Jeremy Howard and the Problem of Making AI Too Easy

On a couch in San Francisco in late 2017, Jeremy Howard watched a machine take a first pass at thousands of movie reviews. It was a rough test of a large idea: could a system trained to predict the next word in a vast collection of ordinary prose carry what it learned into a different job? The answer arrived in minutes. On its first run through the review data, the model judged the sentiment correctly 93 percent of the time. Howard suspected an error. Then he checked. “I had goosebumps,” he later recalled.

The experiment became part of ULMFiT, a method Howard developed with researcher Sebastian Ruder. Their 2018 paper offered an early, influential case for adapting a model trained on general language to a specific task. Today that idea seems familiar because it sits inside the everyday practice of modern language AI. At the time, the result had the quality of a door opening. A model did not need to learn every new language task from scratch.

But the door Howard cared about was wider than a research result. By then he and Rachel Thomas had already founded fast.ai, a project built around a simple invitation: if you could code, you could start working with deep learning. Its lessons were free. The exercises were practical. The welcome was explicit about people whose backgrounds did not look like the usual AI résumé.

A founder who kept changing rooms

Howard's route to the AI classroom had several detours, most of them productive. He studied philosophy at the University of Melbourne in the early 1990s, when the question of whether a machine could appear to understand language was still largely a matter for argument. He spent years in management consulting at McKinsey and A.T. Kearney. Then came companies: Fastmail, the durable email service he helped found; Optimal Decisions Group; and later Enlitic. Between them came Kaggle, where he was a top-ranked competitor and eventually president and chief scientist.

The sequence looks eclectic until one notices the habit underneath it. At Kaggle, the point was to put an idea against data and see what happened. In startups, a product had to work for a person beyond the team that built it. In teaching, an explanation had to survive contact with someone who had not spent years in the field. Howard brought each of those tests into the next room.

A 2017 television interviewer introduced him with the language of a technology celebrity. Howard seemed more interested in the gap between what machines could do and what people were taught about them. He argued that basic data analysis came late in most education systems, if it came at all. The concern was less glamorous than a prediction about superintelligence, and more immediate: who gets enough fluency to take part?

The class after the competition

The fast.ai course turned an inversion into a teaching method. Rather than begin with a thicket of mathematics and promise useful work later, Howard began with models students could train and deploy. The theory followed as a way to understand what they had already touched. He was serious about the theory; he simply refused to treat it as an admission ticket. A programmer with curiosity and some Python could begin on day one.

This approach was consistent with his years on Kaggle. Competitions reward the person who checks a hypothesis, reads the result, and makes the next adjustment. The useful lesson is rarely that a particular model has a magic setting. It is that progress has a shape: try, inspect, revise. Howard's fastai software library lowered the amount of code required to do that work with contemporary deep learning. A free video lesson could carry the method from a university classroom to a desk on another continent.

Thomas brought her own research, teaching and concern for the social consequences of AI. Together they saw concentrated technical knowledge as a practical problem. A small group deciding how a powerful technology worked would also decide whose needs mattered. Their response was educational and concrete: publish the lessons, develop useful software, encourage a community to build things. The fast.ai slogan was half invitation, half refusal to make prestige the measure of competence.

The ULMFiT work gave that mission an interesting twist. Howard had used a general text model to solve a particular problem with far less specialized training data. In the paper with Ruder, the method performed strongly across multiple text-classification tasks and made its code and pre-trained models available. The research and the teaching shared a taste for reuse: do the expensive learning once, then let more people adapt the result.

A breakthrough with an awkward afterlife

By 2023, the language-model revolution was impossible to miss. Howard could see ideas he had helped advance in tools used by millions. He could also see how much of the infrastructure and decision-making had gathered inside a few very large companies. In an interview from Australia, he described the tension plainly. The work had made AI more capable, while his original hope for broad control over it felt less secure.

It is a peculiar position for a teacher. The students arrived. The lessons spread. The technology advanced at a pace that made the old barriers look almost quaint. Yet easy access to an AI interface is not the same as an ability to shape what it does. A person can use a system every day and still know very little about its limits, costs, or assumptions. Howard's worries moved from entry into the field toward agency within it.

There was, at least, room for a small Australian joke. In San Francisco, Howard said, strangers sometimes recognized him from fast.ai and talked about ULMFiT on public transport. Back in Australia he could walk the street without that conversation. “I'm not a fast bowler or whatever,” he said. The line has a modest charm: even a person associated with a major turn in AI can remain pleasingly obscure to most of his neighbors.

The laboratory and the workbench

Howard launched Answer.AI in 2023 with Eric Ries, the author of The Lean Startup. The lab's stated ambition was to connect foundational research with products that people can use. That pairing matters. Research can drift toward elegant demonstrations; products can settle for thin tricks. Howard and Ries described a loop in which building informs research and research informs building. They also put the lab together as a remote team, with a stated interest in ability over pedigree.

The outputs have been varied. FastHTML, which Howard started, offers a way to make modern web applications in Python. The Answer.AI team has worked on smaller software libraries and joined the collaboration behind ModernBERT. fast.ai joined Answer.AI in 2024, bringing the educational project into the lab's orbit. These are different artifacts, but each reflects an instinct for tools people can inspect, modify and use for their own purposes.

The most revealing new project may be Solveit. Developed with colleagues, it combines a course and a working environment built around small, checkable steps. Its premise is almost unfashionable in an era when an AI agent offers to complete a whole task in one burst. Howard and his collaborators want the user to remain close to the problem, to read the intermediate work, and to decide what happens next. The machine can help with the labor. The human keeps the responsibility for understanding.

A slower definition of progress

In a 2025 essay, Howard wrote about software craftsmanship at a moment when teams proudly counted the thousands of lines of code an AI could generate. He called Chris Lattner, known for building programming tools that last, to talk about design and durability. The conversation pointed back to a stubborn truth of engineering: code is a record of decisions. If nobody can explain those decisions, a larger pile of code is a larger mystery.

Howard's position is more interesting because he is no stranger to automation. His career has been an extended effort to remove needless difficulty. He made it simpler to train deep-learning models, simpler to learn the field, and simpler to build certain kinds of software. He knows the value of a shortcut. The distinction he now draws is between a shortcut through drudgery and one through the thinking that makes a person capable of the next problem.

That distinction gives his current work a quiet continuity with the couch experiment. ULMFiT let a model carry general knowledge into a new job. fast.ai let a learner carry one practical success into the next lesson. Solveit asks whether AI assistance can do something similar for a human: leave them with more understanding after the task than they had before it. The measure is difficult to put on a dashboard, which may be one reason it deserves attention.

Howard still builds. He still teaches. Rachel Thomas has returned to work with the Answer.AI team, and the access question they posed a decade ago remains open. Who gets to make use of AI, and on whose terms? The answer will be written in courses, code, products and ordinary decisions at desks far from any research lab. A machine may offer the next sentence. Howard would like us to keep reading it carefully.