Have many others in this forum experimented with artificial intelligence models and online services? If so, do you have any favorites?
Today I installed cuda and ollama and then experimented with three model libraries: Gemma 2, Llama 3.2 and DeepSeek-R1. Gemma 2 and Llama 3.2 could answer questions, might serve in a pinch, but struck me as slower and less capable question-answerers than perplexity.ai, which also provides footnotes. I wondered if I were missing something, if users generally install ollama to create their own models, or for other uses I don’t know about.
(The DeepSeek-R1 model was much more primitive.)
So, this Open Chat question: Why are your AI experiences? Do you have any favorites, tips, recommendations?
shrug I don’t really have an opinion on the matter.
I’ve poked at all of the major players, and I just don’t find them to be any more useful than a search engine, or reading the documentation for whatever it was I was looking for.
That being said, it could just be me, and/or my usecase.
Philosophically, I can’t say that all of these LLM’s and the mistaken moniker of “Artificial Intelligence” being hung on them exactly give me the warm and fuzzies, nor does the wholesale theft of IP by the companies/organizations/individuals behind them, in order to train their models really incline me to give them any benefit of the doubt.
@Tuner I agree with @sfalken on the philosophical point.
I’ve played around with ollama and open-webui running in kubernetes and using the nvidia-container coupled with a Tesla P4, using Llama. Very fast too me in lookups, only seconds in fact…
Now, the use I do see is for you using your own information to create your in-house own models.
I really enjoy using the code completion features of Codestral and the Mistral model in general. These tools have significantly enhanced my productivity and creativity by providing intelligent suggestions and automating repetitive tasks. I find them quite helpful and would recommend them to anyone looking to streamline their workflow.
However, it’s important to be aware that responses from language models can sometimes be hallucinative.
Therefore, it’s always better to ask for sources to verify the information provided.
My current hardware is too slow to run the above mentioned models “in house”.
I need better hardware to run ollama, but I have no idea when the Snapdragon X is working with openSuse or if I should buy an Intel Core or Ryzen XXX AI?
Thanks for the feedback, sfalken, malcolmlewis and SteffNC. Malcolm, I suspect that many do use these offerings to create their own models … which is far beyond me.
In their defense, Perplexity’s answers include links to information sources, so a user can visit the web pages that provided the material. (Can and absolutely should visit, as Perplexity seems to prefer inventing an answer to admitting ignorance, as SteffNC noted.)
I run models locally for in-browser language translation, occasionally epub translation or text-to-speech, interactive fiction and image generation. I have a lot of fun with it. I use kobold.cpp for LLMs and A1111 for image generation. The model I use varies depending on what I’m doing, but I enjoy trying new or experimental ones.