You’ve used AI for years - spam filters, autocomplete, Netflix picks. So why does ChatGPT feel like a different species?
Watch your spam filter evolve:
- 1995: a human wrote rules - “if it says ‘free money’, block it.” Spammers wrote “fr3e m0ney” and walked straight through.
- 2010: it learned from millions of emails what spam looks like - no hand-written rules.
- Today: AI can write the email, reply to it, and summarize the thread.
Each step handed more of the thinking to the machine.
The three types that matter today
Almost everything called “AI” today is one of three things - and they’re nested, each a subset of the one before.
1. Traditional Machine Learning
- What: algorithms that learn patterns from mostly structured (table-like) data.
- How it learns: a human picks the useful columns - “features” like income, age, past purchases - and the model finds the relationship.
- Output: a number or a label - “85% likely to churn”, “fraud / not fraud”.
- Names you’ll hear: logistic regression, decision trees, random forests, gradient boosting.
2. Deep Learning
- What: neural networks with many stacked layers. A subset of ML.
- How it learns: feed it raw, messy data (pixels, audio, text) and it discovers the useful features by itself - no manual feature engineering.
- Output: still usually a label/number, but on unstructured data - “this scan shows a tumor”.
- Cost: needs lots of data and serious compute (GPUs).
3. Generative AI
- What: Deep Learning trained to generate brand-new content, not just label things.
- How it learns: trained on enormous amounts of unlabelled data by predicting missing/next pieces (“self-supervised”).
- Output: new content - text, images, audio, video, code.
- Superpower: one general model handles many tasks it was never explicitly trained for.
One pile of reviews, three jobs
Traditional ML predicts a 1-5 satisfaction score from a table. Deep Learning reads the raw text and classifies each review positive / negative / mixed. Generative AI writes a personalized reply to each unhappy reviewer. Same data - predict, classify, create.
When to use which
Traditional ML
- Data type
- Structured tables
- Who finds features
- You (manual)
- Typical output
- Number / label
- Data + compute
- Low
- Explain the why?
- Often yes
- Best for
- Predicting from tidy data
Deep Learning
- Data type
- Unstructured (images, audio, text)
- Who finds features
- The model
- Typical output
- Number / label
- Data + compute
- High
- Explain the why?
- Hard
- Best for
- Perception tasks
Generative AI
- Data type
- Unstructured, general
- Who finds features
- The model
- Typical output
- New content
- Data + compute
- Very high
- Explain the why?
- Hard
- Best for
- Creating, summarizing, reasoning over language
Pick the right tool
Drag each task onto the approach that fits best.
Traditional ML
Predict from tidy tables
Deep Learning
Perceive raw images / audio
Generative AI
Create & summarize language
0/6 correct
- AI evolved: rule-based -> traditional ML -> deep learning -> generative AI, each handing more thinking to the machine.
- They're nested: GenAI is deep learning; deep learning is ML; ML is AI.
- Match the tool to the job: tidy/explainable -> ML; perception on raw data -> DL; creating content -> GenAI.