OpenAI's Founding Brains: From High-Frequency Trading to Speech

The early days of any research lab are defined by the obsessions of its first hires, and a newly surfaced roster from OpenAI’s formative period reads like a map of where the field was heading—and where it would eventually fracture. The team that Greg Brockman and Jie Tang assembled wasn’t just a collection of PhDs; it was a deliberate blend of hardcore systems engineering, theoretical neuroscience, and a budding, almost prescient, focus on safety.

The list is anchored by names that would become synonymous with the modern AI industry. Dario Amodei, who later co-founded Anthropic, joined after leading Deep Speech 2 and co-authoring the seminal “Concrete Problems in AI Safety” paper. His presence signals that existential risk wasn’t a marketing afterthought—it was baked into the technical culture from the jump. Alongside him, you see the practical, performance-obsessed side of the operation in Scott Gray, previously at Nervana Systems, who wrote assembly-level GPU optimizations for deep learning that were, by the account, the fastest available anywhere. That combination of high-minded safety theory and low-level kernel hacking was a rare alchemy.

The eclectic backgrounds are what make the document so fascinating. Filip Wolski came from the high-frequency trading world, bringing a “practical” modeling discipline forged in hyper-competitive financial markets. Zain Shah was building multimodal GIF search engines and human-machine intelligence systems at Clara Labs. Then there’s the contingent of visiting researchers—Taco Cohen inventing group equivariant convolutional neural networks, and Tambet Matiisen bouncing between cooperative reinforcement learning agents and decoding a rat’s brain activity to predict its location. Jack Clark, who joined as Strategy and Communications Director after years covering the beat at Bloomberg, was tasked with translating this dense technical frenzy for the outside world.

Looking back, the document captures a moment of unusual density. It’s a snapshot of a lab that managed to pull in Catherine Olsson, who built OpenAI Gym’s REST API with a perfect MIT GPA in brain science, alongside Igor Mordatch, who was automating discovery of complex movement behaviors before heading to the CMU faculty. The structure—a core team plus a rotating cast of hyper-specialized PhD students on limited engagements—reflects an academic-mercenary model that proved wildly effective. It’s hard not to read this now and see the seeds of a dozen different futures, many of which have since become rival companies.