Charles K. Fisher and the trouble with being right

In 2012, Charles K. Fisher had good reasons to doubt deep learning. He had been trained to prefer smaller models, worry about too many parameters, and value an explanation he could inspect. When his postdoctoral advisor introduced him to the field, skepticism came readily. The education was working exactly as intended. That would become the problem.

Large neural networks kept producing results that his inherited rules struggled to accommodate. Fisher eventually changed his mind. Years later, he made that revision public, setting out the assumptions he had abandoned. A scientist admitting error is doing his job, of course. Still, a list of one’s own mistaken convictions makes a less comfortable professional accessory than a Harvard degree.

Fisher has both. The Unlearn.AI co-founder holds a doctorate in biophysics from Harvard and a bachelor’s degree in the same subject from the University of Michigan. His career has taken him through academic physics, computational science, virtual reality, and company building. His essays keep returning to a question those credentials cannot settle: when should you trust your intuition, and when should you replace it?

The respectable way to be wrong

The early mistake was understandable. Fisher had absorbed a view of modeling in which restraint conferred reliability. Increasing complexity could mean fitting the peculiarities of yesterday’s observations rather than learning something useful about tomorrow. The smaller model had intellectual good manners.

His encounter with deep learning required him to reconsider that preference. In his account, adding parameters could improve performance on unfamiliar examples. A model trained across different tasks could also acquire useful abilities that a narrow training set would miss. Results were moving faster than the intuitions he used to judge them.

He does not describe an instantaneous conversion. He describes years of adjustment. That duration matters. Knowing that an experiment contradicts your expectations and acquiring a new feel for the subject are different achievements. Fisher’s later emphasis on intuition carries this history inside it: an instinct can be trained, and it can become obsolete.

Three researchers, then a company

By 2017, he and two colleagues, Aaron Smith and Jon Walsh, were ready to start something. They had worked together on the machine learning team at Leap Motion, the virtual reality company. Their earlier training was in mathematics and physics. After leaving, they met once or twice a week with their friend Cami, who would become an advisor.

They shared a conviction that generative models offered a promising route for AI. Much of the visible research concerned images and text. The team wondered what might happen in other kinds of systems, especially those containing noisy observations and missing information. A polished image is pleasing; reality rarely bothers to pose.

Their work combined ideas from restricted Boltzmann machines and generative adversarial networks. They called the resulting architecture Boltzmann Encoded Adversarial Machines, or BEAMs. Fisher later regretted missing the chance to call it BoltGAN Machines. Even mathematical naming has its l’esprit de l’escalier.

DCVC supplied a small pre-seed round. Fisher recalls having neither a business plan nor a dedicated business person at that point. He recruited his friend Graham from Boston to help with the commercial work. The name Unlearn.AI came from “unsupervised learning.” The application would take more sorting out. There had even been talk of weather simulation, which the team never pursued.

The delete key belongs in the laboratory

Fisher’s advice on research is more domestic than the language of discovery usually allows. Begin with a baseline. Sit with the results. Try something. Think again. Resist the temptation to construct a magnificent experimental machine before deciding which experiment deserves to run.

He calls an informal version of this process “graduate student descent.” The joke works because the researcher is doing some of the optimization personally, adjusting an idea through repeated encounters with the result. Speed of understanding matters before speed of execution. Automation can arrive after there is something worth automating.

Then comes his less comfortable instruction: delete the things that did not work. A failed approach left conveniently implemented invites another attempt, and another. Yesterday’s investment quietly starts allocating tomorrow’s attention. Fisher points out that version history can preserve the work without leaving it in the path of every new decision.

There is a modest discipline behind the apparent impatience. Each experiment should improve the researcher’s feel for the problem. The next idea should follow from that understanding. The code that happens to be easiest to run should have less say in the matter.

A physicist with an artist’s vocabulary

In Fisher’s description of scientific work, taste belongs near the center. He writes about imagination, beauty, and the satisfaction of an explanation that feels right. Mathematics supplies a medium for representing the world. The scientist develops a feel for which representation might be worth pursuing.

“Science is a form of art in which one uses mathematics to paint reality.”

Charles K. Fisher

He also insists on an inconvenient obligation for the artist: the model must make useful predictions. Loving an idea cannot excuse its failure. Scientific taste earns its standing through repeated contact with what actually happens.

His suggested exercises are concrete. Imagine an experiment before performing it. Picture the measurements and the shape of the graph. Compare that imagined result with the observed one. Practice explaining a difficult idea simply enough that another person can follow it. In his view, explanation helps form the understanding itself.

Fisher acknowledges that this approach can make collaboration difficult. Researchers may respect one another and still disagree about what a satisfying explanation looks like. He describes having encountered precisely that problem. Shared intelligence does not guarantee shared taste, and an impressive colleague can remain a poor partner for a particular piece of work.

Learning to work with other people

His own route out of academia included research appointments at Boston University and a Philippe Meyer Fellowship in theoretical physics at École Normale Supérieure in Paris. He also worked at Pfizer before Leap Motion. The moves exposed him to organizations where an interesting personal project had to compete with a collective purpose.

For researchers making the same transition, he recommends learning to share intermediate results, communicate clearly, and connect proposals to what the company is trying to accomplish. An intriguing question needs a practical case for the time and resources it consumes.

He is particularly blunt about programming. Academic experience can leave a researcher underprepared to write software that colleagues can maintain. Worse, the researcher may not recognize the gap. An elegant result becomes more useful when somebody else can understand the code that produced it.

These recommendations temper the romantic portrait of the scientist following an inner voice. Fisher gives intuition room to generate an idea. He also expects the person with the idea to make it legible, useful, and available to a team.

The risk of choosing the easy business

That practical streak does not make him an advocate for avoiding difficult technology. His argument for research investment begins with two questions: is the technology differentiated, and do customers trust its value? A new company rarely gets both advantages on its first day.

Fisher favors building something technically distinctive and doing the work to establish its value. The apparently easier alternative, using technology everyone can obtain, leaves a business looking for another source of distinction. Price and service bring their own costs. Research risk can look rather different when the alternative is permanent competition on familiar ground.

Unlearn grew through the research-heavy years. When it announced Fisher’s departure from the CEO role in September 2024, the company said it had secured more than $130 million in funding. Steve Herne would run the business; Fisher would become Chief Scientist. The change assigned different work to the founder after more than seven years in charge.

Turn the amplifier up

Fisher’s prose has fewer reservations than his experimental method. He objects to professional writing that sands away every edge until the reader has nothing left to react to. He recommends a firm position, a speaking voice, and the willingness to break a grammatical rule when the sentence benefits.

His chosen metaphor is an amplifier turned to eleven. There is a theatrical purpose to the volume: he says he may write a position more forcefully than his private certainty warrants. Readers should understand that distinction. The essay is partly a performance designed to provoke thought.

It is an amusing preference in someone who values careful experiments. The experiment must answer to the result. The paragraph, apparently, is allowed to kick the furniture on its way in.

A different kind of approval

By the end of 2024, his writing had also become more personal. In reflections on turning 39, he described chasing status through Michigan, Harvard, and Silicon Valley. Moving among prestigious groups had gradually weakened his belief that the people inside them possessed some exceptional authority over his own judgment.

“You can just do things.”

Charles K. Fisher

He associated greater happiness with reducing the number of opinions that governed his life. Large decisions could give a period of time a direction that ordinary days lacked. Leaving a familiar role belonged to that account of change. He urged acting when a career had drifted away from a personal mission.

EPIC Ventures now lists him as a venture partner. In December 2025, he published his view of AI’s prospects for 2026, arguing for opportunities in fundamental research, new datasets, and different product designs. In March 2026, he publicly announced plans to leave the Bay Area and return to the greater Ann Arbor area in April.

Michigan was already part of the biography. Returning gives it another place in the story. The researcher who once had to revise his rules for judging a model has also been revisiting his rules for judging a life. He has spent enough time among impressive people to want more say over his own next experiment.