TL;DR
Andrej Karpathy built the neural network infrastructure that powers Tesla’s autonomous driving ambitions, then stepped away from the machine learning establishment to start Eureka Labs. He’s betting that the next wave of AI progress won’t come from scaling existing models, but from teaching machines to learn how to learn.
Career Highlights
Karpathy’s path to founding Eureka Labs was unconventional: a man who could have remained comfortably ensconced in academia or at a tech giant instead chose to rebuild himself from first principles. He joined Tesla in 2017 as director of AI, arriving at a company desperate to move autonomous driving from demo to deployment. For six years, he became the public face of Tesla’s Autopilot and Full Self-Driving efforts, publishing research, giving talks, and iterating on the neural networks that would process video feeds from millions of vehicles on real roads.
But by 2022, Karpathy had grown restless. The field of deep learning felt ossified—locked into a pattern of bigger models, more data, more compute. He took a sabbatical to think. He wrote extensively on his blog about the state of machine learning, about the gap between research and engineering, about what he called “the neural network’s implicit bias.” He watched GPT-3 and GPT-4 reshape the industry. And he watched, critically, from the outside.
In early 2023, he returned to Tesla briefly as head of AI, attempting to architect a more fundamental rethinking of how the company approached training. But the pull of his own ideas proved stronger. By late 2023, he departed again—this time to found Eureka Labs. The mission: build AI systems that don’t simply memorize or interpolate, but that actually learn to reason and adapt in novel environments. “I think we’re at the beginning of understanding how learning works,” he has said. “Most of what we call AI today is pattern matching at scale.”
I. The Inflection Point
The inflection point arrived not with a single product, but with a realization. Karpathy had spent years optimizing neural networks for a specific, well-defined task: predicting the next frame of video from a car’s camera, or classifying objects in an image. These networks were getting better. But they weren’t getting smarter. They were getting wider and deeper, yes—but they remained fundamentally brittle, unable to transfer learning across domains or adapt to situations far from their training distribution.
That friction became unbearable in late 2022. Karpathy began publishing essays on his personal blog that read less like the measured output of a Tesla executive and more like manifestos. He questioned whether the industry’s focus on scaling—the trend that would culminate in large language models—was actually solving the right problem. He argued that the field had become enamored with brute-force approaches: throw more data and compute at the problem, and emergence would happen. But emergence, he suspected, was a red herring. The gap between a 70-billion-parameter model’s fluency and actual reasoning ability remained vast and poorly understood.
“The next breakthrough won’t come from making the models bigger,” he wrote in one post. “It will come from fundamentally rethinking how we structure learning itself.”
II. The Build
Eureka Labs is still in stealth as of early 2024, but Karpathy’s stated thesis points to a platform designed around a single core idea: adaptive learning systems that can update their own objectives based on environmental feedback. Rather than training a model once on a fixed dataset and deploying it, Eureka’s approach would enable continuous, online learning—systems that can observe failure modes and reorganize their internal representations on the fly.
The architecture rests on several interrelated components:
- A meta-learning layer that allows models to adjust their learning rate and loss function dynamically
- Active learning mechanisms that identify where a system’s uncertainty is highest and focus training effort there
- A reasoning engine designed to work with structured, symbolic representations alongside neural networks
- Simulation and world-model capabilities that allow agents to test hypotheses before acting in the physical world
- Tools for interpretability—making visible how and why a learning system made a particular decision or adjustment
The strategy is to deploy these tools first in robotics and simulation environments, where the feedback loops are fast and the failure costs are manageable. Success in those domains would position Eureka to offer a fundamentally different approach to autonomous systems than what Tesla or Waymo currently pursue. The long-term bet is that learning systems—not just larger models—are the path to systems that can generalize beyond their training domain.
III. The Person
Colleagues describe Karpathy as restless, intellectually voracious, and impatient with consensus. He codes obsessively. He reads deeply—not just machine learning papers, but neuroscience, cognitive science, philosophy. When he’s interested in something, he tends to disappear into it entirely, emerging weeks later with either a breakthrough or a withering critique. He’s equally comfortable in a Tesla engineering meeting or posting sixty-tweet threads analyzing the state of the field to his 200,000+ followers.
He is not a natural politician. He doesn’t network or self-promote in the way most executives do. His public persona is almost aggressively cerebral—focused on ideas, indifferent to status. This has made him simultaneously respected and occasionally isolated in industry circles. At Tesla, some executives found his directness refreshing; others found it difficult. He has no patience for what he calls “presentation science”—work packaged for marketing rather than designed for truth.
“I think the biggest mistake in AI right now is confusing intelligence with fluency,” he told a small group of engineers in 2023. “A language model is a very good mirror of human text. That’s not the same as understanding.”
IV. The Network & Numbers
Milestones Box
- Founded: 2023
- Last Round: Seed funding (undisclosed)
- Valuation: Undisclosed / Stealth
- Employees: ~10-15 (estimated)
- Revenue: Pre-revenue
Key Relationships
- Elon Musk: Former CEO at Tesla; shaped Karpathy’s thinking on autonomous systems scaling
- Dario Amodei: Anthropic CEO; peer in rethinking AI safety and learning paradigms
- Yann LeCun: Meta AI chief; early influence on Karpathy’s approach to representation learning
- Tesla: Former employer; largest operational deployment of Karpathy’s neural network work
V. The Thesis
Karpathy’s thesis is elegantly simple but radical in implication: the industry has optimized for the wrong objective. The field celebrated the rise of large language models because they demonstrated that scale and data and brute computational force could produce systems that appeared intelligent. But apparent intelligence and actual learning capacity are not the same thing. A system that can predict the next token with high accuracy has not necessarily learned how to reason, adapt, or transfer knowledge.
His bet is that the next tier of progress will require a return to first principles: understanding how learning systems actually construct meaning, how they generalize, how they can update themselves in the face of novel information. This is fundamentally harder than building a bigger transformer. But it’s also fundamentally more important if the goal is AI systems that work reliably in the real world, rather than on benchmarks.
Eureka Labs is, in essence, a bet that someone who helped build Tesla’s autonomous driving infrastructure understands something about the problem that the large model establishment still misses. “Scaling is not the answer,” he posted in late 2023. “Understanding is.”
Factbox
Name Andrej Karpathy Age 36 Location San Francisco, USA
Company & Role Eureka Labs, Founder Funding Stealth / Seed
Most Recent Round N/A Employees ~10-15
Contrarian Belief Scale and compute are red herrings; real progress in AI will come from understanding how learning itself works at a fundamental level, not from making existing models bigger.
