Everyone who follows me knows I'm a big fan of nextcloud. I just added some features to my installation that are pretty neat -- locally hosted AI. It isn't a chat thing, but it can do translations and text generation, summaries, and the like. I installed the Local Large Model app which gave me the option to use freely available large model files, the nextcloud assistant app presents an icon at the top to actually do the generation, and the OpenAI and LocalAI integration that brings everything together.
Since my entire web empire is hosted on parts scavanged from roadside signs the generation doesn't run too quickly, but it does produce meaningful output.
The meteoric rise of locally hosted AI makes me think 2 things: First, I don't know that companies thinking they can gate this off and monetize it are going to be in such good shape since anyone can presently host it and on 5 year old hardware models are getting pretty decent and quick. Second, how soon until it's just a local feature nobody thinks about on standard hardware?
Despite the excellent output from AI image generation and large language models, it is still my stance that the real intelligence at work are the humans who apply algorithms to create the output, and when the output is flawed, they modify the algorithm or decide not to use AI for the purpose.
I've asked several AIs to do what is a relatively simple task: "Write a review of Beowulf in the style of beowulf". I've done it before and posted it here. The thing is, it'll produce something that seems servicable until you start reading it and realize something: It rhymes, and beowulf never once rhymes. It makes heavy use of aliteration rather than rhyming. So next you say "Beowulf doesn't rhyme", and it will come back with an agreement that beowulf doesn't rhyme and instead makes heavy use of alliteration. So you say "So rewrite the review to make use of alliteration rather than rhyming" and it'll present a rhyming review yet again. It cannot be convinced to complete the task correctly.
I've had the same issue with coding problems, where the AI writes an objectively wrong answer to a coding question, and even when given direct instructions about the correct answer, it will continue giving the wrong answer.
Since my entire web empire is hosted on parts scavanged from roadside signs the generation doesn't run too quickly, but it does produce meaningful output.
The meteoric rise of locally hosted AI makes me think 2 things: First, I don't know that companies thinking they can gate this off and monetize it are going to be in such good shape since anyone can presently host it and on 5 year old hardware models are getting pretty decent and quick. Second, how soon until it's just a local feature nobody thinks about on standard hardware?
Despite the excellent output from AI image generation and large language models, it is still my stance that the real intelligence at work are the humans who apply algorithms to create the output, and when the output is flawed, they modify the algorithm or decide not to use AI for the purpose.
I've asked several AIs to do what is a relatively simple task: "Write a review of Beowulf in the style of beowulf". I've done it before and posted it here. The thing is, it'll produce something that seems servicable until you start reading it and realize something: It rhymes, and beowulf never once rhymes. It makes heavy use of aliteration rather than rhyming. So next you say "Beowulf doesn't rhyme", and it will come back with an agreement that beowulf doesn't rhyme and instead makes heavy use of alliteration. So you say "So rewrite the review to make use of alliteration rather than rhyming" and it'll present a rhyming review yet again. It cannot be convinced to complete the task correctly.
I've had the same issue with coding problems, where the AI writes an objectively wrong answer to a coding question, and even when given direct instructions about the correct answer, it will continue giving the wrong answer.
- replies
- 1
- announces
- 1
- likes
- 2