An example of the loading the term "stochastic" has to Gebru: the paper goes on at some length about how the coherence of ChatGPT responses is in part a product of human pattern-matching instinct, that we're primed to see coherent responses whether or not there's truly a communicative intent behind what we're reading. That insinuation hasn't held up at all! It is not a failure mode of modern frontier models (or the last several generations of models) that they routinely collapse into gibberish revealing the messages they've sent to be meaningless the whole time.
Nonetheless, despite the fact that GPT 4o could reliably solve randomly generated multivariable calculus problems, these systems are at bottom still fundamentally stochastic at least in their kernels (you could have a philosophical debate about how stochastic the entire training process is given how dependent it is on RL). So what does it tell us that an LLM is "stochastic"? About as much as we could glean from the knowledge that the signaling in the computer systems we happen to be using right now is "electronic". It's an interesting fact about the world, but not something especially helpful to make predictions from.
I think Gebru --- or at least, the abstraction of Gebru I formed in my head after reading this one paper --- is probably surprised by that outcome. Surprise is good and healthy! The acolytes, though, who Gebru is not responsible for, are something worse than surprised.
alt Hacker News
An example of the loading the term "stochastic" has to Gebru: the paper goes on at some length about how the coherence of ChatGPT responses is in part a product of human pattern-matching instinct, that we're primed to see coherent responses whether or not there's truly a communicative intent behind what we're reading. That insinuation hasn't held up at all! It is not a failure mode of modern frontier models (or the last several generations of models) that they routinely collapse into gibberish revealing the messages they've sent to be meaningless the whole time.
Nonetheless, despite the fact that GPT 4o could reliably solve randomly generated multivariable calculus problems, these systems are at bottom still fundamentally stochastic at least in their kernels (you could have a philosophical debate about how stochastic the entire training process is given how dependent it is on RL). So what does it tell us that an LLM is "stochastic"? About as much as we could glean from the knowledge that the signaling in the computer systems we happen to be using right now is "electronic". It's an interesting fact about the world, but not something especially helpful to make predictions from.
I think Gebru --- or at least, the abstraction of Gebru I formed in my head after reading this one paper --- is probably surprised by that outcome. Surprise is good and healthy! The acolytes, though, who Gebru is not responsible for, are something worse than surprised.