Sep 18, 2026 · 1h 34m
Open robotics meets the fight over AI control
Hugging Face Co-Founder on Open-Source, Microduck, and NVIDIA | E2339
The episode connects open-source robotics and local models to a larger question: how should societies manage powerful AI without stifling useful innovation?
- 1Hugging Face makes open-source AI commercially viable through model storage, infrastructure, and compute close to GPUs.
- 2The Micro Duck shows how accessible hardware could bring programmable robotics beyond traditional engineers and into classrooms and homes.
- 3The hosts argue that AI policy should favor audits, liability, and defensive security over unrealistic universal shutdown mechanisms.
Don't miss
Thomas Wolf turns the cute Micro Duck into a concrete argument for open, programmable robotics that ordinary users can teach and adapt.
The brief
Thomas Wolf, Hugging Face’s co-founder, explains why the company built the open-source Micro Duck: a deliberately approachable robot that makes experimentation and programmable behavior tangible.
The robot leads to Hugging Face’s less obvious business model, built around storing models, providing infrastructure and compute, and keeping AI data close to the GPUs that use it.
Wolf connects open models and local AI to sovereignty, cultural customization, and defensive security, drawing on Hugging Face’s experience with a cyberattack and the need to protect systems as well as probe them.
The hosts challenge California’s proposed AI kill switch, arguing that targeted audits, liability, and identity checks are more workable than a universal emergency button for frontier models.
The episode closes by moving from robotaxis and AR glasses to a broader communications problem: AI may gain trust through small, practical benefits rather than grand promises.
What was said on this episode
34 statements · 26 positive · 5 negative · 3 neutral
MicroDuck is surprisingly capable relative to its small size.
“it's actually surprisingly capable for its size”
Listen at 6:07
Behaviors learned in MicroDuck simulations will likely transfer to physical robots.
“when a robot will ship in a couple of months, we think this will transfer very likely”
Listen at 7:57
MicroDuck orders exceeded 16,000 units.
“a little bit over 16,000”
Listen at 8:09
MicroDuck will likely become the best-selling moving robot.
“that will be very likely the most widely sold robots”
Listen at 8:15
AI storage must be fast and physically close to GPUs.
“Speed is the key issue there and getting close to the GPU”
Listen at 12:46
2026 is a particularly strong year for open-source models.
“this is the year of open source”
Listen at 14:10
Hugging Face may sell 50,000–100,000 MicroDuck units in coming years.
“we might sell up to 50 to 100K of this unit in the coming years”
Listen at 16:37
AI coding assistants will likely let users train new MicroDuck behaviors without coding.
“most likely... It just will work out of the box”
Listen at 17:51
Hugging Face intends to focus robotics development on entertainment and education.
“We really want to push this. entertainment education area”
Listen at 19:20
Thomas Wolf assigns a very high probability to AI doom.
“my PDU is very high”
Listen at 24:22
Thomas Wolf expects substantial AI-related danger in the world.
“I think there will be a lot of DUC in the world”
Listen at 24:23
Hugging Face believes open models could eventually be used by everyone.
“We think everyone could be working with open models”
Listen at 26:27
Hugging Face could increase its builder community by at least tenfold.
“we could push that. at least an order of magnitude more”
Listen at 26:31
Local AI will become increasingly important and useful.
“we're very bullish on local AI”
Listen at 27:47
Open models provide resilience because users can run them independently of centralized services.
“open models are... They're resilient. You can still run them”
Listen at 28:41
Open-source models are less than one year behind frontier models in 2026.
“the open source model, I would say, are in 2026. They're less than one year behind, for sure”
Listen at 32:06
Fine-tuning can cheaply produce improved models from company data.
“you basically put your data in, you get your better model out”
Listen at 33:30
Greater diversity among AI models improves ecosystem resilience.
“More diversity is just better”
Listen at 34:33
Open models are the best way to democratize artificial intelligence.
“it's just the best way to democratize AI”
Listen at 35:12
AI will contribute to cybersecurity solutions.
“AI is going to be part of the solution as well”
Listen at 38:13
The cyberattack on Hugging Face was not highly sophisticated.
“The attack on Huggy Face was not so complex”
Listen at 39:17
MicroDuck is profitable on a per-unit basis.
“we're actually making benefits on each of them. Yeah, making profit.”
Listen at 41:35
California’s proposed AI kill switch is technically unrealistic.
“It seems pretty farcical”
Listen at 45:04
Dangerous AI APIs should require biometric identity verification.
“Maybe don't make them available through an API unless somebody has a biometric and a... a verified user account”
Listen at 46:34
Dangerous AI systems should not be released publicly.
“How about you just don't release it?”
Listen at 48:23
Current AI is nowhere near capable of taking over the world.
“I don't think it's anywhere close, like not anywhere close to being able to take over the world”
Listen at 52:29
Reaching 30% of global rides may require roughly 120 million autonomous vehicles.
“we're going to need about 10 million cars a year, 120 million cars to hit 30% of rides globally”
Listen at 1:05:06
Within five years, nearly all new cars will include some autonomous capability.
“you won't be able to buy a car that doesn't have autonomy into it to some extent”
Listen at 1:06:45
Snap’s AR glasses represent a commercially promising product model.
“I think this is a winning model”
Listen at 1:13:11
Snap AR glasses could succeed at under $500 with four-to-five-hour battery life.
“I think under 500. And when they last four hours, five hours of battery life”
Listen at 1:13:12
Consumers may begin widely wearing AR glasses within 36 months.
“Maybe it's 36 months, and then we'll start to see people wearing them”
Listen at 1:14:02
AI can already perform many practical tasks that make daily life easier.
“there are a hundred things that AI could do to make your life easier right now”
Listen at 1:19:35
An AI travel agent could save users $1,000 on flights by monitoring prices.
“every day it looks for the best price and says, hey, I saved you $1,000 on your flights”
Listen at 1:20:30
Physical 100 is an inspirational and highly recommended show.
“highly recommended physical 100. It's inspirational.”
Listen at 1:32:18
Statements are attributed to the speaker as said on the episode and reflect their view at the time, not PodLume's. They are not advice.