Latest / AI and the Future of Work: Artificial Intelligence in the Workplace, Business, Ethics, HR, and IT for AI Enthusiasts, Leaders and Academics / Rodrigo Liang, SambaNova CEO, shares how he raised more than $1B to launch and grow the first generative AI unicorn

Transcript

Rodrigo Liang, SambaNova CEO, shares how he raised more than $1B to launch and grow the first generative AI unicorn · AI and the Future of Work: Artificial Intelligence in the Workplace, Business, Ethics, HR, and IT for AI Enthusiasts, Leaders and Academics

• 0:02
  Speaker 1: Over time.
• 0:03
  Speaker 1: I think what we saw was the industry started the internet, the industry started segmenting to search and streaming media and right sharing right.
• 0:13
  Speaker 1: There's so many different applications and, within those applications, different use cases that I think it's still too calm for the artificial intelligence.
• 0:34
  Speaker 2: Good morning, good afternoon or good evening, depending on where you're listening.
• 0:37
  Speaker 2: Welcome back to AI and the Future of Work.
• 0:39
  Speaker 2: Thanks again for making this one of the most downloaded podcasts about the topic.
• 0:43
  Speaker 2: If you enjoy what we do, please like, comment and share in your favorite podcast app and we'll keep sharing amazing conversations like the one we have for today.
• 0:53
  Speaker 2: I'm your host, dan Turchin, ceo of PeopleRain, the AI platform for IT and HR employee service.
• 0:59
  Speaker 2: I'm also an investor and an advisor to more than 30 AI first companies and, as you know, a firm believer in the power of technology to make humans better.
• 1:08
  Speaker 2: if you're passionate about changing the world with AI, or maybe just looking for your next adventure, let's talk Now.
• 1:17
  Speaker 2: we learned from AI thought leaders weekly on this show and, of course, the added bonus is you get one AI fun fact each week.
• 1:24
  Speaker 2: Today's fun fact open AI CEO Sam Altman urged the acceleration of AI regulation.
• 1:30
  Speaker 2: in a recent Senate hearing.
• 1:33
  Speaker 2: Altman largely agreed with lawmakers that AI can be dangerous if not regulated.
• 1:38
  Speaker 2: This is a radical departure from the historically antagonistic relationship between Silicon Valley and Capitol Hill.
• 1:44
  Speaker 2: One of Altman's ideas echoes a similar theme we've discussed numerous times on this podcast There should be an agency, federal or otherwise, that issues licenses for the development of large scale AI models, safety regulations and also tests to ensure that AI models that are released to the public behave in a way that's safe.
• 2:09
  Speaker 2: As always, we'll link to the full article in today's show notes.
• 2:13
  Speaker 2: It's an important topic and one we will continue discussing.
• 2:16
  Speaker 2: But now, shifting to this week's conversation, today's guest is one of the original AI first entrepreneurs.
• 2:22
  Speaker 2: Sambanova paved the way for generations of other companies, including today's generation of a generative AI companies.
• 2:31
  Speaker 2: Rodrigo Leong and his team have raised more than a billion dollars from a legendary group of investors including Temasek, blackrock, gv and Walden International.
• 2:42
  Speaker 2: The original vision for Sambanova's chip architecture and software products came from work He and his co-founders did at Stanford's famous AI lab.
• 2:52
  Speaker 2: Today, sambanova has embraced generative AI and is again leading the industry.
• 2:57
  Speaker 2: Before founding Sambanova, rodrigo held seed in your leadership roles at Oracle and Sun after having received his master's and bachelor's degrees in electrical engineering from Stanford.
• 3:07
  Speaker 2: And without further ado, rodrigo, it is my pleasure to welcome you to the podcast.
• 3:13
  Speaker 2: Let's get started by having you share a bit more about your background and how you got into this space.
• 3:20
  Speaker 1: Thanks for having me, dan.
• 3:21
  Speaker 1: It's a real pleasure to be here.
• 3:23
  Speaker 1: Yeah, it's an exciting time that we live in today.
• 3:27
  Speaker 1: We are witnessing the fastest industrial revolution that humanity has seen, and to be in the middle of all of this and for us all to be witnessing all the changes that artificial intelligence is bringing is just amazing.
• 3:45
  Speaker 1: So my background is as you said I was born in Taipei and grew up in Brazil and partly the things that people always ask me about my name and the company name that's part of the reason why we have some references to Samba and things like that But came to the US in my undergraduate work at Stanford and went to work for Hewlett-Packard in the early 90s building high performance systems for enterprise applications, and, after a couple of different startups that were acquired, dan collaborated again with one of my co-founders, kulepo Lukotun, who's a professor, longtime professor at Stanford, and then one other professor, chris Ray, who's just been one of the most prolific individuals when it comes to artificial intelligence innovations, and so came together in 2017 to build this company, and so when we started, we really thought about trying to figure out what is the next wave of innovation going to require when it comes to building platforms that allows everyone to be part of this sea change.
• 4:56
  Speaker 1: We think AI is going to be as big, if not bigger than the internet, and the technology that it brings is allowing us to then create new technologies above it, new ecosystems that would do a bunch of different things, including what you mentioned earlier around regulatory and all things like that, and entire economies that will be powered by artificial intelligence.
• 5:24
  Speaker 1: So we started thinking about this long journey that we as a society are going to go, and how do we create a platform that's one much easier to implement, deploy, much cheaper for people to be able to come in, engage and participate, and much faster for companies that perhaps don't have that deep technical expertise in machine learning to be able to be part and take advantage of the technology.
• 5:49
  Speaker 1: And those are the things that were at the core of what we wanted to do when we started, with someone over, and we're excited to be in the middle of it.
• 5:58
  Speaker 2: A lot of us know Sombanova from the novel chip architecture optimized for AI.
• 6:05
  Speaker 2: Talk us through the evolution of the company over the last six years.
• 6:10
  Speaker 1: Yeah, when we think about artificial intelligence, we think about today.
• 6:17
  Speaker 1: There are a few key things that we think about.
• 6:19
  Speaker 1: We think about GPT models.
• 6:22
  Speaker 1: We think about that.
• 6:24
  Speaker 1: We think about GPUs and the shortage of those and what's required there.
• 6:32
  Speaker 1: We think about machine learning expertise and being able to train these models correctly.
• 6:38
  Speaker 1: We think about data sets and the sizes of data and token counts that are required to train these things properly.
• 6:45
  Speaker 1: We think about integration use cases.
• 6:50
  Speaker 1: One of the things I hear the most is what use case are we going to have for artificial intelligence, which, as a preview of how it's everywhere, is going to be everywhere in everything that we do?
• 7:01
  Speaker 1: But we can talk more about that.
• 7:04
  Speaker 1: But with Sombanova, when we started thinking about those things, we realized that it's too big of a technology transition for the world for us to piecemeal that together.
• 7:16
  Speaker 1: Take a hardware architecture that's 25 years old.
• 7:20
  Speaker 1: Take some software infrastructure that we've morphed over from gaming and graphics over to something else.
• 7:30
  Speaker 1: We take these models that we had to put together, but we didn't have enough data and so we had to glob it together.
• 7:37
  Speaker 1: These are all things that in an evolving industry, in a early development, it's natural.
• 7:44
  Speaker 1: You saw that with the internet.
• 7:46
  Speaker 1: You saw it with clouds, you saw it with various different world transitions.
• 7:50
  Speaker 1: We're now as Sombanova, we're thinking about this.
• 7:53
  Speaker 1: Okay, there will be a second wave, a second wave of things that you have to do in a streamlined, seamless fashion for broad adoption.
• 8:01
  Speaker 1: That's where we are today.
• 8:03
  Speaker 1: You're seeing the technology that was really focused on developers and kind of the first wave of really expert users now branching into a phase where it's a broad adoption For broad adoption.
• 8:15
  Speaker 1: We started thinking about Sombanova in terms of how do we actually encapsulate all these key technologies the hardware, the software, the models, the data sets that you need and actually even the use cases into very simple and easy to deploy packages that allows enterprises to come in and use and own their own model.
• 8:35
  Speaker 1: That's kind of really the evolution that people see a lot about the billion dollars we raised and the data flow chip that people got all excited about a few years ago.
• 8:47
  Speaker 1: But today we don't see ourselves as a chip company.
• 8:50
  Speaker 1: We really see ourselves as a platform company focused for enterprise, where we are actually deploying models like GPT so you can have your own GPT model.
• 9:03
  Speaker 1: Here's an important question that I always tell people you have to think about in terms of AI for your business.
• 9:08
  Speaker 1: Do you see AI as a tool or do you see AI as an asset?
• 9:12
  Speaker 1: And so Sombanova we're really focused on enterprises and companies that feel like AI is something I'm going to invest in for the next 10, 15, 20 years, and so I have to own it and I will have to invest into it, and the value of that model is going to get better and better and better, and you're creating value in that model every year because over that 10, 15, 20 year period, this thing will be a competitive asset for me, and so everything about Sombanova has evolved to the point where we need to enable companies and enable organizations to quickly build, start building this asset and creating value on this asset year after year after year.
• 9:51
  Speaker 1: And how do I do it?
• 9:52
  Speaker 1: as cheaply and as quickly for every company as possible.
• 9:57
  Speaker 2: So you mentioned one of the impediments to the success of the generative AI economy is chip shortages, and you also mentioned that you can't just repurpose chips from gaming high-end gaming PCs and expect to satisfy the insatiable demand for these chips.
• 10:15
  Speaker 2: And yeah, nvidia is doing okay $775 billion market cap.
• 10:20
  Speaker 2: What do you say to those who assume that NVIDIA won the race for AI-first chips?
• 10:27
  Speaker 2: Is there room for anyone else?
• 10:30
  Speaker 1: Okay, i think NVIDIA is not just doing okay, they're doing fantastic.
• 10:33
  Speaker 1: I mean, they've done a tremendous job, being able to leverage assets that they've built over 25, 30 years and really deploying a broad range of applications.
• 10:43
  Speaker 1: And, you see it, they dominate in these various different use cases HPC in gaming and computing, and now artificial intelligence And so they've been able to leverage those assets to build technology that people can then quickly use and deploy, and so there's clear advantage of being an incumbent and driving the technology that people are already using And I do think that there are, you know, this artificial intelligence marketplace, who is still yet to fully develop, is still yet to fully stratify into various different segments and different layers You hear people talk about.
• 11:22
  Speaker 1: Are you in the training on the inferencing?
• 11:23
  Speaker 1: Are you in the data center on the edge?
• 11:25
  Speaker 1: Are you in this or in that?
• 11:27
  Speaker 1: I actually think that the market is still segmenting and there will be different use cases, of which, individually, there'll be opportunities for not just hardware vendors as well as the software vendors to differentiate and provide value.
• 11:41
  Speaker 1: And, like you know, this is true with almost anything.
• 11:43
  Speaker 1: If you find a segment that you're able to come in and differentiate and provide value, there's a play.
• 11:50
  Speaker 1: Like you know, the internet wasn't just hey, this is the internet, and Netscape has one at all or whatever it was, or AOL at one time is one at all, or even Yahoo for search.
• 12:00
  Speaker 1: Over time, i think what we saw was the industry started the internet.
• 12:05
  Speaker 1: The industry started segmenting to search and streaming media and right sharing right Like.
• 12:12
  Speaker 1: There's so many different applications that within those applications, different use cases that I think it's still to come for the artificial intelligence market.
• 12:20
  Speaker 1: And so I do think that NVIDIA is going to continue to be an incredibly powerful player in that market, along with many of the hyperscale companies, the Google's, the Microsoft's and AWS's of the world, and there will be opportunities for new players to come in and carve out a space for themselves.
• 12:40
  Speaker 1: Right, i think history has proven that.
• 12:44
  Speaker 1: You know, there's always room for new players to play if if what you bring is something that's innovative enough and solves a real problem for people.
• 12:54
  Speaker 1: right That, it isn't just about a me too.
• 12:57
  Speaker 1: You got to bring things that solve real problems and bring value to people, and if you do that, and do that consistently, you have a chance.
• 13:06
  Speaker 2: So in Gartner terms, i think as technologists we'd agree.
• 13:10
  Speaker 2: We're going to reach the peak of the hype cycle and we're going to enter a what they call a trough of disillusionment.
• 13:17
  Speaker 2: And I'm curious to get your perspective when we hit that trough, the bottom of the trough of disillusionment.
• 13:24
  Speaker 2: I can envision a bunch of reasons why we get there, and maybe it's a combination of them.
• 13:28
  Speaker 2: But is it going to be because the performance of these systems is poor?
• 13:34
  Speaker 2: We know that it can take upwards of 30 seconds to get a reply.
• 13:39
  Speaker 2: When we get a reply, it's often very expensive, both the compute to train the model as well as the runtime request.
• 13:47
  Speaker 2: The models famously hallucinate.
• 13:49
  Speaker 2: They create facts.
• 13:51
  Speaker 2: They're only built on a corpus of primarily English and only linguistic data.
• 13:56
  Speaker 2: Like I mean, you and I are well familiar with all of the possible reasons why we could reach that trough of disillusionment.
• 14:04
  Speaker 2: We'd love to get your perspective as someone who, like me, is an AI enthusiast but who's also a realist.
• 14:12
  Speaker 2: What do you think is going to lead to that crest when we shift away from the crest of the hype cycle?
• 14:19
  Speaker 1: I think, if I think about the technology itself, the hype cycle becoming waves, and so you can almost say the wave of AI chips hype has already kind of come.
• 14:38
  Speaker 1: There was a time in 2017 where AI chip companies were going to the roof and you haven't seen one get defunded recently, just because the cost structure of doing it and what it takes to actually deliver real value, the reality of that starting to set in, and then you have another wave of ML Ops companies right, And there's so many companies were created, And then you realize, well, for you to really get broad adoption, you have to create real value across a broad range of platforms.
• 15:13
  Speaker 1: So you're seeing these things that get the hype and then they kind of become part of mainstream workflow, in the way that you expect new technology to be delivered.
• 15:23
  Speaker 1: And now you're seeing this Gen AI.
• 15:25
  Speaker 1: Specifically, Gen AI, when it comes to chat, hit this peak And I'll tell you this, there are many additional use cases that are yet to come, tied to AI, that have not yet to hit the peak.
• 15:38
  Speaker 1: Let me talk about image, for example.
• 15:40
  Speaker 1: Right, And think about the excitement around.
• 15:44
  Speaker 1: Chat is nowhere near reflected when it comes to generative imaging, especially high resolution imaging, And then think about this in a streaming fashion, right.
• 15:55
  Speaker 1: You think about all that still to come.
• 15:58
  Speaker 1: And then you have waves beyond that in terms of science and those things that have yet to hit the high.
• 16:03
  Speaker 1: But I think you're going to find that AI is going to be a stream of waves that hit those curves and they will impact us in different ways.
• 16:15
  Speaker 1: But here's what I do think is going to be slightly different with this wave compared to previous ones that we've seen right, Because what happens underneath this, what happens underneath this wave, is at such a fundamental level to a workplace and workforce that it's not really touching just one segment.
• 16:39
  Speaker 1: If you think about the excitement around something that affects one industry, you can say, okay, that industry got excited and then flattened out.
• 16:47
  Speaker 1: Now, if you take this and say artificial intelligence is affecting every industry, every vertical, every department within every one of those groups, the averaging out of all these different groups actually uptaking the technology and then using it actually both tempers the peak of that wave.
• 17:07
  Speaker 1: It also tempers the trough of that wave because different groups are getting excited about it at different times.
• 17:12
  Speaker 1: Right, Our vision of artificial intelligence for the enterprise some of them is an enterprise focused company.
• 17:19
  Speaker 1: We really look at it as developing technology for large organizations to do business work.
• 17:27
  Speaker 1: We talk about AI seriously, so serious artificial intelligence work.
• 17:31
  Speaker 1: We view artificial intelligence as something that will enable every knowledge worker in the world to have an assistant.
• 17:40
  Speaker 1: When you think about this, our model of the world, whether you're in finance, you should have your own personal assistant to look at all the 10 case.
• 17:47
  Speaker 1: If you are in legal, there should be a personal assistant to help you with all the contracts.
• 17:52
  Speaker 1: If you are an engineer, there's a co-pilot that help you with various different things.
• 17:56
  Speaker 1: If you're marketing, you already see people getting a slide generated from that.
• 18:00
  Speaker 1: If you are in sales, you have whatever.
• 18:02
  Speaker 1: Every knowledge worker in the world will have the personal assistant that allows their own productivity to jump 10-fold 10-fold right.
• 18:14
  Speaker 1: When you think about it in those terms, it's not that every knowledge worker will adopt AI all at once.
• 18:20
  Speaker 1: It's not that every industry will adopt AI all at once.
• 18:24
  Speaker 1: What you're going to see is different industries and different departments of those industries adopting different times.
• 18:28
  Speaker 1: What that does?
• 18:29
  Speaker 1: it tempers the peak and tempers the drop.
• 18:34
  Speaker 2: Many who hadn't heard from Sombanova for a bit were reintroduced to the company, at least in my case last week when you introduced Bloom Chat.
• 18:45
  Speaker 2: I appreciate the fact that you and the team chose to address the problem that public LLM's large language models are trained on almost entirely a corpus of English content, yet up to 90% of the world doesn't speak English Talk.
• 19:03
  Speaker 2: Just about the thought process behind picking that problem to solve and the Bloom open source LLM and just the genesis of the work that you did to launch Bloom Chat.
• 19:17
  Speaker 1: Yeah, super excited about this.
• 19:18
  Speaker 1: Just as a recap, what we did was we released jointly with a fantastic startup, togetherxyz It's open source organization, a company that's focused on building these open source models.
• 19:34
  Speaker 1: We partnered with them and really trained these models together with them on the Sombanova platform.
• 19:41
  Speaker 1: What we announced was Bloom Chat at 176 billion parameters for multi-lingual capabilities.
• 19:52
  Speaker 1: Really, the focus there wasn't just arbitrary.
• 19:56
  Speaker 1: it's thinking about our customer base.
• 19:58
  Speaker 1: Our customer base on here is a multi they're usually multi-lingual corporations.
• 20:04
  Speaker 1: Yes, english is predominant in many of these companies and many of our clients, but most of them have businesses across the globe and they all want to be able to operate their businesses with artificial intelligence that touches all the different parts of the world.
• 20:22
  Speaker 1: Some of this is really also driven by our own interest in being able to actually create a community, create an ecosystem, create models that address all the corporate needs across the world.
• 20:36
  Speaker 1: What we did here was release a model into the public domain jointly with Togetherxyz and allow people to take this model and start looking at it Again.
• 20:46
  Speaker 1: this is one of many that we will do over time, but we do think that for artificial intelligence, this is a Linux moment.
• 20:53
  Speaker 1: We're in a Linux moment for AI that, if your listeners recall, 20-some years ago we had.
• 21:03
  Speaker 1: the most popular operating system for developing internet applications 20-some years ago was Solerys.
• 21:11
  Speaker 1: Right, sun Microsystems is a dot in dot com.
• 21:16
  Speaker 1: It produced an amazing operating system.
• 21:19
  Speaker 1: And there are other companies, like you know.
• 21:22
  Speaker 1: so IBM has AI ads, and you had HP, you had HP UX And you had all these different companies investing in the operating system that allow people to create enterprise class applications or consumer grade applications that allowed you to then deploy into the internet.
• 21:38
  Speaker 1: And then there was this tiny little effort around Linux that people that will never be good enough, you'll never be good enough to power enterprise.
• 21:47
  Speaker 1: And here we are, right Turns.
• 21:50
  Speaker 1: you know, what we've proved to the world is the public domain is able to innovate very quickly, very rapidly, because the contributions of the global knowledge is actually baked into it.
• 22:03
  Speaker 1: And someone over has made a commitment to use the open source community as a way to one, accelerate adoption of models, but also leverage what people have done to enable us to deploy these high value type of models to as many people as possible as quickly as possible.
• 22:23
  Speaker 1: And so we're really excited about this And you know, and you'll continue to see us contribute to that community and be able to then bring those models, these open source models that people are creating, inventing and deploying into very well curated, enterprise class products that our customers can then use and deploy within their own organizations.
• 22:48
  Speaker 2: I read the white paper on the performance benchmarks of Bloom Chat versus alternatives amazing progress.
• 22:56
  Speaker 2: Now you're talking to a technical audience and I know you're a technologist.
• 23:00
  Speaker 2: Answer for us a couple of these questions.
• 23:02
  Speaker 2: So I think conventional wisdom is that the bigger the model, the larger the number of parameters, the more it's able to do, or the more sophisticated it is, and then the more tokens it can process, the more useful it is.
• 23:15
  Speaker 2: You took 176 billion parameter model from Bloom and fine tuned it to be able to perform particularly well on multi-language tasks.
• 23:27
  Speaker 2: You talk a little bit about the process of fine tuning so that you took something that was initially, you know, a very credible LLM based on English and fine tuned it so that it's able to perform extraordinarily well on a variety of languages.
• 23:43
  Speaker 1: Yeah, exactly, i mean our, as you know, as someone of a, really we have trained a range of models.
• 23:52
  Speaker 1: We have trained some of the smallest models, the GPT 1.5 to 13, all the way through.
• 23:58
  Speaker 1: You know these very, very large models and we'll continue to push And we do think that, you know, in the world there's still a lot of different views as far as what the future of model will be.
• 24:10
  Speaker 1: Right, there are three important attributes And here, I'm just talking about language right Three important attributes that people are thinking about.
• 24:20
  Speaker 1: One is going to be parameter counts.
• 24:23
  Speaker 1: Okay, is it 13?
• 24:24
  Speaker 1: Is it 65 billion?
• 24:26
  Speaker 1: Is it 175 billion?
• 24:27
  Speaker 1: Is it 400 billion?
• 24:28
  Speaker 1: We can talk about that right.
• 24:30
  Speaker 1: The second attribute that's really important is the number of tokens to train it in order to get to a certain quality right.
• 24:36
  Speaker 1: And so you know, as you train it more and those more and there are, you know, there's enough data out there that shows the depending on the size of your model that you know where those laws curse flat now also will be somewhat proportional to the size of the model that you have, and so pumping more tokens into it doesn't always help you on a certain way Although there's some interesting research around that, maybe, or not as well.
• 25:03
  Speaker 1: And then the third one.
• 25:05
  Speaker 1: The third key dimension that's important is sequence length, right.
• 25:09
  Speaker 1: So if you think about 2K sequence versus 8K, versus 32K and Tropic recently announced their environment of 100K, those are really important, especially for folks that are thinking about enterprise, local use cases, where you have just a lot of context.
• 25:24
  Speaker 1: You want to keep all together and be able to actually optimize and internalize, for those applications to just need bigger and bigger sequence length, okay, And so when you think about that, the first order, it's a little more so.
• 25:38
  Speaker 1: You know my engineers will probably cringe when I say this, but the first order, the first order cost in training those models is proportional to the product of those three things.
• 25:49
  Speaker 1: What's the first order?
• 25:50
  Speaker 1: right?
• 25:50
  Speaker 1: You know parameter count times, you know number of tokens you wanna pipe in times?
• 25:55
  Speaker 1: right, the sequence, first order.
• 25:57
  Speaker 1: So, and there are things that people are doing in order to actually make those runs significantly faster for different things, and more and more people would disclose those over time.
• 26:06
  Speaker 1: And we do set of things that we aren't talking about today, but you know, but there are tricks that people do to actually get you significantly faster training and fine tuning of those models.
• 26:16
  Speaker 1: Now, in order to actually so, what we do is I'm gonna restart with these open domain models that are just.
• 26:21
  Speaker 1: I mean, they're gonna get better and better and better every time, whether that's 175 or 65 or you know, i mean the Lama models and other people creating other models and really small models now with 7B, that are really good as well.
• 26:31
  Speaker 1: And so we start with open models And then we have a set of things that we want to do for our customers that allow us to actually make the models very good, certain things that we wanna enable for our customer.
• 26:44
  Speaker 1: In this particular case, what we focused on was multilingual right For certain geographies, because we knew we needed to actually provide that Over time.
• 26:53
  Speaker 1: What you see is people will continue to pipe in these tokens to actually make those models better, but it's expensive.
• 27:00
  Speaker 1: It's a cost issue.
• 27:01
  Speaker 1: Right, it's a cost issue for people to continue to train because, as you know, these things take and take a lot of GPUs out of the train.
• 27:07
  Speaker 1: Or, you know, in our case, we run 100% on our hardware to train.
• 27:11
  Speaker 1: But as we make them available to the open community, what happens is it gives everybody a hot start.
• 27:17
  Speaker 1: Right Now you're getting something that I've trained, a hot start that allows you to go from there.
• 27:23
  Speaker 1: And people always ask me well, why does that make sense for someone over?
• 27:26
  Speaker 1: Well, it makes sense for two different reasons.
• 27:28
  Speaker 1: Right, it's important for people to understand, kind of, what someone over's model is.
• 27:33
  Speaker 1: Our belief, our belief is most corporations will want their own GPT model.
• 27:37
  Speaker 1: They will want their own, not a share model with somebody, not a private copy of somebody else's, but one that they can own in perpetuity.
• 27:45
  Speaker 1: If I'm going to pipe in my data into it, my model will be knowledgeable about my business.
• 27:51
  Speaker 1: If I'm going to do that, i want to own it.
• 27:54
  Speaker 1: Even if I fired the vendor, even if I get to keep all the bits that were part of that model, i want to own it.
• 28:03
  Speaker 1: So the way that we see the businesses we want to actually contribute to the open community, participate in that, take these open bits that the customers can see all the ways they can do it themselves.
• 28:13
  Speaker 1: But here's what we can do.
• 28:14
  Speaker 1: We can give you a significant acceleration in time, significant improvement in performance and significant lower cost because we can deploy those models faster, cheaper and better than anybody else and train them on their data.
• 28:28
  Speaker 1: As soon as we train on their data your data, your model So it becomes their model.
• 28:34
  Speaker 1: Now, every one of these corporations, they get to have their own GPT model in perpetuity, and so for us, the fastest way for us to enable every organization to have their own model that becomes an asset for them for the next decade or two is for us to get the open community to contribute, create these models that are good enough.
• 28:51
  Speaker 1: They're transparent, open and all those things that then allow us to go in and help these businesses Unlock all of this unstructured data, all of this data that's been hidden behind these firewalls, all the stuff that they don't feel comfortable releasing into open domain, unlocking it so that their business can take advantage of it and create new products, new services, new insight, new whatever, and that's information they've had for decades that you could never reach, and so that's really what we're here trying to do.
• 29:22
  Speaker 1: Yes, participating in the open community is one piece of being able to advance the technology, but as a business, we're trying to accelerate the average organization to get access to their data and leverage that to build more value.
• 29:40
  Speaker 2: So we're in agreement that the future of getting more utility out of these open source models is the process of optimizing them for specific domains or tasks or, like in the case of Bloomchat, optimizing them for multiple languages.
• 29:55
  Speaker 2: I thought the discussion in the performance benchmarking white paper about the process, the rigor that your team applied to making the Bloom open source model essentially perform equally well across I think six languages is really a good example, a really interesting case study in fine tuning, something we haven't talked about a lot on this podcast.
• 30:20
  Speaker 2: Do you mind, can I ask you to go through a little bit of how you approached, or how your team approached, the process of fine tuning, specifically to optimize for those different languages?
• 30:32
  Speaker 1: Well, i think data sets are important.
• 30:34
  Speaker 1: You want to make sure you have the right data set.
• 30:36
  Speaker 1: I think a lot of it is just about being able to actually have quality data sets that allow you to then put a focus on it.
• 30:43
  Speaker 1: As you know, with fine tuning, it's all about taking the model and make sure that you're actually exposing the model to a certain data set And if you give that enough time and enough computing to actually have the machine go through it.
• 30:59
  Speaker 1: And there's some things that we do to actually get faster to achieving the last curves that we achieve.
• 31:08
  Speaker 1: But ultimately it's really important to be able to actually be focused, like you said, about what you're trying to do with it.
• 31:14
  Speaker 1: Right, these open crawls are great in getting you kind of to a certain point, but over time, as you go forward whether that's for languages or whether that's for certain industries or whether that's for certain tasks that you're going to find that the quality level that is required for you to become that system for all those people in those geographies and those tasks can really ultimately be achieved by this very precise fine tuning that you have to go do.
• 31:46
  Speaker 1: That allows you to achieve those results.
• 31:48
  Speaker 1: Right, and so, yeah, it is computing, it's data sets.
• 31:51
  Speaker 1: There are some things you have to do to kind of allow you to get to that quickly enough, right, and then ultimately then being able to then deploy them into environments that allow you to actually get the energy.
• 32:04
  Speaker 2: To the uninitiated these LLMs just seem like magic, and even when you hear even some of the open AI engineers talk about the things that these LLMs are capable of doing that they never anticipated, it's really easy for the public just to say you know what it's artificial intelligence, it's magic.
• 32:24
  Speaker 2: But as a technologist that's a very unsatisfying answer.
• 32:28
  Speaker 2: There's actually a lot of rigor and a lot of process.
• 32:33
  Speaker 2: So, like in the case of multilingual training at least, i read about how large teams of native speakers in these other languages had to take open source tasks or training examples, translate them, and then you had I know it might be not due justice to the full rigor, but kind of double blind, to confirm that the translations were right And then humans had to score the accuracy on these various tasks, maybe just decompose some of the complexity of where it's not magic.
• 33:07
  Speaker 2: It's actually a lot of work and a lot of human effort goes into making these models as effective as they are.
• 33:16
  Speaker 1: Yeah, i mean, you ideally think that this would be something that if I just throw a bunch of data into it, throw a bunch of machines, it just works, right.
• 33:28
  Speaker 1: Well, it doesn't do that necessarily.
• 33:31
  Speaker 1: And we talk about the industry that's getting created around prompt engineering, and how do people do those things better?
• 33:37
  Speaker 1: right, you talk about the fact that these were now there are people now creating models to then train the models right, and so these systems that are created to be able to do this is exactly right, and you know you are going to continue to see the fact that the data set, the raw data set as is, is not good enough for you to just pipe in and get the results you want.
• 34:02
  Speaker 1: You need to curate that data.
• 34:04
  Speaker 1: And then there are companies like Snorkel amazing companies like that that do things that allow us to get better labels and better things on that data set to actually generate better results.
• 34:15
  Speaker 1: So, hands down, there's no doubt about that that if you can do the engineering work to get your data curated in the right way, one we'll generate enough data to get it curated in the right way.
• 34:25
  Speaker 1: You can start by tuning that and then getting better results And then, as you have the model being able to guide the model along the way, right, to do a certain set of things that you wanted to do.
• 34:36
  Speaker 1: The model learns very, very quickly, right, the model learns very, very quickly exactly which way it wants to go.
• 34:41
  Speaker 1: And so, so yeah, the steps, as you know, i'm oversimplifying kind of what the team does, but the step isn't just, you know, just throw a model and just look at the machine's run.
• 34:54
  Speaker 1: You do have to make sure that those pieces that you actually are trying to improve upon is actually being done with the level of detail and methodology that's that is meticulous, right, because otherwise you don't get the right result.
• 35:11
  Speaker 1: And you know, look, there are thousands of people out there trying to train these types of models, right, and, as you can see by just going through down the list, not all of them are very good.
• 35:20
  Speaker 1: Right, the thousands, because it does take, it does take significant expertise to train these things correctly and train them well.
• 35:27
  Speaker 1: And, and from where we are looking at what we need to go, there's still a lot of innovation that's required to bring these models to the level that I think we want to see it, you know, get to, so that you can, like you said, get global coverage.
• 35:42
  Speaker 1: Global coverage, you know as far as all the.
• 35:45
  Speaker 1: You know 176 languages that we need to go and get right.
• 35:50
  Speaker 1: And in not in every language is there today sufficient corpus for us to train a problem right?
• 35:58
  Speaker 1: And so there's again plenty of room for research on what do you do?
• 36:02
  Speaker 1: What do you do in environments where the data sets aren't as robust as what you need to train?
• 36:08
  Speaker 1: And yet you still need the same level of comprehension, the same level of quality for the for business to be able to be done in those environments.
• 36:18
  Speaker 1: What do you do there?
• 36:19
  Speaker 1: And so there's there's a lot of opportunities still to come for, for, for great creative technologies to help us figure out how to tackle those problems.
• 36:30
  Speaker 2: So, rodrigo, we're a bad out of time.
• 36:32
  Speaker 2: This one flew by, but you're not getting off the hot seat without answering one last question for me.
• 36:38
  Speaker 2: Recently, jan McCoon, father of convolutional neural nets and certainly one of the pioneers of deep learning, said where we're at today with large language models and some of these chat capabilities, that we're in the infancy of what we will be able to do with generative AI, and I think he said it to be somewhat controversial or, you know, at least lead a lot of people to question.
• 37:05
  Speaker 2: you know their enthusiasm about the progress we've made, but I tend to agree with him.
• 37:11
  Speaker 2: I think you know it's extremely limited What we can do today with the models versus what I believe we'll be able to do in a couple of years.
• 37:19
  Speaker 2: Let's say you and I are having a version of this conversation in.
• 37:23
  Speaker 2: you know, it's not 2023, it's 2026.
• 37:25
  Speaker 2: What do you think are the things that will be just commonplace with AI?
• 37:31
  Speaker 2: that would seem like science fiction dust today.
• 37:36
  Speaker 1: I think you're going to see everybody coming into work and having most of the detail work that they have that today they spend most of their time be done by an assistant.
• 37:47
  Speaker 1: I just got all these new data for you know, give me the reports, generate the reports for me, do the analysis of the new supply chain for me.
• 37:55
  Speaker 1: Give me all the sales, them, all of those things that we actually use humans to go, analyze it.
• 38:00
  Speaker 1: That's going to take three seconds And you're going to I didn't like that analysis do it with more of this three seconds later and you're going to find that in the workplace, everybody's going to have an assistant that's going to give them 10x, 10x impact improvement, 10x productivity improvement.
• 38:18
  Speaker 1: Right, and the marketing side?
• 38:19
  Speaker 1: oh well, do it with this tone.
• 38:21
  Speaker 1: I didn't like that.
• 38:22
  Speaker 1: Asher wanted to be a little bit more upbeat, a little bit more forward looking.
• 38:27
  Speaker 1: You can generate another marketing campaign, but these are all things that we're not.
• 38:32
  Speaker 1: We're not there yet today.
• 38:35
  Speaker 1: But here's what I think happened.
• 38:36
  Speaker 1: Here's what and I like the parallels with the internet when it comes to this AI transition for the world.
• 38:41
  Speaker 1: I think you know, when Mozilla came out and let's give me all those things were coming out people like.
• 38:46
  Speaker 1: That's amazing.
• 38:46
  Speaker 1: The internet is basically being able to show your company on the web.
• 38:51
  Speaker 1: Right, that's the internet, right.
• 38:54
  Speaker 1: And what happened then was AOL showed up.
• 38:56
  Speaker 1: And AOL showed up a platform that allowed everybody to log in, create their own spaces that allow you to do all these various different things on the internet.
• 39:04
  Speaker 1: Well, ai, today we're in the AOL moment.
• 39:07
  Speaker 1: Right That we've now shown the world that we can do a set of things, you know, for chat.
• 39:14
  Speaker 1: We can do a set of things for document classification, we can do a set of things for imaging and etc.
• 39:19
  Speaker 1: We've shown people what we can do.
• 39:20
  Speaker 1: Now think about, since AOL to now, all the things that internet has entire economies, from the Netflixes to the Ubers of the world, to Airbnb and to like.
• 39:34
  Speaker 1: Think about the economies that are created.
• 39:37
  Speaker 1: That went far, far, far beyond what AOL could offer at that time, right, and so this is where I think we are today that we've now seen at least the potential on language for one thing that AI can really do, and I think over the next two, three years, you can see this explosion of integration of artificial intelligence into your day to day life in a way that you won't be able to undo Like like.
• 40:06
  Speaker 1: None of us can live in a world where there is no internet.
• 40:09
  Speaker 1: I think in five years you said five years or three years, right in your scenario, i think we're going to envision a world that you will not be able to do without the services that artificial intelligence will give you, because your productivity is going to get enhanced 10x Right, and if you don't have it, you will be at a 10x disadvantage compared to everybody else.
• 40:34
  Speaker 2: I want to have you back, Rodrigo, when that version of the world is here.
• 40:37
  Speaker 2: It may not even take three years, but it sure feels like we're just getting started.
• 40:43
  Speaker 1: Absolutely.
• 40:43
  Speaker 1: It's exciting times And, like I said, the fastest industrial revolution the world's ever seen.
• 40:49
  Speaker 1: And we're in the middle of it and really excited to be part and really excited for all of your listeners to participate.
• 40:55
  Speaker 1: And we're just getting going And but it's exciting times, It's really exciting times.
• 41:01
  Speaker 2: On behalf of humanity.
• 41:02
  Speaker 2: We're all rooting for Samba Nova to be successful.
• 41:06
  Speaker 2: Rodrigo, where can the audience learn more about you and the work that your team is doing?
• 41:12
  Speaker 1: Certainly you can go to someone over that AI and the web and there's a lot of things on the on the internet that can follow us, follow us or follow me, rodrigo Leon, on LinkedIn and Twitter.
• 41:23
  Speaker 1: But we are pretty, we're pretty open about sharing kind of what we do on a regular basis, and so folks are trying to learn where AI is going And to figure out kind of what's next.
• 41:36
  Speaker 1: you know, again, when we start the conversation, said language is only the beginning.
• 41:41
  Speaker 1: There's a vision, there's science, all these other things are coming that we're working on as well, that people want to kind of start getting their minds wrapped around what's next.
• 41:51
  Speaker 1: you know, happy to happy to engage and be able to have those discussions with folks.
• 41:57
  Speaker 2: Well, thank you for letting us go way off script.
• 41:59
  Speaker 2: This is a lot of fun.
• 42:01
  Speaker 1: Thanks so much for having me.
• 42:03
  Speaker 2: Well, gosh, that's that is a wrap for this week on AI and the future of work.
• 42:07
  Speaker 2: As always, i'm your host, dan Turchin, ceo of PeopleRain, and of course, we're back next week with another fascinating guest.