Nov. 2, 2022
Read time: 3 minutes and 12 seconds.
tags:We often treat AI like some kind of digital oracle, but it’s not as magical as it seems. It’s like it’s become the new google and everyone wants to swear that it’s got all of the answers. I even find myself saying, hey maybe you should check out ChatGPT. What does it say. It’s a tool, and like any tool, it has its limitations. I can tell you, expecting AI to solve all your problems is a recipe for frustration, and I speak of it first hand, with the launch of Chat GPT and the launch of other similar comparable tools that have come out since.
AI operates on probabilities, much like the stock market. Just like a stock can go up or down, these are based on inputs from the people, the news and the environment around it on which it’s trained. With the stock market you have these expectations… So those inputs generate a specific output kind of like an LLM. LLMs predict the next most likely work based on the input on which the response is encoded. So the response is how the answer is then decoded. So the output is based on what it has seen before, not independent thought. I find it hilarious how the majority of users are trying to use it for generating custom marketing material, even though it’s predominately trained on the majority of the most generic, google accepted websites indexed on a daily basis. This is the common crawl dataset. In case you’re curious, you can see that it’s an extensive dataset, but one with it’s limitations.
So, as much as we can use it to help us use best practices in development or anything like that in marketing, we are really getting the surface level generics that we see at the top of search results before loading a page. LOL
But enter me, trying to get the perfect output just like anyone else, which can be a hair-pulling experience. AI often regurgitates information in a conversational format, which can be misleadingly comfortable. it acts as a friend and tries to give you the same cookie cutter response you find on Google, just with some nice conversational sprinkles. The only difference is that a google search is only influenced by Google whereas you will likely influence the AI to respond or answer a certain way. You’re probably going to get the wrong output or variable outputs in the process… it’s just regurgitating Google with a mix niceness and DEI moderations.
It excels at one-off tasks but struggles with broader, more complex concepts. So if you have a simple ask, the AI is often really good at answering it, providing it in a given context and being your friend if you’re like me and don’t have any. lol. But these broad tasks people make it do give it generic results, not due to the prompt that’s being put in but rather the training data set as the inputs for these types of models. The old mathematical formula:
Garbage In = Garbage Out
AI has to guess a lot, since it’s a deterministic mathematical model trying to predict the likelihood of tokens (or parts of words) are likely to come up as a response without pissing off the user. So, the AI is typically tuned to pay the most attention to the things you say at the very beginning of your prompt and the desired output more than all of the crap in between. The things in between are kind of like the condiments you never asked for and didn’t care to remove, but … probably didn’t want on your sandwich (for example).
AI is a powerful tool for specific tasks, but it’s essential to understand its limitations. Don’t expect miracles, and be prepared to refine your approach. Remember, it’s more about clever word prediction than true intelligence. I’d probably recommend checking out libraries where you can see how others prompt and how people pull their hair out with AI too. Because what would happen to the hat market if we all had hair?