No PriorsNo Priors Ep. 5 | With Huggingface’s Clem Delangue
CHAPTERS
- 0:00 – 2:45
From eBay power-seller to computer vision: the trade-show moment that changed everything
Clem recounts growing up in France, running a prolific electronics business on eBay, and landing an eBay internship. A chance encounter at a trade show—plus a prediction about object recognition replacing barcode scanning—pulled him into ML and a computer-vision startup.
- •Ran a major French eBay seller operation and worked closely with eBay teams
- •Trade shows were dominated by PayPal support complaints directed at the shared booth
- •Met an engineer who predicted object recognition would replace barcode scanning
- •Did follow-up research, validated the person’s credibility, and decided to switch paths
- •Joined a startup doing ML for computer vision ~15 years ago
- 2:45 – 3:59
Hugging Face v1: building a fun “AI Tamagotchi” instead of boring assistants
Clem explains that Hugging Face began as a consumer chatbot product—a playful AI companion inspired by sci‑fi. The team spent nearly three years on this direction and achieved meaningful engagement at scale.
- •Founders wanted an AI project that was scientifically challenging and fun
- •Motivation: Alexa/Siri felt limited to productivity and “weather” use cases
- •Built an AI friend / talking-emoji Tamagotchi-style chatbot product
- •Operated this product for almost three years
- •Reached billions of messages exchanged between users and bots
- 3:59 – 5:37
The pivot to open source: Transformers, BERT traction, and following unexpected pull
The company’s shift toward open-source tooling emerged gradually as transformer models took off. As usage of their open-source components exploded—from a few users to hundreds of companies—the team reallocated effort and raised a Series A around the new momentum.
- •Open source was present early as partial releases alongside the chatbot product
- •Transformers/BERT era triggered a major spike in adoption of their open-source work
- •Progressive internal shift: from one cofounder maintaining it to the whole team
- •Signal-based decision-making: attention and excitement clearly concentrated on tooling
- •Series A funding reinforced the new direction as the primary company bet
- 5:37 – 8:12
Founder guidance on changing direction: balancing exploration vs. exploitation
Clem offers a framework for pivots and ongoing innovation: maintain a deliberate ratio of experimentation to execution both before and after product-market fit. Hugging Face institutionalized exploration, giving experiments time and resources to become major product lines.
- •Common failure modes: too much thrash pre-PMF, too little experimentation post-PMF
- •Aim for a healthy exploration/exploitation split to avoid local optima
- •Hugging Face targets ~30–40% of effort for exploration/long-term bets
- •Experiments need runway and ownership to mature into real products
- •Example: Spaces started as one person’s experiment and scaled to 50k+ demos
- 8:12 – 12:42
Getting the first users: ML Twitter, network effects, and builder-led community
Early growth came largely from Twitter, where the ML community amplified open-source releases. Hugging Face then benefited from marketplace-style network effects between researchers publishing models and industry adopting them, reinforced by a culture where every employee engages the community directly.
- •Initial distribution was driven by ML Twitter and organic sharing/retweets
- •Researchers shared models for visibility, pulling in industry users
- •Industry demand then encouraged more researchers to publish on the Hub
- •No dedicated community/PR hires early; community engagement was everyone’s job
- •Even the main Hugging Face Twitter account is shared across the whole team
- 12:42 – 14:53
Bridging research to enterprise: why ML’s “time to production” is unusually fast
Sarah asks about the gap between cutting-edge research and enterprise readiness. Clem argues ML’s research-to-production loop is dramatically faster than traditional sciences, and that open source is a key reason the virtuous cycle works.
- •Compared with traditional sciences, ML transfer to production is exceptionally rapid
- •Time horizons are shrinking from years to months/weeks/days
- •Open source accelerates iteration: science→production→science feedback loops
- •Concern: proprietary models could slow the ecosystem’s overall progress
- •Open collaboration is framed as core to the speed of advancement
- 14:53 – 18:33
Open source vs proprietary models: coexistence, task-level leadership, and specialization
Clem rejects a single-winner narrative, arguing open and proprietary approaches will continue to trade leads across tasks. He points to the diversity of models and builders on Hugging Face as evidence that specialization, cost, and accuracy drive a multi-model world.
- •Open and proprietary will likely coexist, as in search and databases
- •Leadership varies by task (e.g., proprietary may lead in some text generation; OSS in others)
- •Ecosystem is dynamic—advantages can flip quickly as techniques and releases evolve
- •Hugging Face scale: 250k+ models uploaded by ~15k companies suggests no single “best” model
- •Specialized models are often cheaper, faster, and more accurate for specific use cases
- 18:33 – 19:55
What’s on the Hub: dominant modalities, emerging domains, and model size spectrum
Clem characterizes the distribution of models by task and domain, highlighting NLP, vision/text-to-image, and audio as top categories. He also notes growing activity in time series and in biology/chemistry, spanning model sizes from millions to 180B parameters.
- •Top three task buckets: NLP; computer vision/text-to-image; audio (ASR/TTS, etc.)
- •Emerging growth in time series (e.g., ETA prediction) and financial/fraud applications
- •Increasing presence of biology and chemistry models/datasets/demos
- •Model sizes range broadly: from a few million parameters to ~180B parameters
- •The platform is used for both classic tasks and frontier generative workflows
- 19:55 – 21:10
Why small models still matter: latency, cost, and real-time enterprise needs
Not all use cases benefit from the largest models. Clem explains how production constraints like latency and cost create strong demand for smaller, faster, more targeted models, while some applications still justify larger general-purpose systems.
- •Model choice is use-case dependent rather than “bigger is always better”
- •Real-time products (example: Bloomberg Terminal) prioritize low latency
- •Smaller models can be faster and cheaper while meeting accuracy requirements
- •Larger models fit more general needs when cost/latency constraints are looser
- •Hugging Face observes broad, distributed usage across model sizes
- 21:10 – 22:35
Infrastructure focus: making ML deployment sustainable and cost-aligned
Clem says he’s especially excited about infrastructure—an often overlooked lever for making ML healthy and scalable. He expects the ecosystem to mature by aligning infrastructure costs with real user value and use cases.
- •Infrastructure is a key bottleneck: cost, speed, and operational sustainability
- •The field has often ignored true run costs while chasing capability gains
- •Expectation of clearer cost/performance alignment as the ecosystem matures
- •“Cloud money laundering” critique: disconnect between spend and actual value
- •Better infra economics will enable broader, longer-term progress
- 22:35 – 24:48
Decentralized training and data consent: online learning and opt-in/opt-out datasets
Clem highlights two desired ecosystem shifts: more decentralized/continuous learning infrastructure and stronger consent mechanisms for training data. He cites BigCode opt-out for code and rising opt-in datasets—especially relevant amid text-to-image artist debates.
- •Desire for decentralization in infra and more frequent/continuous training cycles
- •Critique of “train once, refresh in 6–12 months” as archaic and causes staleness
- •Online learning could keep models current (and reduce knowledge lag)
- •BigCode introduces developer opt-out from training datasets for code models
- •Momentum toward opt-in datasets; heightened importance for artist-consent in image models
- 24:48 – 27:05
BLOOM and BigScience: open collaboration as an answer to power concentration and bias
Clem explains BLOOM as the product of BigScience, a large-scale open collaboration to train an LLM transparently. He argues building in the open helps counter AI power concentration and allows broader participation to identify and mitigate bias.
- •BLOOM came from BigScience: ~1,000 researchers across ~200 organizations
- •End-to-end transparency: public runs, discussions, and decision process
- •Compute support included external infrastructure (e.g., French compute provider)
- •Open source/open science framed as solutions to concentration of power and bias
- •Diversity of contributors helps surface harms affecting underrepresented groups
- 27:05 – 28:34
Adapting to new paradigms: RLHF and diffusion-era tooling for the community
Elad asks about RLHF as a post-pretraining step and how it affects strategy. Clem describes Hugging Face’s approach: support new waves early by hosting models and building open-source libraries so companies can integrate emerging techniques.
- •RLHF is treated as an important addition to standard ML pipelines
- •Hugging Face began hosting RLHF-related models before the broader hype cycle
- •They’re developing an open-source library to help companies adopt RLHF workflows
- •Pattern: the company continually adapts to new model generations (e.g., diffusers)
- •Flexibility is positioned as essential for ML startups due to rapid paradigm shifts
- 28:34 – 34:04
Monetization and the GitHub analogy: freemium, enterprise needs, and becoming a compute gateway
Clem outlines Hugging Face’s evolving business model: a freemium platform with paid tiers driven by enterprise security/compliance and infrastructure offerings. In discussing GitHub, he notes that compute/infrastructure monetization may be a key opportunity for platforms that sit at the start of developer workflows.
- •Current shape: ~15k companies use the platform; ~3k pay (freemium dynamics)
- •Paid differentiation increasingly centers on security and compliance for large enterprises
- •Infrastructure monetization: GPU upgrades for Spaces, inference endpoints, optimization help
- •GitHub comparison: a platform can influence downstream infrastructure choices by being the starting point
- •Strategic focus: invest earlier in becoming a ‘gate for compute’ via cloud partnerships
- 34:04 – 37:25
Looking ahead: bio/chem breakthroughs and the rise of full-stack ML-native companies
Clem’s excitement centers on biology and chemistry as high-impact domains for “software 2.0” approaches. He also highlights a new generation of ML-native companies that build the models and the product together, enabling them to challenge incumbents.
- •Bio and chem seen as major next frontiers for ML impact
- •ML as a new paradigm for building technology (“software 2.0” framing)
- •Evidence of momentum: BioGPT, protein work, and growing Hub activity in these areas
- •Clem avoids specific predictions, citing the field’s unpredictability
- •Excitement about ‘full-stack ML’ companies (e.g., Runway, Grammarly, Stability) that build ML natively
