Hello , I am new in programming, I would like to train the program with formulas through machine code and images to recognize examples and processes, could it be done through json files?
If you can give me an example of similar training, I would appreciate it.
2 Likes
Hi agdr90 and welcome to the community!!
GPT-3 (and transformers in general) are amazing at translating text, be it existing languages (including programming languages and math) or made up languages. As long as you think of your problem as text-in text-out, you’re on the right track. I say this, because you mentioned images and json. If you can convert the images to text prior to feeding it to GPT-3, that can work. Json is ideal, I’ve played with going from json to NLP sentences, and vice versa, it works just fine.
Two obstacles I would like to point out:
-
If a “language” exists, for instance French or Java or MatLab, you’ll only need a few examples for the model to recognize the pattern you’re referring to, but if your language is new, you’ll need to teach the model your language, meaning many more examples and also, examples deliberately fabricated to “teach” specific concepts or meanings within the new language.
-
GPT-3 is a language generation model, it’s incredibly good at NLU and NLG, it’s not particularly good at logic. You can, for instance, translate English to MatLab code using GPT-3, and then run the MatLab code to perform the actual math, but you’ll not have much success if you’re trying to solve complicated math problems using GPT-3. I’m also new here, and new to the tech, but I think all us newbies are playing in the playground with ideas to get a feel for what’s possible and what is not possible with this technology.
I know you’re hoping to get a link to read more, my apologies, I can’t help in that regard, just felt like sharing my thoughts.
Enjoy the journey @agdr90 
2 Likes