Shenbo Xu Podcast Transcript
Shenbo Xu joins host Brian Thomas on The Digital Executive Podcast.
Brian Thomas: Welcome to The Digital Executive. Today’s guest is Shenbo Xu. Shenbo Xu is a MIT PhD and member of technical staff and research scientist. Across years of research at the MIT-IBM Watson AI Lab, Shenbo’s work centered on data-driven decision-making under uncertainty for real-world problems, where being wrong has consequences business executives often face.
Shenbo has since built alpha models at Point72 and engaged in training foundational models at Scale AI. That grounding in modern model training and high-stakes prediction shapes the causal engine at the core of Kapnova. He has experience in AI startups, AI research labs, and the buy side, as well as full-stack AI research, data infrastructure, model training, and software development.
Well, good afternoon, Shenbo. Welcome to the show.
Shenbo Xu: Hey, Brian. How we doing?
Brian Thomas: Great. It’s awesome to finally get to meet you, at least virtually here. I appreciate it. You’re in New York, I’m in Kansas City, and I just really appreciate you making the time to jump on the podcast. So Shenbo, if you don’t mind, I’m gonna jump into your first question.
Your path is deeply technical. You’ve got a PhD in AI at MIT with research at MIT-IBM Watson AI Lab, pretty impressive, then building alpha models as a quantitative researcher at Point72 and training foundational models at Scale AI. That’s an unusually wide arc across academia, the buy side, and frontier AI labs.
What drew you into causal interference specifically and how did all those experiences converge into building the causal engine at the heart of Kapnova?
Shenbo Xu: So, I will start with the question that prediction answers. So pure prediction answers what will happen. Causal inference answers what happens if we do something about it.
Those are two different questions, and most AI back then, and still today, is only built to answer- The, the first one. My path is pretty simple. Looking back, MIT taught me that distinction rigorously. 172 showed me what happens when you bet real money on getting this wrong. Then I helped building the kind of infrastructure that could make the right answer to that question scalable.
Kapnova is really just the those three things put into one system the formal discipline from the research side, the real money rigor from the buy side, AI system can actually go run that discipline for you without needing a whole data science team standing behind it. So the short version is what used to take some PhD, a trading floor, a seven-figure budget to just ask what will actually happen if we do this, that’s now something a marketing team can run on their own
Brian Thomas: That’s awesome. And thank you. I appreciate you’ve got just a, a great career background in your education and your training and academia, all that. But a- at the end of the day, you’re doing some really great work in… what I took away here is, how you predict that right answer at scale. And with the tools and the research and the things that you’ve done this allows, like you said, marketing teams to be able to come in and do this stuff.
That was always a challenge in the past, so thank you. Shenbo, your research has centered on data-driven decision-making under uncertainty, where being wrong has real consequences, the same reality business executives face every single day. For a listener who’s heard correlation isn’t causation a thousand times but never really internalized it, why does that distinction matter so enormously when a company is making high-stakes decisions? And what does causal AI actually do differently?
Shenbo Xu: So correlation just tells us two things move together. Causation actually tells you whether pulling one of these levers will move the other one. And a high stake decision is never really something like, did the two things move together in the past? It is, if I do something right now, what will happen in the future?
And mixing these two questions up exactly ex- the, where the most expensive days come from. So, what does causal AI do differently? Instead of just fitting a pattern to the histo- historical data, it builds an actual model of what really driving what. So, before the company spends the money the executive can ask, “What would happen if I did this?”
And get an answer that actually holds up, not just a pattern that happens to hold true in the past. A one example is from a, from retail is Apple stores don’t use discounts, and they sell the most per– but they sell the most per square foot in the world. The CEO of a reta- of a big retailer copied this with no coupons, no sales.
But penny… But customers, but but the retailer’s customers were very different. They expected discounts and the sales crashed. He lost his job in the two years so in terms of what causal AI actually do standard AI is stuck at seeing. Causal AI adds doing and imagining, and those extra layers are what makes the system safe, efficient, explainable, and trustworthy in deployment
Brian Thomas: Thank you. Appreciate that. And you did break apart correlation versus causation. Correlation, you talked about the two items moving together as an example, and causation was if you take one of those away or, or change one of those items does that, does that change the outcomes?
And you talked a little bit about how leaders executives, CEOs, et cetera, that need to make big stakes decisions with this technology prediction scenarios, you’re able to get more of a predictable answer in order to make the right decision. So again, I appreciate your insights. Shenbo, there’s a theme running through your work about decisions where the cost of being wrong is real, not academic.
In a business context, where do you see leaders most often fooled by predictive AI that looks confident but can’t actually explain why something happens? And what’s the danger of acting on a model that only knows correlation?
Shenbo Xu: So, the most dangerous thing a predictive model does isn’t give you a wrong answer. It gives you a confident one. And it has no idea whether that number is is causal. It’s just found a pattern that held up in the past, and it’s compressed in the model and has it backfill you dressed up as a, as insight.
The best example I know of is eBay. eBay actually ran the experiment. They turned off branded search ads, the one that pop up when you literally type eBay into Google just to see what would happen. Their model had said those ads were driving huge returns. When they turned them off, almost nothing changed.
The people clicking were, were already typing eBay, and they were landing on the site either way. The eBay has been paying for years clicks that doesn’t doesn’t do anything at all. And that’s the pattern behind almost every ROAS number you’ve ever seen. Return on ads spent, which is ROAS, justifies the revenue by spend.
If the customer was going to to buy anyway, you’ll get the full credits for the sale. Every dollar looks like it’s working. The metrics is in– built in a way that makes this mistake invisible. In short, prediction answers what will happen. Intervention answers what happen if we act. Those are different questions.
Most business decision, most business decisions or intervention questions. Most business model, however, are prediction models, and th-that gap is where the cost lives
Brian Thomas: Amazing. Thank you so much. Appreciate that. I, I thought it was interesting you said the most dangerous thing about predictive AI is a confident one.
Again, interesting there. You showed that example of eBay example on the search ads. When they did turn those off there was no change in engagement or spend. And you talked about that where traditionally ROAS, right, return on ad spend usually you get that justified budget based on the customer spend.
But again, you proved otherwise, so I appreciate those insights. And Shenbo, the last question of the day, a recent DeepMind result argued mathematically that any AI capable of adapting to enough distributional shifts must have learned a causal model. Some take that to mean causal reasoning is necessary for true general intelligence.
As foundation models and causal inter-in-inference continue to converge, where do you see causal AI heading over the next, let’s say, five or 10 years? And what will it unlock for decision-making that today’s predictive systems simply can’t deliver?
Shenbo Xu: Even that decline is landing on something we believed for years.
Pure pattern matching hits a ceiling. For an AI to actually handle something new it basically has to learn real cause effect, not just the observed patterns. The large language models are great are great at sounding causal and asking one what happens if we do this instead, and it gives you a confident answer.
But it is predicting what a smart sound answer looks like, not actually reasoning through the cause, the real cause and effect in a real world setting. It talks causality, but it, it isn’t causal. What this unlocks testing a decision honestly before you make it, instead of a trusting model built– that, that’s built in the past, which most people ref– which mostly reflect people already hunting for the deal.
The business can ask what happens to everyone if you if if the owner a– if the executive actually wanted to change the price. What this unlocks, knowing if a decision was actually the cause, a model can say a customer working to a high churn looked like the high churn risk. It can tell you if this specific decision is why they actually left.
The gap matters for compliance, for fixing bad experiences, for real accountability. What this unlocks, not breaking every time the world changes. Normal model learn a pattern, then fall apart the moment the world shifts. A model built on real me-mechanism keeps working because it understand why something happens, not just what hap- it happened to historically.
Honest trade-off. This needs you to bring some real structure upfront, some sense of some sense of what causes what, not just only applies to data. And it is harder to start with, but that’s the discipline this is the discipline where the next generation model demands
Brian Thomas: Thank you. Appreciate that. You highlighted or I’ll highlight a few things. You talked about the, that pure pattern matching was kind of the direction we were headed over the years in the past. But again, testing this decision-making this causal inference that you talked about while these models can help, you talked about that gap, and that gap not necessarily means that if a decision was made, it’s gonna you’ll know exactly what the outcome was.
But you– that gap you talked about was y- whether it was a bad customer experience or compliance, et cetera, and that does certainly help in that area. But I appreciate again, your insights today. You’ve got such a great background, and I, again, appreciate what you’ve shared with our audience. And Shenbo, it was such a pleasure having you on today, and I look forward to speaking with you real soon
Shenbo Xu: Thank you so much hels- for hosting, Brian.
Brian-Thomas: Bye for now.
Shenbo Xu Podcast Transcript. Listen to the audio on the guest’s Podcast Page.
