Lewis Tunstall interview – Hugging Face Machine Learning Experts

TL;DR

Hugging Face released an interview with Lewis Tunstall, a Machine Learning Engineer who contributed to the Transformers library, co‑authored the NLP with Transformers book, and helps build the Hugging Face Course and large‑scale model‑evaluation tooling.

Who is Lewis Tunstall?

Lewis Tunstall is a Machine Learning Engineer at Hugging Face. He previously worked on NLP, topological data analysis, and time‑series projects for startups and enterprises. Before entering ML, he was a theoretical physicist. Outside work he enjoys guitar, trail running, and open‑source contributions.

How he joined Hugging Face

  • In 2018 he used the early pytorch-pretrained-bert library for a question‑answering prototype at a Swiss startup.
  • Fascinated by the lack of accessible resources, he and colleague Leandro von Werra cold‑emailed Hugging Face co‑founder Thomas Wolf and, after a positive reply, began writing a book on Transformers.
  • The book, NLP with Transformers, was released after 1.5 years of collaboration and led both authors to join Hugging Face, where Lewis has been employed for roughly nine months.

The NLP with Transformers book

  • Described by Luca Perrozi (Accenture) as “Complexity made simple… destined to dictate the standards in the field for years to come.”
  • Covers both off‑the‑shelf pre‑trained models and training from scratch, performance tuning, and handling missing labels.
  • The cover features a parrot—a nod to the community’s “stochastic parrots” discussion and a playful illustration that the authors consider a metaphor for language models.

Contributions to the Transformers library

  • Focused on production‑ready model export via the ONNX format, enabling conversion from PyTorch to TensorFlow and execution on specialized hardware.
  • Simplified the export process to a single line of code, reducing latency and increasing throughput for real‑time applications such as chatbots.
  • Emphasizes the library’s philosophy of abstracting boilerplate so users can concentrate on their core tasks.

The Hugging Face Course

  • Co‑developed with core maintainers Sylvain Gugger and Lysandre Jik to bridge the gap between software engineers and modern NLP.
  • The free, multi‑part course teaches end‑to‑end model training across tasks and is expanding to cover audio and computer‑vision modalities in its upcoming third part.
  • Community‑driven events have produced practical projects, e.g., a cover‑letter generator that helped a participant land an internship on the Big Science team.

Large‑scale model evaluation

  • Working on infrastructure that lets users evaluate models and datasets directly on the Hugging Face Hub, aiming to surface top‑performing models for a given task without manual benchmarking.
  • Exploring hosted benchmarks that lower the barrier for researchers to submit and maintain evaluation suites, addressing the current reliance on ad‑hoc scripts and paper‑by‑paper testing.
  • Highlights the risk of benchmark over‑fitting: models can achieve high scores by exploiting shortcuts that do not reflect true task performance.

Practical advice for ML engineers

  • Start with baselines. Simple regular expressions or logistic regression often provide a solid reference before moving to state‑of‑the‑art transformers.
  • Avoid over‑engineering. Build end‑to‑end pipelines that integrate with production constraints; a high‑performing model is useless if it cannot be deployed.
  • Contribute early to open source. Early involvement in community projects accelerates learning and can open career opportunities.

Vision for future applications

  • Natural sciences. Combining deep learning with domain knowledge can accelerate discoveries, exemplified by AlphaFold’s protein‑structure predictions.
  • Drug discovery and robotics. ML‑driven chemistry pipelines and the long‑term goal of household robots (e.g., laundry‑folding) are areas of personal interest.
  • Responsible AI. While acknowledging potential misuse (e.g., weaponization, financial market manipulation), Tunstall stresses that the immediate risk is not a “Terminator” scenario but systemic vulnerabilities.

Favorite papers and inspirations

  • Leo Breiman’s 2001 Random Forests paper – praised for its clarity and implementability.
  • A recent DeepMind paper on using ML to prove theorems in algebraic topology – demonstrates that ML can aid fundamental scientific insight beyond prediction tasks.

Personal interests and outreach

  • Regular listener of the DeepMind podcast hosted by Hannah Fry, which he recommends for clear explanations of deep learning and reinforcement learning.
  • Active on Twitter (@_lewtun) and participates in the Hugging Face Forums and Discord community.

Key takeaways

  • Lewis Tunstall’s work exemplifies how open‑source contributions, practical tooling (ONNX export), and educational resources (the Hugging Face Course) together lower the barrier to applying Transformers in production.
  • The interview underscores the importance of baseline models, end‑to‑end system thinking, and responsible evaluation practices for sustainable ML development.

Sources