• 1 Post
  • 14 Comments
Joined 1 year ago
cake
Cake day: October 29th, 2023

help-circle











  • Pretraining = Unsupervised Learning

    Fine Tuning = Supervised Learning

    Human Feedback = Reinforcement Learning

    These three steps produce what are currently calling AI.

    A modern LLM like Mistal7B is made of 32 layers of 4096x4096 transformer nodes

    In pretraining,
    Coherent data is fed thru the network one word at a time (in this case the entire internets text) and the models node-connection-weights are automatically adjusted towards the values such that given a list of words it correctly predicts the next one.

    In finetuning,
    This time Data Pairs are fed thru, (example prompt AND example correct answer) this bangs the model over the head and forces it to respond to our prompt formatting, it’s also where we make it helpful and do what it’s told.

    In Human Feedback,
    (Abbreviated to RLHF) We let the model mutate slightly, having it generate multiple responses with slightly differing internal weights and having actual humans select their favorites, over time this draws the model towards not just generalizing from text examples, but also towards actually pleasing humans with words (what ever that process might entail)

    All intelligence emerges during the pure prediction/pretraining stage, Finetuning and RLHF actually damage the model SIGNIFICANTLY! but working with pure text prediction engines requires much more though than simple prompt engineering.

    There’s a strong mathematical proof that states Modeling == Prediction == Compression == Intelligence

    Saying It’s essentially impossible to get any one of these without also getting the other three,

    Accurate modeling provides Prediction (by simply running the model forward in time), Accurate Prediction provides Compression (by only storing the difference from the prediction)

    And intelligence (I.E. Getting what you want) is simply a mater of using your Compressed Model of the world to Predict what might happen if you performed various actions and selecting the one where you get what you want.

    Modern Open Source LLMs (like OpenChat, DarkSeekCoder etc) are actually superior to GPT4 now…

    The reason they might seem behind to casual users is the RLHF step which GPT4 has received A TON OF!

    This is an expensive step requiring many peoples time, and all open models simply skip this.

    The thing is using techniques (like simply asking a million times) we find the knowledge and skill in these OS models are CLEARLY far and beyond the latest available OpenAI GPT model (its 11/9/2023 now)

    Hope that helps! model LLM based AI is actually extremely simple, we create an intelligence beast using prediction and then we bang it over the head (giving it serious literal brain damage) to behave for us, then we listen closely to it and slap it in the face for the tiniest mistake until we’re happy with it.

    It’s still just an insanely high dimensional word-by-word predictor, it’s just been traumatized by humans to please us.

    This pretty much sums it up

    https://preview.redd.it/pocf96867dzb1.jpeg?width=1000&format=pjpg&auto=webp&s=8447682e05dd1339f563945218d4ccd3f3382a2b

    Enjoy!