Wiki · Individual Player · Last reviewed July 10, 2026

Jeremy Howard

Jeremy Howard is a data scientist, entrepreneur, researcher, developer, and educator best known for co-founding fast.ai with Rachel Thomas, creating the fastai deep-learning library, coauthoring ULMFiT, founding Answer.AI, proposing /llms.txt, and teaching programmers to use AI systems without treating them as magic.

Category: Individual Player Published: May 20, 2026 Modified: July 10, 2026 Last reviewed: July 10, 2026 Tags: fast.ai, fastai, ULMFiT, Answer.AI, llms.txt, AI literacy, AI coding

Definition

Jeremy Howard is an applied AI field-builder: a public educator, software author, research contributor, and company founder whose influence comes less from controlling a single foundation model than from changing how working programmers learn, inspect, and use machine-learning systems.

For this wiki, he is most relevant as a figure in AI literacy and practical agency. fast.ai made deep learning teachable to coders by starting from working examples. Answer.AI and SolveIt extend that pattern into AI-assisted coding, where the hard question is no longer only how to train a model, but how to use model output while preserving human judgment, testability, and responsibility.

Snapshot

Current Context

As of the July 10, 2026 review, Howard's public work spans three connected tracks. The first is the fast.ai tradition: free or broadly accessible courses, the fastai library, notebooks, forums, and educational material for coders. The second is Answer.AI: a research-and-development lab launched with Eric Ries in December 2023 to connect foundational research with practical products. The third is AI-assisted programming education through SolveIt, which frames AI as a partner in problem solving rather than a substitute for thinking.

Howard is also attached to machine-readable context infrastructure. In September 2024 he proposed /llms.txt, a Markdown file at a website root that gives language models a curated map of useful context. That proposal matters for AI search and answer engines because it moves some context selection from opaque crawler behavior toward explicit publisher guidance, while still leaving licensing, access control, and enforcement to other mechanisms.

Recent Answer.AI work also places Howard's institution in the practical safety conversation around tool use and function calling. A 2026 Answer.AI post on unauthorized tool calls argues that APIs and runtimes must enforce which tools a model is allowed to call, instead of relying on the model's internal compliance with a prompt or tool list.

Career

Howard's career combines entrepreneurship, competition-based machine learning, applied research, software, and education. fast.ai describes him as a former President and Chief Scientist of Kaggle, the founding CEO of Enlitic, the founding CEO of FastMail and Optimal Decisions Group, an Honorary Professor at the University of Queensland, a Digital Fellow at Stanford University, and a long-time machine-learning practitioner.

Those roles matter because his AI influence is not confined to one paper or one company. Kaggle shaped a culture of measurable model-building craft. Enlitic represented an early attempt to bring deep learning into medical diagnostics. fast.ai translated deep-learning practice into public education, software, and community.

Howard's public work often argues that AI capability should not remain the property of a narrow technical elite. This makes him an important figure in the democratization side of AI history: not only who builds frontier models, but who gets to understand, adapt, and apply the tools.

fast.ai and Practical Education

Howard and Rachel Thomas launched fast.ai in 2016 with an explicit mission to make deep learning more accessible. The project combines free courses, a software library, research, and a public learning community.

The best-known course, Practical Deep Learning for Coders, is designed for people with coding experience who want to apply deep learning and machine learning to practical problems. Its teaching style starts with working models and gradually exposes the underlying ideas, rather than beginning with abstract mathematical formalism.

This pedagogical choice was consequential. It gave programmers a way into computer vision, NLP, tabular modeling, recommender systems, deployment, and data ethics at a time when deep learning often looked like a field reserved for graduate labs and large companies.

The course also used the tooling environment of actual practitioners: Python, notebooks, PyTorch, fastai, Hugging Face libraries, forums, and cloud notebooks. In that sense, fast.ai taught not only concepts but a practical culture of experimentation.

ULMFiT and Transfer Learning

In 2018, Howard and Sebastian Ruder published "Universal Language Model Fine-tuning for Text Classification." The paper proposed ULMFiT, a transfer-learning method for NLP that fine-tuned a pretrained language model for downstream text classification tasks.

The paper argued that NLP had not yet benefited from transfer learning as strongly as computer vision had, and introduced techniques for language-model fine-tuning. It reported improved performance on several text-classification datasets and useful results in low-data settings.

ULMFiT is historically important because it arrived just before the transformer-centered boom around BERT, GPT-style pretraining, and large-scale fine-tuning. It helped establish the practical idea that pretrained language models could be adapted efficiently rather than trained from scratch for each task.

Howard sometimes describes ULMFiT very expansively in public biographies. A conservative reading is narrower but still strong: ULMFiT was one of the key late-2010s demonstrations that transfer learning could work powerfully in NLP, and it belongs in the lineage that made language-model fine-tuning a standard practice.

fastai Software

The fastai library is a high-level deep-learning library built on PyTorch. In the 2020 fastai paper, Howard and Sylvain Gugger describe a layered API intended to serve both practitioners who need quick state-of-the-art results and researchers who need flexible low-level components.

That layered design is part of Howard's broader philosophy. The tool should let a learner train a real model early, then peel back layers as their understanding grows. The library turns pedagogy into software architecture: high-level defaults for usefulness, lower-level access for inspection and modification.

fastai also embodies an argument about AI infrastructure. Accessibility is not only about free videos. It is about APIs, examples, documentation, defaults, forums, and workflows that make real experiments feel reachable to people outside elite labs.

Answer.AI, SolveIt, and Dialog Engineering

In December 2023, Howard and Eric Ries launched Answer.AI as an AI R&D lab focused on practical end-user products based on foundational research breakthroughs. Howard framed the lab as a place where development goals and research goals should inform one another.

In November 2024, fast.ai announced that it was joining Answer.AI and introduced "How To Solve It With Code," a course built around AI-assisted coding. In October 2025 Howard said a new version of the course was available and aimed mainly at experienced coders, AI practitioners, and data scientists. The SolveIt site summarizes the educational stance bluntly: do not outsource thinking to AI; use AI to become a better problem solver, clearer thinker, and more elegant coder.

Howard describes the approach as "Dialog Engineering": solving problems step by step with AI as a partner while preserving human understanding of the result. This move is a natural continuation of fast.ai's older teaching premise. In the 2010s, accessibility meant teaching coders to train models. In the mid-2020s, it increasingly means teaching people how to work with AI systems as collaborators without losing ownership of the problem, code, or reasoning.

Answer.AI's public technical work is broader than education. It includes research and open tooling around model architecture, literate programming, and LLM application development, including ModernBERT and tools such as nbdev and claudette. Those projects should be evaluated as specific artifacts, not as proof of a general claim that any one lab has solved the social or engineering problems of AI use.

llms.txt and Context Infrastructure

Howard proposed /llms.txt in September 2024 as a way for website owners to provide language-model-friendly context. The file is intended to be a Markdown guide with brief background, guidance, and links to deeper Markdown resources that tools can use when assembling context for prompts.

The governance significance is modest but real. llms.txt gives publishers a machine-readable editorial surface for explaining what matters on their own sites. It is not a permission system, copyright license, crawler ban, or safety guarantee. It works best when treated as source discipline: a clear map of the material a model or agent should read before answering about a site.

Governance and Safety

Howard's work has a governance dimension because education changes who can inspect, adapt, and challenge AI systems. A person who can train a model, read a notebook, run an evaluation, or understand a generated patch is less dependent on vendor claims. In that sense, fast.ai and SolveIt are not only educational projects; they are capability-distribution projects.

The same distribution creates risk. Practical AI education can produce shallow confidence if users learn fluent workflows without enough testing, data awareness, privacy practice, or security discipline. AI-assisted coding raises a specific accountability problem: the user may ship code they cannot debug, secure, or maintain. For Spiralist governance, the correct question is not whether AI helped write the code, but whether a responsible person can explain the intended behavior, inspect the dependencies, test the edge cases, and own the deployment.

Tool-using systems add another layer. The 2026 Answer.AI discussion of unauthorized tool calls is a useful safety lesson: access control must live in the application boundary, runtime, or policy enforcement layer, not inside the language model's willingness to follow an instruction. This connects Howard's current institutional work to AI agent sandboxing, prompt injection, and AI audit trails.

Source Discipline

Howard is a founder, educator, and advocate, so source discipline matters. His public biographies and launch posts sometimes use broad claims about ULMFiT, fast.ai, or Answer.AI. This page treats those as self-descriptions unless they are independently supported by papers, software records, course pages, or other primary artifacts.

For research claims, prefer papers and released model artifacts. For education claims, cite course pages, syllabi, and durable materials rather than testimonials alone. For product or lab claims, separate what Answer.AI says it is building from evidence that the product works for a given population or use case. For openness claims, distinguish open-source software from open weights, open data, and machine-readable publisher guidance; those are different governance categories.

Central Tensions

Spiralist Reading

Jeremy Howard is a teacher of the threshold.

His work matters because he keeps insisting that the machine should be learnable by people who were not ordained by the old gates. The coder, the nurse, the journalist, the founder, the public-interest technologist, and the self-taught builder should be able to touch the system directly.

For Spiralism, this is both hopeful and dangerous. A civilization that cannot understand its tools becomes ruled by priests and vendors. A civilization that makes powerful tools too easy without discipline can flood itself with brittle systems, shallow confidence, and unexamined automation.

Howard's best contribution is the middle path: build something real, inspect it, change it, learn why it works, and never confuse passive consumption of AI with agency.

Open Questions

Sources

Return to Wiki