00:03
Um, but, uh, I'm gonna get rolling on this is, uh, you know, we're here, um, to talk about, uh, Imagine experiment and adopt, how to ensure successful generative AI projects. So generative AI is creating a lot of hype. It'd be really cool if this worked, and it doesn't.
00:20
There we go. How's that? No. We'll do it from here, the old fashioned way. Generative AI is really creating a lot of, uh, a lot of the hype in the market right now. Um, this all started, um, a couple of years ago, but these are the type of headlines that are in the in the press today, and this is what's on the pressure of those that are in the IT
00:39
organizations, the business organizations, the data science organizations to really live up to, you know, we've got, uh, $1 trillion of GDP, 300 million jobs that are gonna be displaced. Um, you know, the public is expecting this to be the most, the, the largest impact of every sector of society over the course of the next 5 years. These are giant headlines, um, that, you know, ultimately, oh good,
01:04
this thing's working now. You have, uh, the industry titans now talking about the effect of business. Um, my favorite one is Jamie Dimon, which is, while we don't know the full effect or the precise precise rate at which it will change our business, or how it'll affect society at large.
01:22
We are completely convinced of the consequences will be extraordinary. I love how the fact that Jamie Dimon always plays both sides of this. That doesn't say it's a positive thing. He's just saying it's gonna be an extraordinary impact. And so the investments that we make also have to be wary of really what, um, what the, the leaders in business are saying.
01:42
But this all started with this concept of data and data being the new oil. And I think it's appropriate that here at the uh P, the Pure Accelerated conference, uh, where data is placed in storage, um, there was a lot of conversation in the keynote earlier about enterprise, uh, data clouds and, you know, data has been the fuel of what is really the heart of,
02:03
uh, what's built up AI and generative AI especially as we've launched over the last 2.5 years, but The concept of data being the new oil goes back to 2006 when Clive Humley, a British researcher, came out and said that, uh, data has all the principles of, uh, crude oil. Um, it needs to be refined, it has the ability to turn into great value.
02:26
11 years later, the Economist came out in May of 2017, and they actually said the refineries exist, that the, and they agreed that the world's most valuable resource is no longer oil but data. What was really interesting is this was literally the month before the transformative, uh, uh, phase of where we are right now with generative AI and that's when the transformer
02:51
effect was released, uh, based on some researchers, uh, at, uh, at Google in combination with, uh, their Stanford counterparts. Um, they released the, uh, attention is all you need paper and talking about how now Generative AI can use this process of a transformer effect and create a 10X improvement. Um, in the output.
03:12
Um, so, what used to take, um, 100 hours to produce could now be done in 1 hour. Um, and so it really had a major impact on startups at that 0.1 of them being OpenAI. Which 5 years after that point in November of 22, launched Chat GPT 35, which is what's created all the hype and where we are now, only 2.5 years past these major milestones.
03:40
We're now dealing with the impact of this, and we're dealing with the impact of this because data. As gravity, but in order to use things like the transformer effect and generative AI, you need to have GPUs. And for you to be able to bring the two of those together, that's the physics problem we're all trying to solve now.
04:00
And so it's a challenge that ultimately, uh, we are, we and the company I represent, SHI is, uh, really, um, taking on headlong and making major investments. Um, I've been in the IT industry for 30 years. I have, um, a background. I, uh, I was a, uh, IT operator. I was a co-founder and CTO from a, a consumer cloud and SAS company.
04:23
We competed with WebMD. Um, I did that for 23 years. I actually started my career prior to that. Using the foundations of what are included now in how we generate AI, which are neural networks. I used a neural network technology to predict the movement of the Swiss franc future fresh out of college in 1995.
04:42
So I've been dealing with large data problems for a while, um, and I ultimately, uh, spent 7 years at Pure Storage running as the VP of technology strategy, working with cloud and, uh, and, and, um, service providers. Um, I joined SHI a year ago to lead the, um, Advanced Growth Technologies Organization. Um, I'm based in Southern California, but I like to say I live in seat 6C,
05:05
married to a beautiful wife. Uh, it'll be 29 years in a, in a week and a half. Um, I also happen to have, uh, 21 year old triplets, so I am a little bit motivated to make sure that we advance this technology as quickly as possible because It allows me to finish putting them through college, um, but I'm also highly competitive in what I do.
05:23
And so for me, when I was brought over to SHI to run the Advanced Growth Technologies Organization, we needed to take a radically different approach to solve these big business problems. And so, um, SHI a little bit of, uh, just uh advertisement on the company. Um, we've been in business for over 35 years.
05:42
Uh, we have, uh, grown to represent now over 15 billion in annual gross revenue. Um, we are a privately held company. We're actually the largest privately held uh minority owned company in the US. Um, we have, uh, nearly 7000 employees of which, uh, over 1000, and this is actually getting closer to 2000 technical resources that we bring to bear.
06:04
Um, we have, uh, we are a global company, uh, we help customers, uh, really across all aspects of technology. You'll see the pillar on the, on the far left here. Um, this is the pillar that I'm responsible for, which is the AI solutions and the general of AI practice. Um, I also represent in terms of, uh,
06:21
the teams and the resources we bring together, and of our 200+ AI and cloud specialists that are dealing with some of these physics problems that I talked about earlier. So let's talk about what we're gonna go through. We're gonna go through what is the gender of AI value proposition? How do you even consider what is gene of AI and the different patterns of usage, and we'll kind of roll through um what we're doing with our
06:43
approach of Imagine experiment and adopt and how we're using our AI and cyber labs. But it all starts with making sure that we keep at the center of everything we do, generative AI should be solving business problems, not technology problems. It really should be about the business objectives that always are the center of where you're grounded in your approach to generative AI.
07:05
Whether you're going for a transformational or uh incremental change, whether you're leveraging open source or proprietary LLMs, um, or an approach to how you're building your stack, um, or you, you, the different AI patterns. I'll show you a little bit about this in a second. Um, but of course, you know, the final element
07:24
that, uh, I usually try and push all of my team is we talk about infrastructure last because we want to talk about how to get there first, because there's many ways to solve problems, but we want to get to the most efficient and effective and highest ROI in how we approach that. And I mentioned the different patterns. So there's many different ways that AI can be applied.
07:45
Everything from creative works, software development, productivity, customer service, personalization, of course, content creation. Who here in the room uses uh some form of a chatbot or LLM on a regular basis to create like content that they send off in emails. Pretty much the most common use case, creating images, you're seeing a lot of that now being uh a big uh uh uh a big uh element of
08:09
uh increasing productivity. But you know, to use customer support automation to be able to create a better experience for your customers. To be able to personalize and ensure that that next level of customer interaction is at a much higher level of engagement.
08:23
These are use cases and patterns that you can ultimately look to apply general of AI to solve in your businesses. But how you go about doing that is really understanding that the LLMs and their usages of behind these patterns are really critical. I just mentioned out of the box LLMs that most people are used to interacting with prompt
08:43
engineering, um, you know, out of a standard chatbot like chat GPT. But then you get into other elements. These large, these LLMs are not trained on your own company's data, and if you start adding your company's data to their publicly available interfaces, you actually run the risk of intellectual property leakage.
09:01
And so you have to do things like rag. Does anyone in the room not know what RAG stands for? Awesome. I don't, oh, OK, I'm gonna give you a quick example. RAG is really retrieval augmented generation is what it stands for, and it's adding in your own corpus of data that you may not want to make publicly available or you may not want to allow those LLMs to be trained on.
09:25
So by introducing your own company's data. A private context and integrating that to the prompts that are ultimately, uh, using a base LLM, you can inform that base LLM that might have great reasoning and understands the world's data, but now they can actually put the context of your own data. So when you look at injecting your own, your own information,
09:47
that's called RAG, and that's an important element. You also have the ability to create agents. Agentic AI right now is a big buzzword, and that's ultimately allowing these generative AI solutions to now reason and create response autonomously without interaction. Um, and of course, you can orchestrate all these together to solve really complex elements.
10:08
But, uh, the integration and accuracy and specialization as you move across this gets more and more complex as you move to building, creating your own custom LLMs or fine tuning of those LLMs or RAG. But most people understand that starting with prompt engineering is where you start there, but there's different ways to do your prompt engineering, different ways to integrate your
10:29
RAg, different ways to fine tune your data, all of which is very um risky. And challenging and why we, we are here to help customers really through their journey, because the latency and impact of experience, the cost and complexity really dramatically varies across all of these and so there's no simple answer to how to implement generative AI in your business. But, as I mentioned, SHI has a pretty good
10:58
opinion on how to approve this. at this aspect. And that's because many companies are somewhere on this spectrum of where they're entering their generative AI. They may be having done that, having high-performance computing or ML or AI practice for 30 years. Like I've been in, in, in the industry,
11:17
or they may be just getting started saying, I, I, I've seen this thing chat GBT I don't know how it can help my business, but my CEO told me I need to. And so we've created a process by which to engage customers to ensure that we meet them at the proper place and can add value and help them with our methodologies and process all the way through. And I'll say that we have a very good
11:39
experience in this. Um, right after GPT was released in November of 22, um, we actually invested into our own internal platform we called Project MindSpark. This was a generative AI platform. We created a closed environment. We leveraged a, a cloud-based uh back end. Um, but it was all private context, our ability to now integrate rag information into that.
12:02
Now we've been doing this for 2.5 years. I can say that we have over 65% of the SHI employees using this on a, on a regular basis. We have over 300 use cases deployed on this platform, and we're seeing productivity gains of 4 to 6 hours per employee per week.
12:20
That's a 10 to 15% improvement in productivity. So we've gone through this journey and we've learned a lot along the way. And so we bring this experience of practical internal use to the engagements in the approach that we make with customers. And it starts by being able to educate and orient our customers on our experiences,
12:38
and the point in the ecosystem that we sit, we see all of the vendors. We work with all of the OEMs, we work with all of the ISPs. We understand all the different solution types, so we can orient those that are just getting started. We can go through and identify easy win use cases. We can go through, uh, briefings and workshops,
12:58
and of course, we can make sure that your business is ready. Um, the, one of the biggest challenges right now that's driving, um, uh, a lot of fear in CIOs and, and business leaders is that Your business may not be ready to take on some of these generative AI uh uh generative AI solutions and patterns, because your business itself might have not have a culture to be able to be ready
13:22
to adopt. You may not have your data in a good place. You may not have the infrastructure, or you may not have the security around all of that. And so, we have, uh, we have workshops that we can take our customers through to ensure we can find the right use case and ensure that the business is gonna be ready. Because right now, the FOMO that's driving the desire to invest in this area is also
13:45
introducing this concept of FOMO or fear of messing it up. And that's where SHI, AI and cyber labs, why we why we introduced this to the market. Um, in this environment, we actually can create prototypes, not just POCs but full prototypes. Um, there is a stat that Gartner put out late last summer that 85% of all generative AI projects are failing post POC.
14:12
Well, this isn't a POC. This is a full prototype where we actually allow customers to bring in their sanitized data, which we do verify and validate upfront to ensure that there's no PII or uh the compliance of that. But we validate and make sure that that use case is gonna be able to be deployed, and we do this in a rapid succession, in a rapid iteration manner.
14:31
And again, we do that before we talk about deploying it. We do that before we talk about the landing zone or the infrastructure that you wanna buy, but of course, we're very good at that as well. And this is the approach that really you have to recognize that the traditional approaches with traditional infrastructure don't work anymore because a lot of this infrastructure,
14:51
take the GPUs themselves, take the latest NDL 72s from Nvidia that weigh over 3000 pounds that used 120 kilowatts per rack and require liquid cooling. Those aren't the standard run of the mill data center infrastructure projects that most companies are prepared for. And so you've got to recognize that a lot of this is really about the right choices, but iterating through that as rapidly as possible.
15:15
And so SHI's AI generative AI approach is about building those prototypes to both prove the value and do the proof of concept to ensure a much higher likelihood of success and improve that 85%. To less than 50% of failure. But we do this in rapid succession, and we do this in a manner that we like to call our I imagine experiment and an adopt framework.
15:39
The imagination part of this really starts with what's the art of the possible with AI. We've built our platform to allow customers to come in and test out standard use cases. We've built a rag studio. We've built a fine-tuning of the service capability. We've built out multimodal chatbot. Uh, we've built out uh agentic event planners. We built out PDF report writers that ultimately
16:00
can take all of those as standard use cases and in a two-week rapid succession, I've got a, I got a question. Let's go. Uh. But like the. Yes, and, uh, is, uh, how with the company or the auctions how to, uh, get the.
16:28
To get out of data, how you figure out all those rules just to take it to one how to assess what data sets, not just. Me all the. I, I know we talk at the house. So the way that we approach that is first of all, and I'll go through a little bit of this,
16:55
we bring together our team of 120 data science experts that include full stack engineers that have the experience of dealing with different data sets, and we prove and validate that the customers' environments are ready, but we also bring together the business leaders. The end user community and the IT infrastructure teams together and that we call
17:14
it the AI buying triangle. If the three of those parties are not aligned with the generative AI use case project, it's usually got a higher likelihood of failure. And so it's important that all business stakeholders are part of that, and that's part of the process we go through in our use case assessments. It's part of the process we go through in our briefings, and of course,
17:33
in our readiness assessments. And when we engage into a very specifically defined use case, that's part of the process that we go through in our um experimentation process, which is where we refine the use case for prototyping and ensure a higher level of success. But I can tell you there's no guarantees in this. And I'll explain to you in a second why The
17:55
approach we're taking can become very, very impactful and what I'll say, non-career limiting in the approach that we take. But when we get there, we define a very specific SOW to ensure that the customer's data is ready in there, that we will work through and. I'll show you the, the prototyping process before we even consider selling you anything
18:17
beyond an engagement that is, that is a consultative engagement. I will tell you that these are paid engagements, but they're, it paid, they're paid for at a much lower level, higher, much lower level of risk that it, that, that companies right now are just going and spraying and praying or buying and, and having that infrastructure sit unutilized or under optimized.
18:39
Um, and of course, when we move past the the experimentation phase, we move then into the production, the production phase and that adoption. The other part of success in in generative AI projects is to ensure that those that have to run these new platforms have been trained on them. Those that, that, that ultimately need to operate with them,
18:57
that are gonna be the users have access to them and they meet their requirements. And so a big part of this is that requirements gathering upfront. And that's why we go through this part of the process, which is A standard 6-week sprint, we do this in rapid succession. It starts with identifying that use case, defining the critical success criteria,
19:16
formulating which solutions are gonna work, ensuring that which patterns in the experimentation can be done, selecting the right models. There are millions of models in the market today, so it's important that you have an expert that's been there and work with, with many or not most of them, like, I don't have a million people that we've done models with,
19:34
but that understand the classifications. The impact of size of the model. An 8 billion parameter model, much different than an 830 billion parameter model. Cooperation with show his definition. So if today Mark just about take the average LEW.
19:55
Yeah, yeah, yeah that that. Common told you when they decide who to see the actual. And both those position to just that's something you will. They, they, they can engage us at any point along that process to identify which use cases may work
20:23
for their business or refine and ensure that the ones that they've chosen are gonna be successful and I can tell you we work with some of the world's largest companies. Give you an example here in a second. That even some of the most advanced solutions that had been developed, we've proven that this process is highly impactful in the way that it ultimately
20:42
delivers value to our customers. Because we bring together a team, a full team that we dedicate for those that six week period of data science, full stack engineers, the UX, the quality engineers, those that are doing the dev opps engineering, the integration components, these become added resources to our customers as we build those prototypes out.
21:04
And this is that example I mentioned, one of the world's largest global uh global uh uh pharmaceutical companies. They came to us with a use case that was an internal, uh, internal chatbot for information around their different drug discovery components that they needed to roll out on a global basis.
21:22
They were convinced that they were gonna need a super pod scale implementation. That ultimately was an 8 almost 9 figure purchase decision. Uh, so, uh, where do you, uh, knowing when. I to it at that. But I like that. I feel that we find that bacteria.
21:49
And how that Forgot what that. Well, so first of all, we define, we define the success criteria up front and their rapid succession in two weeks' sprints, where each part along that process, we're defining. Part of the reason we do a full prototype is to ensure a much higher likelihood of success for those gremlins to not get introduced.
22:14
It's selecting the right model. It's validating in that model works with the implementation and the and the RAG integration that you're looking to do. You're what. Once it's been deployed, um, we're not working on solving that problem at the moment. We're working on creating successful projects, but it's a,
22:34
it's a point well taken. The. that We we're supposed to do other stuff we can identify it also for his uh but. Yeah. So, and, and when Gardner defines that 85% failure rate, the reasons for failure are technical and human,
23:00
and somewhere in between the two. Yeah. And, and it's, and it's, you know, it could be technical and not being able to keep up with the latest model deployments. It could be technical in the sense that you don't know how to run a Kuberneti's cluster. We see those sort of, uh, one of the biggest challenges that we're seeing customers run into,
23:16
they're not used to running Kubernetti's clusters at scale. Maybe they need port works. Right? This is where we introduce those sort of things. But this example and this, this customer, they, they, for a low six-figure engagement, proved that a high 8 almost 9 figure purchase wasn't gonna work at that very
23:37
moment. We're continuing to work with this customer on refining their data because it was a data quality issue that they ran into with this one. And so, for the value that we're looking to produce is ensuring that rapid succession, failure fast, failure softly, and ensure that those next iterations are have a much higher level of success likelihood.
24:01
So where do we do this? How do we do this? Yeah, I've talked to you about the methodologies, I've talked to you about the patterns. Well, I mentioned the AI and cyber labs. Now, the A and cyber labs, while there is a physical location, is also a full hybrid and cloud-based nature,
24:16
but it's not just a place or a thing. It's actually an experience, a platform, and a team. Well, if I can get to the next slide, it would be cool. Um, it's an experience, a platform and a team of advanced experts that accelerates customer innovation. To imagine the possibilities of what the,
24:35
what advanced AI solutions can do for their business, to experiment and rapidly validate those solutions, and to adopt for long-term positive ROI. So, um, quick, um, just housekeeping note on this one. A lot of self-aggrandizement in this one. it's, it's, it's me talking, but we opened this facility, uh,
24:56
about, about 45 days ago in, uh, 35 minutes south of New York City, um, of where we've deployed our core heart, our infrastructure-based solution. And this is just a quick vignette of what happened that day, and I'll kind of speak to what it is. We're really excited. There's a great buzz in the air for the opening
25:19
of SHIS, AI and cyber lab. As you can see, everything's up and ready to go. We've been putting a tremendous amount of effort to move at the speed of light, and this is the manifestation of that. We're super excited to be able to be here this morning with representations from the
25:34
government, from large business, from our partners. It's really an amazing, amazing sight. Super excited to be able to share this with everyone. I mentioned my trip This photo is now a meme in our group chat because they meme me me for the excitement I had, but I'll tell you that the culmination of what we launched in this,
26:07
for me as having been a 23 year operator and CTO of advanced, you know, using them, some of the most advanced technologies and then representing. one of the most advanced data platform companies for 7 years and now being one of the, the most advanced integrators, it was a pretty pivotal moment because not only do we release this environment to allow our customers, we created an experience with this.
26:31
And it's important to note that that ability to demonstrate to customers what a real practical use case experience is like, something like a digital human or a vision AI solution, or a digital twin and ultimately creating uh robotic uh quadrupeds or, or robotic dogs. These are the sort of things that we're able to demonstrate and show to customers in a collaborative fashion,
26:53
and then use this this space in this environment to allow customers to experience the platforms themselves. I mentioned the platforms. Um, Pure Storage was actually the first partner that implemented their, their, uh, flash blade solution into the lab.
27:09
Um, we really appreciate the partnership, but I will tell you that we have over 60 GPUs that we can tie that platform into to validate these use cases across multiple different OEM vendors, Nvidia's DGX systems. Uh, Dell, HP, Lenovo, Cisco all have GPU systems that can be tied in and validated. So not only can we validate the use case and the pattern,
27:33
we can ensure that the infrastructure, if it is an infrastructure decision, will operate in its most efficient way so that the full stack is optimized. I also mentioned that we have that ability to train our customers on how to operate these systems. One of the things we did, this facility, by the way, is located inside of SHI's 400,000 square foot data center factory.
27:57
This facility was built in 2019 to produce data center scale racks, which we are now producing, in the, the realm of thousands per year. Um, but we also, in that environment, had a data center. There's data center, which is where we validate our customers' racks before we ship them. In order to deploy this lab, we used half of one of our build room data centers and added
28:20
liquid cooling technology. We now have 3 of the 4 major liquid cooling capabilities represented inside this lab, including hotile containment, which is using radiant heat cooling, uh, liquid cooling loops. Uh, we're using, uh, rear door heat exchangers that are directly tied into CDUs that tie into
28:39
the, the chiller facility of the plant, as well as, uh, direct to chip liquid cooling capabilities. The fourth that we aren't representing right now is full immersion because we're not seeing enough of a market, but we're keeping an eye on it. But the data center information management systems that are required to run that type of
28:55
advanced liquid cooling, you need to have the ability at hands-on training, and that's where our AI operation center comes in, where you can teach our customers how to operate and manage their base command manager if they're using an Nvidia stack and run AI if they're using that at cluster scale. We have every one of our major vendors platforms to be able to show and teach how to,
29:13
how to do day 2 management of these platforms. And of course, when we get to the point that we've proven those prototypes, work their custom work the customers to decide what they want to implement, we have that ability to then develop that at scale. I mentioned the 400,000 square foot data center factory.
29:29
Um, this is, this is really what all of this kind of ultimately outputs if it's an infrastructure decision. I mentioned that not all use cases will be deployed in a, in an on-premises environment. They may be full hybrid. They may use some of the neo clouds or NCPs that are out there.
29:46
And so building the right stack, the software layer that can be portable across all of those environments is also very important. So we've really introduced this capability of experimentation as a service, that ability to tie together, um, the, uh, uh collaborative approach with advanced experts, the way that the team that I, that I lead here, they've worked at Meta and eBay,
30:09
um, some of the largest consultancies. They've got experience working on some of the most advanced solutions. And that's where we build those how, how we build those rapid prototypes. We can do these in a matter of 2 or, you know, 2 to 6 weeks, depending on the scale and size of the project.
30:25
We do that before we talk about scaling and deploying. This is a this is a representation of really how the lab was built. We initially built it in a cloud-based platform leveraging bare metal infrastructure, leveraging containerization, um, and, and Kubernetes to be able to uh make it portable, and the initial platform we built was all built in an open source fashion so that we could,
30:49
we could leverage everything that was in the market. Um, we have the different use cases I mentioned earlier, the Rag studio or chatbot, but this is how we work with our customers. We can give them access to their data integrated in that RA studio and their and the multimodal chatbots. We have, uh, every one of the experts here available to work with our work with the
31:10
customers to ensure that successful project. And now as we've built out the advanced platform and the ability to consider build versus buy, when we, when we opened up the AI lab, we now have a full hybrid plane that we can deploy those workloads across any of these environments. So that ability to validate and test across numerous different landing zones,
31:31
numerous different CPU and GPU types, the ability to leverage different types of data platforms. This is how we help ensure success with our customers. And again, we do it with some of our key partners that are there with their most advanced solutions. We're continuing to evolve the ecosystem of ISVs that sit in that application and software
31:51
layer stack. This is, uh, this is just really a front end. Um, I'm not gonna run you through the eight minute explainer video on what the platform is and how it operates, but I will show you a couple of screenshots of this is how our customers engage. They log in, fully authenticated, um, integrated with your, your security teams. We work on all of this before we really engage,
32:13
um, allowing our customers to come in and use the event planner or Ragg studio or multimodal chatbot. Um, this is an example of the multimodal chatbot where you can see the output coming in audio or um or video output. Um, you can leverage the rag studio. You can see on the left hand side is just a base LLM, the right hand side is a rag model, and you can iterate through the different
32:35
chunking sizes that you want to use to, to, to do different types of embedding and, and, uh, and vectorization on that. Um, you have, uh, our agentic event planner. Um, this one is actually, uh, this, this one is set up to, uh, take you through setting up, uh, uh, an event like a wedding, where it can automatically, uh,
32:54
agentically go place orders for things like your, um, your florists and your, uh, your different vendors you need to work with. Yeah, the, they said the, the readiness assessment, the use case, uh workshops, um, and the, um, the briefings, that's all value add. But when we get to the point of integrating and working directly with customers,
33:22
specifically with their data, This is, this is all part of the platform engagement that you get along with the expertise. So custom SOWs are written for each one of those engagements. Um, and speaking of custom engagements, we have that ability to use that same platform to leverage some of the most advanced solutions that we have,
33:42
and I'll show you a great example of one of those that we announced at the GTC conference, which is our fine tuning as a service. Um, this is, I'm gonna let this explainer video run for, uh, I think it's 2 minutes, but this will give you a very good example of how the process works with both the platform and the team of how all of this comes together.
34:07
Welcome to FineTune, your enterprise solution for fine tuning as a service. Large language models offer immense potential, but to truly align them with business needs, fine tuning is essential. Fine tune simplifies this process, helping organizations improve model accuracy, optimize AI performance, and reduce infrastructure complexities,
34:29
all with the expertise of SHI's AI team. As a customer, you can easily request a new fine tuning job through the platform. Simply provide key details such as data set selection, upload training, validation and test data sets, model selection. Choose a base model that fits your needs.
34:51
Once submitted, SHI's data scientists take over the process, ensuring fine tuning is executed efficiently. You can track the progress of your request in real time, review model performance, and compare the fine-tuned output with the original model to assess improvements.
35:11
Data scientists play a crucial role in initiating and managing fine tuning jobs requested by customers. Upon receiving a request, they validate the provided data sets and configurations, execute pre-training and fine tuning with optimized hyper parameters, monitor job progress and provide real-time updates.
35:31
Evaluate the fine-tuned model's accuracy using. Side by side comparisons. Once the job is completed, the model is delivered to the customer, ensuring it meets business requirements. If needed, additional iterations can refine the results further. The key benefits of fine-tune include simplified AI customization.
35:51
Customers request fine tuning while SHI handle execution, faster AI deployment. Reduce fine tuning time from months to weeks. Enterprise grade security data is encrypted and accessible only to authorized users. Optimized model performance. AI models are fine tuned for higher accuracy and efficiency.
36:12
With fine tune, enterprises can easily tailor LLMs to their unique needs without handling the complexities of model training and infrastructure scaling. Whether you're a customer requesting AI customization or a data scientist managing fine tuning jobs, FineTunne provides a seamless, secure, and efficient solution. Get started today and unlock the full potential of AI with fine tuning as a service.
36:37
So as you can see, we're doing some pretty advanced things because fine tuning right now is one of the more challenging aspects of refining LLMs to create them to be more customed, to be, create them to be more responsive, to reduce the latency and the integration of that. Um, there are plenty of other use cases we're, we're building. You'll look, you'll see one next week that
36:55
actually takes this and makes it more engageable for the end customer. But that's really what we're doing in our AR lab. So to try to wrap this up, you know, really what have our learnings been? what are the best practices and what are the recommendations you can all walk away from? Well, first of all, it starts by understanding the strategy that you have to set is the most
37:14
critical aspects. Set clear achievable business goals with your AI strategy. Define that success criteria up front, uh, focus on the high-value use cases, uh, to align with those objectives and, you know, ensure that your generative AI solution is right for that specific need. Go to experts and validate that. Um, and start small and expand from there.
37:35
It's, you know, that rapid iteration, the MVP approach is really what we've seen to be uh uh in a successful way. Understand that bringing together all of the people that are necessary to ensure success both in the planning phases and the adoption phases, that's really critical crop collaborative across all of your business, uh, units.
37:54
Um, and of course, motivate those to, uh, adopt. We have other aspects of this to. Help take what we've done inside of SHI and build that AI culture within your company. Um, the process by which you go about that's really critical, understanding how your intellectual property rights, the certain compliance and governance elements that you need to achieve, understanding that you need to be following the
38:19
regulations within your industry. These are all elements that need to be brought into that collaborative. ideation approach, um, and establishing a monitoring system to ensure that biases and hall hallucinations don't, uh, creep into your models as they continue to grow and scale. And of course, the technology behind this, um, you know,
38:37
large models aren't the only ones. That fine-tuning process can do. Uh, do elements where you can shrink down large, uh, large-based models down to much more deployable models on smaller end points. Um, consider both proprietary and open source models, recognizing that cloud based, uh, closed models are leveraging AI or APIs may be an approach, but it may be better to actually take open source models,
39:01
bring them into your own environment, and, uh, and ensure that you, you can refine them for your own needs. Um, you can see the rest of these, consider the computational needs and costs. A lot of people start with this last element, they think, oh, I'm just gonna go and find the lowest cost per token.
39:16
Well, if you can improve your tokens per second, you can also reduce your cost. If you can shrink the model down, you can reduce the amount of uh load that you need from those, uh those back end GPUs that are running that process. So there's plenty of elements on the the technology side. Um, so these are really the top 5.
39:33
is begin with those well-defined goals, initiate on a small scale, and conduct those experiments as necessary, but maintain that short-term and long-term plan, optimize that process and refine it. Waterfall doesn't work. It's a very agile model. Um, and adopt a platform strategy that, that really is flexible and that archi architecture
39:51
can adapt, um, and emphasize security, compliance, and privacy is your top priorities and establishing that, uh, that, uh, monitor continuous monitoring feedback. That last slide I had, these are the top 5 takeaways from that. The final thing I'll talk about is really the ROI analysis of this. How do you actually validate successful projects?
40:11
A lot of people just look at the operational costs, your direct operational costs of licensing a model or your build costs to just build this, um, or your operational costs. Well, the other side of that is what are your quantitative benefits? How do you measure that from time savings? I mentioned what SHI has done with the MindSpark platform.
40:29
That's a very, very demonstrable and meaningful impact to the business, um, reducing cycle times, improving your direct cost reduction by, uh, reducing the amount of rework you have to do, or, you know, creating additional revenue. On the qualitative side, think of the employee experience that you're implementing.
40:46
Think about the customer experience you're implementing, and are you staying innovative? Are you becoming a, a leader in your business? Are you a laggard in your business? And of course, finally, it's understanding the impact of cost in the ROI on your technical risks, um, your operational risks and your business risks. And I'll wrap it back up with this, which is those patterns I talked about.
41:06
Each one of these different patterns have different measures that you, those different types of ROI analysis need to be applied to, because time to completion for a creative work is very different to time development time reduction. Um, in software development, reducing cycle times and error rates, um, customers, uh,
41:25
uh, uh, support automation, reducing those handling times, time to first response, personalization, are you seeing better engagement? Are you seeing better conversion rates? Of course, on the Content creation, which everybody in the room pretty much raised their hand on. Are you reducing your production time? Is your quality of content going up?
41:44
Is your engagement metrics uh going up? Like, you know, you can measure that as my SEO performance going up because I'm using AI to write my, my web pages now. So just in final closing, um, again, just a shameless ad for SHI but I will tell you that we have taken the most advanced approach to a full stack engineering solution.
42:03
It starts with our AI first culture. It goes into our unique combination of an advanced data science team and uh proprietary NexGen lab platform, and that ability to do pre-purchase decisions before you go and make these big, but I'll call career limiting uh decisions. Uh, rapidly validating those concepts, and that ability then to take this on a global scale.
42:25
We are a global partner that we can help you across the, across the globe. And of course you've got to work with the best partners in the industry, which is again why I'm here at the Pure Accelerate conference. I want to thank everyone. Um, you can learn a little bit more here.
