I have a very basic question, I have been working for some years now with traditional ML models(not the modelling part as per say, but the operationalisation via MLops pipelines), and I have recently started deep diving into LLM’s. Is there any scenario where LLM’s can make forecasts based on historical data or it will be limited to text completion/generation, sentiment analysis, code generation etc?
LLM’s are language models.
The only predictions they make are what the next letter should be given a sequence of previous letters.
Yeah that’s my understanding too. I asked it because LLM is being portrayed as something that will end traditional ML models and will become a one stop solution. I am skeptical on that, LLM will be one part of the ML umbrella that’s all.
@gaurshazra Ha ! I’ve been playing too, but I think I agree with you.
At least I feel we should be ‘mindful’.
I don’t see that happening in the near future. The closest an LLM framework has come to AFAIK is to invoke a custom tool to perform a task.
As far as writing code is concerned, LLM can emit code based on its training set which is usually a corpus from github.
Try writing a tool to deal with datasets via something like auto-sklearn and invoke the tool via a prompt. Do turn on logging to ensure that the tool is invoked with the correct configuration.
LLMs (ChatGPT specifically) are great for generative tasks not analysis and forecasting. There is a capability in ChatGPT that allows you to upload Excel files, etc. You can then ask it to analyze the data for you but it usually runs into errors and wont complete basic analysis of a simple file. It would be a struggle to get it to work with complex data sources that need wrangling. You are better off using the normal Python or R packages.
True. They are basically transformers used for NLP tasks.
e.g - ChatGPT is based on a Transformer architecture, specifically a variant of the GPT (Generative Pretrained Transformer) model.
Technically I think these chat tools are LLM’s that have been fine-tuned to answer prompts.
A LLM is just a language model. It can be used as the basis for other types of tools. There is a huge industry forming around creation of such tools.
I agree @gaurshazra . LLM will need guidance from the coder. Currently, you need a coder to give the algorithm/steps to solving the problem at hand.