What is AI, really?
Overall, AI is a predictive process. From an input, we try to predict an output.
Example 1:
An example of this Process is Image Recognition.
Input: Image
Output: Class of animal
Example 2:
Another (more practical) example is House Price Prediction.
Input: Size, neighborhood, number of bedrooms, etc.
Output: Price
How engineers build this
Symbolic AI (Classic)
- Rule-based
- Domain experts program these manually
- E.g. house pricing can be defined by real estate managers using their selling experience
Machine Learning (Modern)
- Pattern-based
- The program is automatically created from input/output pairs
- E.g. house pricing can be automatically learned from thousands of house listings
ChatGPT is not magic! It has been built the same way
It does not "learn from reading the internet". OpenAI used the internet to learn a representation of language, but used 13,000 Q&A pairs to build the InstructGPT (precursor to ChatGPT)
1
Pre-training
Billions of text examples
2
Supervised Finetuning
10s of thousands of Q&A pairs
3
Reward Modeling
Score different outputs and guide the model towards the best ones
Use Case in Manufacturing
Blister Inspection
Detecting missing pills, damaged packaging or incorrect colors can be automated with Computer Vision.
- Input: Blister pack images from production line.
- Output: Pass/Fail decision
Use Case in Retail
Product Recommendations
E-commerces can personalize product suggestions based on customer behavior and preferences, driving more sales.
- Input: Purchasing and browsing history from client A, which is very similar to other clients (let's call them B and C)
- Output: Product suggestions for client A
Use Case in Recruiting
Resume Screening
AI automates the initial screening of resumes, identifying the most suitable candidates based on job requirements, skills, and experience.
- Input: Resume + Job Description.
- Output: Scoring for each qualified candidate.
Use Case in Administration
Email Categorization
AI can automatically analyze incoming customer emails to detect complaints, categorize them by topic, and route them to the appropriate customer service department, significantly reducing manual triage time.
- Input: Email content.
- Output: Complaint detected (yes/no), and recommended department (e.g. sales, tech support, billing).
Use Case in Customer Service
Analytics for Call Centers
An example of 2 or more AI models being used in a chain.
One model can process customer service calls by transcribing the conversation. This transcription is then fed to another AI model that generates a summary with key issues and action items.
- Input 1: Voice call
- Output 1 (+ Input 2): Text transcript
- Output 2: Summary
