Ethan LockshinSales at NVIDIA focused on AI adoption

I asked Scott Dietzen to share his most contrarian view about the future of AI.

Name: I asked Scott Dietzen to share his most contrarian view about the future of AI.
Uploaded: 2025-12-09T15:08:29+00:00
Duration: 1 min 45 s
Channel: Ethan Lockshin
Description: In this interview segment, Scott Dietzen presents a contrarian perspective on the future of AI for enterprises. He argues that while the industry is…

9 Dec 2025 · 2 min

Watch on YouTube

1:45

In this interview segment, Scott Dietzen presents a contrarian perspective on the future of AI for enterprises. He argues that while the industry is currently obsessed with building ever larger, general-purpose models (AGI), the winning approach for businesses will be highly specialized, task‑specific custom models. He predicts a plateau in model performance as compute, data, and algorithms become widely accessible, prompting a shift toward open‑source foundations enhanced by extensive post‑training and distillation. For enterprises, the key will be leveraging retrieval‑augmented generation (RAG) to make internal knowledge—codebases, support tickets, sales interactions—available to these models without directly training on the data. Finally, he emphasizes the need for reinforcement‑learning techniques tailored to this setting, enabling models to better utilize proprietary data to improve customer service and internal processes.

Chapters

What's the most contrarian view you hold about the future of AI or the startup ecosystem that most people disagree with? Yeah, I would say the. It goes back to the point I made about models specializing in the future. Everybody now is chasing AGI. We want to get the largest, most capable model. And I certainly think that's a very interesting pursuit in general. But for enterprises, I think what's going to win out instead are highly specialized custom models, models that are very good at particular tasks. Again, just like we train humans for, they'll be faster and they'll be much easier to train and to post train if we focus on that. So as we start to see models plateau because everyone has access to the same compute, largely has access to the same data, are using the same algorithms, I think that's inevitable, that we're going to see plateaus start to happen. Then the shift, the shift is going to come. It's going to be open source models with a bunch of post training, distillation with a bunch of post training. I think for customers, customers, you know, it's going to be about making sure you can use retrieval, augmented generation to make your knowledge available to these AI models so you can put it to work. So that's, you know, the history of your code base and all your code reviews, or that's your customer support system and sales systems and all the interactions that have taken place. You can bring that knowledge to bear to better serve your customers. I don't think you'll train on that data, but I think that data needs to be available to the models and then we've got to come up with reinforcement learning that works in that setting to make those models better at putting your data to work in order to better serve your customers and partners and make your internal business processes more efficient.