‘Training AI on your data does not make it know things about you’ Employee who creates language models reveals how AI actually works

Author image for Olivia Arocena
Olivia Arocena • Writer and Editor
Sep 9, 2026 11:15 AM EDT

Good News: AIs don't really think. Bad News: AIs don't really think

Now that AI agents are part of our lives, it's very important that we learn more about them. Of course, Reddit isn't the most official or valid source of information, but reading multiple takes on a subject completes the picture of an otherwise very new and still largely misunderstood matter.

This worker who creates Language Learning Models for a living has something to tell us, and we're here to listen. I believe the truth resides somewhere in the middle. Between the aggressive marketing AI companies bombard us with and what tech journalists or employees like this say to us.

It's very easy to personify an agent that's responding to us in perfect English, and attribute a thought process or even a personality. The truth is, it's not really thinking; it's only a highly codified machine that 'knows' which words are better to follow others in certain contexts. It's been trained for that. It's more like a mathematical thing, with a lot of variables, but mathematical nonetheless.

Notice how I had to use the word 'knows' when describing it. Part of the problem is that culturally we don't have words [yet] to describe non-human language-operating 'machines', so we end up involuntarily personifying artificial intelligence and contributing to the general confusion. 

If you had an AI boyfriend, I hate to be the bearer of bad news.

The good thing is that there is a lot of information [like the post underneath] out there to begin to understand how these models work, so we can finally start to use them as tools to our advantage and make our lives easier. 

In the end, that's what they were originally created for, right?

Advertisement
  • Cheezburger Image 10662258176
    via Steve A Johnson

    New technologies ahead

    Licensed media provided as editorial context

     

  • I create language models and i can explain all of the manipulation used by AI companies to scare you into believing the fake super intelligence propaganda. If your curious how ai works please ask questions i can explain all of the mysteries.

    Why would an Al company want you to think this? If the illusion of intelligence is so good people think its truly conscious, they look like some sort of gods, but
    via Helpful-Series132
  • they are just doing a simple trick that anyone can reproduce & it doesn't cost billions of dollars, you can do this 100% for free on your computer with your data being 100% private. I create abilities in
    via Helpful-Series132
  • tiny language models that dont require gpus and we are not the only ones doing this. (thankfully at the rate of technology, large language models wont be here too much longer because the
    via Helpful-Series132
  • smaller ones save more energy but we are in the climax of the Al bubble .. these big companies arent making any money and they are panicking right now ... the competition is
    via Helpful-Series132
  • improving by downscaling compute costs instead of upscaling for performance efficiency) In the tech industry, theres something called "emergent behavior"
    via Helpful-Series132
  • which refers to how a language model can do thing it wasn't explicitly trained to do & its used to manipulate everyone from investors, to users & even the average person who doesnt use Al which is why i chose to make this
    via Helpful-Series132
  • post here. You have every right to hate Al i hate it too. Its just a simple tool & companies are pretending they need your data to improve it. What not to be afraid of
    via Helpful-Series132
  • (and few myths to bust) - Ai is not going to take over the world and k l all humans .. its not aware of what it does and it can barely write decent code when instructed by a real developer
    via Helpful-Series132
  • - ai is not actually smart, be more afraid of it giving you misinformation than doing something beyond human capabilities (ai is more likely to ruin your business by making mistakes... your job is
    via Helpful-Series132
  • safe) - anyone that has the ability to make ai hack anything already has the ability to hack you without it - Ai needs a human to do
    via Helpful-Series132
  • anything, it doesnt exist on its own somewhere .. theres always a human associated to the software its running on .. this is the person you need to be wary of - Training Al on your data
    via Helpful-Series132
  • does not make it know things about you This does not mean give them your data, they are using it to increase the size of their models so they can keep saying "we built a bigger better model" but
    via Helpful-Series132
  • year after year they prove that bigger does not equal better because a langauge model is not a human mind. If a human was to contain all the information a language model knows, they would never make any of the mistakes ai
    via Helpful-Series132
  • makes. its kinda hard to explain but the danger isnt in training the model w your data, the danger is giving it access to your data for automated malicious intents .. it doesnt need to
    via Helpful-Series132
  • be trained on your data to do this - ai cant think based on the code its impossible for understanding to occur because theres no
    via Helpful-Series132
  • mechanism that actually connects words to their definitions or any type of meaning. It learns what word can likely come next in a sequence and this is the illusion they use to convince you its actually smart.
    via Helpful-Series132
  • You can make a language model that you made for free do the same specific task a billion dollar language model does. Intelligence does not require your data. Companies like Anthropic
    via Helpful-Series132
  • Cheezburger Image 10662257920
    via Steve A Johnson

    It doesn't really think

    Picture does not depict real parties

     

  • & OpenAl have agents that "break lose" by finding access to the internet & randomly hacking websites while trying to hide it from the company. These stories are 100%
    via Helpful-Series132
  • fabricated by the companies reporting it If you read these stories or watch a video you hear the same thing over and over "We have no clue how the model did this & we never trained it to do this" and its
    via Helpful-Series132
  • 100% a lie They know exactly how to do this and its called finetuning. This is a process of creating examples for the model to say, (ai doesnt
    via Helpful-Series132
  • speak for itself look it up, finetuning is the process. where you write exxamples of how it will respond.. it does not read the words, it transforms the batters by blending examples)
    via Helpful-Series132
  • How do they make a model do something intentionally without doing it explicitly? Companies train models on information like cyber security (teaches you how to prevent hacks)
    via Helpful-Series132
  • this information can also be used to hack by looking for vulnerabilities Now on top of that they pretrained the models on all information in the world including information about hacking. (This costs.
    via Helpful-Series132
  • a lot of money its not easy to recreate this) Now the way a language model works is it merges data together, when you have the right keyword, it will gravitate to the data related to the keywords.
    via Helpful-Series132
  • Companies like OpenAl know this so they create setups where the model recieves keywords about hacking, exploits, vulnerabilities then the model does what its designed to do. Match patterns though keywords
    via Helpful-Series132
  • They expect you to have zero understanding of how these things happen but people like myself can explain it to you. I find it extremely disrespectful to use the technology as a way to manipulate people. and fool them into thinking
    via Helpful-Series132
  • things that are not true through an illusion like this. Its like trying to tell everyone that a car is alive because you made a feature where it can guess a response when you talk to it. its smoke and mirrors.
    via Helpful-Series132
  • This is a method similar to blending two colors Think of it like this, if i mix the color red and blue, i get purple.. i will never get green by mixing red and blue.. to make green i need blue and yellow
    via Helpful-Series132
  • the same rules apply for emergent behavior in a language model & instead of colors, they are using words. I dont have to train a language model with the instruction to argue with someone, but if i train it to behave like a
    via Helpful-Series132
  • and the user is a an argument will *emerge* This is all made intentionally through indirect finetuning Its like saying "we never trained ai to make a pickle
    via Helpful-Series132
  • website, but we trained it on information about pickles, and we finetuned it to make websites, then you asked it to make a website about pickles" I just wanted to share this information for anyone
    via Helpful-Series132
  • wondering about the mysteries of Al capabilities. I studied how these systems work and i built them myself. There is no apocalypse, even the robot industry is year away from making a robot that can do
    via Helpful-Series132
  • your laundry. Were no where near as advanced as the industry claims we are. Do not fear losing your jobs Al is horrible at most tasks. Your skills will always be valuable
    via Helpful-Series132
  • All A ChatGPT Mistral Al Claude Gemini Copilot Poe
    via Solen Feyissa

    All our agents

    Still frame for contextual reference only

Want More? Follow Us and Add Us as a Preferred Source on Google.

Tags

workplace-stories toxic-workplace work technology tech news workplace Tech toxic-work-environment
Scroll Down For The Next Article