Comment by cadamsdotcom
3 months ago
We should be loudly demanding transparency. If you're auto-opted into the latest model revision, you don't know what you're getting day-to-day. A hammer behaves the same way every time you pick it up; why shouldn't LLMs? Because convenience.
Convenience features are bad news if you need to be as a tool. Luckily you can still disable ChatGPT memory. Latent Space breaks it down well - the "tool" (Anton) vs. "magic" (Clippy) axis: https://www.latent.space/p/clippy-v-anton
Humans being humans, LLMs which magically know the latest events (newest model revision) and past conversations (opaque memory) will be wildly more popular than plain old tools.
If you want to use a specific revision of your LLM, consider deploying your own Open WebUI.
> why shouldn't LLMs
Because they're non-deterministic.
It is one thing that you are getting results that are samples from the distribution ( and you can always set the temperature to zero and get there mode of the distribution), but completely another when the distribution changes from day to day.
What? No they aren't.
You get different results each time because of variation in seed values + non-zero 'temperatures' - eg, configured randomness.
Pedantic point: different virtualized implementations can produce different results because of differences in floating point implementation, but fundamentally they are just big chains of multiplication.
On the other hand, responses can be kind of chaotic. Adding in a token somewhere can sometimes flip things unpredictably.
But experience shows that you do need non-zero temperature for them to be useful in most cases.
I spend $20/month on ChatGPT. I'm not going to loudly anything. Relax and modify your custom prompt. You'll make it through this, I promise.