Comment by skjoldr | Hacker News Reader

Comment by skjoldr | Hacker News Reader

Parent Full thread

The only real difference between what you are describing and GPT-3 is that your dataset from which you draw all your opinions about the world is unique and comes from the sensory input you had experienced in your life, stored directly in memory. Whereas GPT-3's dataset comes from slices of many other peoples' sensory experiences, stored in language. If it was possible for a single human to produce enough words that describe their own experience to rival the volume of GPT-3's dataset, then it would be possible to train a model that would essentially "hold" that person's opinions. But I don't think such consistency is required for intelligence. It is simply arguing that an intelligent being must only experience the world at a single point in space, at a time, while this kind of AI takes in the data about experiences that happened in many different points in space at many different times. It seems kinda arbitrary.

What if AR tech develops to the point where you can experience "being" in other point in space through artificial sensors (say, on a robotic or drone chassis), while also being able to see and hear your real surroundings to some degree? Then you will be able to experience the same real event from multiple vantage points. Will you be able to form two different opinions on what "really" happened? Which one would be the "correct" one?

> The only real difference between what you are describing and GPT-3 is that your dataset from which you draw all your opinions about the world is unique and comes from the sensory input you had experienced in your life, stored directly in memory. Whereas GPT-3's dataset comes from slices of many other peoples' sensory experiences, stored in language.

So is this debate partially just another way of asking if language is the same as reality? The difference between "things I've experienced" vs "stored in languages" seems... not trivial to me? Both in that I think the way biological memory works is non-trivially different from word tokens stored in computer memory, and also in that I think there's more to having experienced something than it just being stored in biological memory (or maybe that biological memory is more than just "memories", but is encoded throughout the body overall -- a scar is a type of memory of an injury, for instance, connected to but still different from my "memory" of having received it).

Yes, it is partially about whether human models of reality (encoded in the relationship between word tokens) can approximate reality itself to such a degree where the quantity of ingested information can substitute for its quality.

Human brains and senses are obviously qualitatively different, but AIs have the advantage of unimaginably massive bandwidth of incoming textual information.

Now, one of the hardest difficulties here is setting up the goal structure for this kind of AI. Just filling in the blanks is obviously insufficient.

Now I'm genuinely curious if anyone tried to stuff some verbose memoirs into GPT to see what it makes out of them.

Reply on news.ycombinator.com