Day 125 MIT Sloan Fellows Class 2023, AI for business 3 "Formula for AI value"

www.youtube.com

Formula

Problem statement

If from <input> we could know <output>, then <someone> could derive a benefit by

Inputs(X)

What are the inputs X?
What does X fail to encode?
What problems might this create?

Outputs(Y)

How is Y constructed? - Automation or Prediction?
How does Y differ from <output>?
What problems might this create?

Training data

Where does the data come from? How was it generated?
Is it representative?
How much do you have? How much do you need?

Errors

Y and Y' will occasionally differ. What impact will these errors have?

Pivot

We reviewed "actuate" example as a case study. This product is about School Security. Actuate protects students by improving police response time at schools and universities.

In the initial plan, their product has

<input>: camera feed from schools
<output>: is there an active shooter?

The problem they have was too scarce data.

Then, they pivoted their position to the following condition.

<input> some rare event→some more common event

<output>Active shooter → gun
<output>Intruder → person not seen before

We can apply this methodology for a lot of similar problems such as...

Hate speech detection
- text/image → person
Officer shooting risk
- officer → office/area
Investing child welfare
- child → family
Train derailment
- maintainance logs → image

Fine-tuning model

Think about what you might want from a generative model.

Truth
Civility
Specialty
Uniqueness

Does “predict the most likely next word” capture those values?

The way to going beyond next-word prediction is "fine-tuning".

Build the best next-word predictor you can using whatever text you have
Collect a much smaller dataset tailored to the particular task you have in mind
1. If you need a legal analyst,<input> = legal question, <output> = answer
Use the next-word predictor as a starting point to build a task-specific model