Clément Delangue

1. Key Themes

Open Source AI Is Structurally Safer Than Proprietary AI

Clément argues that open source models are inherently less dangerous than frontier proprietary models because they tend to be more specialized and lack the concentrated resources needed to develop the most dangerous capabilities. He draws on a striking historical analogy to make this point.

"The most dangerous things usually are not so much developed in open source. Funny thing is that sometimes safety people are talking about the nuclear bomb. Nuclear bomb has never been built in open source. And I think it would have never been built in open source. It's been built in a closed source, proprietary team with billions of dollars of resources." [00:06:47]

Model Routing Is the Next Major Value Redistribution Event in AI

Clément identifies routing — automatically directing queries to the most appropriate model rather than defaulting to a single frontier model — as a structural shift that will move value away from dominant frontier labs toward a long tail of specialized models.

"It's possible that it's going to redistribute a lot of the value capture from frontier models, which have been the case now, right? Like majority of the revenue capture was on frontier models to a more like long tail of models, which in my opinion makes much more sense. It's like AI maturing." [00:19:21]

He cites a Stanford study as concrete evidence of the economic inefficiency of current behavior:

"There was an interesting study from Stanford published last year, end of last year, that was showing that 70% of the queries that people ask to ChatGPT could be accurately answered locally on your laptop." [00:17:54]

The Danger of AI Concentration Is Greater Than Any Competitive Threat to Labs

Clément explicitly inverts the distillation controversy narrative, arguing that the real systemic risk is monopolistic concentration, not Chinese labs copying model capabilities.

"We're heading toward a world where a few companies are completely dominating, concentrating all power, all capabilities, all wealth. And that's much more dangerous than them maybe losing a couple billion dollars of revenue." [00:22:14]

Open Source Has Proven It Has a Viable Business Model

Reaching $100M ARR while explicitly not prioritizing monetization is a signal that the open source platform model for AI is validated, analogous to what GitHub proved for software.

"At this small scale, I think it shows that there's a business model for open source. There's a business model for open source platform. We kind of knew it, right? Because there's been GitHub before and there's been a bunch of open source successful companies. But I guess it's a validation of that." [00:10:44]

Local Models Are a Distinct and Rapidly Growing Category

Clément outlines a clear value proposition for local models that is fundamentally different from API-based AI: privacy by design, zero marginal cost, and suitability for continuous agentic workloads.

"Local models are kind of free, right? Because they're running on your phone or on your laptop. So you don't even, you don't really have to pay for them. So they're much cheaper. They're much more privacy kind of like preserving by design because you don't have to send your data to an API, right? Your data stays on your phone." [00:11:52]

Open Source Provenance Doesn't Matter the Way API Provenance Does

Clément makes a subtle but important distinction: with open weights, the origin country of a model is irrelevant because the user has full control and transparency; with APIs, origin matters enormously because data is transmitted and access can be cut off.

"For open ways and open source, it doesn't really matter where it comes from because it's kind of like you get all the control, you get all the transparency. There's no way to kind of like trick you, bias you, manipulate you, remove your access." [00:14:40]

AI Regulation Risk Is Concentrated at the Frontier, Not the Ecosystem

Clément's core policy argument is that regulatory pressure should remain targeted at trillion-dollar frontier labs — who have the resources to absorb it — and not spread to startups, academia, or open source communities.

"I hope it's going to stay contained to a few frontier journalist models because frankly, I think they're the most dangerous ones. And also these companies are trillion dollar companies with armies of DC people. So they can kind of like deal with it." [00:04:08]

2. Contrarian Perspectives

Distillation Is a Non-Issue and Sympathy for Frontier Labs Is Misplaced

While the media framed Anthropic's accusations against Alibaba as a serious competitive threat, Clément flatly disagrees — both on the merits and on who deserves sympathy.

"If you suck, you suck with or without distillation. It's just kind of like a little bit like accelerating thing, but it's not what makes you good or bad at training models. So if you stop distillation tomorrow, the Chinese labs won't go down and disappear." [00:00:00]

"It's hard for me to say like, oh, poor Anthropic, poor OpenAI, you're getting unfairly competed with when you're like the fastest growing company in the world." [00:00:00]

Restricting Open Source Models Is Practically Impossible

Most policy discussions treat open source and closed source AI as symmetrically regulable. Clément argues they are not — restriction of open weights is structurally unenforceable.

"The way you can restrict it is very different. So for example, if you remove an open way from Hugging Face, then it's still going to be on Modelscope, which is like the Chinese equivalent, or it's going to be on torrent platforms. So restriction looks very, very different for open source than for APIs. You can't really block because it's open." [00:13:45]

More Powerful Open Source Models Won't Necessarily Be More Dangerous

The conventional assumption is that capability scales with danger uniformly across all risk domains. Clément argues you can build a more powerful model that is deliberately not more dangerous in high-risk areas like cybersecurity, by simply not training it on those domains — and that this obvious lever is being ignored in policy discussions.

"I don't think it's automatic that you build a more powerful model that it's more dangerous for cybersecurity. You could build a more powerful model that is not more dangerous for cybersecurity, if you don't train it on cybersecurity, which really people are not really talking about. Why are we not talking about that rather than talking about removing the ability to release?" [00:07:45]

The "Frontier" Is Not One Monolithic Capability

Against the industry narrative of a single race to general superintelligence, Clément argues that frontier performance is domain-specific and uneven, which has major implications for how both safety and competition should be evaluated.

"People right now are talking about frontier as kind of like this general thing. The reality is that the frontier is kind of like jacked between different tasks, different domains. This one model is going to be better at some things and it's going to be worse at other things." [00:10:09]

3. Companies Identified

Hugging Face

Open source AI platform, described as the GitHub of AI models. Crossed $100M ARR while not prioritizing monetization. Hosts the most widely used runtime for local AI workloads (Llama.cpp) and is a central hub for open weight models globally.

"At this small scale, I think it shows that there's a business model for open source. There's a business model for open source platform." [00:10:44]

Anthropic

Frontier AI lab, co-founded by Tom Brown among others. Named in the context of the distillation controversy with Alibaba and government discussions about restricting model releases.

"I wouldn't be surprised if Anthropic used distillation in the past for some of their models, for some of their specialized models using kind of like someone else who's better." [00:20:53]

OpenAI

Frontier AI lab. Named in context of GPT-5 government restrictions and the broader competitive dynamics discussion.

"It's hard for me to say like, oh, poor Anthropic, poor OpenAI, you're getting unfairly competed with when you're like the fastest growing company in the world." [00:00:00]

Mistral

European frontier AI lab. Cited by Clément as evidence that Europe already has labs at or near the frontier.

"You already have some great labs, right? Black Forest Lab, Mistral. All these people are doing amazingly and arguably they're at the frontier." [00:23:54]

Black Forest Lab

European AI research lab. Named alongside Mistral as proof of European frontier AI capability.

"You already have some great labs, right? Black Forest Lab, Mistral. All these people are doing amazingly and arguably they're at the frontier." [00:23:54]

Lovable

AI product/coding assistant. Specifically named as an early example of a company already implementing model routing under the hood — a leading indicator of the routing trend.

"Lovable is starting to do that too, right? Like doing the routing under the hood." [00:19:21]

Modelscope

Chinese equivalent of Hugging Face for hosting open weight models. Named as a concrete example of why restricting models from Hugging Face is practically unenforceable.

"If you remove an open way from Hugging Face, then it's still going to be on Modelscope, which is like the Chinese equivalent, or it's going to be on torrent platforms." [00:13:45]

Alibaba

Chinese tech giant. Named as the company accused by Anthropic of conducting distillation attacks.

"Anthropic accused Alibaba of doing distillation attacks." [00:20:22]

Open Router

AI model routing platform. Named in passing as the host of the "Open Fusion" model — a fusion of multiple models that reportedly surpassed single-model efficiency benchmarks.

"There's a model on Open Router called Open Fusion, I think. And it's this like fusion model of basically a bunch of different models. And that had a lot of advanced capabilities and like surpassed like inefficiency in a lot of ways." [00:16:42]

Google

Named as the origin of the transformer architecture, illustrating how AI ecosystems build on open research rather than isolated breakthroughs.

"The T of transformers obviously is coming from Google, that open source transformers." [00:24:22]

Stanford

Research university. Named as the source of a study showing 70% of ChatGPT queries could be answered by local models.

"There was an interesting study from Stanford published last year, end of last year, that was showing that 70% of the queries that people ask to ChatGPT could be accurately answered locally on your laptop." [00:17:54]

4. People Identified

Co-founder and CEO of Hugging Face. Guest. Central voice throughout on open source AI strategy, policy, and the routing thesis. Reached $100M ARR while deprioritizing monetization.

"It hasn't really been our priority to optimize for monetization and for revenue, given what we're building, which is more kind of like a usage-based platform to reach and empower as many AI builders as possible." [00:10:44]

Tom Brown

Co-founder of Anthropic. Named as someone Clément literally crossed paths with in D.C. during the GPT-5 government discussions — a notable detail about who was in the room during that policy moment.

"We bumped into Tom Brown there, like the co-founder of Anthropic. And we were like, oh, it seems to be like a change of staff or something like that before it was made public that they changed a little bit the people talking to the White House." [00:02:26]

Andrew Trask

Researcher at DeepMind. Named by one of the hosts as having flagged the Open Fusion model on Open Router — identifying the fusion/routing trend from a research perspective.

"Yesterday I was talking to Andrew Trask from DeepMind and he had like a very interesting viewpoint that there's a model on Open Router called Open Fusion, I think. And it's this like fusion model of basically a bunch of different models." [00:16:42]

Anton Leisch

Policy writer. Named as the author of a piece called "The Moonshot" on Substack, arguing Europe has the capacity to build a frontier AI lab if it chooses to.

"Anton Leisch, who's a policy writer, wrote this piece on Substack called The Moonshot, which basically explained how if Europe wanted to do so, they could build a frontier lab." [00:23:13]

5. Operating Insights

Don't Optimize Revenue Early When Platform Scale Is the Moat

Hugging Face reached $100M ARR while explicitly not prioritizing monetization — they optimized for builder adoption instead. The revenue followed the scale, not the other way around. This is a deliberate sequencing choice with major implications for platform businesses in emerging technology categories.

"It hasn't really been our priority to optimize for monetization and for revenue, given what we're building, which is more kind of like a usage-based platform to reach and empower as many AI builders as possible. But at this small scale, I think it shows that there's a business model for open source platform." [00:10:44]

Model Routing Is an Immediate Product Opportunity

Any AI product currently sending all queries to a single frontier model is overpaying and underperforming. Lovable is already routing under the hood. Operators building AI products should audit their model usage and implement routing now — the cost and quality gains are available today, not in the future.

"Lovable is starting to do that too, right? Like doing the routing under the hood. And I think it will, it's possible that it's going to redistribute a lot of the value capture from frontier models to a more like long tail of models." [00:19:21]

Ecosystem Building Beats Individual Company Heroics

Clément's read on what made OpenAI possible — and what Europe needs — is that breakthrough companies emerge from ecosystems of open research and open source, not from isolated will.

"It's more an ecosystem, right? And you see that from OpenAI, right? The T of transformers obviously is coming from Google, that open source transformers. So it's more of a matter of fostering an ecosystem of open research, open source AI." [00:24:22]

6. Overlooked Insights

CAISI (AISI) As the Quiet Institutional Power Center for AI Evaluation

Clément briefly names a government agency called "Casey" (almost certainly AISI — the AI Safety Institute) as doing "amazing work" on model evaluation benchmarks. This is a throwaway mention, but it signals where independent, scientific authority over AI capability claims is being built inside government — and which institution will likely shape the criteria for what gets restricted and what doesn't. For anyone navigating AI policy or building in regulated AI contexts, this is the agency to watch and engage with proactively.

"There's this agency called Casey that is amazing. I think they're doing an amazing job and they're building up this capability to really evaluate and work on benchmark and things like that. And I'm really excited for them to take a little bit more of the workload there and kind of like take a very scientific approach to evaluating these models. And I think when they will, it's going to be really good for the shields." [00:16:05]

Young People Are Skipping the User Phase and Going Straight to Building

In a single unremarked-upon observation, Clément notes that the cohort of young people on Hugging Face has already moved past being AI users and is now actively training models, building datasets, and building products — often in completely non-glamorous domains like climate, biology, and chemistry. This is a leading indicator of where the next wave of AI application companies will come from, and in sectors that are currently underrepresented in AI investment.

"I feel like young people went through the first phase of being users of AI really quick. And now a lot of them, I think, want to be builders. We see a lot of very, very young people going and fine-tuning models and building products themselves or optimizing training models themselves. Not necessarily in the most talked about domains, but also in topics like climate change, biology, chemistry." [00:25:29]