tasks are
• 7:51
  inherently cross team, cross people involving multiple
• 7:55
  person. Even if we if you think of
• 7:56
  coding, there is very few coding tasks that takes a week that
• 8:00
  doesn't involve collaboration even with somebody external, be
• 8:03
  it a product manager, be it a designer or maybe the infra
• 8:08
  person because you have something finicky about the
• 8:10
  infra and so. And if you look at something
• 8:14
  that is outside of coding, then most tasks that takes a week,
• 8:17
  they are extremely cross functional.
• 8:19
  And so that will flip the picture where it's not going to
• 8:24
  be possible anymore for just one human to steer the agent in
• 8:28
  those type of tasks, the agent, even the relationship between
• 8:31
  the human and the agent will, will flip.
• 8:33
  The agent will orchestrate a longer horizon task involving
• 8:36
  humans and the humans will steer.
• 8:38
  But they'll, they'll, I'm quite convinced there's going to be a
• 8:41
  need for many humans to steer those agents beat an agent or
• 8:44
  multiple agents. We don't really care for those
• 8:47
  kind of long horizon tasks. They'll invoke the human agent
• 8:51
  or the human tool when they need to, so it's invoking A SRA tool.
• 8:57
  For that, there's going to be, there's going to be some of that
• 9:00
  for sure. Yeah.
• 9:01
  And now don't you feel like this is just an, it's a UX problem
• 9:06
  more than anything? Because when you tell me about
• 9:09
  the single player mode, I think, OK, but what ways do I have to
• 9:15
  share with my colleagues rather than exporting the chat or
• 9:22
  putting it up with my PR? Things like that.
• 9:26
  It doesn't really feel like I have the tools in order to get
• 9:32
  to bring someone else along this journey with me.
• 9:37
  Yeah, I think, I think the, the, the interaction for pretty much
• 9:40
  remains kind of a, a you with the agent.
• 9:43
  But it's just that the agents will create many of those
• 9:45
  sessions in the to to achieve that task and have will kind of
• 9:49
  collaborate with many humans. And so it is entirely a UX is
• 9:55
  maybe slightly diminishing, but I do think it's not a, it's not
• 10:00
  a, a machine learning problem. It's more of a product problem.
• 10:05
  It means it just means that as an example, if we if we take Co
• 10:10
  work, which is having kind of a real moments these days because
• 10:15
  it's working locally on your computer, it is inherently a
• 10:17
  single player thing. It's very hard to kind of
• 10:20
  collaborate around something through Co work with multiple
• 10:24
  humans. And that means that you need to
• 10:25
  push that harness inside of a Austen environment.
• 10:29
  This is the only way I think and, and there's a lot of things
• 10:33
  that we we are exploring around what is the right products of
• 10:35
  face to enable that collaboration where you have
• 10:38
  multiple humans, multiple sessions around a shared state
• 10:44
  and how that what what it can look like.
• 10:46
  I think this is the frontier in terms of products of interacting
• 10:49
  with agents, and that's the one we're very excited about.
• 10:51
  Talk to me more about that. So one thing we're doing right
• 10:54
  now is, is so we, we, we've, we've, we've released the first
• 10:59
  version of that, which is called really, we call it a pod.
• 11:04
  And a pod is really a shared state with a number of sessions
• 11:11
  between humans and agents. And I'll give you a very simple
• 11:14
  example of where it plays really nice.
• 11:17
  So we have, as with any companies, we have a team weekly
• 11:22
  and the team weekly is mostly made of a set of slides that are
• 11:28
  somewhat always the same with small variations every week and
• 11:31
  maybe some deep dive about subjects every other week.
• 11:37
  And so in the past, the way it would work is that somebody
• 11:40
  would stop doing a slide and and ping the humans and have the
• 11:44
  humans collaborate in those slides, etcetera.
• 11:46
  And the kind of orchestration of all of that would take a some
• 11:50
  amount of time to that human. No comes agents.
• 11:55
  Maybe you can the the initial organization building a slide.
• 11:58
  You can, you can, you can offload to an agent, but you
• 12:01
  still have to ping the humans and organize the, the kind of
• 12:05
  organize the collaboration so that we get to a shared result.
• 12:10
  And so with a pod, which is a group of conversations, group of
• 12:15
  humans and a shared state. No, our team weekly is much more
• 12:19
  driven by an agent than driven by a human.
• 12:21
  So basically on our teamwork is on Tuesdays, Monday morning,
• 12:25
  somebody triggers an agent with a skilled team weekly, the team,
• 12:28
  the agent creates a pod, it creates one session pair slide
• 12:35
  with the owner of each of those areas and maybe one session for
• 12:39
  the people that are doing the deep dive that week.
• 12:43
  And it will will pre walk the slides.
• 12:45
  So going into meta base, pulling the data, keeping up to date
• 12:48
  revenue customers, going on GitHub for the the product walk
• 12:53
  and then going on whatever platform for the support state,
• 12:55
  whatever, you know, prebuild the slides.
• 12:57
  And then it'll ping the the the the person responsible for that
• 13:00
  slide and say, Hey, what do you want to talk this week about?
• 13:02
  And let me let's collaborate on on this slides.
• 13:04
  What do you want to had? Is there anything on the
• 13:06
  highlight? And so that collaboration
• 13:08
  happens in one of that session. Each humans that are responsible
• 13:12
  for each of those subjects are collaborating in different
• 13:14
  sessions. And on Tuesday, let's say noon,
• 13:17
  the agents, the agents wakes up and like, looks at everything,
• 13:21
  probably Ding on the humans that haven't finished their work,
• 13:26
  wakes up again in two hours, like 2 hours before the two
• 13:29
  weekly look at all the, the things that have been built and
• 13:31
  coalesce them into one shed presentation.
• 13:34
  So we, we don't use any presentation anymore.
• 13:35
  We use, we have a frame which is the equivalent of canvas or
• 13:39
  artifacts or really kind of a culturated UIS.
• 13:43
  And so it it takes all of those frames and and and pull them in
• 13:47
  that in that short frame and that thing is ready.
• 13:50
  And really the the weekly, the team weekly, which is kind of a
• 13:54
  a multi days collaborative task is today mostly on auto mode,
• 13:59
  right. And I think that's that's the
• 14:00
  kind of a product surface that are interesting.
• 14:02
  If you don't have that pod, which is that group with that
• 14:05
  shut state feels kind of weird because you have a bunch of
• 14:08
  conversations over the place. You don't have a shuts place to
• 14:11
  see them. All the agents doesn't have a
• 14:13
  place to organize all of that work with multiple humans.
• 14:16
  And so I think that's that's the type of things we're exploring
• 14:19
  and which is very exciting for, for what's to come up for more
• 14:23
  than for what's where we're at. Can you talk to me more about
• 14:26
  this pod? Yeah.
• 14:27
  So yeah, the the pod is really like, technically it's really a
• 14:32
  group of humans, a group of agents, a group of sessions,
• 14:38
  which is the interaction between the human and agents.
• 14:40
  And these are sandboxed. Each of the sessions has a
• 14:44
  sandbox. Each of the session has a, a
• 14:47
  state. So a state, let's go to file
• 14:51
  system. The file system is visible by
• 14:54
  the humans but also mounted on the sandbox basically.
• 14:57
  To make that easy, we're diving into the technical details here.
• 14:59
  We use GCS FUSE. So we back that with GCS.
• 15:02
  So it's very easy to show it in the UI.
• 15:04
  It's very easy to show it as a mounted as a file system to the
• 15:07
  agent. So the home directory of the
• 15:09
  agent in the sandbox is is GCS back basically.
• 15:13
  And so that's the state in one session.
• 15:15
  And then inside of the pod we also have another file system
• 15:20
  which is shared across the sessions.
• 15:22
  And the agents are free to move files between the two file
• 15:24
  systems. So basically they see slash
• 15:26
  session slash blah, blah, blah. And in there they have all the
• 15:30
  all the file system related to the current session.
• 15:33
  And then they have slash pod and the under slash pod.
• 15:37
  That's the shared file system where they can move some of
• 15:40
  those files up or down as they see fit, and so they can really
• 15:44
  collaborate across sessions. You can have collaboration
• 15:46
  across sections through that shut file system at the pod
• 15:48
  level. And when I'm interacting with my
• 15:51
  agent, I can drop into somebody else's slides, presumably also
• 15:56
  it's not just my slide. So if yeah, if the if the slides
• 16:00
  of that also, so somebody has a session with an agent inside of
• 16:02
  the pod is working on a slide, it leaves naturally as a
• 16:06
  starting point on the file system of the session.
• 16:08
  So it's mostly visible to that session and then that person
• 16:11
  will say, OK, the slides are ready, move them to the move
• 16:13
  them to the to the pod, the slides get moved and then
• 16:16
  somebody else will be able to see them or even modify them,
• 16:21
  which is risky in a sense, but it's fine.
• 16:23
  I mean, that's that's in Google Drive.
• 16:25
  We can modify whatever else has been doing in your company as
• 16:29
  well. So I think it's OK.
• 16:31
  Do you have the possibility to look at the history of the
• 16:36
  modifications in case you're like oh control Z?
• 16:39
  Yeah, no, not right now, technically.
• 16:44
  That's why it's risky then, yeah.
• 16:45
  Technically, technically we have it because GCS is version.
• 16:49
  So technically we do have it and that stuff is something that we
• 16:52
  would want. Yes, you can.
• 16:55
  You can. The the though the reality is
• 16:58
  that you do something with an agent, it will.
• 17:01
  What will happen is it's going to read the file and then it's
• 17:03
  going to emit and edit. And so generally they're able to
• 17:08
  pretty easily backtrack their change because they still see
• 17:11
  them in context. The only thing that the agent
• 17:14
  can do that would be bad is like like deleting the whole pod
• 17:18
  sheds file system. That's possible.
• 17:22
  Technically, I don't think it happens much.
• 17:26
  So that's that's kind of fine. And also I imagine you have some
• 17:29
  kind of redundancy, maybe some backup somewhere, just in case
• 17:32
  you. Yeah.
• 17:33
  So again and just yes, everything is version.
• 17:34
  So we could, we could totally restore it and and and yeah.
• 17:40
  So OK, now the shared state and the ability for almost like this
• 17:47
  tiered level of I'm working, I'm bringing it down, not
• 17:52
  necessarily locally, but I'm having my one player experience
• 17:58
  and then I'm pushing to the multiplayer so everyone can
• 18:01
  share and look at it. How are you envisioning that in
• 18:06
  other scenarios? There's, there's so many of
• 18:12
  them. It's, it's one of the, by the
• 18:16
  way, it's one of the very hard difficulty of building an
• 18:21
  Orbiton platform. There's a lot of value into
• 18:23
  building in the northern platform.
• 18:25
  There's a lot of pain in building a northern platform.
• 18:28
  The pain is mostly this one is that the, there is so many use
• 18:32
  cases that's that you, it's, you cannot, I mean, I would love to
• 18:36
  be able to tell you dust is awesome for that, but that means
• 18:39
  we would be verticalized. If we were to be verticalized,
• 18:42
  we wouldn't have the right to really try to equip the entire
• 18:45
  company. And so that's why we accept that
• 18:47
  that tension. Many other ways to use a pod,
• 18:50
  You can, you can use a pod as a, as a, as a.
• 18:54
  So basically we also have in that pod, we also have a list of
• 18:57
  tasks, which is a way to try to organize, walk.
• 19:01
  And so you can create a pod for any project you do.
• 19:04
  And let's say you're on Slack and you're like, oh, I need to
• 19:07
  do, we need to do this, this and that.
• 19:09
  And you call Dust from Slack and you ask it to add those tasks to
• 19:13
  the pod you're working on May, let's say you're working on the
• 19:16
  initiative sandbox. You're working on adding sandbox
• 19:18
  to Dust and say we should do this, this and that the agent is
• 19:22
  capable of interacting with the pod and add tasks itself.
• 19:26
  And then as a human, you can you can the task are kind of a nice
• 19:30
  way to organize the organize the walk and trigger sessions to
• 19:36
  start solving the task with an agent and a human working
• 19:40
  together to solve that task. So you have that kind of a task
• 19:44
  kind of trying to explore the task management part of working
• 19:48
  with agents around the project. There's obviously another usage
• 19:52
  of POD which is which works well is around internal kind of
• 19:58
  internal supports like. So obviously the kind of
• 20:02
  external support, the tier one support will, will is best
• 20:05
  tackled in platform that are specialized for that.
• 20:08
  But when the questions is finicky and it reaches your
• 20:11
  internal support team, a pod is a really great place to share
• 20:15
  the way you solve those questions for people to, to, to,
• 20:20
  to look at and, and, and explore.
• 20:22
  And so I think there's many, many use cases around that we're
• 20:26
  really in the business of creating Lego bricks that work
• 20:30
  well together and and make sure that they are somewhat
• 20:33
  universal. And so we see the very varied
• 20:38
  usages. One principle that I think is
• 20:40
  interesting that we try to enforce is what we call the BI
• 20:45
  directional harness, meaning that we enforce us.
• 20:50
  A very simple rule is that any feature we build, it should be
• 20:53
  equally accessible by agents and humans.
• 20:56
  OK. There shouldn't be anything that
• 20:58
  a human can do in dust that an agent cannot do.
• 21:01
  And this vessel? And how does that play out in
• 21:05
  the time when you're actually building?
• 21:08
  So it, it, it kind of creates a constraint of the things you can
• 21:12
  build. As an example, I mentioned the
• 21:15
  tasks around the pod. This is a product surface that
• 21:19
  is kind of very human centric. Well, it's, it's, it's, it's,
• 21:24
  it's very important that those tasks are, can be listed by
• 21:27
  agents, can be edited by agents, can be created by agents, can be
• 21:31
  the what can be triggered. Another conversation can be
• 21:33
  triggered. The other session can be
• 21:34
  triggered from the stats by agents as well.
• 21:36
  And so it's really kind of a great principle to make sure
• 21:39
  that we, we don't overfit on any of those two users because we're
• 21:43
  really building a product for both those users.
• 21:47
  And we really care about that because we see it's the only way
• 21:51
  to start meshing the work of humans and agents together.
• 21:55
  In our opinion. Only if you have that kind of
• 21:57
  symmetry will you progressively stop making it fuzzy whether the
• 22:04
  work has been done by humans or an agents, which we think is the
• 22:07
  future and is exciting in many ways.
• 22:10
  Other things that there is so much, so much places where you
• 22:13
  really want to build something for humans.
• 22:15
  And when you try to apply and believe strongly in that
• 22:17
  principle, it really makes you rethink what it should look
• 22:20
  like. And so it's a useful principle
• 22:23
  to to apply. But I think most AI products
• 22:25
  today are kind of building for 2/2 users.
• 22:28
  And so it's important to build for them both.
• 22:30
  Yeah, it's a It reminds me of when someone told me you can
• 22:35
  create documentation that humans can read, but agents can't.
• 22:41
  But you can't do it vice versa. If you create documentation for
• 22:44
  agents, a human is going to understand it.
• 22:46
  So you should default to agents being able to understand your
• 22:51
  documentation first and foremost, and humans will by
• 22:56
  default also understand. Now, when you were talking about
• 23:00
  the idea of the pods and also how work gets done over longer
• 23:07
  time horizons, it made me realize, and I had the vision of
• 23:13
  data moving through like a DAG and how a finished piece of work
• 23:19
  is not dissimilar to that. You have different nodes where
• 23:25
  work gets done and then it goes to the next node.
• 23:28
  And whether that node and that piece of work is being done by a
• 23:33
  human or it's being done by an agent, it doesn't always have to
• 23:37
  be the same human that is doing it.
• 23:39
  And I think that's the huge unlock that I'm understanding
• 23:41
  with you is that a human is an expert in different parts of
• 23:49
  that process. So it shouldn't be the same
• 23:53
  human that owns it all the way through.
• 23:55
  We hear a lot of talk of how well now PMS and designers are
• 23:58
  shipping code, and that's great. But they have their unique skill
• 24:06
  set where they're very good at things.
• 24:08
  And isn't it almost more efficient if they can do what
• 24:14
  they're best at in that note of the DAG and then it gets shipped
• 24:20
  to someone who can make it battle hard.
• 24:22
  Maybe it's the dev OPS person and the security person.
• 24:25
  They can harden it and then it goes to production.
• 24:27
  And then you have that feedback loop where it will go back
• 24:30
  around and you're constantly iterating and updating.
• 24:33
  But those longer time horizon events.
• 24:37
  Now when you talk about multiplayer, I'm understanding
• 24:40
  what you're saying there. Yeah, I think, I think that's
• 24:45
  that's exactly that. And and and the grail is the
• 24:52
  grail is really try to create a place where it's really easy to
• 24:55
  end off. And so to really create an edge
• 24:58
  and an A next node and end off to the the the next processing
• 25:02
  unit. And the next processing might be
• 25:03
  humans, might be an agents. And it should be really easy to
• 25:06
  root towards one or the other and and and and and and, and,
• 25:11
  and that's, that's the thing. And, and to make that a reality,
• 25:15
  the, the complex, the, the, the, The Dirty bulk that needs to
• 25:19
  happen is that those agents has to be very contextful.
• 25:23
  And like it's super interesting into trying to think what is the
• 25:29
  right context for those agents to be useful agents of work.
• 25:34
  Is it just, I mean, today the basic answer is a long agency
• 25:39
  group and the gazillion MCP calls to rebuild the context in
• 25:43
  every session, the kind of a second stage and Sir is is turn
• 25:49
  everything into a skill and eventually the context will be
• 25:52
  rebuilt. Is that part automated?
• 25:55
  Is that part on autopilot? Is that part manual?
• 25:58
  These are all the kind of a question at the frontier of what
• 26:00
  it means to create that that right context for agents, which
• 26:03
  is quite interesting. Yeah, how are you seeing success
• 26:08
  with that and breaking down those barriers?
• 26:11
  So it is more fuzzy when you have the hand off and you can
• 26:14
  still. I am also understand there's
• 26:17
  moments where there's two different people, humans that
• 26:22
  are working on the same thing and then the agent in the
• 26:24
  middle. And so you almost have like this
• 26:26
  Venn diagram of you have the product manager, you have the
• 26:29
  engineer, you have the agents all looking at one node, if
• 26:34
  we're going to continue the metaphor of the DAG.
• 26:40
  I, I, it's, I think we are still, I don't have a crisp
• 26:44
  necessary answer to that. I think we are, again, I think
• 26:48
  we are still very early on, on this one because as we said it
• 26:52
  as, as we said at the very beginning, the capabilities of
• 26:54
  the agents are still kind of jagged.
• 26:58
  And so that makes that whole process not super smooth at all
• 27:02
  time, but we start seeing light at the end of the tunnel.
• 27:04
  And so our goal is really to discover that with our user to
• 27:07
  give them the right Lego bricks. That means that the, the, the,
• 27:11
  the within, within our user base, there's always going to be
• 27:14
  kind of a AI operators, tinkerers that will see the
• 27:19
  bricks and will want to assemble them in kind of new ways.
• 27:22
  And it's so it's super exciting to see what people do with the
• 27:25
  bricks. You give them sometime they come
• 27:28
  up with stuff that are just mind blowing.
• 27:30
  They just rebuild entire systems, all of that.
• 27:34
  And it's, it's exciting to see, But I think it's a, it's a Co
• 27:36
  discovery. It's a Co discovery thing.
• 27:38
  So it's a would be very, would be very.
• 27:41
  I mean, on top of that, it's so dependent on the capability of
• 27:45
  models which is moving so fast and and which is still kind of
• 27:48
  somewhat unpredictable in many ways.
• 27:51
  So it would be very adventurous of of me of trying to to have a
• 27:56
  very crisp answer on this one. Slant a flag right now and then
• 28:00
  it's like by the time this podcast airs.
• 28:03
  Exactly, exactly. Going back to the alignment, I
• 28:09
  kind of seated that like at the very beginning.
• 28:13
  Going back to the alignment, this is one of the major
• 28:15
  challenge I see into trying to reproduce the kind of conditions
• 28:19
  of Stripe early on. I think it's the alignment part
• 28:22
  because when you're operating in that space, the technological
• 28:25
  substrate is moving under your feet in a way that wasn't true
• 28:29
  before. For the past 20 years when we
• 28:31
  are building tech companies, you are building on JavaScript and
• 28:34
  Postgres was really stable. Like there's nothing more
• 28:39
  stable. I mean, those things evolve, but
• 28:41
  it's like stable, stable, concrete type stable.
• 28:46
  And today we're building on the on models that are changing
• 28:48
  every weeks. And so every time we try to
• 28:51
  build a very crisp picture of, of what we could be constructing
• 28:55
  the year, we'd got it completely wrong because within that time
• 28:59
  frame, the models shifted the place we wanted to build that
• 29:03
  nice city of yours kind of went undersea and the big mountain
• 29:06
  appeared right next to it. And so now you want to be on top
• 29:09
  of the mountain, obviously. And so that creates the that
• 29:13
  creates, that creates a real challenge for creating a crisp
• 29:17
  alignment of the team because you, you, you, you're subject to
• 29:20
  the fog of AI. And it's very hard to see past
• 29:24
  six months, let's say. I like this term, the fog of AI.
• 29:29
  The fog of AI, yeah, that's a, that's a, that's a pretty,
• 29:32
  that's one of an interesting challenge of building a building
• 29:34
  an AI company, I guess. And now just going back to dust
• 29:39
  and how you're building. Do you allow folks to interact
• 29:45
  and be multiplayer in all different types of scenarios
• 29:50
  like Slack, like in their web browser or in GitHub or they
• 29:55
  throw like a JIRA or linear issue at it?
• 29:59
  Is that kind of the vision where you're saying we have all these
• 30:01
  tentacles? Yeah, we try to build as much
• 30:04
  tentacles as as possible. The big one is obviously.
• 30:07
  So there's always A and there's always an interesting tension is
• 30:10
  that when you build one of those connection like Slack is one of
• 30:14
  the one of the big one. It's the very natural one.
• 30:16
  It's a chat system. And so that's the place where
• 30:20
  you might want to trigger work with agents.
• 30:24
  It is also an acquisition engine internally within work within
• 30:28
  companies because people are on Slack and they see other people
• 30:32
  using a system, so they discover it and they might start using
• 30:35
  it. Obviously the, the tension is
• 30:38
  that the, you're constrained by the interface of the platform
• 30:41
  you're connected to. And so the experience,
• 30:44
  everything we just said is that is very hard to do in Slag
• 30:47
  because you, you're bound by the interface.
• 30:50
  And Slag has been doing a lot of work there and, and they
• 30:53
  actually you can create pretty smooth interface today in Slag.
• 30:55
  But let's say GitHub as an example.
• 30:58
  It's just a text, text, I think. You don't see the syncing
• 31:03
  tokens, you don't see the tools being used, all of that.
• 31:06
  And so I, we really see those places as entry points more than
• 31:12
  place where the collaborations or the collaboration with agents
• 31:16
  really happens because you always want to, you always try
• 31:20
  to drive the user to go open the session inside of those because
• 31:23
  that's where all the richness happens.
• 31:26
  I like that. So it's the gateway into it and
• 31:29
  you're pushing them to go and look and inspect on dust so that
• 31:35
  you can get that richer experience.
• 31:36
  Yeah, some actions that agents take for the for obvious reasons
• 31:40
  of some admins mark as high stakes.
• 31:44
  So it means it requires a user confirmation before the action
• 31:47
  is taken. And a good example is sending an
• 31:49
  e-mail and. And so in Slack you can actually
• 31:54
  build a pretty good tool approval experience.
• 31:58
  But in GitHub you cannot. You can't send the next message
• 32:02
  on the issue. Say you need to click here to
• 32:04
  approve the tool. That wouldn't make sense.
• 32:06
  Well, I'm also thinking about even things like MCP apps that
• 32:11
  make the experience so much richer.
• 32:14
  You don't have that ability if I'm not mistaken.
• 32:17
  You could you could, you could imagine that Slack rebuilds
• 32:22
  supports MCP app and you managed to forward the MCP app payload
• 32:25
  to Slack. That could be, but again, GitHub
• 32:28
  wouldn't work, Zendesk, all all of that.
• 32:30
  Pretty complicated. Yeah, we're there's so many
• 32:34
  areas or surfaces that you touch that do not support that.
• 32:39
  And it would be very nice to have that richer data
• 32:43
  experience, like even at JIRA or Linear, you might want to see
• 32:48
  some of these data points or data exploration that happens
• 32:52
  with the agent before you commit to a quote UN quote Sprint,
• 32:57
  because I don't know what happens in sprints these days.
• 33:00
  It's an absolute mess. Maybe that's a thread we can
• 33:04
  pull on, on how sprints work now, but don't continuing on
• 33:08
  this MCP apps. You almost want an area where
• 33:12
  you can use all of the data that you're having happen in these
• 33:19
  different spots, like the slacks or the Jiras or the linears.
• 33:23
  And then put them in something to allow folks to have that
• 33:30
  richer experience. So I can see that being a huge
• 33:33
  value prop. Yeah.
• 33:35
  But we, we really see those platforms as I mean we try to
• 33:37
  insert ourselves in a discos system.
• 33:39
  We don't know, we don't want everything to happen on dust.
• 33:42
  We have walked some some amount of walk happen on dust, but we
• 33:47
  sometimes that walk starts from Slack, continues on dust and
• 33:50
  ends up in GitHub as a as an issue.
• 33:54
  And that's really fine. It's really important to be able
• 33:56
  to read from them. It's really important to be able
• 33:57
  to write back to them. I think that's the that's the at
• 34:01
  this opposition on this one. Yeah.
• 34:03
  And it goes back to the whole fuzziness of the handoffs and
• 34:06
  how those work right now and how that actually is is being done.
• 34:11
  Because you've got some folks that are working in Slack or
• 34:13
  you've got decisions being made in Slack and then an agent will
• 34:18
  execute on that decision that's being made, which then becomes a
• 34:21
  GitHub issue or APR. And then somebody else has to
• 34:26
  review it, but they are using an agent to review it.
• 34:30
  So you've, you've got all these different ways that it could be
• 34:37
  siloed or it probably is being siloed right now.
• 34:41
  And I know when we had our lunch and learns, one of the big
• 34:45
  things that folks asked for was how are teams working with
• 34:50
  coding agents? How are you figuring out which
• 34:53
  skills should be part of the whole company?
• 34:57
  Or should it be team wide skill? How are you making sure that
• 35:02
  sessions can be reproduced or shared?
• 35:04
  Or are you committing your whole chat history to the PR?
• 35:10
  Like there's these questions that are coming up and we don't
• 35:13
  have the tools for them right now.
• 35:15
  No, at least in the in the coding agent space, we don't
• 35:19
  have them yet. But that seems a I mean it seems
• 35:22
  pretty obvious that at some point your codecs dash dash
• 35:28
  session and that ID or cloud dash dash session and that ID
• 35:31
  will be hosted in the cloud and you'll be able to tie it to the
• 35:34
  PR and somebody will want to re restart from there.
• 35:38
  That seem pretty obvious. For.
• 35:41
  Them. Now I don't want to Passover
• 35:46
  this very big important thought of finding alignment is very
• 35:53
  tricky in the age of AI because you have the fog of AI, right?
• 35:57
  Yep. Are there things that you have
• 36:00
  done or seen work, or just like a principle that you're
• 36:05
  operating on in dust so that you can not necessarily predict the
• 36:10
  future but iterate quickly when it does happen?
• 36:13
  Yeah, exactly. I think, I think the the, the
• 36:16
  failure mode of that is to refuse to paint a picture.
• 36:22
  I think you have to operate with some form of conviction within
• 36:27
  uncertainty. It has to be clear to everyone
• 36:30
  that we are operating in a very uncertain environment and that
• 36:34
  whatever we say will be subject to change because the
• 36:37
  environment change, but that we create conviction on the actual
• 36:42
  current direction we're taking. And so that means people, I
• 36:48
  mean, everybody wants to rally on a vision or rally on a
• 36:51
  direction. And so if you cannot give a one
• 36:54
  year and two year direction, you have to give a six months
• 36:58
  direction, that's fine. And whenever you give a six
• 37:00
  months direction, by month 3, it's, it's going to be updated.
• 37:03
  And so that means that you, you, you need to have that kind of
• 37:06
  muscle of updating it and being, providing a lot of conviction
• 37:10
  in, in, in that process, but really being very explicit each
• 37:14
  time in doing it because it's wise.
• 37:17
  You might fall in the trap of, of not not creating that, that,
• 37:21
  that shared at least speed vector that you're trying to, to
• 37:26
  have for the group. And that's the worst, that's
• 37:29
  the, the, the, the worst part. Basically the longer the speed
• 37:32
  vector, the best it is, the shorter it's, it's worse.
• 37:38
  Three months is so that's we're mixing time with speed here, but
• 37:43
  it's fine. Three months, what would be what
• 37:45
  would be anyway, three months definitely swing starts.
• 37:48
  Three months is better than 0 because then you don't know the
• 37:51
  direction. Yeah.
• 37:52
  And I imagine how it plays out in practices.
• 37:55
  Every day folks are showing you new things saying hey, should we
• 37:59
  maybe go in this direction and you have to make those hard
• 38:03
  decisions of is it worth course correcting for this?
• 38:08
  Have we seen something that fundamentally changes where we
• 38:12
  want to go? Yep, yeah, this.
• 38:15
  So what we're trying to do is we're trying to really give a
• 38:18
  good picture of what we want to achieve within three to six
• 38:21
  months. But then we also, so that's,
• 38:24
  that's kind of a, a photo like it's a, it's a snapshots and we
• 38:30
  try to update it, but it's kind of at the horizon of big O of
• 38:34
  months. And then we, we have what we
• 38:37
  actually doing. And so we call it our stack
• 38:40
  ranks. That's pretty easy.
• 38:41
  It's a list of projects that we actively working on and we try
• 38:44
  to create it very clearly so that the rest of everybody
• 38:47
  knows. So engineering generally knows
• 38:49
  what they're working on, not always true, but most of the
• 38:52
  time it's better if true and enables everybody else within
• 38:56
  the company to know what we are changing on the product right
• 39:00
  now. And and then that's, that's
• 39:02
  stack rank. We really try to keep it very
• 39:05
  dynamic. So that means that we, we have a
• 39:08
  process which we call stack rank update that is open to anybody
• 39:11
  in the company to update the stack rank.
• 39:14
  And so obviously it gets windy at the top.
• 39:17
  So it's not an easy process. It's not a, that hard process
• 39:21
  either, but basically a button you can push and both founders
• 39:25
  appear in the room and we chat about how we should update the
• 39:27
  stack rank And everybody was in the company has access to that,
• 39:30
  to that. But generally we try to, we
• 39:33
  expect, we, we have a high expectation for the input to
• 39:36
  that process depending on the on the size of the, the kind of
• 39:41
  appetite that is traded with the updates.
• 39:44
  But but I think it's very nice because it gives a policy to
• 39:46
  everybody to be a bit of a product person within the
• 39:49
  company and we're trying to push everybody to be a bit of product
• 39:53
  person inside of a company. Let's change gears and talk
• 39:57
  about something. Before we hit record, we were
• 39:59
  mentioning tokenomics and I think that is.
• 40:04
  Very hot topic these days because the general narrative
• 40:08
  you hear from the Internet is, oh, the era of subsidies is over
• 40:13
  now folks are trying to go public.
• 40:15
  We're not going to get these cheap tokens.
• 40:17
  I'm not sure that I fully buy into that.
• 40:20
  Whenever I hear something from the masses, I that automatically
• 40:24
  gives me like red flag. However, I do see that the costs
• 40:31
  are going up a lot and it is really easy to spend a lot of
• 40:38
  money and not create anything of value.
• 40:41
  So there's like that. Is.
• 40:45
  I can burn a whole hell of a lot of tokens.
• 40:47
  That doesn't necessarily mean I'm good at my job, right?
• 40:51
  And so I do understand that you were mentioning there's the cost
• 40:54
  of tokens, but then there's also the cost of inference.
• 40:57
  And so let's maybe center the next 5-10 minutes around that.
• 41:03
  Well, I think the the way the way I we think, or at least I
• 41:07
  think about that is we know the end states this we are subject
• 41:11
  to gravity, this technology as a sailing, it will converge and
• 41:17
  when it does, it will be commoditized and tokens will be
• 41:20
  cheap and plentiful. That's the end state.
• 41:23
  The question is, and maybe we'll have a side by them, but
• 41:27
  whether, whatever is the state where it reached, it will
• 41:30
  plateau at some point because it, and there's a limited amount
• 41:34
  of energy and stuff, you know, and, and I think it when, when
• 41:37
  it plateaus, it'll get communitized and it'll be pretty
• 41:41
  cheap, very close to the price of, of, of power.
• 41:44
  Basically. The question is when that
• 41:47
  happens, maybe it opens in a in a year, maybe it happens in five
• 41:50
  years, maybe in 20 years, maybe it happens in 100 years, who
• 41:53
  knows. And so the question is, is
• 41:56
  trying to think about what's the transition state between now and
• 42:00
  then. But I guess the end state is
• 42:02
  plentifulness, cheap intelligence from from
• 42:05
  electricity. And then, well, the one thing
• 42:10
  that I've kind of was a new learning for me.
• 42:13
  So I'll, I'll share it because I think it was a learning for me
• 42:16
  in the past couple weeks months is that I've always supposed
• 42:21
  that if models performance plateau, we would enter a very
• 42:26
  rapid communication phase. And I think I've been coming
• 42:29
  back from that, that assumption because the the demand has
• 42:36
  increased so much. That's the pressure on
• 42:39
  inference. So just the the fact of serving
• 42:42
  the tokens is so high that the actors that have pre bought that
• 42:49
  inference capacity will be able to maintain high margin due to
• 42:54
  that pressure. Which means that at least my
• 42:58
  learning is that even if I suppose that today is the best
• 43:01
  we'll have a get in terms of models, which is probably not
• 43:03
  true. But even if I make that that
• 43:04
  sort of experiments, that doesn't mean that in six months
• 43:07
  the tokens will be close to 0. Because there is so much build
• 43:11
  up demands And some of and some of that demand is might not be
• 43:14
  the best demand because as you said, it's easy to to be token
• 43:18
  maxing on some pure AI slot bullshit.
• 43:21
  But but this the demand is still there and people are still ready
• 43:26
  to pay for it. And so there's a there's kind of
• 43:28
  a nozzle pressure in the market for around in France that will
• 43:31
  maintain the price of token high.
• 43:33
  But but anyway, it's very hard to to know.
• 43:38
  There's a there's also the other thing that you can save that is
• 43:43
  very easy to say that makes sense is we're getting in the
• 43:46
  world where does the those models on some tasks are PhD
• 43:50
  level, if not better. We need the PhD level to fill in
• 43:55
  Salesforce. No one.
• 43:59
  But the counter argument to that there is 2 is we need the
• 44:03
  MacBook M2 to do spreadsheets. No one but everybody wants 1
• 44:11
  because it's the best thing. And the most serious arguments
• 44:15
  is even if the model is APFC level, if you apply to task like
• 44:19
  feeling self force, you might gain 19 of of reliability.
• 44:24
  And whenever you gain 19 of reliability, some funky stuff,
• 44:27
  interesting stuff happen. You go from I look up the I look
• 44:31
  up what's going on. I'm checking the work to to I'm
• 44:36
  checking every to This thing is fully entirely automated and I
• 44:39
  will never ever in the story of humankind feeling a Salesforce
• 44:45
  card ever. And so I think that.
• 44:50
  And so I think there's there's still an argument for using the
• 44:52
  best frontier model even for tasks that seems pretty not so
• 44:56
  PhD level esque because of that kind of added nine of
• 45:00
  reliability on the task. Well, if the demand right now is
• 45:06
  the main constraint on what's making this plentiful, like you
• 45:11
  were saying, because 4.8 dropped recently and I've played around
• 45:18
  with it, I haven't seen that explosion of wow, this is so
• 45:21
  amazing and I need to use it all the time.
• 45:26
  There's a lot of talk about how Methos is going to drop and it's
• 45:29
  going to be a game changer, etcetera, etcetera.
• 45:32
  We'll see you when that happens. But assuming in this thought
• 45:35
  experiment that the models are what they are and we have this
• 45:40
  pent up demand, the question in my mind is how long does this
• 45:46
  pent up demand stick around? Does it grow over time?
• 45:50
  Does it diminish over time? Is it something that stays
• 45:54
  stable? And so we just need to keep
• 45:57
  bringing more energy, more inference online to service that
• 46:02
  demand. Well, I think there's a, there's
• 46:05
  a everybody, I get presumed everybody along the, the, the
• 46:08
  chain, the, the value chain will be happy to provide more offer,
• 46:14
  right? The, the ecstasy companies in
• 46:18
  every countries will be happy to build more nuclear plants.
• 46:21
  And then the, the cluster building companies will be happy
• 46:26
  to build new cluster in the place of old industrial places
• 46:30
  that are not used anymore. And the GPU providers mostly 1
• 46:35
  today, but probably more competition tomorrow, we'll be
• 46:38
  happy to provide the GPUs there. And so I mean, it will have
• 46:42
  economically, we'll have to accelerate and, and we know
• 46:46
  where the I mean, it's with those type of technology, it's
• 46:51
  always, it always equilibrates to towards commensuration.
• 46:55
  It's like your, it's like your, your, your mobile phone, your
• 47:00
  mobile phone service service. You, you pay 10 bucks, 20 bucks
• 47:06
  a month or maybe a bit more in the USI don't know.
• 47:09
  But it's, it's so cheap compared to the value it provides you.
• 47:13
  It's just because it's in fully commuted eyes and it's
• 47:15
  equilibrates to cost plus margin.
• 47:20
  It is fascinating to think about that.
• 47:22
  And right now though, where we're at living today, I imagine
• 47:28
  that's another piece that makes the ground underneath you very
• 47:31
  shaky. And so trying to build in going
• 47:35
  back to this metaphor of the fog of AI, knowing that the costs
• 47:40
  aren't really stable, locked in, reliable.
• 47:44
  And so then you're going out there and talking to users and
• 47:49
  you have your own product pricing models and all of that
• 47:53
  fun stuff. And then the users are already
• 47:56
  having their own thoughts about the tokens that they're spending
• 48:00
  and how they're spending it. That makes for some interesting
• 48:04
  conversations. Yeah, definitely.
• 48:06
  I think these are it's the fog of pricing.
• 48:10
  I guess this time around we interestingly add dust.
• 48:16
  We so we started with the with the flat price and the, the
• 48:22
  thinking was, was the following. The thinking was you want a flat
• 48:27
  price because you want to encourage usage and value
• 48:29
  creation. So you want kind of a Chinese
• 48:31
  Buffet type of pricing because they you pay once, can use as
• 48:36
  much as you want and let's go for usage, let's go for for for
• 48:40
  value creation. And it was also built on the
• 48:44
  assumption that eventually, even if the eventually was very
• 48:48
  loosely defined, eventually cost would go down.
• 48:52
  And as it happens with the emergence of the models, I mean
• 48:58
  with the awakening of the models at the end of last year, it is
• 49:04
  not sustainable anymore for us. And so we have to refactor our
• 49:07
  entire pricing to move to what is the industry send out today.
• 49:10
  And I think it's the only sustainable pricing unless you
• 49:13
  are willing to go extremely negative margins.
• 49:19
  You, you, you, you have to move to a crate based pricing because
• 49:23
  you don't control those costs. We don't know what's going to be
• 49:25
  made of tomorrow. We don't know what's maybe mitos
• 49:28
  will be really a banger and it will be maybe really expensive
• 49:33
  at the same time. And that's great.
• 49:35
  That's great if that's the case. That's great if you are in a
• 49:38
  crate based pricing wall because you're like people just go for
• 49:41
  it. Use meadows if you want.
• 49:43
  It's an awesome model. It's going to cost you a lot.
• 49:45
  But we I mean, it's fine if you are on a kind of a flat price
• 49:50
  like a typical SAS pricing you just you just like if a model
• 49:56
  drops, awesome, very powerful, but super expensive, just can't
• 50:00
  serve it. Just can't serve it because it's
• 50:02
  the the the economics don't work anymore.
• 50:05
  So. Oh, sorry, go.
• 50:07
  Ahead, yeah, no, sorry. So I think the uncertainty and
• 50:10
  kind of the time horizon for the evolution of pricing makes it a
• 50:15
  pretty much an unstable position to not have a crate based
• 50:18
  pricing in the market today I think.
• 50:21
  Yeah, it is very dangerous to go per seat.
• 50:25
  And then especially if you're abstracting away the cost of the
• 50:30
  providers, there's some creative things you can do like you were
• 50:33
  talking about earlier of using smaller models, but that at the
• 50:38
  end of the day is not going to work, especially if folks are
• 50:44
  wanting these the best. Yeah, exactly.
• 50:46
  Models all the time. For, I mean, maybe there is a
• 50:51
  set of task where I'm happy to use this model though, but for
• 50:55
  most of the stuff I do with agents I I just want the best,
• 50:58
  yeah? Well, and you don't.
• 51:01
  It almost takes more time to figure out where to insert the
• 51:05
  small model, which slows you down and you just, it's like,
• 51:10
  well, I know I could save a few cents here, but let's try and
• 51:15
  use it because it's going to be a lot easier for me then if you
• 51:21
  set up and, and that's one thing that has changed a lot recently
• 51:24
  too, is how we don't necessarily set up by hand each step of this
• 51:30
  graph. We let the agents, we kind of
• 51:32
  throw the problem at the agents and let them figure it out.
• 51:36
  And so if the agents have the ability to choose the smaller
• 51:41
  models or kick off tasks with smaller models, that's great.
• 51:45
  I'm not going to go in there and like use a smaller model for
• 51:49
  these specific tasks when you are going and doing them in the
• 51:52
  plan. You know, maybe there are people
• 51:55
  that do and those folks, wow, hats off to you for now.
• 52:00
  I imagine you could. Optimize micro micromanagement
• 52:04
  of and. You probably could get more lift
• 52:07
  or more cost savings just by optimizing certain context
• 52:11
  window tricks or the ability to make sure that your agents
• 52:18
  aren't constantly ingesting files.
• 52:21
  Maybe it's like cashing tricks or things like that that are
• 52:24
  going to be much, I'm not going to say like much cheaper, but
• 52:29
  they are going to potentially give you more bang for your
• 52:32
  buck. So there's there's all that fun
• 52:36
  stuff. Now I have one thing that has
• 52:39
  been going through my head that I've been wanting to ask you,
• 52:41
  and I'm sure you hear it every day that you talk to customers
• 52:45
  and it's well, yeah, cool dust. I see this vision, but can't I
• 52:50
  just do that with Co work? As we were talking about
• 52:53
  earlier, Co work is having its moment.
• 52:55
  What is the unlock on this? Like single player to
• 52:59
  multiplayer? I can share my slides with
• 53:02
  people or I can share my sessions with people from Co
• 53:06
  work. So where do I go with dust?
• 53:10
  Yeah, I think there's, so there's, there's three pillars
• 53:13
  to our different situation with cohort.
• 53:15
  It's so first on the on, on that multiplayer, there is some use
• 53:18
  cases that you can explain for, you know, as I did for the team
• 53:21
  weekly where I think you can paint a picture with you.
• 53:24
  You it's very hard to do with cohort that kind of a
• 53:28
  collaboration around a unit of work that needs multiple people.
• 53:32
  The, I mean, you can obviously do it, but it creates more, more
• 53:36
  kind of manual work and it's more, it's less integrated.
• 53:39
  I think there's in terms of enterprise, there's the
• 53:43
  governance part, which is very important.
• 53:47
  Having the ability to, we, we try to do a better job than code
• 53:52
  on that. Then be able to distribute in a
• 53:55
  control way the different MCP servers, different tools, the
• 53:59
  different skills limits the access to some data to some
• 54:03
  users have that things transitively managed so that's
• 54:06
  people can discover the the agents and the skills they can
• 54:09
  use. But in a way that's is aligned
• 54:11
  with other what are the restrictions of the workspace?
• 54:14
  I think cohort with all the things turning locally, the
• 54:18
  skills being shared through kind of out of out of bands to some
• 54:22
  extent as of today is a bit a bit more of a of a far West when
• 54:27
  you want to have kind of a slightly more governed
• 54:31
  deployment of AI with a new company.
• 54:33
  So that will talk to different type of company obviously.
• 54:36
  And finally, the third pillar, which is a pretty obvious 1 is
• 54:38
  like you don't want to get stuck in with one model provider.
• 54:43
  Especially as we're talking about tokenomics.
• 54:45
  Exactly. You don't want to be stuck with
• 54:47
  one token provider because the next best model you don't know
• 54:50
  where it's going to know where it's going from and you surely
• 54:52
  want to be have access to it. So I think it's kind of a mix of
• 54:56
  those three. We try to be very innovative on
• 54:59
  the products in terms of multiplayer AI and that's a
• 55:04
  differentiation. That's something that we hope
• 55:07
  will awaken and become stronger and stronger as we go.
• 55:11
  We want to be better at the governance and kind of the
• 55:14
  enterprise readiness as much as possible in at the product
• 55:19
  layer. Obviously we're not talking
• 55:21
  about we're not talking about FD ES trying to implement stuff in
• 55:25
  your company, but really as a product, be more enterprise
• 55:27
  ready. And then the sub pillar is that
• 55:29
  kind of a, is that kind of a, is that kind of a a freedom of of
• 55:33
  freedom of choosing your model and being able to test open
• 55:37
  source models and keep your finger on that and have all of
• 55:40
  that in one central place. So I think it's really that's
• 55:42
  kind of a those three factors that makes us still a, a
• 55:46
  compelling option in the time where everybody, everybody talks
• 55:51
  about cohort. But I'm I mean, we've lived
• 55:53
  through other hype micro hype cycles.
• 55:56
  We've had obviously tragedy. We've had we've rolled enough.
• 55:59
  We've we've rolled enough for for having lived as a company
• 56:02
  the emergence of tragedy Imagine and we've lived through the
• 56:07
  glean moments like it was all about glean 2 years ago.
• 56:11
  Glean what? Happened Glean.
• 56:12
  That's true, Glean. Well, I think it's still a very
• 56:15
  powerful enterprise search and AI answer products and I think
• 56:20
  they're doing actually really well.
• 56:21
  I think they're just they're just I've seen, I don't know if
• 56:24
  it's official or not, but I've seen on on newspapers, I don't
• 56:26
  know if it's true number, but they've reached 200 me in
• 56:29
  revenue. It's it's a very healthy growing
• 56:32
  company at that. At that time, it was the IT was
• 56:35
  really darling. Yeah.
• 56:36
  The be all to end all. I remember those days in 2023 or
• 56:40
  2024. Exactly, exactly.
• 56:42
  So there'll be there'll be other hype cycles, that's fine.
• 56:45
  I think we what we want to provide our users is really a
• 56:48
  platform where there is there is really building and and helping
• 56:53
  them being at the forefront of what can be done with with
• 56:56
  agents. So at the risk of jumping all
• 57:01
  over the place, Yep, file systems, Yep, you, we got into
• 57:08
  this earlier. Are you in a place where you
• 57:12
  feel like file systems can be updated for more agentic work
• 57:19
  And or, or add to this longer question, cause the way the
• 57:25
  agents work and the way that they will almost fan out and do
• 57:28
  things. Yep.
• 57:30
  And do things very compute intensive things and then just
• 57:34
  hang out for a minute and be like, hey, a human, I want I
• 57:38
  want your feedback. And maybe it's a minute, maybe
• 57:41
  it is an eternity in computer time.
• 57:44
  Are there other areas where you feel like the current
• 57:48
  infrastructure that you have to build dust, you need new tools?
• 57:57
  Well, on the file system, it's true that we can predict a
• 58:02
  moment where. So right now we discuss about
• 58:06
  it. We have a file system mounted on
• 58:07
  the pod, we have a file system mounted on the sandbox.
• 58:09
  Some of the file system in the pod are shared.
• 58:11
  You can have multiple agents working with that file system.
• 58:13
  As of today, it's mostly still human, collaborative driven.
• 58:17
  And so the kind of concurrency on that file system remains
• 58:22
  within typical human bounds. Yeah, but you could totally
• 58:25
  imagine like an army of agents cooperating on a project.
• 58:29
  And then the file system becomes the wrong abstraction.
• 58:34
  Oh, I mean, it's not necessarily the wrong abstraction, but
• 58:36
  there's going to be there's going to be a concurrency
• 58:38
  pressure on it. And it doesn't have any
• 58:40
  primitive for that. I mean it does file locking, but
• 58:46
  who uses file looking? Agents will.
• 58:50
  Agents will, agents might, but I think but to some extent this,
• 58:56
  to me this has been sold as there's a couple of companies
• 59:00
  that are building kind of a git infrastructure.
• 59:03
  Because if you want to scale, providing GitHub repository to
• 59:08
  agents using GitHub is not very convenient because there are
• 59:11
  rate limits that are meant for humans on the platform.
• 59:14
  And so there's a bunch of companies, they're pretty
• 59:16
  exciting that that builds Git infra.
• 59:20
  And Git infra seems to be a fair, fair solution to the to
• 59:25
  the problem because at least you have that resolution first
• 59:29
  resolution step when you merge back to Maine.
• 59:31
  That makes kind of some sense. But at the same time, yeah, the
• 59:37
  but I think the as was as was chatting.
• 59:41
  I think one, one step before that, that excites me a lot is
• 59:45
  the, is the kind of a stateful sandbox, the sandbox that stays
• 59:49
  live because I'm super excited about kind of a hyper
• 59:54
  specialized business apps. Many of our users are already
• 59:58
  doing that with those frames that we mentioned.
• 1:00:01
  But those frames, they're not backed, they're not backed by a
• 1:00:04
  particular sandbox. So they, they, they, they kind
• 1:00:07
  of, they kind of they're snapshots, which means that if
• 1:00:11
  you want to update it, you need to ask the agent to go fetch the
• 1:00:14
  data again and go update the frame.
• 1:00:16
  So what you really want is you want those frames to be backed
• 1:00:19
  by a sandbox that stays around. Because then you can have inside
• 1:00:22
  of a sandbox, you can have a SQLite database and you can
• 1:00:25
  rebuild the CRM and ERP. You can rebuild the whatever
• 1:00:29
  things you need to do your job. And the frame is the UI for the
• 1:00:34
  human. The sandbox is accessible by
• 1:00:37
  agents and so you have an API for you have a way to walk with
• 1:00:40
  agents or directly as a human. Every human can create a
• 1:00:43
  different view of the the app. If let's say we are talking
• 1:00:47
  about feedback system for the product where you automatically
• 1:00:53
  routes to different owners. And that's something that you
• 1:00:56
  cannot manage to do well in GitHub or linear.
• 1:00:59
  And so you build your own thing. Maybe the manager will want a
• 1:01:02
  different view than the IC, which will want a different view
• 1:01:04
  than the person that report the feedback.
• 1:01:07
  And that's having that habit is extremely exciting.
• 1:01:11
  And in that world, the kind of a de facto way to build that I
• 1:01:15
  think that is very easy is to is to do a shared nothing type of
• 1:01:19
  approach where the sandbox is the state and you just, you just
• 1:01:23
  let the agents do whatever they want, use whatever technology
• 1:01:27
  they want and you just Snapchat the sandbox.
• 1:01:29
  We can go back in time easily. And there is nothing more to it.
• 1:01:33
  There's no external database, there's no external Git
• 1:01:35
  repository. I think that will that's kind of
• 1:01:39
  a very nice convenient way to package those kind of things for
• 1:01:42
  agents. The only problem is like if you
• 1:01:44
  imagine that you have hundreds of agents collaborating on the
• 1:01:47
  shut sandbox, then it starts failing.
• 1:01:49
  And maybe then you need the codes to be managed by GitHub.
• 1:01:53
  Maybe you need the the data to be managed by your hosted
• 1:01:56
  database, which so we'll see. So it works until you get to a
• 1:02:01
  certain scale, which it doesn't feel like a lot of these
• 1:02:05
  workloads are at yet. Oh yeah, I mean it's so the
• 1:02:11
  moment we're so imagine you are you are a small SMB in whatever
• 1:02:18
  countryside of of Germany and you and you do cookies and
• 1:02:22
  you're paying your you're paying your ERP20K or 30 KA year to use
• 1:02:27
  three features of an ERP that has 10,000 features.
• 1:02:31
  That's that's the state of SAS today.
• 1:02:33
  And and so you may want to rebuild that, that that that
• 1:02:38
  process, which is get me the dog, get me the missing.
• 1:02:42
  And then here are the the comments.
• 1:02:43
  And it's pretty easy one inside of one of those kind of a
• 1:02:48
  specialized app. And so there indeed, there is no
• 1:02:53
  that's going to be a few people interacting with it.
• 1:02:56
  Maybe it's for agents, maybe directly through UIS.
• 1:02:59
  And I think having that that's that that shared nothing
• 1:03:01
  approach really works. The moment where it would fail
• 1:03:04
  is where you start growing and you have thousands of humans.
• 1:03:08
  If they interact through the UI that would the model would still
• 1:03:11
  work fine. The model would fail is like,
• 1:03:13
  you know, 10,000 people and and although you're 10,000, you're
• 1:03:18
  10 times bigger. And no, you need to make that
• 1:03:22
  system evolve and you have agents working with it and
• 1:03:25
  changing the codes at the speed of hundreds of agents at the
• 1:03:28
  same time. We can imagine that, Paul, but
• 1:03:31
  that's that's where it's going to fail.
• 1:03:32
  But it's so far away that's it seems like a fine to not over in
• 1:03:36
  sure for it. And even the current state of a
• 1:03:41
  large company, you don't necessarily have the majority of
• 1:03:45
  your projects being hit hundreds of times a minute by agents.
• 1:03:52
  So it's almost like I see a 208020 principle here where you
• 1:03:57
  could have that working for a lot of your different projects,
• 1:04:02
  but then those projects that really need scale and they need
• 1:04:09
  that type of limiting or unlimitedness to it, you would
• 1:04:17
  take those or you would architect those in a different
• 1:04:20
  way. Exactly, so that is stateful
• 1:04:28
  sandboxes. Are there any other hang ups
• 1:04:32
  that we might encounter if we tried to do them?
• 1:04:35
  It feels like it could get expensive.
• 1:04:37
  If I constantly have a sandbox running and I only use it once a
• 1:04:41
  week, it could be wasteful. Oh, that, that, that, that we
• 1:04:45
  have really great companies that gives you the nice privacy
• 1:04:47
  there. You you just shut them down.
• 1:04:49
  The time is boot up time is 100 milliseconds.
• 1:04:54
  You don't have if, if it's sure nothing.
• 1:04:56
  The beauty of that is that you don't have to care about where
• 1:04:57
  is the DB. Do you have to shut down the DB?
• 1:04:59
  No, everything is in the sandbox SQLA database.
• 1:05:02
  It's a file. It's 1 megabytes to do Most of
• 1:05:04
  any companies would do with anything serious.
• 1:05:07
  It just works. I think it's what I really see
• 1:05:09
  as the prolongation of the Excel spreadsheets.
• 1:05:12
  Today we have the Excel spreadsheets, No man's Lands and
• 1:05:16
  SAS. And it's kind of the, it's kind
• 1:05:19
  of a way to it's, it's, it's, it's a capture of a lot more
• 1:05:23
  than it can do with a spreadsheet that you were forced
• 1:05:26
  to do with SAS. And so that's funny because if
• 1:05:28
  you really believe in that vision, it feels like Wall
• 1:05:31
  Street was right way in advance for dangling the SAS businesses.
• 1:05:39
  And that's, it's pretty rare to think that Wall Street was
• 1:05:41
  really, really right in advance. But but that's yeah, that's the,
• 1:05:47
  the fun, the fun thoughts. But yeah, I think there's a
• 1:05:50
  massive opportunity to have anybody create the, the app they
• 1:05:54
  need to do their job that is equally accessible.
• 1:05:58
  The state is accessible by agents and the, the UI is
• 1:06:00
  actually can be create. They can be customized to every
• 1:06:02
  human basically. That's that's something very
• 1:06:05
  exciting over there. Yeah, especially if you're
• 1:06:09
  interacting with this through the chat ability or I, I guess
• 1:06:16
  where my mind gets caught up and where I've never fully bought
• 1:06:19
  into this generative UI principle, although I'm coming
• 1:06:23
  around to it a little more. The more and more that I use
• 1:06:27
  things like the HTML that the agents will create on the fly.
• 1:06:32
  It's like this is, this is actually kind of that idea of
• 1:06:35
  generative UI. But what I would get hung up on
• 1:06:41
  is that a lot of times we have learned and we are comfortable
• 1:06:46
  with certain tools because we know those five things that we
• 1:06:50
  need to do. We know how to do them and where
• 1:06:52
  to do them. If we now are being asked to
• 1:06:56
  create our own tool that does those five things and we have to
• 1:07:00
  spend the time creating it, even if it is going to give us a lot
• 1:07:04
  of success on the tail end because it's going to be much
• 1:07:07
  faster. It's going to be more custom.
• 1:07:09
  We don't have to spend as much money.
• 1:07:10
  It's a pretty hard ask to get get it set up.
• 1:07:16
  Yeah, but it's going to be as with any content economy ever,
• 1:07:20
  it's a 1.1% producer, 99% consumer, which is true of of of
• 1:07:25
  spreadsheets as well. There are people that knows that
• 1:07:30
  know how to do a great spreadsheets and they're not
• 1:07:32
  that many of them and they are the ones building a
• 1:07:35
  spreadsheets. And then there are the rest of
• 1:07:37
  us consuming, and there's the rest of us consuming consuming
• 1:07:40
  them. So you're thinking there will be
• 1:07:45
  the builders of these. It's just not.
• 1:07:49
  We're not all going to be expected to build it.
• 1:07:52
  There's going to be the specific builders and they're going to, I
• 1:07:55
  like that idea of extending the spreadsheet.
• 1:07:58
  You're now going to get this micro SAS that is very specific
• 1:08:03
  to your business. Yep, and anybody can build if
• 1:08:07
  they want. It's just that we know how it
• 1:08:10
  wet lens. It's like some will build, some
• 1:08:12
  will consume, and. Yeah.
• 1:08:15
  The the one place that I think that is for sure having it's
• 1:08:22
  questions is the no code low code type of tools because that
• 1:08:28
  is 100% in the crosshairs of what we're talking about.
• 1:08:33
  However, I did just see a tweet of somebody laughing, saying I
• 1:08:39
  am constantly surprised at how folks will set up with their
• 1:08:46
  agents a whole workflow to automate something that costs
• 1:08:52
  like $0.90 to run, when with Zapier you could do it for five
• 1:08:57
  cents, yeah. Yeah.
• 1:09:00
  So it goes back to that thing of like what it costs versus
• 1:09:03
  convenience type of thing. And it is much more convenient
• 1:09:06
  to explain what we want, especially if you're using some
• 1:09:09
  dictation tool, you just mind dump the agent organizes it and
• 1:09:13
  says, all right, let's try it. You test it, it works, it looks
• 1:09:16
  good. Ship it.
• 1:09:17
  Exactly, we'll see. It's hard to it's hard to
• 1:09:21
  predict where it's going to land, but it's surely there's
• 1:09:24
  truly a ton of ton of exciting stuff to explore.
• 1:09:27
  Agreed that the kind of on the fly dynamic UI, it's hard to see
• 1:09:32
  it fly right now just because of the latency.
• 1:09:34
  But here it's kind of slightly more.
• 1:09:37
  I think it's like you build ones use many times, but you can
• 1:09:41
  build very custom and agreed. That's all of that is doable
• 1:09:45
  with there was many no code platforms out there and it's
• 1:09:51
  it's very hard to understand why they why they didn't conquer the
• 1:09:56
  world. Yeah.
• 1:09:58
  Well, there there's also the other idea of just spitballing
• 1:10:03
  with you on product here because you're taking the approach of 1
• 1:10:09
  sandbox that will have different views that people can drop into
• 1:10:14
  and they can look at. However, what I wonder and, and
• 1:10:21
  I see it as the use case in my mind is you have a lot of
• 1:10:25
  customer calls. Like I'm sure you're on calls
• 1:10:27
  all day long talking with customers or talking with
• 1:10:29
  potential customers. And some of the time of that
• 1:10:34
  call or the same call can have value for a salesperson, it can
• 1:10:38
  have value for APM, it can have value for your engineering team.
• 1:10:42
  And so you want to take that transcript and you want to
• 1:10:44
  display the valuable parts to each party that's looking at it.
• 1:10:50
  And so going back to this analogy of, all right, we've got
• 1:10:54
  data and we want to display it to different stakeholders in
• 1:10:57
  different ways. Wouldn't it be, or is there a
• 1:11:01
  world where I can as APM have my sandbox and everything gets
• 1:11:07
  piped into that one place? Or maybe that is just like you
• 1:11:12
  don't need to do it at the sandbox level because you are
• 1:11:16
  doing it in all these different ways and you have something
• 1:11:19
  that's on top of all the sandboxes that's coagulating all
• 1:11:24
  of that data. Yeah.
• 1:11:26
  I mean some use cases will be leverable by just having the
• 1:11:30
  data and having agent just just process the data.
• 1:11:34
  There's a lot of that happening in dust today.
• 1:11:36
  Like we, we have a, it's a gun connector.
• 1:11:40
  And so companies that have connected to gun, they have all
• 1:11:43
  those gun calls that gets transcribed and recorded.
• 1:11:47
  And that enables interesting use case where indeed product teams
• 1:11:50
  can have launch an agent, do a deep dive on all of the customer
• 1:11:54
  calls for the past months. Tell me everything that they
• 1:11:57
  said about that feature or that thing.
• 1:11:59
  And it's also completely possible today.
• 1:12:04
  That is an exciting that that's something that's possible today.
• 1:12:07
  The kind of a micro SAS view of that would be you have a call,
• 1:12:15
  you dump this transcript. There's a it triggered in in a,
• 1:12:18
  in a, in AUI. It triggers an agent that will
• 1:12:21
  create stuff for the sales team, create stuff for a proc team,
• 1:12:25
  auto assign tags for the product surfaces that are concerned,
• 1:12:29
  auto assign tags for the geography.
• 1:12:32
  And maybe you have a way to say, OK, looks good and submit and
• 1:12:36
  then it gets stored in different places.
• 1:12:39
  I mean there are two ways of tackling the same problems.
• 1:12:42
  But I do think that like there is many cases, many use cases
• 1:12:50
  for kind of ADB backed apps that will that that could be
• 1:12:57
  customized for special way of doing work.
• 1:13:00
  I have I have many of my own add dust on that could change a
• 1:13:04
  little bit. The way we operate, we kind of,
• 1:13:06
  we can always kind of, you know, you, you bend the way you want
• 1:13:11
  to operate to the tools that, that, that you use, the way you
• 1:13:15
  share goals for your week or your days at, at dust happens in
• 1:13:21
  Slack. So it has a certain fun factor
• 1:13:24
  that is banded compared to the ideal state, the way we make
• 1:13:29
  decisions. We do that on GitHub, which is
• 1:13:31
  interestingly a fun place to do because it helps kind of a
• 1:13:34
  asynchronous collaboration. And sometimes it's done on Slack
• 1:13:37
  and it's bended in very different ways depending on the
• 1:13:39
  platform you use. And so there's so many processes
• 1:13:43
  and ways to ways to work that we that are shaped by our, by our
• 1:13:46
  tools and the other way around. And so that's what I find that's
• 1:13:49
  pretty exciting. So then you would have the
• 1:13:53
  creative control to do what you want and not be beholden to what
• 1:13:58
  the tools forced on you. I could see that and and the my
• 1:14:08
  imagination is going wild as I think through what you could do
• 1:14:12
  in those instances. I like this like database backed
• 1:14:17
  apps or just a sandbox with a database that's stateful is a
• 1:14:22
  very fun thing to ponder. Have you felt like the way that
• 1:14:27
  agents use databases, we need a new type of database or the way
• 1:14:35
  that agents are constantly going around and trying to collect
• 1:14:40
  data? And the reason I bring this up
• 1:14:42
  is because I was just reading that quote UN quote, old paper,
• 1:14:47
  I think it was from December 2025.
• 1:14:50
  It was like supporting our AI overlords.
• 1:14:54
  And it's written by, you know, some of the greats, I think
• 1:14:58
  Matteis Arya and Jan Stoika and Shreya Shankar.
• 1:15:03
  And so they, they mentioned how you have this, these agents and
• 1:15:08
  what they called it was like agent speculation.
• 1:15:11
  So the agents will go and they kind of have a theory and then
• 1:15:15
  they'll try and find the data to support that theory and execute
• 1:15:18
  on what they need to do. But a lot of times if they're
• 1:15:23
  trying to get data from a database, it's not that
• 1:15:27
  efficient to pull the whole row or create this join from these
• 1:15:33
  gigantic tables or whatever it may be that you're like, can we
• 1:15:39
  have agents interact with data in a more efficient way?
• 1:15:45
  Yeah. I mean there surely is many
• 1:15:48
  things to invent there, but I'm also, I'm a bit of a pragmatic
• 1:15:52
  system, those questions. And at the end of the day, when
• 1:15:55
  you think about that technology, it's trained on the Internet,
• 1:15:57
  it's trained from reinforcement learning, from human feedback.
• 1:16:01
  It's trained on, on, on human traces being built by many labs
• 1:16:06
  extra. And so my point is that it's a
• 1:16:08
  technology that is extremely, extremely, extremely
• 1:16:11
  anthropomorphic. And so it and it feels it, it
• 1:16:16
  actually we had that weird moment where we needed rag.
• 1:16:21
  Yeah. And it was kind of a new way of
• 1:16:24
  presenting the data made for agents, made for the constraint
• 1:16:27
  of the context size that was rather small.
• 1:16:31
  And with the context size augmenting and the the very
• 1:16:34
  entropometric nature of that technology.
• 1:16:35
  The what's the best way to do search in company data today is
• 1:16:40
  just let's do what people call agentic search, which is
• 1:16:43
  basically just let's use the tool that the human use.
• 1:16:46
  If we can do it, they can do it as well.
• 1:16:48
  And we're ready to pay the tax of the latency shaded because
• 1:16:52
  it's just simpler to just give them the tool and let them let
• 1:16:55
  them do the work. And so I'm, I think I'm, it
• 1:16:59
  feels like as the context is getting longer, the, the, the,
• 1:17:02
  the, the, the way those, those, those agents use tools is
• 1:17:06
  looking much more. I mean, it's looking very much
• 1:17:09
  like we use tools. And so if we don't have the need
• 1:17:13
  in the past, why would they? Interesting question.
• 1:17:16
  And they've been RLD on what the humans will expect of it.
• 1:17:20
  So if you, and they've been railed on all these traditional
• 1:17:25
  tools also. So if you're now creating a new
• 1:17:29
  tool that you are expecting the agents to use, well, it may take
• 1:17:35
  a few cycles. I I'm fascinated by that space
• 1:17:38
  just because I feel like you've got this whole new way of
• 1:17:43
  interacting with machines through agents and the
• 1:17:47
  infrastructure that we have. It's fun to think about.
• 1:17:51
  Is it good enough? Can we do like you were saying,
• 1:17:54
  can we explore and not be beholden to what tech currently
• 1:18:00
  constrains us to? Or is there, like the other
• 1:18:04
  piece, the other side of the coin that you're saying, be more
• 1:18:07
  of a pragmatist and say, look, I think if we can use it this way,
• 1:18:13
  the agents can use it this way? It's it's obviously going to be
• 1:18:18
  always a bit of a bit of a bit of a bit of both.
• 1:18:21
  But even in the way we use tool, we use them very differently.
• 1:18:26
  It's like we use UIS, they use AP is turned MCP, turned CLI and
• 1:18:33
  eventually it's the CLI. And so I would love to use every
• 1:18:37
  SAS in the world with the CLI myself.
• 1:18:39
  So in the sense at the end of the day we merge back into the
• 1:18:43
  most efficient place. I might use, I might use the the
• 1:18:47
  CLI that agents have been that have been built for agents for
• 1:18:50
  using using the SAS tool myself. I'm a big of a fan of the
• 1:18:54
  terminal, but. I just saw a tool, I can't
• 1:18:57
  remember what it was. It was like turn any SAS into
• 1:19:00
  ACLI. That was their whole tool.
• 1:19:02
  I can't remember what it was, but that is.
• 1:19:07
  It's exciting to me, yeah. It's yeah, it's a different
• 1:19:14
  level that we're playing at. Well, man, this has been great.
• 1:19:18
  Is there anything that you want to mention before we jump?
• 1:19:21
  Is there anything that I didn't ask you about?
• 1:19:23
  No, I think this was great. We covered, we covered the the
• 1:19:26
  main subject. I think we on the verge of going
• 1:19:28
  towards multiplayer AI. We don't quite know yet what it
• 1:19:31
  look like, but surely won't look like what we're doing today.
• 1:19:34
  And so that's, that's why I, I think it's in the, in that wall
• 1:19:38
  where there are two black holes being created in the market.
• 1:19:43
  I think it's always an interesting questions of are
• 1:19:47
  they not going to going to take it all?
• 1:19:50
  We, we, we, we really attached to the, to the, to the thinking
• 1:19:54
  that even if the technology again, plateaus, even with
• 1:19:58
  current model shape, the way we're going to walk in 10 years
• 1:20:02
  is probably nothing compared to the way we work today.
• 1:20:04
  And so there's many, many, many opportunities to build for for
• 1:20:07
  everyone. I will also finish with dust is
• 1:20:12
  known throughout the world for having the absolute best in
• 1:20:16
  person event. So if you see a dust event
• 1:20:19
  happening near you, I highly recommend go for it.
• 1:20:23
  Yeah that is for sure. They are awesome events.
• 1:20:27
  Anybody that has been to one will tell you the same.
• 1:20:30
  So thanks for doing this dude. I really thank.
• 1:20:32
  You very much. It was great.