Types of AI
HMMs, logic, machine learning, deep learning, reinforcement learning, transformers, and RLVR. A taxonomy of the ideas that built modern AI.
"AI" is a family of techniques that evolved by replacing one another's weaknesses. When people talk about AI, they usually mean one of these specific things. Being able to name them, and roughly when each dominated, is the difference between vibes and understanding.
Symbolic and logic-based AI
The original approach: encode knowledge as rules and use logic and search to reason over it. Prolog, expert systems, theorem provers. Powerful for well-bounded domains, brittle everywhere else. This was the dominant paradigm from 1956 into the 1980s, and its collapse triggered the second AI winter.
Hidden Markov Models
HMMs model sequences as a chain of hidden states with probabilistic transitions and observations. They powered speech recognition, part-of-speech tagging, and bioinformatics for decades (roughly the 1970s through the 2000s). They are the "before deep learning" answer to sequence problems, and they matter historically more than practically now, though they survive in niche tools.
Machine learning
The statistical turn: instead of writing rules, learn a function from labeled examples. Classical algorithms like logistic regression, SVMs, decision trees, and random forests solved structured-data problems and dominated from the 1990s until deep learning displaced them on unstructured data. This is still the right tool for many tabular problems today.
Deep learning
Neural networks with many layers, trained with backpropagation on GPUs. AlexNet's 2012 win started the modern era. Deep learning learned representations rather than hand-engineered features, and it wiped out classical approaches on vision, speech, and much of NLP.
Reinforcement learning
An agent learns by acting in an environment to maximize reward, trading exploration and exploitation. Rooted in 1950s dynamic programming and Q-learning (1989), it exploded with DeepMind's DQN (2015) and AlphaGo (2016). RL now runs inside modern LLMs as a fine-tuning stage.
Transformers
The architecture behind every frontier model since 2017: self-attention over tokens, trained in parallel at scale. Transformers made scaling the main lever, and scaling is why 2023-to-2026 models keep getting measurably better.
RLVR and reasoning models
RLVR (reinforcement learning with verifiable rewards) is a major technique in the reasoning-model era. Instead of human preferences (RLHF), the reward is a checkable fact: the correct math answer, a passing test, or the right code output. It is especially effective on tasks with a reliable grader, and it helps explain why models can now spend more compute on problems with objectively checkable answers.
| Approach | Time | Space | When |
|---|---|---|---|
| Logic / symbolic | 1956–1980s | Expert systems | Brittle but explainable |
| HMMs | 1970s–2000s | Speech, tagging | First sequence models |
| Classical ML | 1990s–2010s | Tabular, spam, ranking | Great on structured data |
| Deep learning | 2012–today | Vision, speech, NLP | Learns representations |
| Transformers | 2017–today | All frontier models | Scale is the lever |
| RLHF / RLVR | 2022–today | Alignment + reasoning | Makes LLMs useful |
Quick check · Which training technique rewards a model for correct, verifiable outputs rather than human preferences?
Sources and further reading
Premium
Unlock the rest of this guide
Premium unlocks every pattern deep-dive, every problem breakdown and solution, the practice sandbox, and verdict feedback on your practice runs.
- AI is a family of techniques, not one thing: symbolic logic, HMMs, ML, deep learning, RL, and transformers.
- Each wave solved the weakness of the one before: brittle rules became learned statistics, learned features became learned representations.
- RLVR (verifiable-reward RL) is the 2025-era technique behind today's reasoning models.
- The full article, complete and uninterrupted
- All pattern deep-dives and problem breakdowns
- Practice sandbox and verdict feedback