Probably not you, but on this very forum there are plenty who will argue that LLMs / AI are entirely deterministic and that given enough time, a chisel, and sufficient clay tablets, any AI output can be calculated by hand.
I don’t know enough about anything to determine which opinion is correct.
Is AI output not just matrix multiplication and a random seed?
I don't disagree necessarily on the fundamentals; I am slowly catching up on what LLMs do and don't do but that sounds right to me.
But what I would observe is that the healing brush does not have a random seed. It will always do the same thing if applied at the same pixel.
(I am actually less sure if content-aware fill randomises; I always got the impression it did not)
This makes it both incredibly powerful and occasionally frustrating.
Because on the one hand, you can learn to apply your judgement to precisely control what it will do, and change the radius or position if you learn it is likely to fail, which becomes instinctive. I absolutely love using it to fix scratches in film scans; it's a quick, precise, controllable tool that can be used in a way that is amazingly convincing, and it ends up quite a "zen gardening" thing as a result. It'll sell you on the cheapest wacom pen once you know how efficient it can be.
On the other hand there are situations where it simply cannot work the way you want because it will always find a pattern you don't want it to.
(You can sometimes use the clone brush tool first, to manually break up the pattern that patchmatch will find)
> But what I would observe is that the healing brush does not have a random seed. It will always do the same thing if applied at the same pixel.
Given a model architecture that supports it, greedy decoding + the same inputs + prompts, that's true for most LLMs today too, I don't think people consider them less/more AI because of that.
>Probably not you, but on this very forum there are plenty who will argue that LLMs / AI are entirely deterministic and that given enough time, a chisel, and sufficient clay tablets, any AI output can be calculated by hand.
Any particular AI output is deterministic (and can be calculated by hand even, if you're immortal).
But it's not deterministic in the sense that (a) it's a black box on the Cloud, (b) repeated AI prompts don't give the same results.
LLM inference can be implemented in a way where nondeterminism depends only on the random seed, but that's not common. It ends up being more efficient/easier to implement kernels whose exact results depend on how many other prompts are being processed in parallel. See https://thinkingmachines.ai/blog/defeating-nondeterminism-in... for a pretty extensive exploration.
Even if you were a "point" (an endpoint assigned to the node) you still had to set up the software and (in the mid-to-late 90s at least) set up a modem to call your node to upload/download. And sometimes you had to set up repeated dialing until you got through because the node could be busy (some nodes doubled as BBSs), or connection could be bad and it'd had to retry etc. Wasn't an easy task, so it served as a sort of a filter so that most people on there were geeks.
Later on of course some nodes started distributing over the Internet so setting up a node became much easier (and I think there was a way for the node to allow multiple users read/write without even setting up a node/point at all).
As an outsider, it’s not clear to me who you are referring to as tribalist regressors here.
Would you mind clarifying?
reply