What is it?
The problem with AI assistants, like ChatGPT, Gemini, Claude, and the like is that they are good at mimicking and giving correct information with no immediate way to confirm their correctness. LLMs by their nature are not capable of discerning truth from falsity and are not capable of returning absolutely correct information. What they are good for is predicting one word after another with reasonable plausibility and vaguely following instructions.
This means that if they weren’t trained with enough data on a subject you want to know of, or simply do not have a correct answer for it, they will just make up something sounding vaguely right without knowing if it’s right or wrong. It would be great if, in situations like these, it would advise the user to look for answers elsewhere instead. Enter: NotGPT!
Not only will this assistant advise against using AI for information gathering, it will straight up advise the user to instead search for answers in search engines like Google or Bing!
Obviously, this is more of a joke than anything useful, more of an “interactive performance”, but the premise is very real. These dangers of LLMs are largely ignored by users who are more content with getting reasonably good-sounding information quickly instead of correct information with some research.
How does it work?
The answer-generating component is developed by Microsoft called Phi3, a very small, high-performance LLM, that’s just enough to follow a system prompt with a user prompt to achieve this goal. Running on a free Oracle Always Free VM with 4 ARM cores gives just enough performance so it looks like it’s doing something intensive but still fast enough that the user doesn’t leave the page.
To get an API for the LLM inference I’m using Ollama which is a very nice and easy-to-use starting point for LLMs if you want to try any out. Coupled with Open WebUI you can have a ChatGPT clone up and running in no time.
Instead, I just cobbled together some basic UI with Bootstrap that looks a bit like ChatGPT and ran it with Deno along with a basic TypeScript API which calls Ollama.
You can check the project out for yourself at ai.doctor0710.xyz!