Silicon Valley GirlBig Job Disruption in 5 Years — Hugging Face Co-Founder on How to Stay Ahead
CHAPTERS
Meet Thomas Wolf: Hugging Face’s mission and the coming job disruption
Marina introduces Thomas Wolf at VivaTech and frames the conversation around rapid AI-driven change. Thomas sets the tone with a clear warning: major job disruption is likely within five years, so people need strategies to stay relevant.
- •Thomas Wolf’s role: Hugging Face co-founder and Chief Science Officer
- •Theme: future of work, staying ahead amid industry disruption
- •Time horizon: “big job disruption” within ~5 years
- •Focus on practical skills and adaptation, not just hype
Hugging Face explained for non-technical users: models, datasets, and an AI “app store”
Thomas breaks down Hugging Face’s core offerings and clarifies that much of the platform is developer-oriented. He highlights the most accessible part for non-technical users: Spaces, a community-built directory of runnable AI apps.
- •Three pillars: models, datasets, and apps (Spaces)
- •Open-source alternative to closed providers (e.g., OpenAI/Anthropic) for owning the stack
- •Spaces as an AI app store with searchable, task-specific mini-apps
- •No-code-ish interfaces enable quick experimentation (e.g., background removal, TTS, 3D generation)
Using Spaces in practice: run in the cloud or clone and run locally
They discuss how users can try AI apps directly on Hugging Face (with hosted compute) or copy them to run on their own hardware. Thomas emphasizes that model size and compute requirements determine whether local use is practical.
- •Use directly on the platform with provided compute
- •Clone (Git) a Space to run offline/on your own machine
- •Popularity signals (likes) help choose between many similar apps
- •Local execution depends on model size and hardware capability
Vibe coding and local LLMs: building apps without sending prompts to external servers
Thomas mentions a popular “vibe coding” Space leveraging the DeepSeek model. He stresses the appeal of running models locally to reduce privacy concerns—especially around where prompts are sent and stored.
- •Vibe coding tools let non-technical builders prototype quickly
- •Example: DeepSeek-based Space for coding assistance
- •Running locally means prompts don’t need to go to remote servers
- •Privacy/sovereignty concerns are becoming a practical product feature
Who owns AI-generated output in open source? Licenses, credit, and commercialization
Marina raises ownership concerns: if you generate a character using open-source tools, who owns it? Thomas answers by analogizing to open-source software licensing—MIT/Apache and beyond—where reuse is allowed but attribution and commercial terms depend on the license.
- •Open source: share code with a license defining reuse rights
- •Common licenses: MIT, Apache—generally permissive with attribution norms
- •Some licenses add commercial restrictions or payment requirements
- •Ownership and credit become clearer when licensing is explicit
The “$15 photo” problem: value creation vs. creator compensation in open source
They explore the tension between massive downstream value and minimal compensation for creators—common in both media and software. Thomas cites foundational projects (Linux, NumPy) and explains why many contribute for mission, while others build businesses around open source.
- •Creators often capture only a tiny fraction of business value built on their work
- •Examples: Linux, NumPy powering broad ecosystems
- •Motivations: mission, accessibility, public benefit
- •Open-core model: keep a free core, monetize business-grade features (security, integrations)
Coding’s future: more builders, different learning paths, not less technical
After a sponsor break, Marina asks whether coding becomes less technical as AI tools improve. Thomas predicts both: more non-technical people will build apps, while new learners (including kids) will still need to understand fundamentals when AI-generated code breaks.
- •Non-technical entrepreneurs now prototype with AI coding tools
- •Kids learn by generating first, then debugging to understand
- •Learning sequence shifts from “theory-first” to “results-first,” but depth still matters
- •Overall developer pool likely grows rather than shrinks
Raising kids for an AI world: creativity as a timeless edge over “most likely” outputs
Thomas explains how he thinks about teaching children in an AI-rich future. He argues creativity and willingness to make novel, non-obvious choices will stay valuable because LLMs optimize for probable outputs, not truly original ideas.
- •Creativity as durable leverage in a world of capable AI
- •LLMs produce “most likely” continuations, not inherently novel concepts
- •Encouragement over perfectionism helps reduce self-censorship
- •School systems vary: Sweden cited as creativity-forward; France as more rigid
How to encourage without forcing: expanding kids’ horizons with exposure
Marina presses on parenting tactics—do you push coding directly? Thomas suggests a lighter touch: expose kids to conferences, tools, and possibilities so they self-select interests instead of resisting parental pressure.
- •Over-pushing can backfire (“that’s your thing, Dad”)
- •Bring kids to real-world environments (conferences, demos) to spark curiosity
- •Show tools and let them explore independently
- •Goal: broaden horizons rather than prescribe a single path
Robots as the next platform: physical AI plus an open-source “app store” for behaviors
Thomas shifts to robotics, noting Hugging Face acquired a robotics company and wants accessible, developer-friendly robots. He envisions robots that gain new capabilities via shared open-source apps—like a growing marketplace of behaviors.
- •Robotics as a major focus; Hugging Face acquisition in the space
- •Robots give AI physical presence and more intuitive interaction
- •Vision: upload/share robot skills on Hugging Face like apps
- •Open ecosystem means robots improve over time beyond the out-of-box features
When household robots arrive: capability is near, but price and regulation gate adoption
Marina asks the practical question: when does every family get a household robot? Thomas says prototypes are close, but consumer adoption depends on affordability and safety/regulation, especially since robots can cause physical harm.
- •Technical demos may be feasible as soon as next year
- •Main blockers: cost and regulation/safety standards
- •Early capable robots may cost “a car,” with two-handed systems around ~$15–20K
- •Subscription/rental models could emerge for expensive home robots
Privacy and safety in home robotics: local execution and offline reliability
Marina raises surveillance risks: what if open-source robot apps spy inside homes? Thomas argues open source enables local, offline operation—cut Wi‑Fi and the system can still run—improving privacy and robustness compared with cloud-only systems.
- •Open source advantage: download, run locally, and disable internet
- •Home robots need strong privacy guarantees due to always-on sensors
- •Offline behavior matters: robots should fail safely if connectivity drops
- •Cloud-only systems inherently require data transmission and trust
AI in 5 years: agents, photorealistic media, embedded models, and AI-for-science breakthroughs
Thomas predicts three major shifts: web-based agents automating complex computer tasks, growing real-world robot capabilities, and photorealistic synthetic media that blurs reality. He also highlights a hopeful frontier: applying AI training techniques to science—materials, weather, fusion—beyond “just chatbots.”
- •Agents automate multi-step work on computers; early signs already visible
- •Robotics capabilities expand steadily from prototypes to broader utility
- •Synthetic video/audio becomes indistinguishable, increasing premium on in-person trust
- •Smaller high-performance models + better chips embed AI into everyday devices
- •AI-for-science: materials discovery, climate tech, weather prediction, fusion research
Mass unemployment risk and how to stay ahead: master the tools, then redesign your work
They close on unemployment concerns: Thomas sees entrepreneurship becoming more accessible, but acknowledges many people don’t want that path. He worries about disruption in long-training professions (e.g., law/IP) and advises individuals to actively use AI tools and identify the remaining meaningful parts of their jobs—while noting society-level solutions may be needed.
- •AI lowers barriers to building products, potentially increasing entrepreneurship
- •Not everyone wants to be an entrepreneur—creates a policy/societal challenge
- •High-study professions with support functions (e.g., law/IP) face strong disruption
- •Personal advice: adopt AI tools early; become proficient rather than ignore them
- •Reassess which parts of work remain fulfilling; consider pivoting if needed
- •Broader ideas raised: UBI, entertainment-centric society, and what we lose by automating (GPS analogy)
