Explaining Explainable AI
Transcript
all right let's try and explain explainable AI this idea that you understand at least as best as you can what the AI does you have a input into the AI and you're going to get a consistent output and if you're not getting that consistent output you know why you're not getting that consistent output and you can go back and adjust so let's talk about you know the value and the problem statement and how we solve for it you in theory can put all of your data
into AI let's say it's open AI we can create a agent and we can use rag to upload all of the data all of our data set into a platform like open AI it can then tell you oh you're not getting improved Twitter engagement because you're not tweeting about uh software enough or you're not talking about customer experience how do you test that like how do you on that platform go in and say well if I were to change the amount of software conversation I'm having
then shouldn't I see a increase in engagement so you should be able to tweak that you should be able to ask chat GTP or whatever model your using chat Che DP may not be the best model for this but you should be able to tweak that input and then get the expected outcome it should Max its prediction should then say Oh if you talk about software development or customer experience 20% more then you should have a outcome that maxed your diagnosis of what your original problem
is if you don't get that then you have some tweaking to do whether it's weights biases or the data that you're inputting or the query or parameter that you're uh that you're asking that prompt needs to change so that it's consistent and you're getting the AI results that you expect years ago I attended this Tech Field Day event in which a software provider said hey trust us we can improve your SQL performance my initial feedback was wait SQL runs my most important applications how can I
trust you what are the weights and biases you're using they responded oh you know it's the black is of boxes basically we created a Twitter handle to to make fun of it needless to say that company didn't stay in business because their AI model wasn't transparent you don't have have to give away the keys to your IP when you're creating models but you have to make them as explainable as possible even if you're not creating models you're making models available your internal customers need to know
the difference between one model and a different model and the B bias weights in one model versus the bias and weights in another model and how to adjust the inputs and out outputs in the different weights and biases in the model explainable AI if you thought this was an good explanation of explainable AI comment like if you didn't think it was all that great and you know like for me to just go away also comment don't dislike the video though
