Jeremy Howard is an Australian entrepreneur, researcher and educator known for co-founding fast.ai, a widely used deep learning education platform and open-source library, and for helping create ULMFiT, an early and influential Transfer learning method for Natural language processing.
Before moving fully into AI, Howard was president and chief scientist of Kaggle, the machine learning competition platform, and founded Enlitic, one of the first companies to apply Deep learning to medical diagnostic imaging. He co-founded fast.ai in 2016 with Rachel Thomas, building both a free online course, "Practical Deep Learning for Coders," and an accompanying software library, with the explicit goal of making state-of-the-art deep learning accessible to programmers without requiring an advanced mathematics background.
ULMFiT and transfer learning for NLP
In 2018, Howard and researcher Sebastian Ruder published ULMFiT (Universal Language Model Fine-tuning), a method for Pretraining a language model on a large general text corpus and then fine-tuning it cheaply on a specific downstream task with comparatively little labeled data. ULMFiT demonstrated that the Transfer learning approach already standard in Computer vision, where models pretrained on ImageNet were fine-tuned for other tasks, could work just as well for text, predating the wave of pretrain-then-fine-tune language models that included BERT and later GPT-style models. The paper is widely cited as an important precursor to the modern Large language model paradigm.
llms.txt and later work
In 2024, Howard helped propose llms.txt, a simple text-file standard placed at a website's root intended to give AI systems a concise, machine-readable summary of a site's content, drawing an explicit analogy to the long-standing robots.txt convention for search crawlers. The proposal was part of a broader push around making websites more legible to AI agents and chatbots as a growing share of web traffic and citations began coming from AI systems rather than direct human browsing, an area sometimes described as Generative engine optimization.
Throughout his career Howard has emphasized practical, code-first teaching and open tooling over theory-first academic approaches, and fast.ai's courses and library have been credited with bringing a large number of practitioners into the deep learning field who might not otherwise have had access to graduate-level machine learning training.