Kenneth O. Stanley is an AI researcher whose work includes neuroevolution, novelty search and open-endedness. These lessons distinguish research findings from his broader proposals about discovery. — UCL DARK — Novel Opportunities in Open-Endedness.

Infographic for "Lessons from Kenneth O. Stanley".

On the Illusion of Objectives

Stanley and Lehman argue that ambitious objectives can misdirect search when useful intermediate steps are unknown. — Novelty Search and the Problem with Objectives.

  1. Allow unexpected discoveries: Picbreeder’s collaborative exploration produced meaningful images without a shared predefined target. — Woolley and Stanley — Objectives and Representation.
  2. Question the compass: An ambitious objective may reward apparent progress that leads into a dead end. — Novelty Search and the Problem with Objectives.
  3. Preserve unlikely stepping stones: Intermediate discoveries need not resemble the eventual solution; rejecting them can obstruct the search. — Woolley and Stanley — Objectives and Representation.
  4. Keep nearby objectives: Stanley distinguishes reachable, understood goals from ambitious discoveries whose intervening steps remain unknown. — The Jim Rutt Show — Why Greatness Cannot Be Planned.
  5. Notice unexpected prerequisites: Stanley uses vacuum tubes and computers to illustrate how inventions can become useful beyond their original purpose. — The Jim Rutt Show — Why Greatness Cannot Be Planned.
  6. Examine objective-driven learning: Stanley and Lehman question whether fixed-objective search limits machine learning’s discovery potential. — Novelty Search and the Problem with Objectives.
  7. Question institutional metrics: Stanley and Lehman challenge reliance on measurable goals in research funding and education. — Why Greatness Cannot Be Planned.

The Power of Novelty and Interestingness

Novelty search, developed by Joel Lehman and Stanley, selects for behavioral difference rather than proximity to a specified target. — Lehman and Stanley — Abandoning Objectives.

  1. Let discoveries branch: In Picbreeder, publishing an image lets other users continue from it, creating further branches of discovery. — Woolley and Stanley — Objectives and Representation.
  2. Reward behavioral novelty: Novelty search rewards different behaviors rather than improvement toward a chosen target. — Lehman and Stanley — Abandoning Objectives.
  3. Learn through exploration: Stanley argues that continued novelty can require accumulating knowledge about the environment. — The Jim Rutt Show — Why Greatness Cannot Be Planned.
  4. Follow genuine curiosity: Stanley and Lehman advocate playful exploration rather than requiring a predetermined payoff for every pursuit. — Why Greatness Cannot Be Planned.
  5. Use experienced intuition: Stanley argues that domain experience can make subjective judgments of interestingness useful, without making them infallible. — The Jim Rutt Show — Why Greatness Cannot Be Planned.
  6. Collect promising possibilities: Stanley describes recognizing interesting stepping stones before their eventual use is clear. — Idea Amplification and Generative AI.
  7. Evaluate breadth of discovery: Producing diverse interesting artifacts is different from reliably reproducing one specified target. — Woolley and Stanley — Objectives and Representation.
  8. Value exploration before polish: Stanley values research that opens new possibilities even before it excels on familiar benchmarks. — Brain Inspired — Ken Stanley on Open-Endedness.
  9. Consider indirect search: In their maze and walking experiments, novelty search outperformed the compared objective-based approaches—not every possible alternative. — Lehman and Stanley — Abandoning Objectives.
  10. Distinguish novelty from interestingness: Stanley treats measurable novelty as an imperfect proxy for the harder problem of formalizing what is interesting. — The Jim Rutt Show — Why Greatness Cannot Be Planned.

Open-Endedness: The Grand Challenge

Stanley presents computational open-endedness as an important unsolved research challenge. — Why Open-Endedness Matters.

  1. Seek continuing invention: Open-endedness asks how computational processes might keep generating interesting new discoveries rather than converge once. — Why Open-Endedness Matters.
  2. Study natural evolution: Stanley treats evolution’s continuing diversity as a model for artificial open-endedness, not a process aimed at humans. — Brain Inspired — Ken Stanley on Open-Endedness.
  3. Look at human invention: Stanley describes civilization’s branching inventions as another example of open-ended discovery without a single shared endpoint. — Brain Inspired — Ken Stanley on Open-Endedness.
  4. Separate constraints from destinations: Stanley interprets survival and reproduction as continuing viability constraints rather than a final target. — Brain Inspired — Ken Stanley on Open-Endedness.
  5. Recognize control trade-offs: Stanley sees tension between unrestricted creativity and predictable task completion; compromises remain possible. — Brain Inspired — Ken Stanley on Open-Endedness.
  6. Treat open-endedness as a challenge: Stanley argues that formalizing open-endedness deserves attention alongside artificial intelligence as a major computational research challenge. — Why Open-Endedness Matters.
  7. Avoid assuming a direct route to AGI: Stanley uses intelligence’s unplanned evolutionary origins to question exclusively target-driven AI research, not to prove an alternative succeeds. — Brain Inspired — Ken Stanley on Open-Endedness.
  8. Study the artifacts agents leave: Stanley highlights houses, cars and other persistent artifacts as underexplored influences on civilization-like open-ended systems. — UCL DARK — Novel Opportunities in Open-Endedness.

On Representation and How You Get There

In Stanley’s research, similar-looking outputs can conceal different internal representations and discovery histories. — Woolley and Stanley — Objectives and Representation.

  1. Inspect the representation: Simpler Picbreeder targets were sometimes reconstructed with larger internal networks than the original discoveries. — Woolley and Stanley — Objectives and Representation.
  2. Consider future exploration: Stanley proposes that internal structure may affect later learning and creativity; the broader implications remain unresolved. — UCL DARK — Novel Opportunities Talk.
  3. Keep cognitive extrapolations tentative: Stanley speculates that discovery paths could shape human ideas, but does not establish psychological lock-in. — UCL DARK — Novel Opportunities Talk.
  4. Do not judge only by benchmarks: Stanley warns that target-performance tests can overlook algorithms’ capacity for broader discovery. — UCL DARK — Novel Opportunities Talk.
  5. Examine how results were reached: Failure to reconstruct previously discovered Picbreeder images shows that discovery conditions matter in that experimental setup. — Woolley and Stanley — Objectives and Representation.
  6. Connect exploration to the environment: In Stanley’s robot-and-door illustration, doing something new requires discovering how the surrounding world works. — The Jim Rutt Show — Why Greatness Cannot Be Planned.
  7. Use AI to amplify ideas: Stanley argues that generative AI can help realize ideas previously blocked by limited execution skills, while acknowledging originality and data concerns. — Idea Amplification and Generative AI.
  8. Encode regularities: Stanley describes CPPNs as development-inspired representations that combine functions to generate patterns with symmetry and repetition. — UCL DARK — Novel Opportunities Talk.

Practical Learnings for Life and Innovation

Stanley and Lehman extend their argument to education, research and creative life. These are proposals, not guarantees of success. — Why Greatness Cannot Be Planned.

  1. Accept an uncertain destination: Exploration may produce an unexpected accomplishment rather than solve the specific problem that first motivated it. — The Jim Rutt Show — Why Greatness Cannot Be Planned.
  2. Acknowledge the risk: Stanley says exploratory choices can fail and require resources sufficient to tolerate that possibility. — The Jim Rutt Show — Why Greatness Cannot Be Planned.
  3. Listen to reported frustration: Stanley reports that people from varied organizations welcomed alternatives to rigid objectives; this is anecdotal, not a population survey. — Brain Inspired — Ken Stanley on Open-Endedness.
  4. Do not reduce a career to salary: Stanley argues that maximizing immediate pay can miss promising career directions; a pay cut offers no guaranteed reward. — The Jim Rutt Show — Why Greatness Cannot Be Planned.
  5. Allow exploratory research: Stanley argues for some funding of justified expert exploration even when its eventual outcome cannot be specified. — Brain Inspired — Ken Stanley on Open-Endedness.

Learn more:

  1. UCL DARK — Novel Opportunities in Open-Endedness: Abstract and Speaker Bio
  2. Woolley and Stanley — On the Deleterious Effects of A Priori Objectives on Evolution and Representation
  3. Lehman and Stanley — Abandoning Objectives: Evolution Through the Search for Novelty Alone
  4. Lehman and Stanley — Novelty Search and the Problem with Objectives
  5. Stanley and Lehman — Why Greatness Cannot Be Planned: The Myth of the Objective
  6. The Jim Rutt Show — Transcript of Episode 130: Ken Stanley on Why Greatness Cannot Be Planned
  7. Brain Inspired 086 — Ken Stanley: Open-Endedness
  8. Kenneth O. Stanley — Novel Opportunities in Open-Endedness (UCL DARK Talk)
  9. Why Open-Endedness Matters
  10. Kenneth O. Stanley — Idea Amplification Is the Really Exciting Potential of Generative AI