dariozancaThe following is a summarized and AI-curated English transcript of the podcast episode. While it aims to capture the key points and essence of the conversation, this is not a verbatim transcription. Some sections may be rephrased for clarity and readability.
Host: Dario Zanca
Guest: Alessandro Ferrari
Dario: This is my conversation with Alessandro Ferrari, founder and CEO of Argo Vision and professor at the University of Pavia. With Alessandro, we talked about numerous topics related to Computer Vision from research to the state of the art, to the pros and cons of open source, to the future prospects of this discipline.
Dario: Let’s start from the top. Computer vision is delivering amazing results in areas like medical diagnostics, self-driving, and image generation. Are we at the point where this tech is “solved”?
Alessandro: Not quite. It’s true that computer vision has matured a lot, especially in the past two decades. We’ve seen exponential growth in both performance and efficiency. But the field is huge, with hundreds of different tasks. While some are close to human-level performance—what we might call “human-ready”—others are still very much in development. It’s definitely not a solved problem.
Dario: Right. And even when results look impressive, they’re often cherry-picked from very specific experimental conditions.
Alessandro: Exactly. In research, it’s common to test on clean, curated datasets. But in the real world, data is messy, and models often don’t generalize well. That’s where we step in at Argo Vision—we try to bridge the gap between cutting-edge research and practical, industrial deployment.
Dario: What are the biggest hurdles in getting from a published model to a real-world solution?
Alessandro: Licensing is the first big one. Just because a model is open source doesn’t mean it’s free for commercial use. Then there’s domain adaptation—models trained on one dataset rarely work out-of-the-box in a new setting. Data collection, model retraining, and efficiency optimization all become necessary. Especially when working on the edge, like we do, computational constraints are a big deal.
Dario: That brings us to Argo Vision. Your company is known for focusing on edge computing and low-cost AI. How did that vision come about?
Alessandro: We started Argo in 2016, right as deep learning was taking off. We realized most companies—especially small and medium-sized ones—didn’t have access to the computing resources needed for these big models. So we developed a lean, proprietary computer vision stack designed to run efficiently on embedded systems and edge devices, not massive cloud infrastructures.
Dario: That approach feels tied to the broader conversation around the democratization of AI. How do you see that playing out?
Alessandro: Democratization is a double-edged sword. On one hand, open source has enabled a lot of innovation. On the other, there’s a kind of illusion that just publishing something makes it truly accessible. In reality, maintaining and supporting open source projects is expensive. We chose to keep our stack closed because it gave clients confidence and helped us stand out in a skeptical market.
Dario: Do you think AI will become a true commodity, like cloud storage?
Alessandro: To an extent, yes. We’re already seeing it. The barrier to entry is lower—people can download models from GitHub, fine-tune them a bit, and go to market. But this also leads to standardization and commodification, which can flatten innovation. We’ll need to find a balance between accessibility and quality.
Dario: Let’s talk about talent. The field’s evolving fast. What skills are most important right now?
Alessandro: The basics are still essential—understanding math, algorithms, and how to read research papers. But the mindset is increasingly important. It’s not just about knowing a tool, it’s about knowing how to approach problems. And then there’s resilience. You need to stay curious and adaptable but also avoid getting swept up by every hype wave.
Dario: Makes sense. Before we wrap up, I showed you an AI-generated image where I asked for “a completely empty room with no chairs,” and it gave me… a room with a chair in it. What do you make of these kinds of AI “failures”?
Alessandro: They’re actually fascinating. They reveal how AI models process information—often statistically, not semantically. They’re powerful, but not perfect. And that’s okay. AI is a tool. It can be better than humans in many tasks, but it still needs oversight. Especially in critical applications.
Dario: Looking ahead 10–20 years—what do you think will feel like science fiction today but totally normal tomorrow?
Alessandro: Integration between biology and technology—hybrid systems where neural networks work with biological data storage, like DNA. Also, specialization in AI hardware. Today, we’re running everything on general-purpose GPUs, but I see a future with chips designed specifically for training or inference. And maybe quantum computing will play a role. It’s still early, but the potential is huge.
Dario: Alessandro, thank you so much for this deep dive into AI and computer vision. And for your famous LinkedIn feed—genuinely one of the best sources for keeping up with research!
Alessandro: Thanks, Dario. That means a lot. I’ll keep waking up early to read and share papers as long as I can!