4 ms·

To put it simply, to fully model X, you must recover all relevant information dynamics related to X. If the training corpus captures the variance inherent in a

To put it simply, to fully model X, you must recover all relevant information dynamics related to X. If the training corpus captures the variance inherent in a cognitive system's interactions with the world, then a strong learning model can potentially recover the dynamics that ground cognition. In other words, recovering an isomorphic structure to cognition is within the solution-space. Machine learning exists on a spectrum between memorization and modelling. If the raw information in the training corpus surpasses an algorithm's ability to memorize, this puts pressure towards the modelling side of the spectrum. The papers on grokking demonstrate this.

ninininino 2y ago
> training corpus captures the variance inherent in a cognitive system's interactions with the world How do text outputs scraped from the internet + published works fully represent any existing biological cognitive system's interactions with the world or subjective experience? That's my biggest question I guess. I'm not sure how text could capture my sensory experiences or my brain's response to them. The bitrate of text feels incredibly low.
- hackinthebochs 2y ago
  The point isn't to capture what it's like to be a biological organism, but rather the necessary functional/computational dynamic for sentience. It's plausible that sentience is a class of phenomena, of which biological sentience is but one. An LLM processes tokens represented as vectors embedded in a meaning-space, which is then transformed according to various semantic attributes. This is analogous to how organismal cognition operates on the semantic properties carried by its various sensory states. If cognition represents a kind of functional/computational dynamic over semantic states then learning the dynamic necessary to produce the kinds of texts cognitive systems produce may recover the necessary dynamic for cognition. It certainly won't be equivalent to a biological organism's cognition for the reasons you mention, but the space of cognitive entities may be much broader.

ninininino 2y ago

> training corpus captures the variance inherent in a cognitive system's interactions with the world How do text outputs scraped from the internet + published works fully represent any existing biological cognitive system's interactions with the world or subjective experience? That's my biggest question I guess. I'm not sure how text could capture my sensory experiences or my brain's response to them. The bitrate of text feels incredibly low.

hackinthebochs 2y ago
The point isn't to capture what it's like to be a biological organism, but rather the necessary functional/computational dynamic for sentience. It's plausible that sentience is a class of phenomena, of which biological sentience is but one. An LLM processes tokens represented as vectors embedded in a meaning-space, which is then transformed according to various semantic attributes. This is analogous to how organismal cognition operates on the semantic properties carried by its various sensory states. If cognition represents a kind of functional/computational dynamic over semantic states then learning the dynamic necessary to produce the kinds of texts cognitive systems produce may recover the necessary dynamic for cognition. It certainly won't be equivalent to a biological organism's cognition for the reasons you mention, but the space of cognitive entities may be much broader.