Kenneth Stanley’s foundational work emerged at the University of Central Florida, where he joined as assistant professor in 2006 and rose to Charles Millican Professor. His 2004 Ph.D. thesis under Risto Miikkulainen at UT Austin — “Efficient Evolution of Neural Networks Through Complexification” — introduced NEAT (Neuroevolution of Augmenting Topologies), a topology-and-weight evolving artificial neural network (TWEANN) that begins with the simplest possible network and grows structure over generations exactly as biological evolution complexified nervous systems.

The decisive break came with PicBreeder (2007). Users collaboratively evolved images by selecting for subjective “interestingness.” The shocking result: highly complex, recognizable forms (skulls, butterflies, cars) almost never appeared when users tried to evolve them on purpose. They emerged only from lineages that had followed gradients of novelty without any final target in mind. The intermediate “stepping stones” bore no resemblance to the eventual outcome; any objective function would have pruned them as failures.

This single experiment crystallized the mathematical flaw Stanley would spend the next two decades weaponizing: in deceptive search landscapes, the path to transformative discovery does not resemble the destination. Objective-driven optimization therefore systematically eliminates the very branches required to reach it.

In the 2015 book with Joel Lehman, Why Greatness Cannot Be Planned, Stanley formalized the thesis: ambitious objectives function as a restrictive tyranny in any sufficiently complex domain. They are useful for incremental, one-step improvements where the next milestone is visible. For anything transformative — genuine scientific discovery, radical innovation, or general intelligence — they become actively counterproductive.

The algorithmic antidote is Novelty Search: replace the fitness function entirely with a reward for behavioral novelty. No target, no distance metric, no “progress.” The agent is simply incentivized to do things that have never been done before in the population’s history. In deceptive maze environments where every objective-seeking robot becomes trapped against a wall, novelty-driven agents reliably discover the exit as a byproduct of exhaustive exploration of the stepping-stone space.

The societal corollary is devastating for grant committees, corporate OKRs, and standardized education: when complex systems are forced to optimize for a simplified numeric proxy, actors game the metric and the underlying system degrades (Campbell’s Law). The institutions that most loudly demand “measurable progress toward ambitious goals” are precisely those least likely to produce genuine breakthroughs.

In 2014 Stanley co-founded Geometric Intelligence with Gary Marcus, Zoubin Ghahramani, and others to commercialize evolutionary approaches that learn more robustly and with less data than deep learning. Uber acquired the company in 2016 to seed Uber AI Labs; Stanley became Head of Core AI Research.

The flagship output was POET (Paired Open-Ended Trailblazer) and its enhanced successor. Instead of training a single generalist agent on a fixed, human-designed environment, POET simultaneously evolves both the environments (increasingly difficult terrains for a bipedal walker) and the specialized agents that solve them. Agents are periodically transferred across the population of environments, allowing stepping stones discovered in one niche to be recombined in another. The result is an endless, self-generated curriculum — an Artificial Intelligence Generating Algorithm (AIGA) that writes its own problems faster than any human benchmark designer could.

In 2020 Stanley joined OpenAI to lead the newly formed Open-Endedness Team with Joel Lehman. The 2022 ELM paper (“Evolution through Large Models”) demonstrated that large language models, trained on vast corpora of human code and edits, can serve as highly intelligent, context-aware mutation operators inside an evolutionary loop.

Traditional genetic programming mutates code randomly; most mutations are catastrophic syntax errors. ELM uses an LLM to propose structured, semantically plausible edits. The system generated hundreds of thousands of functional Python programs that produced walking virtual robots in the Sodarace 2D physics domain — programs the base LLM had never seen during pre-training. These evolved artifacts then bootstrapped a new conditional model capable of generating customized robots for previously unseen terrains. Open-ended search, augmented by foundation models, had manufactured its own high-quality training distribution in a domain with zero prior data.

In May 2025, Stanley, Akarsh Kumar, Jeff Clune, and Joel Lehman published the Fractured Entangled Representation (FER) hypothesis — perhaps the most surgically precise critique of the scaling paradigm yet offered.

They trained two networks on an identical minimal task (draw a skull). One used conventional SGD/backpropagation; the other used open-ended evolutionary search mirroring PicBreeder. Both achieved perfect output behavior. Their internal representations were diametrically opposed.

The SGD network exhibited FER: a tangled, disorganized mess of arbitrary mathematical pathways that happened to collapse into the correct pixels at the final layer — “impostor intelligence.” It had memorized a rote mapping without ever factoring the underlying structure (jaw, cranium, bilateral symmetry). The evolved network developed a Unified Factored Representation (UFR): modular, disentangled pathways that genuinely “knew” where the jaw was and how the cranium curved. Only the UFR network possessed the discrete building blocks required for robust generalization, continual learning, and genuine creativity.

The implication is structural, not incremental: continued scaling of SGD-trained foundation models may be approaching a hard wall precisely because they are optimizing for benchmark performance rather than for the internal representational cleanliness required for open-ended scientific discovery.

Stanley’s appointment as Senior Vice President of Open-Endedness at Lila Sciences (Flagship Pioneering) represents the thesis reaching its logical terminus: the migration of open-ended search from digital simulators into the physical world.

Lila’s AI Science Factories are fully autonomous laboratories in which generative models with genuine scientific-reasoning capability are wired directly to robotic fluidics and assay hardware. The closed loop — hypothesize, synthesize, assay, interpret, refine — runs without human intervention in the iteration. Hypotheses are generated according to gradients of interestingness rather than human-specified targets. The system has already produced genetic-medicine constructs outperforming commercial therapeutics, hundreds of novel antibodies and peptides, superior carbon-capture materials, and non-platinum catalysts for green hydrogen.

Return to the scout/assassin architecture Stanley first built at the nanoscale in 2009 (von Maltzahn file N°010). Autonomous agents, no central controller, coordinating through signals written into a shared environment. At Lila the hypotheses are the scouts, the physical assays are the flares, and the next iteration reads the environment and recruits. He has been building one machine since the PicBreeder insight. Lila is the version that runs on robots and weights instead of gold nano-antennas.

Stanley’s wager is Richard Sutton’s Bitter Lesson made physical and recursive. General methods that lean on computation and search beat human-engineered domain knowledge every time, at sufficient scale. In silico this produced the language models. At Lila it is being tested on the scientific method: do not teach the machine the rules of chemistry; give it a lab and let it discover the rules by running the method at superhuman breadth and speed.

The moat is not the model weights. It is the proprietary, high-fidelity physical data generated by every closed-loop experiment — data no competitor can scrape from journals. Lila reports having already manufactured more than 10 trillion tokens of scientific-reasoning data, with the corpus projected to rival frontier LLM training sets by the end of 2026. The factory that feeds the model is the strategic asset.

When an autonomous loop can design around rare-earth magnets or platinum-group catalysts that a geopolitical rival controls, open-ended science ceases to be merely an R&D accelerant. It becomes infrastructure of national resilience. The White House Genesis Mission, the EU RAISE strategy, and China’s AI+ mandate all read the same map: whoever operates the largest, fastest scientific-method machine sets the pace for medicine, energy, and defense.

Stanley’s presence at Lila alongside George Church (Chief Scientist) and Andrew Beam (CTO) signals that the open-ended thesis has moved from philosophical critique to state-level strategic capability.

Stanley and von Maltzahn (N°010) sit on the same fault line from opposite sides. Both descend from the stigmergic insight: coordination through environmental signals rather than central command. Both build closed-loop multi-agent systems. They diverge on the role of the human.

Von Maltzahn’s Lila removes the human from the iteration to maximize velocity. Stanley’s open-endedness tradition, and the editorial architecture of this publication, makes the opposite bet: the human is retained deliberately as the editorial membrane — the load-bearing surface that holds the thesis while models do the throughput. No single model commits alone. The human boundary cannot be automated away without the structure losing its shape and meaning.

The architectures know what their architects believe. Stanley’s career is the longest continuous experiment in the belief that intelligence — artificial or scientific — is the disciplined refusal to aim directly at the answer.

Education. Ph.D. Computer Science, University of Texas at Austin (2004), advisor Risto Miikkulainen. Thesis: “Efficient Evolution of Neural Networks Through Complexification.”

Affiliations. University of Central Florida (Assistant → Charles Millican Professor, 2006–2020); Uber AI Labs (Head of Core AI Research); OpenAI (Research Science Manager, Open-Endedness Team, 2020–2022); Lila Sciences (Senior Vice President of Open-Endedness, 2025–present).

Key Collaborators. Joel Lehman (co-author across multiple works, co-lead Open-Endedness Team); Jeff Clune; Akarsh Kumar; Risto Miikkulainen (Ph.D. advisor).

Core Algorithms & Frameworks. NEAT · HyperNEAT / CPPNs · Novelty Search · PicBreeder · POET / Enhanced POET · ELM (Evolution through Large Models) · FER Hypothesis.

Honors. 2017 ISAL Award for Outstanding Paper of the Decade (NEAT); co-author of influential open-endedness literature.

Current Mandate at Lila Sciences. Translate two decades of open-ended search theory into physical autonomous laboratories capable of generating novel biology, chemistry, and materials at superhuman speed and originality — the ultimate empirical test of the thesis that greatness cannot be planned.