MIT-Forscher entdecken Simulationen der Realität, die sich tief in LLMs entwickeln, was auf ein Sprachverständnis hindeutet, das über einfache Nachahmung hinausgeht
https://news.mit.edu/2024/llms-develop-own-understanding-of-reality-as-language-abilities-improve-0814
2 Kommentare
„Researchers from MIT have uncovered intriguing results suggesting that language models may develop their own understanding of reality as a way to improve their generative abilities.
The team first developed a set of small Karel puzzles, which consisted of coming up with instructions to control a robot in a simulated environment. They then trained an LLM on the solutions, but without demonstrating how the solutions actually worked. Finally, using a machine learning technique called “probing,” they looked inside the model’s “thought process” as it generates new solutions.
After training on over 1 million random puzzles, they found that the model spontaneously developed its own conception of the underlying simulation, despite never being exposed to this reality during training.“
I’ve been saying this is a pretty obvious prerequisite to the behaviors exhibited by LLM for generations now. The only and the optimized way to do much of what they do is for mid layers to abstract to world models.
This is in turn has some hair-raising implications for the question of self awareness. The most plausible substrate for any theory of consciousness is for a system to have not just a world model but a model of others‘ world models and of ones‘ self as perceived by others. Such modeling is a prerequisite for understanding intention and probable behaviors, and much of language in use is confirming or signaling the correctness of such models so as to establish a shared world of references: language is the medium whereby we synchronize and then postulate observations about such shared models.
Regardless of whether they persist over time or incorporate new information or have a sensorium, there is reason to believe that vestigial notions of self viz world are incorporated already into large models.
The most efficient way to speak as if they have such things within them—is to have them.
We live in times I didn’t imagine I’d see.