Jeremy Howard

wiki
, , , ,
1

Australian data scientist, serial entrepreneur, and educator, known as the co-founder of fast.ai, the creator of ULMFiT, and a leading voice for democratising deep learning.

Profile
Born 13 November 1973, London, England (raised in Melbourne, Australia)
Nationality Australian
Current Institution(s) fast.ai (founding researcher); Answer.AI (co-founder); University of Queensland (honorary professor); Stanford Digital Economy Lab (Digital Fellow); CSIRO Data61 (Adjunct Science Fellow)
Research Areas Deep Learning, Transfer Learning, NLP, AI Accessibility, Medical AI
Education BA, Philosophy — University of Melbourne (no formal CS degree)
Website jeremy.fast.ai
X / Twitter @jeremyphoward
GitHub jph00
Google Scholar Jeremy Howard

Overview

Jeremy Howard is one of the most consequential non-credentialed figures in the history of modern AI. Without a computer science degree, he built careers in management consulting, insurance analytics, and email infrastructure before discovering machine learning and becoming — by competitive standing — the world’s top-ranked data scientist on Kaggle in 2010 and 2011. He founded Enlitic, the first company to apply deep learning to medical diagnosis, and later co-founded fast.ai, whose course has been watched by over six million people and whose fastai library has over 25,000 GitHub stars. His ULMFiT algorithm (ACL 2018), co-authored with Sebastian Ruder, established transfer learning as the standard paradigm for NLP and directly influenced the training methodology of GPT, ChatGPT, and their successors. He is also the co-founder of Answer.AI, a practical AI R&D lab launched in 2023, and led the global Masks4All movement during the COVID-19 pandemic, co-authoring a landmark evidence review in PNAS. His career is defined by a consistent philosophy: that powerful AI tools, rigorously understood and openly shared, should be accessible to anyone.

Early Life & Education

Howard was born in London on 13 November 1973 and moved to Melbourne, Australia with his family in 1976. He attended Melbourne Grammar and studied philosophy at the University of Melbourne. He holds no formal degree in computer science or mathematics — a fact he has repeatedly emphasised as proof that deep learning can be learned by anyone willing to invest the time. His early programming involvement included contributions to open-source infrastructure projects: he contributed to the Perl programming language as chair of the Perl6-data working group, and to the Cyrus IMAP server and Postfix SMTP server.

Career

Management Consulting (early 1990s–late 1990s)

Howard started his career in management consulting at McKinsey & Co and AT Kearney, remaining in consulting for eight years before becoming an entrepreneur. At AT Kearney he led the Leveraging Customer Information group, an international machine learning and data analytics practice. He describes himself as the first analytical specialist at McKinsey in Australia in the early 1990s, at a time when only two such specialists existed globally at the firm.

FastMail & Optimal Decisions Group (late 1990s–2000s)

While in Australia, Howard founded two successful startups: the email provider FastMail, which he sold to Opera Software, and the insurance pricing optimisation company Optimal Decisions Group (ODG), which he sold to ChoicePoint. At ODG he invented profit-optimised pricing for insurance, a method now used by nearly all home and auto insurers, and created the Drivetrain Method to building data products.

Kaggle (2010–2013)

Howard first became involved with Kaggle, the data science competition platform, after becoming its globally top-ranked participant. The competitions he won involved tourism forecasting and predicting the success of grant applications. He then became President and Chief Scientist of Kaggle. In December 2011, Wired Magazine ran a piece on Howard, calling him “The Accidental Scientist.” He left Kaggle in December 2013.

Enlitic (2014–2016)

In August 2014, Howard founded Enlitic to use machine learning to make medical diagnostics and clinical decision support tools faster, more accurate, and more accessible. He raised $15 million within two years of creating the company. MIT Tech Review ranked it #14 on their list of the Smartest Companies in the World, just ahead of Facebook and SpaceX. Enlitic was the world’s first company focused specifically on deep learning for medical imaging — predating the wave of AI health startups that followed by several years.

fast.ai (2016–present)

Fast.ai is a non-profit research group focused on deep learning and AI. It was founded in 2016 by Jeremy Howard and Rachel Thomas with the goal of democratising deep learning. The organisation’s flagship output is the Practical Deep Learning for Coders course (course.fast.ai), which requires no prerequisites beyond Python knowledge and has been described by The Economist as a model for mass AI education. The course has been watched by over six million people and is the longest-running and most widely-used deep learning course in the world, and has launched many careers in research, industry, and startups.

The fastai software library, co-developed with Sylvain Gugger, provides a layered API over PyTorch that enables rapid prototyping and state-of-the-art results with minimal code. It has over 25,000 GitHub stars. Howard and his students also set record-breaking training speeds on the Stanford DawnBench benchmark in 2018, demonstrating that a small team could match or beat major technology companies on cost and time to train ImageNet classifiers.

At fast.ai, Howard also created nbdev, a literate programming tool that allows developers to write Python libraries directly in Jupyter notebooks, dramatically simplifying documentation, testing, and package distribution.

Wicklow AI in Medicine Research Initiative / University of San Francisco

Howard served as Distinguished Research Scientist at the University of San Francisco and as founding chair of the Wicklow AI in Medicine Research Initiative (WAMRI), building partnerships with UCSF, Harvard, Stanford, and many other academic medical centres, which resulted in significant published medical research breakthroughs.

COVID-19 & Masks4All (2020–2021)

Early in the pandemic, Howard became a leading public advocate for universal masking. He led the largest evidence review of masks, published in the Proceedings of the National Academy of Sciences, and became the most read paper of all time on preprints.org. He wrote op-eds in the Washington Post, the Guardian, and The Atlantic, and appeared on major US television networks including Good Morning America and Nightline.

Answer.AI (2023–present)

Howard co-founded Answer.AI with Eric Ries as a new kind of AI R&D lab that creates practical end-user products based on foundational research breakthroughs. The lab’s philosophy is to remain small, ship quickly, and focus on genuinely useful products for individual users — explicitly contrasted with the “agentic AI” trend of large companies. Early notable releases included a system based on FSDP and QLoRA enabling 70B model training on two consumer GPUs, and tools aimed at helping developers and writers grow their own capabilities using AI assistance.

Key Contributions

  • ULMFiT — Universal Language Model Fine-tuning (ACL 2018) — Co-authored with Sebastian Ruder. Proposed an effective transfer learning method applicable to any NLP task, reducing error by 18–24% on the majority of benchmark datasets, and matching full-dataset performance with only 100 labelled examples. Cited over 5,000 times and credited as the technical predecessor to the GPT family’s training approach. When published, Howard and Ruder encountered substantial scepticism: the community widely dismissed fine-tuning as unprincipled, and ULMFiT’s vindication took a year.

  • fast.ai course — The world’s longest-running online deep learning course, teaching a distinctive top-down, code-first pedagogy that has introduced millions of practitioners to the field without prerequisites.

  • fastai library — A layered API over PyTorch enabling state-of-the-art results with significantly fewer lines of code; its underlying design principles were published in the peer-reviewed journal Information (2020). Over 25,000 GitHub stars.

  • nbdev — A literate programming system for developing Python software in Jupyter notebooks, including automated test generation, documentation, and package publication; widely adopted across the fast.ai and open-source Python communities.

  • Enlitic — The world’s first deep learning for medicine company (2014), demonstrating that deep neural networks could match or exceed radiologist-level performance on medical imaging tasks.

  • DawnBench speed records (2018) — With fast.ai students, set records for the fastest and cheapest ImageNet training, beating Google and Intel teams and demonstrating the practical power of the fast.ai training methodology.

  • Deep Learning for Coders with fastai and PyTorch (O’Reilly, 2020) — Co-authored with Sylvain Gugger; rated five stars on Amazon and praised by Peter Norvig as one of the best sources for programmers to become proficient in deep learning.

  • PNAS Masks Evidence Review (2021) — Co-led with 20 authors across multiple institutions; became the most-read paper on preprints.org at time of publication and directly influenced public health policy in multiple countries.

  • TED talk — “The wonderful and terrifying implications of computers that can learn” — Over 2.5 million views, translated into 26 languages; one of the most-watched public introductions to deep learning’s societal implications.

Awards & Recognition

  • MIT Technology Review Smartest Companies (2015, 2016) — Enlitic ranked #14 globally, ahead of Facebook and SpaceX.
  • Wired “The Accidental Scientist” (2011) — Profile recognising his rise to global top ranking on Kaggle.
  • World Economic Forum Young Global Leader — Invited to speak at Davos 2014 on “Jobs for the Machines.”
  • Honorary Professor, University of Queensland (2022–present) — Formal academic recognition of his educational contributions.
  • Digital Fellow, Stanford Digital Economy Lab (2023–present) — Affiliated fellowship at Stanford HAI.

Key Relationships

  • Rachel Thomas — Co-founder of fast.ai; mathematician, ethicist, and educator who co-created the course and directed the data ethics curriculum. The partnership made fast.ai one of the most balanced AI education organisations in the field.
  • Sebastian Ruder — Co-author of ULMFiT; NLP researcher then at the Insight Centre for Data Analytics, now at Google DeepMind. The collaboration bridged academic NLP and fast.ai’s applied approach at a pivotal moment.
  • Sylvain Gugger — Core contributor and co-maintainer of the fastai library and nbdev; co-author of Deep Learning for Coders. The engineering depth behind many of fast.ai’s practical tools.
  • Eric Ries — Co-founder of Answer.AI; creator of the Lean Startup methodology. Howard has cited Ries’s influence on his thinking about iterative product development and useful technology.
  • Peter Norvig — Google’s Director of Research, who publicly endorsed Deep Learning for Coders as one of the best deep learning books available; the endorsement carried significant weight in validating Howard’s non-credentialed approach.

Personal Style

Howard is a deliberate contrarian on questions of AI pedagogy and research culture. His fast.ai course inverts the conventional sequence — starting with applications and working backwards to theory — on the explicit belief that motivation precedes understanding, and that the gatekeeping function of mathematical prerequisites has historically excluded the people most capable of applying AI usefully. He has been equally contrarian in research: ULMFiT was developed and published when the NLP community was deeply hostile to fine-tuning, and his DawnBench results deliberately challenged the assumption that only large organisations with large compute budgets could advance the state of the art. His Answer.AI philosophy extends this: explicit opposition to “agentic AI” that automates humans out of the loop, in favour of tools that give human practitioners more leverage over their own work. He writes and speaks accessibly, without jargon, and has a long record of engaging directly with learners — his AMA on Reddit from 2014 remains widely linked as an introduction to his thinking.

References