Latest / Agentic Conversations (formally mlops.community) / Fireside Chat with LLM Startups // LLMs in Production Conference 3

Transcript

Fireside Chat with LLM Startups // LLMs in Production Conference 3 · Agentic Conversations (formally mlops.community)

• 0:04
  Keynote #2. This is exciting because we've
• 0:08
  got a very special panel coming up and I'm going to invite my
• 0:13
  friend Paul back onto the stage. This is not his first time on
• 0:16
  the stage. Awesome talk earlier and now I'm
• 0:19
  going to hand it over to you for this next panel.
• 0:25
  For today, what we wanted to do, as many of you are out there are
• 0:30
  are you know, building products, some of you might be building
• 0:32
  companies and we get the pleasure to talk to many
• 0:37
  companies out there, the founders that are building
• 0:39
  interesting companies in this new space.
• 0:41
  We thought to actually bring some of the ones that we've
• 0:43
  spoken to recently together to discuss this.
• 0:46
  And joining me today to moderate this panel before I introduce
• 0:49
  the actual panellist is Sandeep, the joining us from London, I
• 0:53
  think. Sandeep, can you hear me?
• 0:55
  Yep. Yeah.
• 0:57
  Hey, Sandeep, thanks for joining us.
• 1:00
  Great to see you. So Sandeep, you had our
• 1:04
  investments team in Europe and also look at the at the at North
• 1:09
  America. We get the pleasure of working
• 1:11
  together a lot and and talking to interesting founders out
• 1:14
  there. So thanks for for joining us.
• 1:16
  Do you want to maybe add something?
• 1:18
  Today, what's on your mind? Ah, no, look extremely proud and
• 1:23
  privileged to be part of this, this, this amazing session that
• 1:26
  you've put together. And yeah, we work in close
• 1:29
  conjunction and and wanted to share you know some of the
• 1:33
  really exciting companies that we have the the pleasure of
• 1:35
  speaking with on on nearly a day-to-day basis.
• 1:38
  And hopefully the audience enjoys hearing from them.
• 1:42
  So perhaps we'll pass it over to the people that the people that
• 1:45
  are here to see. Excellent.
• 1:47
  So let's do that. So we we we asked the three
• 1:50
  founders to join us. These are technical founders
• 1:54
  that. It had basically chosen to found
• 1:57
  their own companies in this space, which is probably one of
• 1:59
  the most exciting spaces to be building your company in.
• 2:03
  And so let me just introduce them one by one.
• 2:07
  So I think I see here Pietro Darlington from from Canada.
• 2:12
  Good morning to you. Hey, good morning, Paul.
• 2:15
  Good morning, everybody. Great to see you again.
• 2:18
  Pietro, do you want to maybe describe what you're building at
• 2:21
  transitional forms and the new type of media you've you've been
• 2:26
  you know working on? And also maybe for the folks who
• 2:29
  are listening, tell us a little bit about the stage you're in
• 2:31
  with the company and and sort of the size.
• 2:35
  Yeah, for sure. So we started as a creative
• 2:39
  machine intelligence company back in 2018.
• 2:43
  My background is in entertainment innovation.
• 2:45
  So I came from a company that I had sold to Entertainment One.
• 2:50
  And so I had this, you know, foresight that different AI
• 2:54
  generators were going to do all these amazing things and unlock
• 2:58
  human imagination for us. And so I started a company
• 3:01
  called Transitional Forms a number of years ago and we've
• 3:04
  we've done a bunch of experimentation and different
• 3:06
  types of AI to create you know. Rate of content from
• 3:11
  reinforcement learning to LM to Music Transformers and we're
• 3:15
  about 10 tenish people now on any given day where all
• 3:19
  different hats all the time. And we've created a TV platform
• 3:23
  that allows you to generate shows, but TV shows by a click
• 3:27
  of a button. Awesome.
• 3:29
  Thanks Peter. Thanks for the introduction.
• 3:31
  I think you also shared a video that sort of illustrates some of
• 3:35
  the work that you guys have been doing.
• 3:36
  So maybe we can play that video to bring bring this the some of
• 3:40
  the stuff you're building to life in a visual way.
• 3:44
  Attention fellow content creators, here's a sneak peek of
• 3:48
  our new Maker app where real time AI generated TV is just a
• 3:52
  click away. Choose from 2D or 3D templates
• 3:56
  and simply edit your titles and premise prompts.
• 3:59
  Or customise further with story, beats, music, themes and sound
• 4:04
  effects. And camera philtres.
• 4:08
  You can even swap entire sets in a matter of seconds, customise
• 4:13
  your cast and even change their back stories for more dynamic
• 4:17
  results. To prove the power of this
• 4:20
  patent pending tech, we developed a live stream
• 4:23
  generative variety show called Robots Make TV.
• 4:26
  Robots. Remember.
• 4:29
  That if it's not broken, you must break.
• 4:32
  It Someday you'll be an alien, but only if someone farts on you
• 4:35
  in our space. You both suck lol, and these
• 4:39
  outputs are all real time, fully autonomous and never the same
• 4:43
  twice. Because no reality is too weird.
• 4:48
  Connect with us to bring your story worlds to life.
• 4:52
  Great. Thanks for sharing that theatre.
• 4:54
  I remember when we spoke to the first time thought that came to
• 4:56
  mind was you know is this what Pixar looked like in the early
• 5:00
  days this sort of new you know way of of of creating
• 5:03
  entertainment, this interactive movie making.
• 5:05
  So really great to to to share that Maybe then over to to Lars
• 5:11
  who's the CTO and cofounder at Cordy dot AI Large.
• 5:14
  Thanks for joining us. I think from New York if I'm not
• 5:17
  mistaken, maybe you want to tell us a little bit about the
• 5:20
  problem you're solving at Cordy Dot AI and and also where you
• 5:23
  are with the company which stage and and and just to give the
• 5:27
  folks. Idea.
• 5:29
  Yeah, of course. Thanks so much for inviting me.
• 5:32
  So my background is actually in terms of models.
• 5:35
  I've been researching and generative models and I've been
• 5:37
  doing a lot within that field and I've been for many years
• 5:40
  trying to explain to people why it's so cool and then suddenly
• 5:44
  outburst a lot of generative AI and now it's called Genie AI and
• 5:47
  everybody knows why it's cool. So that's amazing.
• 5:50
  And basically what we've tried to emphasise with our company is
• 5:52
  that then we want to to solve the big pain that is the medical
• 5:56
  conversation. So when someone is having a
• 5:58
  conversation with a doctor and nurse paramedic and they really
• 6:01
  wanted to want to be understood and that conversation.
• 6:04
  Be difficult to be understood because not everyone knows how
• 6:07
  to actually explain their symptoms in the right way and so
• 6:10
  forth. And if if you're misunderstood
• 6:13
  then you can end up in a dire situation.
• 6:16
  We started out with the very acute field, so the number one
• 6:18
  call when someone calls them one number in the US or 112 around
• 6:22
  mainland Europe and then they're in dire need of health and to to
• 6:26
  get some assistance from the person, the paramedic that is
• 6:29
  receiving that call can be difficult.
• 6:31
  And so we have built this AI that taps into that medical
• 6:34
  consultation, that medical conversation and assist the
• 6:37
  healthcare professional in making the right choice.
• 6:40
  In in a matter of in a in a matter of seconds.
• 6:43
  So the I listen again is understanding what is being said
• 6:45
  and then it's giving prompts to the healthcare professionals for
• 6:47
  the healthcare professional to to work, work their magic for
• 6:51
  the patient and for the better of the patient.
• 6:54
  We are right now in the company, approximately 100 people which
• 6:56
  just raised our Series B where process is one of the leading
• 6:59
  investors in that round and then very happy to join this.
• 7:03
  Ohh, great. Thanks for letting us be part of
• 7:05
  the journey. And I remember when when we
• 7:07
  first spoke, Lars asked you, so you know what's your background,
• 7:11
  Can you send some material I've been working on and you send a
• 7:13
  couple of publications also at nerfs, if I remember.
• 7:15
  And you know that's that's the kind of founder.
• 7:19
  I think we, we, we enjoy working with the deep tech background
• 7:22
  and I think not only that for yourself, an incredibly
• 7:25
  important problem for for all of us.
• 7:27
  So we're rooting for you. And then maybe an introduction,
• 7:31
  last introduction. Josh, thanks for joining us.
• 7:34
  I think you're, if I'm not mistaken, in that Philadelphia,
• 7:37
  is that correct? San Francisco, right?
• 7:40
  San Francisco Alright, even further away.
• 7:43
  So thanks for joining early in your day.
• 7:46
  Maybe the same question to you. Yeah.
• 7:48
  So what problem are you solving and where are you at the stage
• 7:51
  of of Martian at the moment? Yeah, we're trying to make using
• 7:57
  LM more like using electricity, right?
• 7:59
  Right now there's many different LE providers.
• 8:02
  It's hard to figure out which one should you use.
• 8:04
  Most people default to GPD 4, which is the most expensive one.
• 8:07
  We want to make it as easy as plugging your your application
• 8:11
  into an LM, just as you would plug a electrical socket into an
• 8:16
  outlet, right? Such that you don't have to
• 8:18
  worry about where you're getting your intelligence from, just
• 8:21
  you're getting enough of it as cheaply as possible.
• 8:23
  We're already able to do this today.
• 8:25
  In fact, on Open AI's own E valves were able to beat GBT.
• 8:29
  More level performance at a lower cost by routing your
• 8:33
  request to the right model, right?
• 8:35
  So you give us your request, we find the best model to run it
• 8:38
  on. Best meaning the highest level
• 8:40
  of performance at the lowest cost, and then we give you the
• 8:42
  result. Yeah, that that's incredible.
• 8:45
  And also there, you know, I remember when I heard about what
• 8:49
  you guys are doing at Martin and we talked about what we're
• 8:51
  building and inside the group, one of the things being plus
• 8:53
  one, the first thing I thought was I want to use what you guys
• 8:56
  are building. We need something like this.
• 8:58
  So again, very excited to to have this conversation.
• 9:03
  So maybe, Sandeep, I'll let you do the honours of the first
• 9:07
  question of. Course, of course.
• 9:09
  So look, thanks so much for for sharing the introduction.
• 9:12
  And we heard about the myriad use cases that each one of you
• 9:15
  are solving through through through your company.
• 9:18
  But perhaps just to start out, Pietro.
• 9:22
  Would love to understand, you know in the the creative content
• 9:25
  space who do you believe to be the incumbent in the space if
• 9:28
  anybody and why do you feel like you're best position to to to
• 9:33
  solve this problem. It's one of the one of the
• 9:36
  common questions that we ask in the in the venture world.
• 9:38
  So would love to hear your perspective.
• 9:40
  There are so you know there there are incumbents and and and
• 9:45
  various forms of competition. We see generative creativity
• 9:49
  emerging all day that seems like every day there's a new tool you
• 9:53
  know runway ML and and mid journey and stable diffusion and
• 9:56
  all of these amazing tools that we use these tools as well in in
• 10:00
  our pipeline. But at the same time, and this
• 10:03
  isn't a good thing, there's not a lot of people that are taking
• 10:07
  exactly our approach. You know I'm a I'm an old VR
• 10:12
  pioneer umm. So real time interactive
• 10:16
  immersive possibilities for immersion have been very
• 10:19
  important to me from the beginning.
• 10:20
  So versus these like hard ML approaches with with gangs and
• 10:26
  diffusion models and and whatnot, we're going with a real
• 10:29
  time approach in in a game engine and using Unity.
• 10:33
  So that's our that's our angle right now and and there's not a
• 10:37
  tonne of competition like I said that's not necessarily a good
• 10:39
  thing. You kind of need competition to
• 10:41
  prove the future of what you're trying to build.
• 10:44
  But yeah, that, I think that's that's our angle.
• 10:47
  Super interesting and I might, I might pick it up then with with
• 10:51
  Yash, obviously. I'm sure there's a number of
• 10:54
  people in the audience have tried to change utility
• 10:56
  providers, be it electricity or Internet and whatnot, and trying
• 10:59
  to figure out how to get the best performance for the lowest
• 11:01
  cost. Who do you believe to be the
• 11:03
  incumbents in your space and and and how are you thinking about
• 11:05
  your positioning? Yeah.
• 11:08
  So right now most people when you use an L, we're going to be
• 11:11
  using a single model, right. And so it's, it's kind of
• 11:15
  getting people to switch away from using that single model to
• 11:18
  having a kind of multi model architecture where they're able
• 11:21
  to switch between many different providers.
• 11:23
  We can actually do routing for them, right.
• 11:26
  I think the reason that you know you sort of your, your
• 11:28
  incumbent, you're incumbents here for these LM providers are
• 11:33
  not going to get the best level of performance is because
• 11:35
  they're going to focus on on their own models.
• 11:37
  They're never going to route you away from their model to one of
• 11:41
  their competitors. Right.
• 11:42
  And so if their competitor is doing something better, you're
• 11:45
  never going to see that if you use only their platform.
• 11:48
  There and and makes makes a tremendous amount of sense and
• 11:52
  and last but not least Lars, you know the the medical space.
• 11:56
  How should we think about, you know, the the the medical space
• 11:59
  solving this critical problem? How should we think about the
• 12:01
  incumbents in your space? But I could talk for hours about
• 12:06
  this but but but I'll try to keep it brief and to the point
• 12:09
  here. I think that the what we're
• 12:10
  seeing a lot, we're seeing a lot of companies that are trying to
• 12:13
  to to embark in the healthcare space.
• 12:15
  It's a difficult space to be in. So I'll I'll stick to the point
• 12:18
  of the ones that we see out there when we go out to the
• 12:20
  customers right And and I hope you catch my drift in terms of
• 12:24
  that right, because there is a big difference here and the ones
• 12:27
  that we see out there are are some of the bigger tech
• 12:30
  companies. I think that one of the ones
• 12:32
  that I want to emphasise is is Nuance that has been there for a
• 12:35
  long time. There is a dictation service
• 12:38
  normally. They have also gone further than
• 12:40
  that in order to respect, but are mostly known for the
• 12:43
  dictation software for doctors etcetera and that's the ones
• 12:46
  that we're competing the most with what we're we're having a
• 12:50
  slightly different route towards providing value to the client
• 12:53
  where we are going a little bit higher up the value chain in
• 12:56
  terms of predicting outcomes for these conversations.
• 12:58
  So we're actually also having deployments where we are
• 13:01
  deployed next to new ones. So it's, it's, it's, it's a
• 13:04
  finicky space in terms of that and then that's how we try to
• 13:07
  get in and we try to be smart on manoeuvring our way into the
• 13:10
  client. Awesome.
• 13:13
  Yeah, I recall when we were spending time, I think the one
• 13:16
  thing that we we also found when we were learning more about
• 13:18
  Cordius is that for the doctors in the medical space, pen and
• 13:21
  paper still speeds and seems to be the predominant one.
• 13:24
  But, but Paul, let me pass it back to you for for the second
• 13:27
  question. Yes.
• 13:28
  So you know for let's say all of the change we've seen in the in
• 13:33
  this space, companies that are labelling themselves as a I
• 13:37
  companies. There's a big difference between
• 13:38
  the companies that when we say it comes they're really building
• 13:41
  their own AI core technology. Could be models, could be data
• 13:45
  sets, could be processes. Whatever.
• 13:47
  And those that aren't. And of course you know the
• 13:49
  reason you're here is because we believe you've built some of
• 13:52
  those core elements that make you different or set you apart.
• 13:57
  I'd love to sort of understand maybe in your view what you
• 14:00
  know, what are the things that that set you apart because
• 14:04
  you've built some AI that without that you know your your
• 14:07
  company wouldn't be able to do what it what it promises to do,
• 14:11
  maybe starting with with Pietro. Sure.
• 14:16
  Yeah. You know from the beginning we
• 14:19
  we've never really labelled ourselves as an AI company.
• 14:22
  We we've, we've positioned ourselves as an entertainment
• 14:26
  innovation company using AI as the future of of of to to you
• 14:31
  know unlock these new formats that are coming.
• 14:33
  So we we went through, as Paul knows, we went through a big
• 14:37
  life cycle of prototyping using all sorts of different we didn't
• 14:40
  know where things were going to land, you know using music
• 14:43
  Transformers and reinforcement learning we put into a game and
• 14:47
  a film, large language models, etcetera.
• 14:50
  So we don't we don't build core AI ourselves but we were the
• 14:56
  glue that holds it together and translates from one generator to
• 15:00
  the next. It's like you know it's art
• 15:03
  meets science. We're using the science to
• 15:05
  generate the art. I think that's a Pixar quote.
• 15:07
  Even letting the art. No letting the tech inspire the
• 15:11
  art and the art challenge the tech.
• 15:14
  So yeah, that's that's been our approach is kind of like
• 15:16
  creating that entertainment Lego using different different models
• 15:20
  and and that keeps us fresh too because there's always something
• 15:22
  new. Yeah.
• 15:25
  And I think to some extent when you say you're not building Cory
• 15:27
  I technology, but I remember Inter interacting with your
• 15:30
  interactive assets whether it's games or the director studio.
• 15:34
  It's very clear that you're you know, the way you connect these
• 15:37
  large language models through the Unity engine and then bring
• 15:39
  it to live in a visual way is part of your your AI assets,
• 15:45
  right. It might be if somebody says not
• 15:47
  your own model, but I think that entire framework and making that
• 15:50
  come alive in a real time way, right.
• 15:51
  I mean you can see some of your things live on twitch as you
• 15:55
  make these movies and people interact during bananas and so
• 15:57
  on. I think that you're.
• 16:00
  Yeah, yeah, I think you're understanding the AI.
• 16:02
  Is this the stuff, you guys? Well, thank you.
• 16:04
  I should plug to our our reinforcement learning game
• 16:07
  where we're letting people train neural networks right within the
• 16:09
  game itself. That's a world first.
• 16:11
  So, yeah, we, we, we certainly are creating core AI, but just
• 16:14
  maybe a different approach than than some of these other AI
• 16:17
  companies. Thanks, Jethro and Lars, how
• 16:23
  would you think about your sort of core AI components for that
• 16:28
  you know, that make the way you solve the problem at Corey
• 16:31
  different from others? Yes.
• 16:34
  So, so we need to go through the entire stack from from
• 16:37
  understanding what's being said inside the audio and to yeah,
• 16:42
  speech representation learning, speech recognition learning,
• 16:44
  then natural language processing and then clinical predictions.
• 16:48
  That's basically what we have built upon since the beginning
• 16:50
  of our of our journey here. And we'll build it from a
• 16:53
  process of saying hey, we need to be in the frontier of
• 16:55
  science. So we need to also take whatever
• 16:58
  paper is out there and replicate that model inside our own stack
• 17:01
  and reproduce the results of that of those models and go from
• 17:06
  an outline. We we call it, internally we
• 17:07
  call it. A kill your darling strategy
• 17:09
  here. So, so as soon as someone is
• 17:11
  coming up with a model like when the transformer was coming out
• 17:13
  in 2018, we found out that we need to kill some of our LSTM.
• 17:17
  Maybe they're coming back again but that's a different more
• 17:19
  technical discussion. But then but so so basically we
• 17:23
  have tried to learn as much as possible from the research
• 17:25
  environment and then try to put some of these components inside
• 17:29
  our own modelling paradigm. And then so our core I basically
• 17:33
  the entire value chain from the audio all the way to the
• 17:35
  clinical predictions and be able to do that in real time.
• 17:38
  That's quite difficult because audio is difficult and we do
• 17:41
  that in like predictions in less than a second.
• 17:43
  And we do that with quite complex input sources and also
• 17:46
  prediction sources. Great.
• 17:48
  And maybe let me just before I, I, I, I go to to to Yash.
• 17:53
  One thing that I think you know we had discussed Lars on this
• 17:56
  front is yes of course you've got because of your use case the
• 17:59
  real time need to be able to do this and you know emergency
• 18:02
  centres and so on to do the transcription and the coding and
• 18:05
  support. So that's one thing that you
• 18:08
  guys really need to have built in house to make sure that
• 18:10
  works. And we also looked at, OK,
• 18:12
  what's the, what are the implications of some of these
• 18:15
  large powerful models that are coming out, let's say GP4 that
• 18:18
  you could actually ask through the tasks you guys are you know
• 18:21
  building in your into your commercial products.
• 18:23
  Can you comment a little bit because you shared some of the
• 18:25
  valuations and we went back and forth on that, how you see your
• 18:28
  ability to within that specific task within your domain based on
• 18:31
  the data sets you have to sort of build something that is more
• 18:35
  reliable, safer and higher quality for that specific use
• 18:37
  case? Yeah, I think that there's a
• 18:40
  perfect site way to what, Yeah, she's building, right.
• 18:42
  So if we had your your tool a little bit before then that then
• 18:47
  we could have used life like quite, quite heavily.
• 18:50
  But we've done a lot of that plug.
• 18:53
  In a lot of those outlets to use the arts and culture and in
• 18:56
  house, we were for believers in these more vertical models
• 18:59
  especially for virtualized fields like ours because we can
• 19:02
  get a lot out of that. You can see some of the papers
• 19:05
  in terms of not supporting too much on that, but some of the
• 19:07
  papers on being able to answer multiple choice, medical QA
• 19:11
  questions. When the markets really good on
• 19:14
  that, they decline on some, some other things.
• 19:16
  And that's basically our finding as well on our domain is that we
• 19:19
  can actually verticalized these models and then they become
• 19:21
  really, really good at a cost of something else, right.
• 19:24
  We know that there will never be a pizza delivery called coming
• 19:26
  into our services and. As it comes in then we don't
• 19:29
  need to help on that, right. And so, so basically we can, we
• 19:33
  can, we can specialise these multiplied heavily.
• 19:35
  And what we basically do have been doing is trying to take a
• 19:37
  lot of the limbs out there, specialising them, fine tune
• 19:41
  them, restructuring them and so forth.
• 19:43
  And also restructuring them in terms of the input modalities
• 19:46
  into four hours about that as well in terms of actually
• 19:48
  building that nicely so that we get something that is.
• 19:51
  At a high precision accuracy and the and ensuring that that can
• 19:55
  also run fast at A at A at a low cost.
• 19:58
  Right, so. Awesome.
• 20:01
  Thanks for sharing that and then Yash, uh you know you know you
• 20:06
  you you. I think I've hypothesised and
• 20:09
  we've talked about sort of the future where we'll have hundreds
• 20:12
  of thousands of MLM's and and and you're so Martian being the
• 20:15
  you know the the gateway to make that easy to choose which one.
• 20:18
  How are you you know building AI into providing that capability
• 20:22
  to to for example ourselves right.
• 20:24
  That are trying to make these choices on a day-to-day basis,
• 20:27
  which model for which use case and so on.
• 20:30
  Yeah, this is, this is a question I love diving into the
• 20:32
  kind of like, you know, how are you actually doing this routing
• 20:36
  that because the, the fundamental sort of piece in
• 20:39
  routing is understanding these models, right?
• 20:43
  In order to say, OK, we're going to send this request to this
• 20:46
  model, you have to know what each model is good or bad at.
• 20:50
  And that is often quite idiosyncratic, right?
• 20:54
  You know, these models hallucinate.
• 20:55
  These models have all sorts of unknown behaviours, right.
• 20:57
  They're really, really intricate machines.
• 21:00
  And So what you have to do if you want to route well is really
• 21:04
  understand what those machines. Doing right.
• 21:07
  You have to basically understand them at a mechanistic level,
• 21:09
  right, how they work internally. And so the core technology we're
• 21:12
  developing is actually understanding these transformer
• 21:16
  models, right? You train them on a bunch of
• 21:18
  data, you get this black box. We are opening up that black box
• 21:22
  and we're doing this through a process which we're calling
• 21:24
  model mapping, right. The idea with model mapping here
• 21:28
  is to turn a transformer into a more interpretable
• 21:31
  representation to say, OK, Transformers are hard to study
• 21:34
  these big black boxes. If we can turn them into
• 21:37
  something else first, something which we can understand better,
• 21:40
  then we can actually. Study them in more depth.
• 21:43
  And that in depth study lets us do routing better than any other
• 21:48
  company that possibly could, right?
• 21:50
  The reason we're able to do this, actually part of the
• 21:52
  reason we started this company is that if you look at almost
• 21:55
  any other company, they're often think of themselves as a kind of
• 21:57
  model provider, right? They're saying cool, we're going
• 22:00
  to provide these models. We're in competition with other
• 22:02
  people. We're going to be bigger and
• 22:03
  bigger models. So when they raise funding, it
• 22:05
  goes towards building bigger models, right?
• 22:08
  Our approach is very different to say, let's really understand
• 22:11
  these models deeply so that we can use them in more
• 22:13
  fundamentally interesting ways, including, for example, routing.
• 22:16
  And so that's the core technology we're building, the
• 22:18
  ability to understand models rather than simply treat them as
• 22:21
  black boxes. Yeah.
• 22:23
  And that's I think incredibly powerful.
• 22:26
  You've you've built your own models to do this in fact,
• 22:29
  right. It's a little bit meta in that
• 22:31
  way. But you know the theme of this
• 22:33
  conference LM's in production and you know, we know from a lot
• 22:36
  of the work we've done that you need to have this kind of
• 22:38
  intelligent model selection to go into production.
• 22:41
  And going for larger, always larger models isn't the either
• 22:44
  cost effective or or better best for the user experience.
• 22:47
  So thanks for sharing that. Over to you, Sandeep.
• 22:52
  No, perfect. Well, just to to change change
• 22:55
  gears a little bit. One of one of the questions
• 23:00
  given the fact that the audience is predominantly you know people
• 23:03
  in that are ML practitioners and in the in the field.
• 23:07
  Just to share an anecdote from somebody in the venture capital
• 23:09
  community. I heard that I know some junior
• 23:12
  people that are reaching out to every single person on LinkedIn
• 23:15
  at the moment that have anything to do with a IRL in their
• 23:18
  LinkedIn bio. So perhaps just wanted to get a
• 23:21
  flavour from from from the three of you around.
• 23:25
  Any tips or advice you can share with the the folks that are in
• 23:29
  the audience at the moment on how best to engage with the
• 23:32
  venture community? To engage selectively, not at
• 23:35
  all. In in in detail, perhaps
• 23:38
  starting with with Lars, given, perhaps given given the the, the
• 23:42
  scale of your business. Yeah.
• 23:46
  Yeah. So it's it's difficult question
• 23:49
  right. So let me try to at least might
• 23:52
  might take another is it can be difficult to to to to manage all
• 23:56
  of the inbound. I think that after this the like
• 24:00
  like you have different ways of AI right now follow some of the
• 24:02
  waves and I can always feel like the on the inbound on LinkedIn
• 24:06
  that that that that that that you feel the wave coming and I'm
• 24:09
  of course entering all of those in don't know no of course not
• 24:13
  but I think that the. I think that the.
• 24:18
  What one of the things that that that I see as a really good way
• 24:22
  of engaging with the venture capital community is is actually
• 24:24
  listening for a lot of the a lot of the competitive advice
• 24:27
  because a lot of the venture capital community have a really
• 24:29
  good finger on the pause in terms of the competition that
• 24:32
  they see out there in the field and also the competition on your
• 24:35
  business. I I think that's just talking
• 24:37
  for myself but others can might feel that my idea is better than
• 24:40
  anyone else's right or at least that that feeling you know end
• 24:44
  up with right and actually ping pong a little bit more on the
• 24:47
  venture capital community and and the venture capitalist in
• 24:50
  terms of actually finding out whether your idea is that.
• 24:52
  Brilliant and and getting some insights by asking enough
• 24:55
  questions back again in terms of actually finding out where you
• 24:57
  compare your competition is. That's something that I've
• 25:00
  really felt is is a good investment for my side and it's
• 25:03
  something where we can actually get a lot of sanity cheques
• 25:06
  right because a lot of things that are happening out feel
• 25:08
  everything can feel so amazing what people are doing out there
• 25:11
  and you can look at LinkedIn and everyone our success right.
• 25:14
  But scratching a little bit underneath the surface then you
• 25:17
  can also feel that other people are sitting with the same pains
• 25:19
  that yourself and it's good to, it's good to use that the
• 25:22
  venture capital community in, in that sense because a lot of
• 25:25
  people are actually using that as the 100%.
• 25:27
  Your day job and finally in terms of finding out, scratching
• 25:30
  a little bit beneath the surface in terms of finding out where
• 25:32
  you are, are at this moment. Right, so.
• 25:36
  And Yash, anything for you to add being in in San Francisco
• 25:39
  and Silicon Valley? Yeah, it's a good question.
• 25:43
  I think there's always this, this temptation, especially for
• 25:46
  newer founders to, you know, sort of jump at the exciting
• 25:49
  things, right. And one of the things exciting,
• 25:51
  we're building companies when other people start seeing, you
• 25:53
  know, what you're building and being interested by it, seeing
• 25:56
  your vision. And so, you know, when VC are
• 25:58
  reaching outcomes like, oh, wow, that's like, you know, the
• 26:00
  really cool thing, and it is, it is you should be excited by it,
• 26:03
  right. At the same time, you know, the
• 26:05
  core of your business is to like go and build a business, right.
• 26:08
  And so it's worth talking to folks who are, you know, really
• 26:11
  high intent, who really, like, have conviction in what you're
• 26:13
  doing, and also getting to know folks.
• 26:16
  More generally, but I mean I know that like like we're we're
• 26:20
  really excited to have worked with process because you guys
• 26:22
  are so high conviction, right. If I feel like it gives us the
• 26:25
  ability to like go and focus on on building, right.
• 26:29
  So I guess as a as a kind of like general advice which of
• 26:33
  course would vary if you know we had more specifics about any one
• 26:37
  person situation. But as some kind of general
• 26:39
  advice I'd say you know building is the most important thing.
• 26:42
  It's good to sort of get to know people in the field and and
• 26:45
  build relationships with people, but really focus on the folks
• 26:48
  who are highest conviction. Thank you, Petra.
• 26:51
  Anything that you'd like to add from your perspective?
• 26:54
  Sure. Yeah.
• 26:55
  Yeah. I love that that that that word
• 26:57
  conviction that that's that's kind of my perspective too.
• 27:00
  Although we're, you know we're early on our journey like we
• 27:03
  raised some precede 2 years ago. And so I'm here to listen as
• 27:07
  much as I am to preach. But I would say, you know, try
• 27:10
  to engage with future facing people if you have a future
• 27:13
  facing idea and try not to get discouraged and find find that
• 27:18
  conviction, You know, shout out to Paul and and Euro when when
• 27:21
  we first sat down with them, we were selling through an idea
• 27:25
  that might be five years out in a problem that didn't even exist
• 27:28
  yet and they. They saw potential in it and
• 27:30
  that was, you know, not everyone's going to see
• 27:33
  potential in your idea, especially if you're solving a
• 27:36
  problem that's further out. So you know, stay, stay the
• 27:40
  course, stay with your your vision and with the right people
• 27:44
  you know they can make the the magic happen.
• 27:47
  Awesome. Paul, back to you.
• 27:49
  Yeah, given given where we are on time and and also you know
• 27:54
  we've got I think today a few thousand people listening.
• 27:57
  Most of the folks in the in the conference with the event are
• 28:00
  and all practitioners. And I'm of the fundamental
• 28:03
  belief that in the next wave of AI companies we're going to be
• 28:07
  in need of founders like yourselves that have a strong
• 28:10
  technical background, that have strong intuition with what you
• 28:13
  know, where's the right time and space to you know deploy this
• 28:16
  technology in a Safeway and impactful way.
• 28:19
  So I'd actually given you have you know, maybe similar
• 28:22
  background to many of the folks listening.
• 28:23
  Maybe ask each of you for one piece of advice you wish you had
• 28:27
  known before starting your AI company that you'd like to share
• 28:30
  with folks here that hopefully will help them if they ever take
• 28:33
  this step in the same direction. So maybe starting in the in the
• 28:37
  same order large you want to want to share some words.
• 28:41
  One piece of advice, if you have more, feel free.
• 28:45
  Ohh yeah yeah I I think it's it's it's difficult to to get
• 28:50
  that right. I think that the follow your
• 28:53
  conviction is is really good. I learned from it from you guys
• 28:56
  here. Don't be discouraged and and and
• 28:57
  stay sane. There's one piece of advice.
• 29:00
  Uh, split three, right? Thanks for that, Lars Yash.
• 29:07
  Yeah, I would say that because there's so many cool things
• 29:10
  going on in the field right now, it's really important to focus
• 29:12
  on what gets you internally motivated, right?
• 29:15
  So instead of looking at sort of, oh, external kind of of
• 29:18
  motivations, look at what really gets you going internally that's
• 29:24
  I think that's important for everybody probably and
• 29:27
  especially if you're going to dedicate such an important part
• 29:29
  of your life to to a mission like founding a company and
• 29:33
  Pietro, how about you? Yeah.
• 29:35
  When I look back, I I kind of wish I knew what terms were
• 29:39
  going to come to be and so I could use those instead of
• 29:43
  taking a paragraph to explain it.
• 29:45
  And then now Jen AI boom, that's the term.
• 29:48
  So I would say like as you're inventing things, really try to
• 29:51
  latch on to what what culture is calling, what you're doing.
• 29:54
  Don't get stuck in in, you know, the way you used to explain it,
• 29:57
  especially if you're inventing something.
• 29:59
  But yeah, ultimately follow your conviction.
• 30:01
  I guess I'm just going to reflect what's been said before.
• 30:06
  Great follow. Do what you love and and don't
• 30:09
  get distracted by by anyone else.
• 30:12
  Change what your conviction should be.
• 30:13
  That's fantastic. Thank you guys for joining.
• 30:15
  And Sandeep, also thanks for joining the conversation and
• 30:19
  taking the time. I know it's either very late or
• 30:21
  very early for some of you, so I hope we get the conversation
• 30:24
  very soon again. And now I'd like to maybe hand
• 30:28
  it over to Demitrius is going to change the topic to something
• 30:32
  related to Mystic Wizards like Gandalf.
• 30:37
  Ohh yeah. Now shall not pass.
• 30:41
  Yeah. I'm Immanuel Mason, Machine
• 30:47
  Learning engineer at Stripe and author of Building Machine
• 30:50
  Learning Powered Applications. And if you don't want your
• 30:53
  machine learning models to explode, well, you should
• 30:57
  subscribe to this podcast.