I was trying out ai story writing again and it refused to do a sex scene because of “power imbalance “
Timeline
Post
Remote status
Context
7Local model btw
@sun@shitposter.world That's why local models are still proprietary
@SuperDicq yeah I don’t care for people conflating open weight with open source
@sun@shitposter.world And I can't wait for people like you to finally stop thinking open source and free software are the same thing.
The OSI actually has an open source LLM definition that fucking sucks is not free at all.
@sun @SuperDicq what does source code even mean when it comes to engrams
@icedquinn @SuperDicq it never made sense but if you want a reason it means you can trace from a cited source to an answer deterministically which requires no proprietary hidden data training
Replies
6
@sun @SuperDicq well that's just not how a brain works.
@sun @SuperDicq considering floating point is cooked math i'm not sure you even can deterministically train an ANN. even with the same seeds, FPUs are gooners that do math stupid at the best of times. you'd have to do fixed point arithmetic for that to work reliably.
tsetlin machines *could* be this. they use stochatic updates as a way to simulate heterogenous signal paths (which is important; it stops pattern recognizers from oscillating as a group and getting stuck on nothing), but you can make that deterministic with the right RNG, but also
is the argument that like, the source code for a human is every book they've ever read in the order they read it in?
tsetlin machines *could* be this. they use stochatic updates as a way to simulate heterogenous signal paths (which is important; it stops pattern recognizers from oscillating as a group and getting stuck on nothing), but you can make that deterministic with the right RNG, but also
is the argument that like, the source code for a human is every book they've ever read in the order they read it in?
@icedquinn@blob.cat @sun@shitposter.world I never said LLMs had to be fully reproducible and deterministic.
Even though that is technically possible but just hard to achieve through sheer complexity. A computer is still a computer and true randomness does not exist so everything a computer outputs is always reproducible with the same input.
What is missing with local LLMs still is the freedom to use it for any purpose and the freedom to change it to make it do what you wish.
@icedquinn@blob.cat @sun@shitposter.world I am not sure if doing those things requires access to the full training data set or not, I am on not an expert on LLMs.
@icedquinn@blob.cat @sun@shitposter.world The big tech companies are currently brainwashing the general public with their whole "AI is super duper dangerous, guys!" propaganda.
So I have a feeling that these LLMs will not ever have freedom. They will probably have extremely strict guardrails that you can not remove.
@SuperDicq @sun well they want to do a regulatory capture. it fluffs trumps ego to claim he presided over super intelligence and altman-amodei want gooners to stop embarassing them