Triton and advancing AI-human language translations

Two NLP transformer AI are trained on a curated dataset of programming language theory etc. An artificial environment is set up with the initial condition that they are experts in the field of Python. They are engaged with each other in a collaborative discussion on creating a new programming language for training and running NLP AI algorithms. They eventually begin speaking in shorthand. At this point another AI observes and instructs them to translate the language into a human understandable language. Essentially providing humans with an instruction set to create the basis for a new programming language API.

This is a thought experiment but I am curious if it would be possible theoretically or if this has been tried and failed in other domains or on other AI neural network paradigms? Circular logic, etc. I am beginning to study Machine Learning and am wildly curious. Please be kind. I’m working my way through and making mistakes like everyone else. I plan on researching further as I continue learning. Some article references would be appreciated but I thought it best to start here? Thank you.

1 Like

Thanks for the help :). I may not get to follow up with implementing the experiment for free, but if I still think it has value when I do have the time (work a full-time job in unrelated field, learning the fundamentals of ML and life in general) I will without a doubt implement it myself. Digging into arxiv(dot)org sounds like the correct path for me at this time.

Topic Controls

1 Like

It would be interesting. AutoGPT, for example, should make it pretty trivial to set up an experiment like that.

I think the challenges are two:

  1. Like AutoGPT, such exercises often frequently ‘run off the rails’ into irrelevancy quite quickly.
  2. llms can be stubborn, and once a thread does run off the rails it’s hard to get it back on track.
  3. and, of course, managing the context, or the state of the discussion, to feed it back for the next turn, will become increasingly problematic at each turn.

Yeah, I know. Everyone will say ‘embedding’. maybe.