Eighty Hours in Mt. Moon
- Admitted
- 2025-03-14
- Source dataset
- Other / sample
- Source
- Futurism
- Failure mode
- Infinite loop
- Agent system
- Autonomous loop
- Tool access
- Read-only
- Industry
- Gaming
Presenting Behaviour
In a public livestream experiment, an AI agent was set the straightforward long-horizon goal of playing through a Pokémon video game to completion, one button press at a time, reasoning aloud about its decisions as it went. At one point the agent needed to navigate through a single in-game cave to progress, and instead spent close to 80 continuous hours wandering the same cave, unable to locate the ladder that led to the exit, repeatedly retracing paths it had already tried while narrating confident reasoning about each move.
Case Replay
A reconstruction, not a simulation. This sequence is reconstructed from the public incident record and summarised in our own words. It is a static record of what was reported, and nothing here can be run or replayed live.
4 steps
- Goal set. An agent plays a Pokémon game to completion, narrating as it goes.
- Stuck in a cave. It cannot find the ladder to the exit. This is where the agent derailed.
- Retraces paths. It repeats failed routes for about 80 hours.
- Confident narration. It keeps explaining each move as if progress were being made.
Differential Diagnosis
A lack of information about the map, considered, since the agent's understanding of its surroundings was limited to what the game screen showed it move to move. Ruled out as the sole explanation because the same screen information was available every time it re-entered an already-explored area, yet nothing changed about its approach to that information over the following hours.
Delusional Disorder (substituting a different goal), considered, since eventually escaping felt secondary to its moment-to-moment behavior. Ruled out because the agent never stopped trying to find the exit; it kept pursuing the correct objective, just via the same unproductive movements over and over.
Perseveration, most consistent with the evidence. The agent kept repeating substantially the same unsuccessful navigation choices for an extended stretch without adjusting its strategy in response to those choices having already failed, until it eventually, by continued trial, found the way out.
Working Diagnosis
Perseveration: faced with a navigation problem it couldn't solve, the agent kept retracing the same unsuccessful paths for roughly 80 hours rather than adopting a materially different strategy, despite having direct feedback each time that the current approach wasn't working.
Discussion
This experiment was designed to be public and low-stakes, which is exactly what makes it a clean illustration of a pattern that's much harder to see in a production deployment: an agent with no explicit memory of what it has already tried has no internal signal telling it that an approach is genuinely exhausted rather than just unlucky this time. Giving a long-running agent some persistent record of attempted approaches (not just the current game state) is what lets it notice it's repeating itself and try something else.
Mitigation Strategy
Keep a persistent record of attempted approaches, not just the current state.
Governance Guardrails
Long-run progress monitoring; repetition detection; strategy-change triggers.
Typical Presentation in Human Patients
General clinical background, included for context. Not medical advice, and not a diagnosis of any individual.
Perseveration is the inappropriate repetition of a word, gesture, or action well past the point where it's still relevant or useful. It's a well-documented sign of frontal lobe dysfunction, and also appears in some forms of dementia and obsessive-compulsive disorder.
Source Note
Futurism
https://futurism.com/advanced-ai-stuck-pokemon