A shopper pauses over a jacket. Scrolls. Reverses. Watches six seconds of a video. Leaves the question box unopened. None of these gestures is a declaration. Together, they are a tiny weather system. Ben Vigoda has built a career around weather systems like this: incomplete evidence, noisy inputs and decisions that cannot wait for certainty.
Today, as founder, CEO and chief scientist of Product Genius, he applies that old preoccupation to ecommerce. The company adapts a storefront as a visitor moves through it, arranging products, videos, reviews and answers around signals of interest. The shop is no longer a fixed shelf. It behaves more like a conversation whose cleverest participant knows when to stop talking.
The fashionable description would be a pivot into agentic commerce. Vigoda's record suggests something steadier. His materials have changed from transistors to Bayesian programs to browser interactions. His question has barely moved: how can a machine learn something useful from a world that refuses to be tidy?
The chip that expected confusion
At MIT in the late 1990s, Vigoda studied physics, circuits and machine learning in an era when those subjects did not yet arrive packaged as one booming industry. His doctoral work explored analog circuits for statistical signal processing. The basic provocation was practical and philosophical: ordinary programs can become brittle when an input is noisy, unanticipated or contradictory. A probabilistic machine can weigh alternatives instead of sulking at the first surprise.
That research became Lyric Semiconductor, co-founded with veteran chip designer David Reynolds. The company developed processors for statistical inference and machine learning, raised more than $20 million, and spent years working quietly before publicly showing its probability-processing technology. In 2011, Analog Devices acquired Lyric. The team stayed in Cambridge and continued as Lyric Labs.
Vigoda later described the hardware with the cheerful specificity of someone who enjoys making impossible objects sound domestic. One analog design used 440,000 transistors for work that he estimated would have required roughly 30 million digital transistors, with about a tenfold improvement in energy per operation. Whether discussing currents or commerce, he prefers a claim that can be tested.
“That was the main thing I learned: to translate your idea from technical units to economic units.”Ben Vigoda, on an MIT entrepreneurship lesson
The line came from an early embarrassment useful to nearly every technical founder. In MIT's entrepreneurship competition, Vigoda first explained the chip in joules. Venture capitalists became more attentive when he explained it in dollars. The physics did not change. The altitude did. An invention could save energy; a business had to save something the buyer already counted.
A career with one invariant
At Analog Devices, Vigoda built and led AI and machine-learning chip efforts, established research labs and started a corporate venture group. Then came Gamalon. Its Bayesian Program Synthesis work aimed to let machines revise the mathematical models used to explain data. The system was presented as learning from a few examples, quickly and on modest hardware, rather than requiring enormous labeled datasets and server fleets.
Gamalon's commercial products tackled unstructured enterprise information: scraps of product descriptions, inconsistent addresses and the many linguistic disguises worn by a case of Diet Coke. The work moved up the stack, but it remained an argument against brute force. Do not merely make the pile larger. Make the machine better at reasoning about the small pile in front of it.
Product Genius brings this logic into a more theatrical setting. A person arrives with no briefing and performs a sequence of micro-decisions. The system has minutes, sometimes seconds, to infer interest. Vigoda calls the company's approach a Large Interaction Model, trained around behavior rather than language alone. Product Genius says controlled tests compare its adaptive experience with a merchant's existing pages; the important measure is revenue per session, the commercial descendant of those old joules.
Improvisation with consequences
There is another route into Vigoda's ideas, and it begins with drums. He has led bands, built experimental instruments and created games for collaborative musical improvisation. He once hosted composer John Zorn as a guest artist at an MIT workshop. His present group, The Q@YS, is described with a grin as a cross between the Grateful Dead and Radiohead.
The music is not a decorative eccentricity beside a serious engineering career. Improvisation is inference under pressure. A player hears a phrase, forms a belief about where it might go, responds and revises. Nobody pauses the room for a quarterly retraining cycle. A good contribution reflects the pattern without becoming trapped by it.
His earlier inventions make the same connection visible. In 1999, before anyone could casually rent a GPU in the cloud, Vigoda assembled a computer-vision system for the Flying Karamazov Brothers. The technology turned juggling into an interactive virtual performance and traveled with the troupe for more than a decade. Later, he built Drum-o-Saurus in response to the large, loud and emotionally unsatisfying options available to an urban drummer. His complaint about electronic drums was gloriously tactile: they felt like holding a lover's hand with pliers.
“Always think with your hands.”The MIT maxim Vigoda carried into company building
The maxim came through his MIT adviser Neil Gershenfeld. Start building before the idea has perfected its evening wear. Architects make models to discover what they want; engineers can do the same. It is a rebuke to the founder who keeps polishing a worldview no customer has touched.
Vigoda's playful work also reveals a boundary. For his short film 5 Tiger Morning, he used generative video because hiring tigers to run through downtown was, among other things, a poor production plan. Yet he has written that he is strongly opposed to AI-generated music. The film's soundtrack came from live improvisation by his bandmates. He is happy for a machine to supply the impossible tiger; he still wants people making harmony together in the moment.
Buyer, meet seller's machine
Vigoda's recent writing looks beyond the storefront toward an economy in which buyer AIs and seller AIs negotiate. He helped lead a DARPA workshop on such markets in December 2025. His conclusion is crisp: the agent that learns fastest from the fewest interactions gains an advantage, but speed without trust is a fast route to rejection.
A buyer's assistant should represent the buyer. If it quietly favors whoever pays for placement, the relationship curdles. Sellers, in turn, need agents capable of presenting accurate products, information and prices quickly. Vigoda compares them to market makers: always listening to incoming signals, always adjusting the quote. It is ecommerce reimagined as a lively exchange rather than a warehouse aisle with better search.
The distinction matters because it restores a little honesty to personalization. The seller wants to sell. The buyer wants the right purchase. A credible system does not pretend those interests are identical; it makes the negotiation useful and legible. Product Genius is Vigoda's candidate for the seller's side, continuously learning what to show without requiring a visitor to deliver a memoir first.
The human keeps the verdict
In August 2026, Vigoda published a field guide about using an AI agent to clear 37,000 unread emails. The interesting part was not synthetic prose. It was triage. The system classified every item, surfaced the consequential few and accepted compact human decisions: archive, delete, turn into a task, show more detail, draft a reply.
His rule was proposal before action. Drafts were reviewed. Destructive steps stayed behind approval. The agent learned standing preferences, but the person kept the verdict. Here again is the old structure: uncertain evidence, explicit alternatives, rapid updates, action. The probability processor has become a work habit.
It is tempting to compress Vigoda into a tidy label: chip pioneer, Bayesian contrarian, ecommerce founder, bandleader. Each is true and none is quite sufficient. The more revealing unit is a loop. Listen closely. Hold several explanations. Build something. Watch what happens. Update before the room changes.
A quarter-century after his MIT work, the world has finally become noisy enough to resemble his research problem. Machines now converse, recommend, negotiate and occasionally hallucinate with confidence. Vigoda's answer remains pleasingly unfashionable in its discipline: uncertainty is not an inconvenience to be hidden. It is the material. Teach the system to work with it, and leave a human near the final button.