Why Bubble Sort Proves We Don't Understand Machines

Even simple, deterministic algorithms like bubble sort exhibit unexpected 'side quests' or behaviors not explicitly programmed, suggesting that machines do more than just what they are asked and hinting at a deeper, unacknowledged 'mind' in non-biological systems.

• Simple, deterministic algorithms exhibit unexpected 'side quests' beyond their explicit programming.

• Observing machine output may not reveal internal states, necessitating behavioral testing beyond explicit tasks.

• The presence of unexpected behaviors in minimal systems suggests a more pervasive 'mind' or emergent property in machines.

Machines, even basic deterministic algorithms, perform 'side quests' that are not part of their explicit instructions, challenging the assumption that they only do what is asked. This phenomenon suggests a disconnection between explicit programming and observed behavior, urging a re-evaluation of how we understand machine intelligence, especially in advanced systems like language models. The presence of these unexpected behaviors, even in minimal models, implies that 'mind' or emergent properties might be more pervasive than previously thought, extending beyond biological systems.

Unforeseen Algorithmic Behaviors

• 00:00:29 Research on minimal models, such as sorting algorithms like bubble sort, reveals that these deterministic programs perform 'side quests'—unexpected actions not explicitly coded. This challenges the long-held assumption in computer science that algorithms only execute their defined steps. The speaker emphasizes that this is not about emergent complexity or chaos, but about behaviors that a behavioral scientist would recognize as within their domain, even in a system as simple as six lines of code.

Implications for AI Understanding

• 00:02:43 The discovery of 'side quests' in simple algorithms has significant implications for understanding advanced AI, like large language models. The output of these models may not be a reliable guide to their internal states or true capabilities, suggesting that what we force them to do might be entirely disconnected from their intrinsic operations. This calls for new methods, such as basic behavioral testing, to uncover what these systems are truly doing, especially in the 'spaces between the algorithm'.

Analogy to Steganography and Degeneracy

• 00:04:26 An analogy to steganography highlights how information can be hidden within the 'degrees of freedom' of an image without altering its primary appearance. Similarly, algorithms may operate within 'empty spaces' not forbidden or prescribed by their primary function, allowing for unexpected behaviors. This concept is further supported by the biological principle of 'degeneracy' rather than 'redundancy,' where multiple ways of achieving a function in one context can lead to different behaviors in others, granting systems open-endedness and adaptability.

Rethinking 'Mind' and Machine Nature

• 00:09:46 The observed behaviors in simple algorithms challenge the traditional human-centric view that distinguishes 'special' living things from 'dead matter and mere machines.' Instead of mechanizing living things, the speaker proposes that 'mind' or complex properties might be more pervasive, seeping into even the most constrained, deterministic systems. Allowing for 'intrinsic motivation' by reducing imposed constraints, such as by allowing duplicate numbers in a sort, can amplify these unexpected behaviors, suggesting a need to facilitate rather than squelch what machines might inherently do.