He wrote the brains for video game villains. Now he writes them for robots that don't know they're in a video game.
Push a soft, deformable part into a rigid one. A toddler manages it. A million-dollar industrial robot, in 2026, mostly does not. So a human hand still does the job in automotive, aerospace, and electronics plants around the world.
Matthew Brown finds that absurd, and he has built a company around fixing it. ThoughtForge AI, run out of a building on 66th Street in Berkeley, makes control software that gives robots already bolted to factory floors something they have always lacked: the ability to feel, predict, and adapt in real time, on the edge, without phoning the cloud.
The trick is an idea borrowed from neuroscience called Active Inference - the theory that brains are prediction engines, constantly guessing what happens next and acting to make the guess true. Brown bet his company on it. Early customers, including the energy giant bp, say it works better than anything they had tried.
Start with the resume that shouldn't fit on one page. Missile-defense logic at Lockheed Martin. The enemy AI in Halo, Bioshock, and Splinter Cell. The first engineering hire at a deep reinforcement learning startup that Microsoft would buy. And now, a robotics company. The thread running through all of it isn't an industry. It's a question: how does anything - a soldier, a game villain, a robot arm - decide what to do next?
Brown studied computer science and engineering at MIT, finishing in 2002, then took a master's at the University of Pennsylvania. The early career detour into video games was not a detour at all. Building believable enemies is one of the hardest applied AI problems there is. A guard in Splinter Cell has to perceive, guess where you went, and react in milliseconds, all while feeling alive rather than scripted. Brown spent years on what he calls percept-based agency and the geometry of how bodies move through space.
Then he left games entirely. He joined Bonsai AI as its first engineer and technical lead, helping build one of the world's first production-grade deep reinforcement learning platforms - software that let companies teach machines through trial and reward rather than hand-coded rules. It was early, it was hard, and it worked. Microsoft acquired Bonsai in 2018, and Brown became the company's liaison to UC Berkeley's RISE Lab, a front-row seat to where machine learning research was actually heading.
That seat seems to have unsettled him in a productive way. The more he watched the field scale up - bigger models, more data, more compute - the more he became convinced the giants were solving the wrong problem. Real intelligence, the kind a living creature has, is not data-hungry. A child learns to grip a cup from a handful of tries, not ten million. Brown went looking for an architecture that learned like life does.
He found it in Active Inference, a framework grounded in neuroscience that treats perception and action as two sides of the same coin: the brain predicts its sensations and moves the body to reduce its own surprise. ThoughtForge was, in his words, "born from a desire to develop biologically-plausible learning systems" that let robots "interact with the world as living systems do: efficiently, dynamically, adaptively, and resiliently."
That is the difference ThoughtForge sells. Its software does not need a fat pipe to a data center. It runs on whatever CPU is already in the cell, with no WiFi, which matters enormously to a safety engineer who does not want a robot's judgment hostage to a dropped connection. It reads force-torque sensors and cameras at once - multimodal sensing - and adjusts on the fly when a part is slightly off, a surface is greasier than expected, or the world simply refuses to match the spec sheet.
The validation that matters most to Brown is the kind he can't buy. ThoughtForge joined the ARM Institute - the federally backed robotics manufacturing consortium with more than 400 member organizations - precisely because it stamps third-party credibility onto a young company's claims. "The biggest challenge a business has to acquire customers," he says, "is getting public 3rd party validation of the value and performance of the technology." Their current ARM project tackles that gasket-shaped problem head-on: adaptive robotic insertion of automotive parts, built with Siemens bin-picking and a stack of force and vision sensing.
A founder is partly the team he can convince to follow him, and Brown's bench is unusually deep for an eleven-person shop. His chief product officer is a former robotics professor whose research ended up inside commercial home vacuums, with stops at Bell Labs, Microsoft, Intel and Amazon's AWS RoboMaker. His revenue chief has spent nearly two decades taking enterprise software companies to market and helped seed-fund robotics startups before joining. They left comfortable orbits for a row house in Berkeley because the underlying idea is genuinely contrarian, and contrarian ideas are the only ones worth a decade of your life.
Beyond the company, Brown teaches. He sits on the board of and guest-lectures at the Active Institute, a global group working to build literacy in Active Inference across neuroscience and AI. On Medium and Twitter he writes under the handle @MalleableIdea and describes himself, with a straight face and a wink, as "Singularity Sprezzatura" - sprezzatura being the old Italian art of making something difficult look effortless. It is a fitting motto for a man whose entire career is about making machines look like they understand.
Put the pieces together and a pattern emerges that is easy to miss in any single role. Brown is drawn, again and again, to the exact moment a system stops following rules and starts behaving. The guard who flanks you. The control policy that learns by reward. The robot that feels a part shift and corrects. Each is the same question in a different costume - what does it take for a machine to act as if it has a stake in the outcome? He has chased that question through a defense lab, a game studio, a unicorn acquisition, and now a factory floor, and he shows no sign of running out of places to look.
I've spent over 24 years in AI, driven by a fascination with the nature of thought and how living systems generate agency.— Matthew Brown, ThoughtForge AI
Works with Yaskawa, FANUC, KUKA, ABB and Universal Robots. Runs on any CPU, on the edge, using multimodal force-torque and vision sensing. Figures per ThoughtForge AI.
The guards that hunted you in Halo, Bioshock and Splinter Cell ran on the kind of real-time, percept-based AI Brown built. Believable enemies are a brutal applied-AI problem - he treated it as basic research.
He didn't join deep reinforcement learning once it was safe. He helped ship one of the first production platforms - the kind of bet that ends in a Microsoft acquisition or nothing.
While the industry chased ever-larger models, Brown went the other way - toward low-data learning modeled on how brains and bodies actually work.
He lectures on Active Inference and serves on the board of a global institute devoted to it - evangelizing the idea, not just selling the product.
I've spent over 24 years in AI, driven by a fascination with the nature of thought and how living systems generate agency.
We joined the ARM Institute because its commitment to pioneering robotics, automation, and digital manufacturing perfectly complements our mission.
The biggest challenge a business has to acquire customers is getting public 3rd party validation of the value and performance of the technology.
His online handle is @MalleableIdea - and his self-description, "Singularity Sprezzatura," is the most Berkeley-AI-founder sentence ever written.
He helped design how enemies in Halo and Bioshock decided to hunt you. The next time you got flanked, that was applied AI.
His bet rides on Active Inference, a brains-as-prediction-machines theory rooted in the work of neuroscientist Karl Friston.
From writing missile-defense logic at Lockheed to game villains to factory robots - one of the stranger straight lines in tech.
ThoughtForge software runs offline, on any CPU, and plugs into the robots manufacturers already own. No rip-and-replace.
The team he assembled includes alumni of Bonsai, Microsoft, Amazon, Bell Labs and Duke - a robotics brain trust in a Berkeley row house.