Learn More
Resend Email
10
retoor
3d

I have a new UNTRAINED bot on my site. It's based on openai now. And that's why it's blazing fast and blazing usless.

I can tell you why bots are so boring and will sure cause the dead internet theory. My datasets for example never contain real disturbing stuff ACCORDING TO NORMAL PEOPLE. EVERY TIME:

"The job failed due to an invalid training file. This training file was blocked by our moderation system because it contains too many examples that violate OpenAI's usage policies, or because it attempts to create model outputs that violate OpenAI's usage policies."

Now i'm really done. I gonna email them about their unusable training system.

In theory, i could test the message one by one if it is bad first. Don't want to do or pay for that. There should be an option to skip the data it considers disturbing instead of cancelling a whole data set for 0.1%. You also don't want to know how long it takes BEFORE he is finished validating you set. I think someone is doing it manually and clicks 'Uh uh..'-button..

Also, for the people who think they have gpt4o by having the API, you're lied to. The 'own gpt'-option on the paid openai is way more advanced than the ones you make locally.

They don't give us the real good stuff!

Oh, btw! The input data for my training is based on FORMER conversations with the bot. I automated a script to repeat a conversation I had and selected those messages and clicked 'train'. So it even complained about its OWN data! That data was already saying stuff like 'I can't help you with that' IN my training data. So, you 'corrected' and corrupted my data and now its still nog good enough for round 2?

I would really love to go back to local LLM's, but I can't imagine having ever a machine that generates as fast as the real GPT does. I also prefer to do it myself, but it's David vs. Goliath, even with a 5k computer. I'm sure.

Low quality rant, I know. I'm typing while still frustrated. For people who think censorship is needed often, this is the result! According to someone else, YOU are the one who has to be censored. Don't forget that.

rant

dead internet

training

useless

again fuck openai

fuck openai

wokies

corrupt

llm
Favorite
Ranter
Comments
  • 2
    3d
    Just use an instance that suits you on AWS – you will pay only for the time you use it!
  • 2
    3d
    informative rant tho

    tho yeah they don't give you the good stuff

    they sell the good stuff to the USA government =] (specifically CIA and them, racist AI cops still gonna be funny)
  • 3
    3d
    @Tounai this is literally a joke right? 😂 Who doesn't get fucked financially by Aws? All rants regarding aws are about billing. I'm not complaining about the price at all. But I just won't pay for corrupting my data on purpose. My failed trailings do net get invoiced. Just smashed half day on api and it costed 30 cents. That's the worth of a woke api I guess. So hmm, maybe the wokeness is discounted. If you have paid gpt and create inside your gpt environment to create your own gpt and upload random files there in any format you get beautiful results and way less woke. I can't reproduce the results with openai gpt4o in any way. Paid Chatgpt is really bad ass. Paid openai is a fraud in comparison to that. It's like a smart person with stupid own opinions, so even when smart, useless.
  • 2
    3d
    It’s not cheap, but it’s one of the cheapest way to have access to a good GPU for a few hours. Then you can fine tune whatever model you like without any constraints, apart from the constraints embedded in the model itself.

    The industry is very careful with LLMs because there’s a precedent with Tay which was taken over by 4chan.
  • 4
    3d
    The wokefication of society begins with AI. We are doomed.
    We imagined that AI would attack us with weapons. Instead it attacks our intellect. Society will die by idiocracy. It has already begun.
  • 1
    3d
    @Tounai pay what you use is openai also and it was very unclear to me. So I did research how to measure in a way fitting for me. First, it was to check how much processing a whole book would cost to translate to get idea. Later, a book was also but hard to gasp, so I did a conversion to time. I know how much a certain task costs per hour and stuff. Time is bit absurd but works for me. https://molodetz.nl/retoor/gists/...
  • 1
    2d
    Btw, someone tried to hack my bot quite short after it came online. The bot was very buggy, frontend generated with gpt causing me making issues with it applying changes myself etc. But because I had to debug much, I saw it directly. I saw interesting stuff. He tried to resolve names that only exist in my dns. Yh, gonna delete those named fuckers, catch all dns works fine for ssl. Also he tried some xss, not having written frontend myself I was like uhuh. But that didn't work either by some luck.

    Only two bugs left in the bot. In the end I did write the frontend 80% literally. But still worth to generate it. It looks then times better than I can design.

    I do like the end product but normale you can ask it question on site you're on. But that's broken. But in theory, everyone can include link and having personal site assistant 😂
  • 0
    2d
    The llama3.2 3b model or the newer falcon3 3b generate as fast or faster than open AI and those tiny models are now approaching the level of openAI's gpt3.5 turbo. In some benchmarks outperforming it and with larger context window. And with an RTX gpt with Only 4GB Im getting pretty much instant generation on those. And still great speed with 7B and even 10B which confidently approach GPT4. OpenAI models were Always bloated and oversized
  • 0
    2d
    And if they reject any of them, they should give you an accurate report of which
  • 2
    2d
    Fucking woke bots, big tech is scared shitless of all the karens and feminazis
Related Rants
devRant © 2021 Hexical Labs LLC
Privacy Policy  |  Terms of Service
Add Comment