Comments on Profile Post by Vtory

Notes

1. Applicable to entire popular LLMs incl. gpt3.5turbo or gpt4.
2. The larger LLM is, the harder it fails.
3. Very likely because these AIs do NOT understand semantics but pretend to understand through correlations.

Inspired by this: https://arxiv.org/abs/2305.15507

AIs are the new MBAs. Very eloquent with the presentation but lacking precise understanding of just about anything they touch. And that's not a bad thing either cause if AI were precise on a very niche topic with say n=1 data points, it would be equivalent to copying that verbatim and essentially revealing where it learned things from, which is a big data privacy nono.

And programming languages are designed for precision work, where the interpreters/compilers/semantics really are designed hand-in-hand with the syntax (some subset of English words and math notations), so GPT will never beat an interpreter at, well, interpreting code. It might still beat inexperienced programmers at many things though.

The cool part I think is that AI is even "softer" than current software, where you can ask it a question, argue with it, curse at it, etc...in any language and it will give an approximately apt response, especially in English where there's most data. If you've ever worked in customer support, translation, or writing, etc, having a 5-year-old polyglot bookworm assistant is still quite useful for certain things.