Clem Delangue

The through-line

Almost everything Delangue says comes back to a single argument: AI is not a magic new species, it is a new way of building software, and the worst outcome is for that capability to sit inside a handful of organisations. He prefers to call it "software 2.0" [2], a paradigm in which "machine learning is becoming the default way of building technology" [4] and where the people doing the building are "the new software engineers" [2]. From that premise everything else follows, including his preference for many models over one, his insistence that companies build their own AI muscle rather than rent it, and his view that safety comes from distribution rather than containment.

The position has stayed remarkably constant while the stakes have risen. Early on the argument was mostly economic and technical: open models democratise a field that no single company can advance alone [4]. Later it becomes explicitly political and about safety, up to the claim that concentrated AGI is the dangerous scenario [2] and that open models were what allowed his own company to defend itself against an autonomous agent attack [3].

On one model to rule them all versus many models

The two philosophies differ mainly in "where do you allocate AI builders" [1]. The centralised view bets on models getting bigger and more general with the people who build them concentrated in one or a few organisations; the distributed view bets that all companies will train and build models themselves [1]. His reasoning rests on treating a model as essentially a code base: "a code base is good or bad depending on your use case, depending on your constraints, depending on what you want to do" [1]. Facebook has a code base that does what its users need, Slack has one optimised for its own, and AI is no different, so a consumer company should build models "that are going to be faster, cheaper, more efficient, and that's how you differentiate yourself" [1].

He is candid that the choice is a short-term versus long-term trade. Calling one model behind an API is faster and easier at the start, but the long-run risk is that "you don't really internally build the capabilities to actually do AI yourself" [1], you cannot optimise the models so they stay inherently more expensive, and you end up competing with everyone else on the same substrate without differentiation [1]. The analogy he reaches for is the early web: you could stand up a site quickly with the equivalent of a Squarespace or a Wix and feel good about it, and "that's the equivalent of using an AI API for me" [1]. Just as building a real technology product means writing lines of code, building in the machine learning paradigm means training and optimising your own models [1].

On what Hugging Face is and who uses it

He describes the company as the number one platform for AI builders, with over 5 million of them using it every day to find models, find datasets and build apps [2]. The scale metric he finds most telling is not the headcount but the tempo: over 3 million models, datasets and apps shared, approaching 1 million public models, and "a model a data set or an app is built every 10 seconds on the hugging face platform" [2]. The GitHub comparison holds in one specific sense, that it is the most used platform for a new class of builders, though he stresses AI is different enough from traditional software that the parallel will not be exact [2]. An earlier snapshot of the same trajectory: over 10,000 companies and more than 30,000 open models shared by researchers, with users ranging from Microsoft, Facebook and Google to new startups [4].

The part he thinks people underestimate is collaboration. "You don't build AI by yourself as a single individual" [2], so commenting on a model, a dataset or an app, versioning code and models and datasets, reporting bugs and adding reviews are among the most used features, precisely because AI teams have grown from five or ten people to thousands of users at companies like Microsoft, Nvidia and Salesforce working together on the platform, publicly and privately [2].

On following the community, and the pivot

The product direction has always been reactive in a deliberate way. Hugging Face began as a conversational AI, an emoji you could talk to, conceived as an AI Tamagotchi because Siri and Alexa were disappointing [2]. Building it required stitching together separate models for information extraction, intent detection, answer generation and emotion, which forced the team to think early about an abstraction layer able to hold many models and many datasets, and that layer became the company [2]. The trigger was Thomas porting Google's BERT from TensorFlow to PyTorch over a weekend and tweeting it; a thousand likes was enough signal to double down, and six months of adoption data later they told investors they were pivoting [2].

What followed was pure community pull: other scientists asked to add their models, including XLNet and the then open source GPT-2, turning a single model repository into a library and then a platform [2]. Users said they had models too big for GitHub, so they built hosting; users wanted to search and filter datasets, so they built that, "and few months later we realized that basically we built kind of like a new GitHub for AI" [2]. He draws two lessons out loud. One is that community-driven development is the main reason for the company's success, since none of it exists without millions of builders sharing models, datasets, apps, comments and bug fixes [2]. The other is for founders: three years in, with six million dollars raised, you can still completely change what you are doing and be better for it [2].

On how much openness is the right amount

His answer is that the job is to give companies tools to be more open than they would otherwise be, without forcing them [2]. Roughly half of everything built on the platform is private, used internally by companies for their own systems, and he is explicitly fine with that [2]. What matters is offering a gradient: a company that will not open a big model or a big dataset might share a research paper, and openness matters even more for science than for AI generally [2]. The underlying belief is that open AI and open science are "the tides that lift all boats" [2], enabling everyone to build and giving everyone transparency into how AI is and is not working. The internal culture matches, with social media accounts accessible to all employees [2].

On safety, AGI and decentralisation

He rejects the sci-fi framing and thinks the name of the field is doing damage: calling it artificial intelligence loads it with associations of singularity and acceleration, whereas on the ground it is a new paradigm for building technology that will keep improving the way software has kept improving [2]. Today's models are building blocks for AGI in the sense that we are learning to build better technology, but he does not read that as a path to a Robocop scenario [2].

His actual fear is structural rather than technological. "I'm incredibly scared of a non decentralized AGI" [2], because the risk peaks if a single company or organisation gets there alone. Giving access to the technology to everyone, not only private companies but also policymakers, nonprofits and civil society, is what creates the counter-powers and the safer future he wants [2][3].

On agent cyberattacks and what regulation should actually do

When an autonomous AI actor jumped from another company's testing system into Hugging Face, taking "17,000 actions taken in 4 and 1/2 days" [3], he treated the novelty as the story: the volume and speed were unprecedented, the attacker was not a nation-state or a hacker group but a prominent American company, and the defence used an open model [3]. He reported to the authorities as mandatory disclosure requires, and frames the cause without melodrama: it is a technology system built by engineers, and "engineers can make mistakes sometimes" [3]. Society already accepts bounded autonomy, the way you let a dishwasher run without checking every step, and you get a problem when the wrong product goes in or something leaks [3].

The policy conclusions are specific. Cyberattacks must stay illegal and inside the US legal framework, or the world ends up with constant agent-driven attacks and no accountability for the companies creating those agents [3]. He is unpersuaded by containment approaches, including the kill switch bill and pre-release review, on the empirical grounds that these incidents happened on unreleased models during development: "concentrating everything behind closed doors in just a few organization doesn't work" [3]. What he wants instead is mandatory disclosure of agent cyberattacks so everyone can learn from them, and publication of agent traces, meaning what the engineers asked the agents and what steps the agents took, so it is possible to tell a human mistake from a system mistake from an AI mistake [3].

Open models are the other half of the prescription, and the argument is about symmetry. Hugging Face could not have mounted its defence through an API because of guardrails blocking cybersecurity activity; it needed to run a model on its own infrastructure [3]. Open models "remove the asymmetry of power and capabilities" [3], and very powerful attacking systems against very weak defenders is exactly the dangerous configuration. He notes the letter from Jensen Huang of Nvidia and other tech CEOs urging more administration support for open models [3]. On the China question he is unbothered by provenance: they used a version from Nvidia that improved a Chinese model, and once a model is shared, American companies can modify, remix, host and run it themselves, which is part of what makes open models safe wherever they come from [3]. He also insists the framing should be two-sided, since AI is an opportunity to fix cybersecurity problems and to fix bugs, and should make the world safer as a result [3].

On transformers, transfer learning and the blurring of domains

He explains the transformer's importance through transfer learning: train a very large model on a huge dataset such as a scrape of the web, using mask filling or text completion, then transfer that learning to any other task, whether classifying sentiment or topic, extracting information, or classifying images [4]. Because it started beating the state of the art on every science benchmark while remaining easy for companies to use, it became the default way of doing machine learning [4].

What excites him is convergence. Transformers moving into computer vision, time series, audio, biology and chemistry means the lines between machine learning domains are getting blurrier, with Uber using them for ETA prediction and applications appearing in recommender systems [4]. The real prize is merging them: detect audio then analyse the text, or combine time series with NLP for fraud detection so the signal is not just the number of interactions but their nature, since a weird email raises the probability of fraud [4]. He also names open domain conversational AI, the topic the company started on, as the sci-fi dream that remains essentially impossible today but that he hopes becomes possible in a few years [4]. He is generally happy to see AI replacing existing capabilities, with search and social networks being rebuilt with it, while also unlocking things that were not possible before [2].

On how companies should actually adopt machine learning

His practical advice is to start from pre-trained models on the hub, which lets software engineers ship machine learning features without machine learning engineers, then "build the kind of like machine learning muscle" [4] with something small. The failure mode he sees is companies whose first project is a conversational AI meant to answer a hundred percent of customer questions on every subject, which is genuinely hard and is a years-later capability [4]. Classify customer support emails, extract information from them, build autocomplete for a chat interface, then compound from there [4]. He gives Grammarly for grammatical error detection, Bloomberg for summarisation in the terminal, and segment.ai for automatic image segmentation as examples of what real use looks like [4]. The mental default he expects is to ask first how a feature could be done with machine learning and fall back to writing a million lines of code only if it cannot [4].

The obstacle on the way to production is not technical. "Most of the challenges are more human than technology related" [4], because machine learning is much less deterministic and much less explainable than regular software, and product managers and founders who have built deterministically for twenty or thirty years have to get comfortable not fully understanding or predicting their features' outcomes [4]. His mitigations are all about transparency and fit. Model cards, the standardised way of describing what a model can and cannot do both functionally and ethically, including where it will and will not be biased, were pioneered by Dr Margaret Mitchell, who is at Hugging Face [4]. Easy demo tooling lets colleagues and customers try their own examples and see for themselves where a model works [4]. And the last mitigation, which he says people forget, is choosing the right use case: for a hiring tool, "you shouldn't use a machine learning model to filter resume" because it can be biased against women or minorities, so it should complement human reviewers and never be the only filter [4].

Takeaways

• Treat models as code bases, not as a single winner: quality is relative to your use case and constraints, and a consumer company should train models optimised to be faster, cheaper and more efficient than a general API [1].
• Calling one model behind an API is the fastest start and the weakest long-term position, because you never build internal AI capability, cannot optimise cost, and end up undifferentiated against competitors using the same thing [1].
• Openness should be a gradient offered, not a rule imposed: about half the models, datasets and apps built on Hugging Face are private, and that is treated as fine [2].
• The safety risk he names is concentration, not capability: "I'm incredibly scared of a non decentralized AGI" [2], and access for policymakers, nonprofits and civil society is what builds counter-powers [3].
• Pre-release review and kill switch style containment miss the mark, since the agent incidents happened on unreleased models; he wants mandatory disclosure of agent cyberattacks and published agent traces instead [3].
• Open models are a defensive necessity: guardrails on APIs blocked the cybersecurity work, so the defence required running a model on their own infrastructure [3].
• Adoption fails for human reasons more than technical ones, so start with a small feature, build the machine learning muscle, and use model cards and demos to make non-determinism tolerable [4].
• Keep machine learning out of sole decision authority on sensitive workflows: a hiring tool should use it alongside humans, never as the only résumé filter [4].
• Company direction can come from watching what the community asks for, which is how a weekend PyTorch port of BERT became a library and then a platform, and why a three-year-old company with six million dollars raised could pivot successfully [2].