Qantm AI CEO on AI Strategy, Open Source Models and What’s to Come

Seth Dobrin, the founder and CEO of Qantm AI and former chief AI officer of IBM, joins the AI Business Podcast to discuss how organizations can set themselves up for success when deploying AI, as well as the true cost of open source and the idea of technological colonialism. He writes a monthly column for AI Business, called 'AI iQ.'

Listen to the podcast below or read the edited transcript.

In your November column for AI Business, you had an interesting premise: Companies should ask the right questions in AI deployments. Have they been asking the wrong questions? What are the right questions to ask?

This whole thought process started when I joined IBM in 2017 and it has not changed. At IBM, part of my job was transforming a Fortune 500 company using data and AI. And they wanted me to share my experience, not just with IBM, but with IBM customers. I spent the better part of the first two years flying around the world and I authored an article on my findings titled ‘Data Science is a Scam.’ And that is because the leaders and companies were asking the question, ‘How do I use AI?’ And not asking what I pose is the right question, which is, ‘What are my strategic business objectives? And how does technology help me accelerate those or meet those?’

Data, AI and today, generative AI, are going to be part of the solution. When people ask me for help implementing generative AI, the first thing I talk to them about is their business strategy and strategic objectives. Let's figure out how you use technology to shore up your moat and accelerate the strategic objectives around that defensible position.

Can you share with us any examples where companies looking to implement have gotten their goals right?

Let's look at WPP, one of the largest media companies in the world. One of their businesses, Wunderman Thompson Data, had this massive dataset that they had never been able to use. Previously, Wunderman Thompson Data built custom marketing campaigns for clients using a small number of features or elements of data. They wanted to use the entire dataset, which included 25,000 features or data elements. Wunderman Thompson Data amplified the business model around their moat, which was this massive dataset to create truly custom marketing plans.

Another example was my former company before I joined IBM, Monsanto, where our moat was decades of agricultural data. We used testing data from all the crops that we had grown to transform the process growing crops using DNA markers, data science or AI to accelerate things like disease resistance and drought tolerance. At Monsanto, we also looked at accelerating yield using conventional breeding technologies. The strategic objective of the company was to feed the world but there were more strategic objectives, such as to improve yield, and by improving things like disease resistance, we were able to improve overall yield.

Take us through step by step how companies would align these processes with the business goals. Is it specific for each company? Or is there a blanket approach?

I’ve been honing this approach to a strategy over the last two decades. It starts with sitting down with the C-suite and finding out their core business objectives. And then understanding how technology is going to help accelerate those. So, taking the business objectives, looking at the current state of technology, and which of those business objectives can technology help accelerate, and building out a plan at a very high level of where we want to focus the technology.

The next level is going down the lines of a business leader that sits in the C-suite and understanding what are the use cases that are going to help accelerate those business objectives, then mapping out those use cases, and setting up for the actual data science teams to then take those and execute. And you align those use cases to KPIs that matter to the business.

The C-suite doesn't care how many models you have, the C-suite doesn't care about recall or your precision. They care about how much money you're saving, how much money you're making, and maybe cost avoidance in some cases. So those are your KPIs. And they need to be done in agreement with the CFO so that when they roll up, they matter.

There are some other pieces involved, one is AI governance - what are the policies, controls and frameworks that are needed to enable that? So how do you make sure that when the teams are building systems, they understand what the guardrails for your organization are, what's within bounds, and what's out of bounds. AI governance done right does not inhibit innovation, it accelerates it because it focuses within your organization where people can innovate

And then education, which needs to happen from the mailroom to the boardroom. Everyone needs to understand, especially in today's day and age with generative AI, or AI in general, what it is, how it can be used and how it's potentially going to impact them.

Would you say the C suite is more engaged now considering they're interested in capitalizing on generative AI? How have you found their involvement now compared to two years ago?

Yes and no - it's kind of spotty. In private companies, particularly private equity-owned companies, there is a lot of focus at the C-suite level on this and that is being that's driven by the private equity firms because they have a lot at stake. They are expecting a certain amount of return and there is a potential risk to their investment, so they want to understand the risk and want to address that risk head-on. Some conversations are being mandated by some private equity firms, at the C-suite, and even at the board level of these medium to large companies, not Fortune 500 companies, for the most part, which are happening.

There are a few public companies that are having these conversations. I would say a small number of Fortune 500 companies are having these conversations at the board level. And then the smaller public companies are starting too as well. I've gotten some inquiries about board seats in C-suite conversations at some smaller public companies, but for the most part, if you look at these Fortune 500 companies, those conversations are not happening. And this needs to be taken as seriously as cybersecurity. And it's not.

In Europe, boards are evaluated so people on the boards are evaluated. But in the U.S., there's not that requirement. A lot of times boards are there for cronyism, there are customers on boards who don't add value, aside from kind of quid pro quo, you're on my board, I'm on your board or you're a big customer, we want to keep you close. That's value, but it's not in my mind the way to run a company. They are not bringing people onto boards that understand AI or data that are going to push the C-suite to have these conversations and challenge the C-suite in this area.

You wrote in one of your earlier columns that the industry focuses on the big players. But are there any wider considerations for the smaller firms looking to implement AI?

It is a cost and talent problem for small to medium-sized businesses (SMBs). These models are not cheap to run. Talent is expensive and scarce. Smaller equity-owned firms have an advantage which is their backers can choose to put money into them to help fund these efforts.

There are some commodity tools smaller businesses can use. Even open source – yes, it’s “free” but open source is not free. Open source has never been free. There are a lot of support costs, compute costs and you are not going to run your open source models as efficiently or effectively as Microsoft or Google. Open source is going to cost you more to run it than it probably would be for you to just buy the license.

SMBs need to look at commodity tools, they need to focus on addressing challenges that they can buy off the shelf and then build on top of those and add their own IP into those. But they need to make sure when they add their data, it stays proprietary to them, especially in smaller niche industries, because their data is their IP and it's their differentiator. The other thing that I always tell companies when they're looking at these tools is that there is a lot of risk that's baked into these tools and a lot of unknowns around regulations.

Companies need to make sure that they are indemnified. If you're buying a model or tool from some startup and are using a custom model or an open source model, are they indemnifying you against it? If not, I wouldn't touch it. It's worth the extra cost to go to one of the bigger players. As much as I like supporting startups, if they're not indemnifying my business and it's a core part of my business, I'm going to be hesitant about using it because I don't want to inherit any of that risk.

Are there any open source tools or benchmarks that could work for businesses then?

There are some open source models that you will want to keep an eye on. But there are only two truly open source models – BLOOM and Falcon, because they open source the model and the data. You can go in with your legal and privacy teams and inspect the data and understand the data that's in there, which minimizes your risk because you can make decisions about that.

The model I’m excited about is the Falcon models built by the Technology Innovation Institute (TII). I'm also excited about the Gemini models that came out from Google. They're truly multimodal models, meaning they're trained on language, video, audio and images, not just text.

Also, Apple’s paper published last December where these models are really small and can run on your phone: Taking some of this incredible power that models have and slimming them back once they're trained, so they can run on your phone. I wear hearing aids and these models are being made small enough so they can run on hearing devices. Imagine being able to do language translation directly in a pair of hearing devices. That's going to be possible in the near future from things like that.

What Meta is doing is pseudo-open source. You can use it up to a certain amount of usage, once you make a certain amount of money, you have to pay them. But on top of that, their data is not open source. They are not giving you access to their dataset so you can't make evaluations of the risks that you're inheriting in those models.

In terms of benchmarks, there are two benchmarks you want to keep an eye on. And both of those benchmarks are from private companies. One is the Hugging Face benchmark, and that looks at the performance of models. The other benchmark comes from a company called Galileo. And they published a hallucination index where they've ranked models but also developed some methods specifically to reduce hallucinations.

What is to come in the next six to 12 months? You wrote a piece reminiscing about 2023 being a crazy year for AI and the Gen AI wave. What are you expecting to come?

From a technology perspective, 2024 is going to be a year of verticalization. You're seeing all this investment in chips. If you look back at compute, we went from specialized compute to generalized compute to specialized compute looking at GPUs. And now we're going get to a very specialized compute where each of these companies that are building their models are going to have their own specific architecture for their foundation model. This will make them more performant, it will make them more energy efficient, because there will be more performance, which will reduce the overall cost of training and running them while reducing the environmental footprint.

Smaller models – we’re going to see big models being trained and then being made small. We’re also going to see more multimodal models being developed this year. Google has released Gemini but I imagine we’ll see something from Meta, TII and OpenAI will release GPT-5 and it will be truly multimodal. Those are going to be the big things to happen between now and the AI Summit London in June (held by the AI Business’ parent division, Informa Tech).

The six months after that we're going to see new types of generative AI models. The current architecture of these models is not sustainable, so I think we'll see new types of generative AI. There's this whole concept of agent-based generative AI - imagine a bunch of humans trying to generate stuff and solve a problem. You bring together various language models all trained on different things, generating content and arguing with each other about what the best content is. Then there's this whole other field called neuro-symbolic learning, which is where AI truly reasons. There's the opportunity to combine that with generative AI.

The other thing that I hope we start talking about this year is the concept of technological colonialism. There's a lot of talk about addressing bias specifically and this is just one part of technological colonialism in these models. …

Since these models are primarily being built in China and the West, … with any foundation model, which is a pre-trained deep learning model, we are essentially inflicting our social ideals on the rest of the world, specifically, those south of the equator. We are then forcing them to adopt our social constructs of bias, which don't necessarily impact them, and ignoring their social constructs of bias.

We need to make sure that they are involved in the development of these models, they are involved in a definition of bias. And maybe we need to rethink if we want to control bias in these models, or if we want to look at something called disparate impact, which is where you actually address negative impacts in the outcomes. That's an important distinction that we need to think about that. We often focus on the model and we don't care what's in the model, we care what's in the output.

The other aspect of this is the socio-economic difference. These models have the opportunity to accelerate divides between the haves and the have-nots, whether it's the global digital divide between north of the equator and south of the equator, or it's socio-economical divides within society. This technology can either help minimize those if we include these people in the development and use of these technologies, or it's going to create a whole lot of social strife and further divide these communities. We need to be very mindful about how we address this whole issue of technological colonialism.