Machine Learning Experts Interview with Sasha Luccioni – Highlights on Ethical AI and Climate Impact

TL;DR

Sasha Luccioni, a research scientist at Hugging Face, explained how she measures AI’s carbon footprint, leads the Big Science carbon‑footprint working group, and builds tools that make large‑language‑model development more transparent and environmentally responsible.

Background and Motivation

Sasha’s path to Hugging Face began with a linguistics undergraduate degree, a Ph.D. in computer science with a linguistic focus, and a stint in finance. She left a high‑paying finance role to pursue “socially positive AI” research, eventually joining Hugging Face after meeting founders Thom and Clem, whom she trusted to align with her values.

“I quit my job because no amount of money was worth working on AI for AI’s sake; I wanted to do more.”

Current Projects that Matter

Big Science Transparency Initiative

  • The Big Science project aims to turn large‑language‑model development from a “black box” into a “glass box.”
  • Sasha chairs the carbon‑footprint working group, which expands impact assessment beyond GPU hours to include manufacturing costs and even the carbon cost of an email.
  • She also contributes to data‑curation work‑groups that prioritize multilingual data, mindful curation, and the removal of undesirable internet content.

Carbon‑Footprint Tools

  • ML CO₂ calculator – an online tool where users input hardware, training duration, and location to receive an estimated CO₂ emission.
  • code‑carbon – a Python package that streams real‑time emission estimates during model training.
  • Research on proactive emission reduction, e.g., scheduling training in low‑carbon regions (hydroelectric Quebec vs. coal‑heavy Australia) can cut emissions by up to 100×.

“Choosing a low‑carbon cloud region is a simple nudge that can save massive emissions.”

Why Geography Matters for Emissions

Sasha emphasizes that electricity generation mixes differ worldwide (e.g., nuclear France, hydro Canada, coal Australia). Default cloud‑instance selections often ignore this, leading to hidden environmental costs.

Practical Advice for ML Teams

  1. Measure first – use the ML CO₂ calculator or code‑carbon to surface emissions.
  2. Choose low‑carbon compute – many providers now tag regions with carbon intensity.
  3. Change defaults – set low‑carbon options as the default to leverage nudge theory.
  4. Connect to real‑world workflows – ensure model outputs are interpretable and actionable (e.g., satellite‑based wildfire alerts must trigger alerts, not just predictions).

Machine Learning for Climate Solutions

Sasha co‑authored the “Tackling Climate Change with Machine Learning” paper (≈100 pages, 25 authors) covering:

  • Renewable‑energy forecasting – ML predicts short‑term energy demand to reduce reliance on diesel generators.
  • Battery‑charging optimization – anticipatory charging based on weather forecasts.
  • Remote sensing – automated detection of deforestation and wildfires from satellite imagery.

These examples illustrate how ML can offset rather than merely add to emissions.

Overcoming Data Scarcity

  • In a Montreal soup‑kitchen hackathon, Sasha discovered NGOs often lack any data. She organized digital‑literacy workshops to teach basic data collection (spreadsheets, tracking food waste). After a year, the kitchen adopted a simple inventory system that reduced waste without heavy AI.
  • The key lesson: start with data hygiene and education before deploying sophisticated models.

Common Pitfalls

  • “AI solves everything” mindset – teams may build high‑accuracy models but neglect integration (e.g., missing UI, email alerts), causing the solution to stall.
  • Lack of interdisciplinary outreach – successful climate‑impact projects often require networking beyond academic circles (e.g., aviation industry contacts for emission‑reduction algorithms).

Meaningful Work and Personal Philosophy

Sasha follows the Japanese concept of Ikigai (intersection of what you love, what you’re good at, what the world needs, and what you can be paid for). This guided her move from finance to Hugging Face, where open‑source model sharing satisfies all four pillars.

Vision for AI Accessibility

  • Democratizing AI for non‑experts (journalists, doctors, bakers) is a priority. She cites a bakery friend who wonders if AI could improve bread quality—an illustration of untapped domains.
  • Projects like AutoTrain and simple web UIs could enable biodiversity researchers to classify camera‑trap images without deep ML expertise.

Favorite Research Directions

  • AI ethics – papers by Abeba Birhane on how researchers’ values (accuracy vs. efficiency) shape model development.
  • Holistic evaluation – Sasha’s current work expands model evaluation beyond accuracy to include efficiency, carbon emissions, and fairness across data sub‑populations (e.g., gender‑shade style analyses).

Rapid‑Fire Insights

  • Math foundation – Sasha wishes she had embraced mathematics earlier; a solid math base accelerates model‑building intuition.
  • Hands‑on learning – Start with notebooks and personal datasets (e.g., flower images) rather than generic benchmarks like MNIST.
  • Community impact – Small‑scale projects (butterfly classification app for a Montreal insectarium) can translate research into tangible public experiences.

Personal Interests

Outside AI, Sasha is cultivating a home garden, growing tomatoes, peppers, and cucumbers with her children—a grounding contrast to her computational work.

Where to find Sasha: Twitter @SashaMTL and her personal website www.sashaluccioni.com.

If you need a dedicated ML expert to accelerate your project, explore Hugging Face’s ML Experts program at hf.co/support.

Sources