Time to average down ?, page-75

Lets work out how LLM's work in laymans terms.

1. Accumulate 10 terrabit of data (thats basically the whole internet, henmce Nvidia going crazy as the power LLMs need)
2. Next step is to develop parameters for the data - parameters are basically a line of code that highlights a weight and bias on each string of text.
3. Create an alogrithm to give the AI personality or what the use case may be. - what this means you can give each parameter a weight towards comedy or horror for example.

An example on how many parameters an LLM needs - GPT 3.5 has 175 billion paramters GPT 4 will have 100 trillion parameters.

Issues arising with this method is that with the initial 10 tb of data the AI hallucinates as its just a mess and no mater how good the code is it will hallucinate as data is scrambled

The other problem is with data dumps you literally need to spend billions on AI chips to handle the power load.

..and we have not even touched images and video yet so in other words the 10 tb of text the size of the AI increases 200 fold from GPT 4 10 trillion parameter model.

If one thinks managing 10 trillion parameters is easy with a lump of messy data they are mistaken. Punters need to take a deep breath and understand despite Big tech losses it still bought in $400m revenue FY23. AI data is big business thats only growing