Joe Green (TechForge):

Hi, I’m joined here by Jerome Gabryszewski. I hope I’ve pronounced that the right way, here in our little corner of the TechEx North America show. Jerome’s joining me today from some company called HP! And Jerome, your title is AI and data science business development manager?

Jerome Gabryszewski:

Yeah, you nailed it.

Joe Green (TechForge):

…which means you’re the technical face of HP who can explain to, let’s say, enterprise customers, some of the more technical code-based networking elements of the HP offering, but also present it in such a way that it’s business first. But before we get into this area of the business and HP. Let’s talk a bit about you. So basic life story. Were you an engineer? Were you a sales guy? Where did you come up and how come you ended up at HP?

Jerome Gabryszewski:

I ended up at HP actually by a happy accident, kind of. So I double majored in econ[omics] and finance in college

Joe Green (TechForge):

Which school did you go to?

Jerome Gabryszewski:

I went to a really small liberal arts school called Fort Lewis College in Durango, Colorado. Not a lot of people have heard of it, but I went to school there. I did a trimester in Germany, studied out there for a little while in Regensburg about an hour north of Munich.

Joe Green (TechForge):

Which part of Germany? I know, sort of, Southwest, Rheinland-Pfalz.

Jerome Gabryszewski:

Yep, beautiful area. I studied there for a while. I came back, finished my degrees and I was actually going to be an actuary at a law firm. I had a job lined up and my grandfather got sick and so I had to move back to take care of them. HP was the best opportunity that I could find in the area I was in, and it’s been that way ever since. I’m not a formally trained data science person. I entered as an HP as a sales resource. And then during COVID, HP helped pay for me to upskill myself. So I learned Python. I learned about data science. I learned before AI was the big thing; machine learning and deep learning. I learned how to code and program.

Joe Green (TechForge):

A bit of networking as well?

Jerome Gabryszewski:

Yeah. Still learning that actually. I’m in CompTIA+ plus certification training right now.

Joe Green (TechForge):

OSI layers and so on?

Jerome Gabryszewski:

So I’m doing that currently. But yeah, auto-didactically taught myself how to do a lot of the stuff, and HP has supported me along the way. I went from econ finance, to sales, to here; a very technical transition.

Joe Green (TechForge):

It’s like you’re late blooming into pure geekdom!

Jerome Gabryszewski:

Yeah, kind of. I guess in a way. Finance and economics, in a lot of the ways, where the math is concerned, I guess you get a little bit of a quantitative analysis and stuff like that. It all transfers over. But it was learning the syntax and all the rest of the comp-sci stuff, it was a bit daunting, but we got through it.

Joe Green (TechForge):

Okay, that’s great. So we can talk a bit technically. Now the only reason why I’m limiting that and saying “a bit” is because it’s my limit!

Jerome Gabryszewski:

Sure.

Joe Green (TechForge):

Obviously this is a trade show we’re at and so there are a lot of ‘big brush’ messages. But I think what’s quite interesting is I’ve got someone here in front of me who knows their technical chops, you know, has made major technical bones. A bit late in the day, Jerome, but you’ve made your technical bones! And what has come over from my research behind HP and what you guys do is that you are really quite concerned, not about necessarily deploying artificial intelligence, but it’s the stuff around it that you’re saying, “AI’s great, but wait. There’s all this detail, there are all these other things you’ve got to do and actually they can be a bit mundane.” You know, things like governance, like data sanitisation. Why these things? I mean surely AI is exciting, let’s just get on with that? Why are these peripheral items important?

Jerome Gabryszewski:

Well, because functionally the AI doesn’t work if you don’t lay the groundwork first. As far as data cleansing and processing, if you have a ton of data silos and you have information that isn’t communicating cross department. These are actually some of the problems at HP that we’re working on solving actively. But if you have instances where your finance department isn’t communicating with your supply chain department and those data systems can’t communicate, it’s really, really difficult for you to get to that.

[This is] the next phase – of agentic AI, right? Because this is a cliché statement that’s been going around for a long time, but if you have the whole age-old garbage in, garbage out, right? Now you’re using that same logic, but you’re allowing agentic systems to act autonomously. So if they’re acting autonomously on bad information or incomplete information, that’s a huge problem.

Joe Green (TechForge):

Yeah, I mean it’s an old issue. That sounds disparaging, “it’s an old issue.” What I mean is, it’s an ever-present issue and has always been thus. And you know, I think we used to call them data silos. We’re talking about data ownership. And these problems have existed, you know, since the pre-ERP stage of computing history. ERPs and other monolithic systems came about in order that “all data shall belong in one place, for thus is it written.” That didn’t necessarily work out in every case and so therefore software tended to split out again and we got the whole SaaS boom. So we’ve had more or less constantly, this separation of data, and you’re saying this is still going on to a certain extent. But now it’s particularly important, particularly critical, because we’re giving software its own mandate to go off and do things?

Jerome Gabryszewski:

Where everyone’s pushing very hard for automation is the big key, right? And so agentic systems automating workflows and making you more efficient, which is great in concept. But if those things aren’t built properly and they’re not using data that’s current or correct or relevant, you’re going to have a lot of issues, right? Because you’re going to have these things. The idea of agentic AI, eventually, is to have humans out-of-the-loop in some cases, right? And so if you have those human-out-of-the loop instances and the data’s bad, that’s a pretty scary road.

Joe Green (TechForge):

And we’re talking as well [about] governance, [which] isn’t just overseeing data cleanliness and making sure that all the information in an organisation is available where it should be. It’s also a case of data ownership. I’ve worked with some fairly big businesses and there’s almost a territorial aspect to it. I mean, I can understand that approach from, for instance, the finance department because they don’t want people to know their colleagues’ salaries. They don’t want the employees to know the investment plans of the company. So I can see why there’s a certain amount of protectionism going on. But I think that’s still that still happens to a certain extent in larger companies does it?

Jerome Gabryszewski:

Yeah, 100%. And I think also there’s a warranted reason and [around] a lot of what you just highlighted, there should be some protectiveness with your data. Your data in a lot of cases is your livelihood. You should be careful and protective with it. But having places where you’re not doing interdepartmental sharing is where you’re going to start to have issues. And maybe that is, you inter-departmentally share in the C-suite of your organisation, and then you do some sort of restricted base access control: your job level only. It allows you access to a certain level of information, but the information is still shared cross-vertical or cross-industry inside of an organisation.

That’s really, really critical. And I think to the point about governance are the times when people like me nerding out and building software and then deploying it out, or over the time of stakeholdership, is very much in now. You have to get everybody in the boat with you before you develop something or it’s not going to work. I’ve seen that. I’ve seen so many companies fail really good AI or machine learning projects because they had a small group of people that developed it and they didn’t check with legal, and they didn’t talk to these people and they didn’t talk to cybersecurity, and they didn’t involve these folks. And by the time you get all those people in the boat, it’s over. I say, get them in before.

Joe Green (TechForge):

That’s practical advice from an expert at HP. Now, in an interview that we did with Jerome before this particular show, we were talking a lot in a really geeky way, which was great, about kit, about hardware. And of course, HP’s stand here in San Jose is full of some really mouth-watering kit there on show, which is fantastic for the likes of me. There’s a lot of emphasis coming from you guys. Obviously you’re a hardware company: Do you see The Future, capital T capital F, do you see The Future of AI as being, let’s say, locally hosted. I’m not necessarily talking about four servers in the broom cupboard. It might be a private data centre. But it’s you who owns or leases the hardware. You run your stuff, your AI models, which are right there. And the likes of Claude and the big cloud based services hosted on AWS, wherever. That’s not really where it’s going to be at. Is that HP’s bet?

Jerome Gabryszewski:

No. I don’t think that we’re necessarily saying that the frontier models of the world are going to go anywhere. That, I think, would be a foolish statement for us to make. What I think, though, is you’re going to start to experience some pretty drastic increases in how much it costs to get into these frontier models, I think very quickly you’re gonna start to see these companies… I think you’re already seeing OpenAI start to explore ideas of IPO. And right now we’re in the honeymoon phase, I would say, with these frontier models, where you can get a subscription for about 20 bucks a month or something like that. Those days are limited because right now they’re all getting venture capital money and if they’re not profitable, it’s not really that big of a deal. But when they get to a place where they’re going to go to a public offering, you’re going to have a pretty substantial cost increase. So at HP, what we’re going to clients and customers and saying, “Hey, we understand that you guys are doing a lot of this work in the cloud and we’re not saying that you should stop.” What we’re saying is, let us help you build a better hybrid compute infrastructure to where you have options and use points to pull on, when you want to use or reserve your precious frontier cloud resource tokens, right? So basically our stance is things that you’re doing that are what I call science experiments. So anything that you’re ideating and you’re doing a lot of iteration on, maybe you’re fine tuning a model multiple times, so that you can get it correct – those are activities that you should avoid doing in the cloud unless it’s absolutely necessary. You should try your best to offload that work to an edge computer of some kind. And edge computers have come a long way. Workstations have come a long way with technology to support local AI development.

Joe Green (TechForge):

I can understand why the big AI companies – that’s essentially Anthropic and OpenAI – I can understand why they started out with a subscription model. You know, it’s how you pay for SaaS, it’s a subscription. I think what I’m seeing in the industry at the moment is from the enterprise tiers down, we’re moving towards a literally token-based payment model. And I think as that goes down through the different ranks, even down to the consumer level, I think all of a sudden your $200 a month has really got to be worth $200 a month to you because the chances are, it could be $200 a month, it could be $2,000 a month. There just isn’t going to exist particularly in the near future. And so therefore looking towards hybrid models and hosting things yourself is probably good, safe. It’s a good, safe way of getting insurance. Now you mentioned, you’ve got some real cutting edge stuff, but it’s going to be iterative. It’s going to burn a lot of tokens. Let’s not put that on the cloud.

What are the use cases for saying, “Okay, I’ve got these big racks, now of HP kit. Absolutely fantastic. What else am I going to run on the things that I own or lease?”

Jerome Gabryszewski:

Any of your training, fine tuning of models, you can host, and you can do some of that. That becomes more of a, “how many concurrent users do you have to support before you have to start making some judgement calls?” Of, “Do I burst this to the cloud or I continue to invest in my own infrastructure?” But we have people doing all kinds of stuff. We have anywhere from race car teams that have burned through all of their cloud tokens because they only get an allotment of so many, and so they’re relegated to having to use some sort of local instance. We help them do that.

Healthcare is a huge space for us because, for them, a lot of their data can’t be put in the cloud, even if they have a private instance or something like that. There’s policy in the US that prevents them from doing that, and so most use cases there. And then I would say downstream of that would be, basically, all of our federal government customers, anyone that’s touching secure data or working in a secure facility. They quite literally don’t have the option to use a cloud-based model. And so they’re having to work using local hardware, local data centres, local AI models. And what we’ve actually found in a lot of cases is local models will never beat a cloud or a ChatGPT, being a Jack of all trades, right? And then they’ll use a standard by which that they test models. So give it tons of physics questions, math questions, science questions, and then they grade the performance of it based on how it answers in all of those categories. So, in general level understanding, cloud will always be better because it’s larger. However, what we have found, what I’ve even found in my own personal use, is if you can fine tune a local model or something like that, and you can actually train it on data that you wouldn’t have been able to put into the cloud. Like those specific narrow domain specific tasks are actually better served using local models than from [the cloud].

Joe Green (TechForge):

Yes, this is a recurring theme I’m hearing actually, this fine-tuning for a particular niche. And of course, it stands to reason, doesn’t it, that the first big niche, if you like, that’s made even the mainstream news is coding? Because, of course, the people who develop AI models are coders – it’s going to be the first area of specialisation, isn’t it? Because that’s what they know. And actually, what’s going to be interesting next, I think, is which of the next specialist areas or specialisms is going to see a lot of activity and get that massive burst of productivity that coders have seen. What’s your bet, if you had to put money on the next niche, where fine-tuned, very carefully-constructed data inputs are all put together? What sector is your best bet on the next one to explode in the way that coding has?

Jerome Gabryszewski:

So I guess next future the one I feel a lot of investment is going is into agentic robotics. So actually building physical AI is where most of the venture capital, I feel, is heading in this area. Where we’re physically at now, I’d say that’s probably on the forefront. It will take I think more time than… it won’t be next year, right? We’ll have this but I think we’ll start to see the outset of that and then I think there’s…

Joe Green (TechForge):

Sorry to interrupt you. This is going to be, I’m assuming, in factories and engineering works and that kind of thing?

Jerome Gabryszewski:

But I think you’re even going to get to a place where you’re going to start to see a lot of those. You’re gonna start to see humanoid assistant robots that are gonna be tangible to the average individual, to be like buying a car. And that’ll be really weird and really cool and maybe scary and all [that] in one thing. But I would say, outside of that, there are other things that AI has already started to do a really good job at. Like, we have people that use local models to do real-time monitoring of CNC machines. Active failure fixing, anomaly detection, and research for computer vision language. I work with some really advanced biological researchers at a few universities where they’re generating petabytes of data a day. It’s not humanly possible for them to physically look at each one of those images, but they need to find anomalies for cancer.

Of that petabyte of data, 95% of it is going to be normal and fine and they can throw it away. But 5% maybe or less will be anomalies that they really need to take a look at. And how do you do that without using some sort of machine learning or AI system? And so I think even beyond coding, AI has done such a good job at taking tasks that were mundane, or took a crazy long time, or just weren’t tangible, and we’ve made it so that we can augment and become way more productive in those endeavours.

Joe Green (TechForge):

Is progress on these new areas, is that still going to be contingent on these really rather mundane issues like, you know, “Are all the zeroes zeroes, or are they capital O’s? Are there spaces before the numbers? You know, data cleansing, and governance, and privilege systems, and authorisation; these rather more mundane areas. Are they still going to be affecting these new sectors, do you think, as they start to emerge?

Jerome Gabryszewski:

Yeah, I think so. I do.

Joe Green (TechForge):

I’m trying to tease something controversial out of you! So is it a case of, do you think, that there’s a great deal of enthusiasm for AI but the actual reality on the ground is more complex than most people realise?

Jerome Gabryszewski:

Yes, 100%. I think two things on that. One, I feel like every business is taking what was classical machine learning or maybe even just old school Python scripted automation and just putting the words AI in front of it to extend it. So you have that camp. And then you have the other camp where the AI work is legitimate. But you have stakeholders who want to just inject the AI into places where maybe it doesn’t belong. And you have people who are jumping ahead to building the agentic systems without doing the back-end work of making sure that their data is of quality and relevance and usable.

Joe Green (TechForge):

This sounds to me like HP are grounding people, that you’re in some ways trying to say, “just calm it down a bit! Let’s get realistic about this. Let’s look at each problem in turn and work on each one.” Which I think is a good approach. It’s the way software should be built.

Jerome Gabryszewski:

And I think also the way that people follow the trend is you have all of these companies like Anthropic and OpenAI and these large frontier models, Mistral and all of these, who are in this never-ending race for artificial super intelligence. They want to be the first one to get there. And I feel like a lot of businesses are trying to take big bites out of AI in the same way as those folks are. And I think that that is maybe the wrong approach. I think what you should be doing, especially if you’re new to the world of AI, actually implementing it in your business, I would say go the opposite way and find purpose-built mundane things that are a little bit easier to tackle and give yourself a couple of quick and easy wins so that it becomes a little bit easier to get the stakeholdership to do the larger things.

Joe Green (TechForge):

So this is a sales pitch from a huge company saying keep it small, at least to begin with?

Jerome Gabryszewski:

Yeah, keep it small, keep it simple. I was at a federal conference a few weeks ago where I was talking to this exact topic of, “You guys can do this and you can do it locally. Start with easy things. You guys have probably tons of engineering documents in these labs, right? It’s probably a nightmare to query and search all that stuff. That’s an easy lift for an AI system using retrieval-augmented generation that’s been around for years. It’s a proven, tested thing. You guys can implement that in a couple of weeks, probably.” That’s a really, really underrated, very valuable thing that you can do that’s going to provide a huge win and a roadmap for you to continue on.

Joe Green (TechForge):

Okay, Jerome, I’m going to bring things to a close here. We were chatting earlier on before we started this interview, [and] Jerome is one of the few people on the HP stand today, just as circumstances fell out this way, who’s the technical voice on the show floor for the entire booth. So, his voice is getting a bit strained. I’ll take pity on you there. Jerome, thanks ever so much for joining us.

Jerome Gabryszewski:

Yeah, that was fun. good time.