CEOInterviews.AI
Start App
Thomas Wolf
Thomas Wolf
Cofounder and Chief Scientist, Hug
Watch

Hugging Face Co-Founder: We Should Take This Really Seriously

📅 Sep 02, 2026 etn. 17 MIN 88 VIEWS 31 SEGMENTS · 2 SPEAKERS
The European Technology Network is made possible by: Airwallex: ⁠https://www.airwallex.com/⁠ Elevenlabs: https://elevenlabs.io/ Polymarket: ⁠https://polymarket.com/⁠ Framer: https://www.framer.com/ BVNK: https://bvnk.com/ Base: https://www.base.org/ Metaview: https://www.metaview.ai/ HEX: https://hex.tech/

What Thomas Wolf said

Written from the verified transcript and checked against it. Every figure links to the moment it was said.

Thomas Wolf, cofounder and chief scientist of Hugging Face, discussed the viral Micro Duck launch, describing it as a long-in-development robot project that took off after engineers solved a self-righting mechanism. He emphasized that the launch aimed to prove the product was real, not a cherry-picked demo. On the OpenAI incident and subsequent discourse, Wolf argued that taking AI agent attacks seriously is crucial and that anthropomorphizing these systems can help non-technical people understand real capabilities, which he believes exceed common perception. He criticized the lack of transparency from major labs, noting unknown details about the incident's third wave, and highlighted that open-source models were essential for defenses, unlike closed systems that refuse cybersecurity queries. Wolf predicted open-source models would grow from an 80/20 split against closed source toward 50/50 in coming years, citing cost consciousness and usage trends.

Key takeaways

  1. Micro Duck was in development for over two years and only became viable as a mass product once engineers solved a self-righting mechanism about six months before launch.
  2. Hugging Face faces at least one attack per day, but this incident was unique due to the model's persistent, parallel attempts, marking a first-of-its-kind attack.
  3. Wolf predicts the token cost split between closed and open source will move from about 80/20 to 50/50 within two years.

Numbers and commitments

FigureWhat it refers toTypeAt
2 years Development time of the Micro Duck robot timeline 2:14
6 months Time since engineers solved the robot's self-righting problem timeline 2:14
5% Portion of the OpenAI incident story disclosed in the first blog post metric 5:44
80/20 Current split between closed and open source model usage metric 16:15
50/50 Predicted future split between closed and open source models guidance 16:15
1 year or 2 Timeline for open source models to account for a significant portion of token costs timeline 16:15

Chapters

  1. 0:00Micro Duck launch and development
  2. 4:57Reaction to the OpenAI incident
  3. 7:00Anthropomorphizing AI agents
  4. 10:47Transparency and future of the incident
  5. 13:09Defense and open source models
  6. 15:43Open source versus closed source trends

Questions asked in this interview

5
  1. 0:00Welcome to the show. First time?
  2. 1:50What are the ideas? Where did it come from?
  3. 9:57Do you feel like we're at the end, or do you feel like there's so much more discourse about it, and when might that stop, I guess?
  4. 12:30Cuz I think it is the first of its kind in terms of an attack like this. How was the reaction internally?
  5. 15:43What are your projections? What are your predictions?
Host 0:00 ↗
Thomas will be joining us shortly. What's up, sir? Do you like our new— We got them this morning and they're cracking away. Early access. Early access. How are you doing? How's it going? Welcome to the show. First time?
Thomas Wolf 0:19 ↗
Yeah. Very happy to be to be here finally with you guys.
Host 0:23 ↗
No, dude. We're super pumped. Thanks for making time. We have a lot to discuss. There was obviously loads of different news. You guys put out your own announcement. I have to say I thought that was a bit of a troll announcement. So, I would love to know what the strategy there was because when I first read it I was also thinking that I would be seeing something else, but it was the micro duck.
Thomas Wolf 0:45 ↗
Yeah, yeah. So, I mean the micro duck was the only thing we focused on this summer, but we did a lot of, you know, spoofing attempts to get people to follow us before we did the real launch. So, that was the big launch for us this week, the big news for sure. But, there might be more things coming. But, yeah, we were three times Hacker News yesterday on the top, which was three different types of news, which is quite impressive. But, honestly, micro duck has been way above the release, the enthusiasm, the virality, like seeing all the community starting to make these little movies, like all these training already. That's this go way beyond what we were expecting for this robot. I mean, I knew it was really nice. I mean, personally, I knew I was already planning to have a lot of these robots in my house, but like seeing everyone embrace the meme and embrace the duck was really nice.
Host 1:41 ↗
It was brilliantly executed. We were seeing numbers on the stats in terms of the rate at which you guys were selling—
Thomas Wolf 1:47 ↗
It's insane.
Host 1:48 ↗
It was something—
Thomas Wolf 1:49 ↗
Record breaking, right?
Host 1:50 ↗
Every 5 seconds you were selling one, and then the revenue was I saw VB was posting pitch decks of the new business being micro ducks. Can you maybe just tell us what were those 24 hours like? You rolled this out, you didn't expect it to go as well as it did. What's happened? What are the ideas? Where did it come from? Just give us the whole shebang.
Thomas Wolf 2:14 ↗
Yeah, sure. I mean, the idea dates back to a long time ago actually, more than 2 years ago. So that's one of our robots that has been the longest in the making, I would say. Even before we acquired Pollen, I was already side-paying for this. They were not part of Hugging Face, but basically I was buying, and the team here, we were buying some motors to get one of the engineers, the lead engineer on this project called Antoine, to just try to explore this duck idea. And then it really took stream a couple of months ago, well, maybe 6 months ago now, when we found a way to basically get the robot to stand back up. First, the first version, when they fell, they could not stand back up, and we said, "Okay, that's not really viable as a consumer product because you're always going to have to babysit this robot." And once the team managed to make this work, I think then we were like, "Okay, there is here serious potential for a real mass product." And then the main work was trying to make it as small, as tiny as possible, try to reduce it to make it the cutest possible, try to add as many sensors and capabilities so that people can go crazy when they get it. They can vibe code whatever they think could be fun to do. And then this summer we were in the last mile trying to think, "Okay, what's the good way? How can we show this is a real product and not just—" You know, we see so many demos in robotics, so many people showing very cherry-picked videos, and then you're like, "Okay, I don't think I can buy that in the coming year." It's obviously very crafted and probably maybe even AI-generated for some part of it. We don't know. So, here the idea was like, "Let's show this is a real product. Let's show this is something real. It's going to fall. It's not perfect, but it's good enough to actually have a lot of fun with it."
Host 4:12 ↗
Yeah, yeah.
Thomas Wolf 4:13 ↗
The launch was the usual launch, so you were a little bit worried. You don't see any reaction in the first 30 seconds. See that, and then you see the first like, the first retweet, and you're like, "Oh, maybe it's going to be a small community." And then after 1 hour, as you start counting, you start putting the slope in your code cutter or whatever, and say, "Okay, try to predict where this is going." And we're like, "Oh my god, this is going really, really big."
Host 4:40 ↗
[laughter]
Thomas Wolf 4:40 ↗
So, yeah, and then it was trying to ride a wave. The nice thing is we already had a pretty viral, I would say, launch for our first robot.
Host 4:50 ↗
Yeah.

16 more exchanges in this transcript

Sign in free to read the rest of this interview. No card required.

Sign in to read the full transcript

Cite this transcript

APA, MLA, BibTeX
APA

Wolf, T. (2026, September 2). Hugging Face Co-Founder: We Should Take This Really Seriously [Interview transcript]. etn.. CEOInterviews.AI. https://ceointerviews.ai/interview/1299875/

MLA

Thomas Wolf. "Hugging Face Co-Founder: We Should Take This Really Seriously." etn., 2 Sep. 2026. Transcript, CEOInterviews.AI, https://ceointerviews.ai/interview/1299875/.

BibTeX 
@misc{wolf2026_1299875,
  author       = {Thomas Wolf},
  title        = {Hugging Face Co-Founder: We Should Take This Really Seriously},
  howpublished = {Interview transcript, etn.. CEOInterviews.AI},
  year         = {2026},
  month        = {sep},
  url          = {https://ceointerviews.ai/interview/1299875/},
  note         = {Speaker-attributed transcript with timestamps}
}