|
The thing that always confuses me about "learn AI" in reference to Chat-GPT is... What exactly is there to learn? It's a web terminal that you type queries into and it returns results. Learning how to coerce it into giving you what you want is certainly an art, but no more esoteric an art than using Google to correctly return a useful result to your query - it's a lot of trial/error and simple logic about how to get it to respond with something useful. In the case where people think that they either already are or can rapidly become experts in NLP, language modeling, or other topics related to Machine Learning... Well that's just the Dunning-Kreuger effect on full display. While you may be able to develop rudimentary tools based on machine learning w.r.t. NLP, none of these lifetime React devs with "interest in the AI space" are doing anything close to that, or if they are, it falls amazingly short of a project like OpenAI's. At this point, AI is the new "data science" of 2023, a handy buzzword for laypeople to invoke when they want to gesture towards "high tech", and one that's frequently divorced from an advanced technical understanding about how training/using a service like ChatGPT works. As a closing anecdote, a friend of mine (would be startup founder) recently showed me her business plan for a "AI powered music generation service" where users could use a ChatGPT like interface to compel the computer to give them a "lofi beat to relax and study to" or an "upbeat electronic track for a space exploration video game stream" for YouTube videos or producing other forms of license-free music. When I started digging into how this was going to be done, the furthest we got is: > "You need to train a neural network on a tagged and curated list of music samples, and develop instruments to allow for additional human training/tagging and re-processing, in addition to developing a suite of MIDI -> audio tools (essentially a headless/distributed DAW) to actually produce music. " Their response? > We'll build the web interface first then we can iterate from there. |
Even just chatGPT in its current form is more than that. It is a full-fledged virtual assistant - but an esoteric one. Using it as better search is one use case, but how about to write a novel? It isn't mature enough to write _good_ prose sufficient for the the whole length of a publication. But it is good enough to generate ideas for the plot, edit, identify continuity errors, flush out backstories, generate maps, new ideas, write starter content that you modify. It could seriously change how productive one is. Similarly for writing code, or putting together home remodel plans or whatever. Learning how to have an AI "assistant" is a new thing because the intuitions about what it will be good and or bad at, especially for folks less familiar with the underlying way LLMs are built is not something that you get automatically.
> AI powered music generation service
Your friend does sound a bit niave about the complexity of such a service. But even right now, you can ask ChatGPT to generate coord progression, or representative example pieces of music that match the prompts you gave that would could use with some downstream system without a full "music model" along the lines you described.
Plenty of us are finding dozens of new applications every day. But they don't all work 100% out of the box. Most don't. So understanding the limitations and what kind of integrations are needed is really key to leveraging the value.