Written by the Samaipata Editorial Team
This is the first episode of "No-Filter Talks", an interview-based video series that invites our Operating Partners to speak openly about their experiences, offering valuable lessons and sometimes hard-to-swallow truths for our founders’ success.
In the rapidly evolving world of Artificial Intelligence and Machine Learning, it takes a seasoned expert to cut through the noise and deliver unvarnished truths.
Julien Simon, the Chief Evangelist at Hugging Face and a valued Operating Partner at Samaipata, is precisely such an authority. His extensive background includes 6 years of experience at Amazon Web Services as the Global Technical Evangelist for AI & Machine Learning and 10 years in leadership roles, such as CTO and VP of Engineering, at various large-scale startups. With a wealth of experience in AI and a keen eye for the practical applications of Machine Learning, Julien is here to share his insights and challenge preconceived notions in the tech industry.
In this episode, Julien doesn't hold back as he delves into the common pitfalls and strategic imperatives of implementing AI within businesses. From the necessity of unique data to the critical importance of clear business objectives, he shares invaluable advice for anyone looking to leverage AI for a competitive advantage. He also tackles the tough topics of startup culture, hiring practices, and the alignment of interests between founders and their teams. This interview is not just a conversation, it's a wake-up call to the industry, urging leaders and innovators to think more critically about the role of AI in their ventures.
Whether you're a startup founder, a data scientist, or simply an AI enthusiast, this is one conversation you won't want to miss.
Samaipata: How can a founder prove or show that they are building something defensible on top of AI?
Julien: At the moment, tons of companies are looking at AI to improve workflows or improve the user experience or automate things in the company. The first question I ask those customers and those organizations is: "Is it really worth it? You know, do you have enough unique, high-quality data to build something that will give you a competitive advantage?" Let's say you are in the automotive industry. You have tons of manufacturing data: supplier data, etc., and everybody loves their data... but doesn't your competitor have similar data, or potentially the same? So if everybody ends up using very similar data to train AI models, where is the competitive advantage?
So I think in order to be defensible, in order to have the moat that everybody's looking at, you need to find use cases where you have data that is very difficult or possibly impossible for your competitors and your industry to find and use.
And there are obvious examples. Bloomberg, the financial company, has tons of unique data that no one else has. So they go and train models, right? Makes sense. Disney has the rights to Star Wars and the Avengers, and there's only one Luke Skywalker and there's one Captain America, no one else has this intellectual property. So if they go and train models on that, it's going to be unique. So that's what you need to do. You need to look, you need to find the Luke Skywalkers and the Captain Americas in your data, train models, and that will create a competitive advantage.
Samaipata: How can a founder assess the value of integrating AI into their products and business models? How can they calculate the ROI on it?
Julien: A lot of the customer conversations that I start with want to do AI. Fine, and sometimes customers already have a list of use cases that they think makes sense, but in a lot of cases, they haven't thought it through. And I can kindly challenge them on what's the business metric that will show success. Let me take an example. Every website, every retail website out there, every insurance company, every bank website for the last five years has had some kind of chatbot. You know, the thing goes blip in the bottom right corner. What do we all do? We close it. Why do we close it? Because it's useless. So what's the ROI on that? Zero.
Unfortunately, it's very easy to go into projects that look sexy and exciting and disruptive without actually thinking it all the way through. What are we even trying to achieve here? Are we trying to automate something to save time and money? Are we trying to create a better user experience that increases engagement, which could be more clicks, more conversations, etc., whatever the metric is? But what is that metric? So my advice is, for every single machine learning or AI project, define clearly what question you're trying to answer, and that should be one sentence, non-technical, that you can write on the whiteboard saying, "We want to increase user engagement by 5%," or "We want to automate customer support and reduce answering time to four hours," and then define one or two business metrics that will actually show success. And that's what you give to your Data Science team.
So that's the thing: if for the project, you can't state the business objective in one sentence, and you can't put KPIs and track them all the way through the project, don't do machine learning.
Samaipata: What are the top three pieces of advice you would give to an experienced founder in order to create a successful product out of your experience?
Julien: Creating a startup is, you know, a hugely difficult thing, and there are a million things you'll need to look for. And I've worked in startups for many years, and I'm back in a startup now. I think the biggest mistake—on top of, of course, finding product-market fit and what you will read in the books and probably learn in school now—is hiring the wrong people. I've seen companies literally die from that, or teams completely crumble because of that. And it's a harsh thing to say, but every job interview I've done and every job interview I'm doing now starts with no. My answer will be no, prove me wrong.
Then obviously, I'm asking questions to the candidate, and I'm trying to find things that make me want to say yes, but it has to be a strong yes. If I have the slightest doubt, it's a no.
I've managed big teams, and I've always preferred being massively understaffed than having to deal with head cases and underperforming people. The second thing is you need to automate immediately. A lot of companies think they can live with manual processes. It's like, "Yeah, we'll do it later," but no, if you don't start doing it now, you won't have time later, right? Because later means, hopefully, if you're successful, more users, more traffic, more customers, more business, and the problems only get bigger.
You need to build automation immediately, and that's also why AI is important. Nobody should do translation manually. Nobody should do summaries manually; nobody should do all that stuff manually. Models are here. Use them from day one.
The last piece of advice, and this is a little more controversial, so apologies: don't pretend that your interests are aligned with your team's interests. It's not true. Every single startup I've worked in, the founders will tell you we're all in the same ship, and we're all aligned. Yeah, well, hopefully, we're aligned on being successful and building cool stuff, but the founder's personal interests are never aligned with the team's interests. So be clear about it because I've been around the block a few times, so I know that, but younger folks, engineers, staff—they can buy into that.
Then obviously, when hard discussions need to take place because the company is being acquired or because it has to close down, there's going to be a lot of hard feelings. So be clear about what your interests are, up to a point, and be clear that you obviously involve your teams. Give them stock options and bonuses. Keep them happy, keep them involved, but don't pretend your interests are aligned. It's not true.