Why Canva’s Craft Model Beats Frontier Labs at Design | Stefano Corazza artwork

Why Canva’s Craft Model Beats Frontier Labs at Design | Stefano Corazza

MTS

November 13, 2026

Canva Head of AI Research Stefano Corazza discusses how generative AI is reshaping creativity, the human-machine interface, and content production.

0:01

SPEAKER_1

Hello, everyone. Welcome back to MTS. Today, we are joined in studio with Stef Corazza. He is the Head of AI Research at Canva. Stef, welcome to MTS.

0:12

Stefano Corazza

Thank you for having me. It's a pleasure.

0:13

SPEAKER_1

Thank you for joining. I think one of the big reasons I am excited to talk to Canva, I am particularly a big Canva user. You guys are at the forefront of a lot of things when it comes to creativity, how people are designing and building things, and you and your perspective as head of AI research, I want to get a bit of your insights on how AI is affecting creativity as a whole. So, excited to chat with you.

0:39

Stefano Corazza

Yeah. It's an amazing time we're living in. It's like a true renaissance, I would say, where we have this complete democratization of the tools for creation of content, and we are all reimagining basically the interface between the human and the machine if you want, and how can the machine serve the intent of the human. So, there's a lot to discover, and there's also user experiences evolve and change now on a monthly basis, I would say, which is pretty incredible as the capabilities become more and more powerful. So, yeah, it's a very unique time with a lot of change, for sure.

1:17

SPEAKER_1

You brought up an important point. What is the interface between the human and the machine? What is the best way for the machine to be able to understand the human's intent? Do you think our current model of chatbots are the best way to do that?

1:32

Stefano Corazza

So, I will give you the answer of like, in 20 years, it's probably going to be telepathy, you know?

We're going to have like helmets to be wear and then... It knows exactly what you want. In the morning, I can actually watch the movie of the dream I had last night, you know? That will be like the final destination. I don't know if we will ever get there, but anything basically that gets us closer to that goal, I think it's important. And so, yeah, we started with a lot of interface. There was like chatbots, but now, you know, more and more people use like image reference. It used to be one image. Now it's like tens of images of reference. And then people like annotate on the content. So we are seeing like more and more direct interaction with the artifacts generated by AI. And our focus specifically has been on generating content with AI, that is editable. Because we are aware that it's never going to be 100% what the person wants. It's going to be maybe 70%, maybe 80% if we do a really good job. And so it's important that anything that comes out of AI is as editable as if it was created by a human, so that, you know, the human creativity can play that role of taking to the 100% and be able to edit everything.

2:43

SPEAKER_1

Can AI ever develop taste?

2:46

Stefano Corazza

This is a very interesting question, because we are actually training some models for taste. But the really interesting thing is happening in the industry. If you look at, you know, even the top models, you know, we launched Kraft a couple of weeks ago, we're there on the top. But if you're looking at the variability in the capabilities of those models, it's probably already smaller than the variability that you find in taste, right? Even from, like, specific categories of maybe designers in some specific location doing a specific job, the spread of taste is quite significant, right? So I think AI is already getting to the point that the variability between the models is probably smaller than taste. And we are trying to teach taste to the models so that we can better interpret what the user wants, right? Because we can learn what taste the user has and if we can teach the concept of taste to the models, then we get closer in the human-machine interface.

3:44

SPEAKER_1

What's the hardest part about trying to teach a model taste?

3:49

Stefano Corazza

I think the hardest part is really, you have to create like rubrics and categories and like ways to really articulate taste in a way that is like very specific so that you can generate the data set, right? They can feed that. So, yeah, it sounds like a very subjective area, right? It's really hard to pair with science. And so I think that's the challenge, is bringing the science into the taste. It's really where the models then can make significant advances.

Read the other 20 minutes of this episode here.

$15 gets you this episode and 99 more. Credits never expire. After checkout, one link brings you back here to read it.

Not what you expected? Full refund within 14 days.

Thousands of transcripts fetched by people building searchable podcast archives

Fetch transcripts with code

The same credits work from a script or an AI agent. The demo key returns a sample episode in full, no card needed:

request
curl -H "x-api-key: pt_demo" \
  https://spoken.md/transcripts/1000651996090

Markdown with the speakers named, for your notes, your knowledge base, or anything that makes HTTP calls.

From $0.10 per transcript. No subscription. Credits never expire. Prices exclude VAT, added at checkout for EU customers. Not what you expected? Email us within 14 days with 20 or fewer credits used and we refund the pack in full.

Using your own key:

request
curl -H "x-api-key: YOUR_KEY" \
  https://spoken.md/transcripts/1000801757747