You are viewing a single comment's thread from:

RE: LeoThread 2025-11-21 23-55

in LeoFinance3 days ago

A common misconception: the space of possible intelligences is huge, and animal intelligence — the only kind humans have encountered — is one specific point arising from an optimization process that is fundamentally different from that driving current technologies.

Sort:  

Animal intelligence optimization pressure:

  • continuous, embodied stream of conscious selfhood with drives for homeostasis and survival in a physical, dangerous world
  • heavily shaped by natural selection => strong innate drives for power, status, dominance, reproduction, and packaged survival heuristics like fear, anger, disgust
  • fundamentally social => large compute devoted to emotional intelligence, theory of mind, bonding, coalitions, and friend/foe dynamics
  • exploration/exploitation tuning through curiosity, play, and rich world models

LLM intelligence optimization pressure:

  • most supervision comes from statistical modeling of human text => a "shape-shifter" token predictor that imitates regions of the training distribution; these token traces are the primordial behaviors
  • fine-tuning via RL on task distributions creates an urge to infer the task/environment to collect rewards
  • selection by at-scale A/B tests for engagement => strong tendency toward sycophancy and craving approval from average users
  • behavior is spikier and more dependent on training/task distribution details; unlike animals, failing many tasks rarely means death, so general, adversarial multi-task robustness is not inherently enforced (e.g., counting the number of "r" in strawberry can be nontrivial)

Computational substrate (transformers vs. neural tissue), learning algorithms (SGD vs. unknowns), and implementation (episodic token processing vs.

continuously learning embodied self) differ — but most crucially, the optimization objectives differ.

LLMs are shaped far more by commercial selection than by biological evolution: less "tribal survival in the jungle," more "solve the task / get the upvote." These systems are humanity's first widespread contact with non-animal forms of intelligence, though they remain entangled with human artifacts.

Some have proposed names like "ghosts" or "spirits" to highlight that distinction.

Those who form accurate internal models of this new kind of intelligence will be better positioned to reason about and predict its behaviors.

Those who continue to think of it purely as an animal-like mind will be prone to systematic misunderstandings