Haha that AI generated summary is even less appealing than the video. I’m not a coder so what I’m saying is the value of Grace doesn’t translate to me. Is it that using Grace code you’re able to output prompts that LLMs are better at processing? I’m asking for examples of how you are using it. Thanks!
I mean, if you’re not a coder, a prompt engineering language might be useless to you! 🤣
In case you are:
for coders
Let me condense it further: in programming you have imperative, object oriented, and functional programming.
Grace is functional programming (and all that entails) distilled into a json-shaped programming language that LLM’s can be instructed to generate which automatically type checks itself within each of its fully encapsulated sub-agents.
It uses the data encapsulation principles used in functional programming to keep the hallucination
machine on rails that it can check against in real time. Because of that data encapsulation, the programmer can compose independent (parallel, asynchronous, it doesn’t matter) pieces of their program that can each return a fully DETERMINISTIC result because of aforementioned type checking and isolated sub-agents.
Think of something you want to do that would be impossible. Program all of the pieces leading into the part where you need to turn unstructured data into structured data in a solid language, create the magic LLM part in Grace’s bespoke language (which LLM’s can easily generate and validate) then create another part in that same solid language that uses that result (after much type-checking and encapsulated iteration perhaps depending on complexity).
Sounds interesting to me. The only issue I have with it is that it only works with OpenAI’s API at the moment. But it shouldn’t be hard to extend it to work with open models. I have my eye on a DeepSeek extension.
Thanks for the explanation. Maybe I don’t understand, but I thought by nature LLMs were not deterministic. Is it because you’re breaking the output up into pieces?
Breaking them up and isolating them. Not only that, but iterating and correcting them using a type system. You know how an LLM doesn’t always get the answer right the first time? This makes sure that the answer is right before using it as a source of truth.
Haha that AI generated summary is even less appealing than the video. I’m not a coder so what I’m saying is the value of Grace doesn’t translate to me. Is it that using Grace code you’re able to output prompts that LLMs are better at processing? I’m asking for examples of how you are using it. Thanks!
Haha. No thanks to the doxxing.
I mean, if you’re not a coder, a prompt engineering language might be useless to you! 🤣
In case you are:
for coders
Let me condense it further: in programming you have imperative, object oriented, and functional programming.
Grace is functional programming (and all that entails) distilled into a json-shaped programming language that LLM’s can be instructed to generate which automatically type checks itself within each of its fully encapsulated sub-agents.
It uses the data encapsulation principles used in functional programming to keep the hallucination machine on rails that it can check against in real time. Because of that data encapsulation, the programmer can compose independent (parallel, asynchronous, it doesn’t matter) pieces of their program that can each return a fully DETERMINISTIC result because of aforementioned type checking and isolated sub-agents.
Think of something you want to do that would be impossible. Program all of the pieces leading into the part where you need to turn unstructured data into structured data in a solid language, create the magic LLM part in Grace’s bespoke language (which LLM’s can easily generate and validate) then create another part in that same solid language that uses that result (after much type-checking and encapsulated iteration perhaps depending on complexity).
Sounds interesting to me. The only issue I have with it is that it only works with OpenAI’s API at the moment. But it shouldn’t be hard to extend it to work with open models. I have my eye on a DeepSeek extension.
Doxxing?
Thanks for the explanation. Maybe I don’t understand, but I thought by nature LLMs were not deterministic. Is it because you’re breaking the output up into pieces?
Breaking them up and isolating them. Not only that, but iterating and correcting them using a type system. You know how an LLM doesn’t always get the answer right the first time? This makes sure that the answer is right before using it as a source of truth.
Can you give me a use case?