Jordan Fisher and the problem with easy levels

Jordan Fisher and TJ Lutz spent eight hours trying to beat a level that a computer had designed. Fisher had been working on a mathematics PhD when he began the experiment. He wanted to see whether an AI could make Mario-style levels. Once the program worked, he showed it to Lutz. They turned the difficulty up, and there went the day.

It is an appealing origin story because the machine was already doing something useful: furnishing its makers with an excuse to avoid other work. The experiment became Cloudberry Kingdom, a platform game built by their studio, Pwnee. Fisher would later co-found Standard AI, help bring autonomous checkout into convenience stores, and move into AI alignment research at Anthropic. The path sounds peculiar until you look at the questions underneath it.

How do you make a system that can cope with changing conditions? How do you know its output is any good? And who gets to decide what “good” means? Those questions look different when the output is a difficult jump, a shopping receipt, or an answer from an AI assistant. Fisher has spent his career moving among them.

A kingdom built during a detour

Fisher’s academic training was in applied mathematics. He studied at the University of California, Santa Barbara from 2006 to 2011, earning a doctorate. His public description of that work concerns numerical solvers and the immersed boundary method. It is a considerable distance from a cartoon hero dodging obstacles, although both require an understanding of how things move.

The game began with curiosity rather than a finished commercial plan. “It seemed like a cool math problem,” Fisher recalled. The challenge was to generate levels that were surprising and demanding while remaining possible to complete. Anyone can scatter hazards across a screen. Giving a player a fighting chance requires more thought.

Fisher described three requirements for his generator: feasibility, interesting design, and appropriate difficulty. The first asks whether the player can finish. The second asks whether the player wants to. The third asks whether the challenge suits them. A perfectly feasible level can be dull; a spectacular one can be useless if every possible route ends badly.

The game used a computer player to verify that a level could be beaten. That also supplied an elegant answer to the frustrated human: the AI could show a successful run. The apparent impossibility had a witness against it. Whether watching that witness improved the player’s mood is another matter.

The project took roughly a year of work on the algorithm alone, Fisher explained in a 2013 interview. It also passed through changes in its visual style. The studio brought in artists, sound designers, and musicians to turn the underlying machinery into a game people could recognize and enjoy. A clever generator still needed a world around it.

Ubisoft published Cloudberry Kingdom in 2013. Its levels accommodated different character abilities and altered physics, making variation part of the design. Fisher’s early experiment had become a product, with all the extra decisions that word entails. The math problem could no longer keep its own company.

The next level had shelves

Fisher’s career also took him into financial regulation. From 2014 to 2017, he served as Development Branch Chief at the U.S. Securities and Exchange Commission. His work there involved machine learning. Video games and securities regulation make an unusual pair on a résumé, but each gave him a different setting in which to apply computation.

Standard’s early work began before the company’s formal founding. The team dates its beginnings to a Texas garage in 2015, where former SEC engineers experimented with computer vision for shopping. Standard Cognition, later called Standard AI, was founded in 2017 and joined Y Combinator’s summer batch that year.

Fisher was part of a founding group that included Michael Suswal, David Valdman, John Novak, and other technical collaborators. The project required several kinds of expertise: learning systems, hardware, and the business of persuading retailers to put unfamiliar equipment into familiar spaces. A shop does not become a laboratory simply because engineers would find that convenient.

The proposed experience was easy to describe. A customer would enter, select products, and leave without stopping to scan them at a register. Cameras and computer vision would track the shopping activity needed to assemble the transaction. For the person buying a drink, simplicity was the point. For the people building the system, simplicity was the assignment.

A shop is a stubborn piece of software

In October 2021, Standard’s technology became available to customers at a Circle K store in Tempe, Arizona. The location had been retrofitted, rather than built from scratch around the system. It was smaller than a typical convenience store while carrying the same assortment, making product density a particular challenge for the ceiling cameras.

The team tested the technology while the store continued operating. Selected shoppers tried it before the wider rollout. Conventional checkout remained available. That detail gives the scene its practical shape: the future arrived alongside an existing way to pay, in a shop that still had to serve customers.

Fisher wanted installations to become faster. The ambition was to move from lengthy projects toward deployments measured in months, then weeks, then days. “It shouldn’t take a year to launch a store,” he said. Speed mattered because every new location brought another physical installation. Scaling meant returning to the ceiling.

“It shouldn’t take a year to launch a store.”

Jordan Fisher, 2021

Earlier that year, Standard announced a $150 million Series C led by SoftBank, at a valuation above $1 billion. The funding gave the project considerable financial backing. It also helps explain the scale of the expectations surrounding autonomous checkout. The numbers were large; the unit of progress remained a working store.

A store installation has a stubborn physicality. Someone must fit the equipment, keep it working, and support the customer using it. Software that occupies a ceiling cannot reach another hundred locations through a download alone. For Fisher, the retail years made the distance between an impressive capability and an everyday service unusually visible.

By 2023, he was speaking openly about the difficulty of growing the market. The work had stages: prove the system works, make deployments repeatable, and make the economics acceptable. Retailers might move substantial amounts of money through their businesses without having much available to spend on new hardware.

This is where a founder’s original fascination meets someone else’s priorities. The retailer needs a better store, a manageable installation, and a price that makes sense. The shopper needs an experience they can understand. The algorithm has to satisfy all of them while quietly doing its own difficult job.

What the ceiling could see

Fisher’s ambitions for computer vision extended beyond the moment of payment. At Collision in 2023, he discussed AI’s capacity to improve experiences in physical retail. He also challenged the idea that collecting data was sufficient. “Data is dead,” was his deliberately provocative shorthand; the emphasis was on making information useful.

The point matters in a shop. A sale tells a retailer that something was purchased. It leaves many questions about what happened beforehand. Understanding movement and product interaction creates another kind of information about the space. The cameras installed for one task may help reveal patterns relevant to other decisions.

Standard’s account of its own development follows that widening view. Building technology to map a customer’s journey from entrance to exit led the team toward a broader analytics product. The company compares the opportunity to what web analytics brought to online commerce: a way to examine behavior as well as outcomes.

In 2024, Standard announced its move toward vision analytics and a change in leadership. Angie Westbrock, who had been chief operating officer since 2021, became CEO. David Woollard became CTO. Fisher remained on the board as chairman. His role changed as the company’s focus changed.

He publicly welcomed Westbrock and Woollard into their new roles and said he looked forward to supporting the team from the board. There is no need to turn that transition into a tidy moral. A founder’s work can take several forms, and a company’s next phase asks for its own decisions. Fisher’s subsequent work would take him back toward research.

A different kind of difficult

By June 2025, Fisher was leading an AI alignment research team at Anthropic. At Y Combinator’s AI Startup School, he opened with an admission: “I’m extremely confused.” He described a shrinking ability to predict where technology would be years ahead and invited founders to reconsider their assumptions about strategy, products, and teams.

It was a candid opening from someone with experience selling a vision of what came next. Confusion, in his account, could signal the beginning of an interesting scientific question. His talk asked what would endure as AI changed the cost of building software, and how trust would work as systems took on more responsibility.

His research role also brought him into conversations about institutions. In a 2025 session with Stanford students, Fisher and Anthropic’s Ashley Zlatinov discussed adapting models to organizations’ particular needs. Anthropic later acknowledged Fisher among those who gave feedback on Claude’s Constitution, its statement of intended model values and behavior.

Look back at the platform game and the resemblance is suggestive, rather than exact. A machine could produce an elaborate level, but someone still had to decide whether it was possible, interesting, and suitable. More capable systems enlarge the territory in which those judgments must be made. They do not make judgment disappear.

Fisher began one chapter by asking a computer to make his leisure time more difficult. He has since helped put AI above the shopping aisle and returned to questions about its behavior. The settings keep changing. Somewhere in the work, there is still a person inspecting the result and asking whether anyone should have to play this level.