Foundations Of AI
Types of AI
AI systems get categorized along two largely independent axes — confusing them is a common beginner mistake, so this chapter treats them separately before conne
Jr Codex AI Notes
Level: Beginner Prerequisites: Chapter 2 Time to complete: ~20 minutes
Table of Contents
- Two Ways to Classify AI
- By Capability: Narrow, General, Super
- Artificial Narrow Intelligence (ANI) — Where We Are Today
- Artificial General Intelligence (AGI) — The Goal, Not Yet Reached
- Artificial Superintelligence (ASI) — Speculative
- By Approach: Symbolic vs Sub-Symbolic AI
- Why Both Classifications Matter for This Curriculum
- Summary & Next Steps
1. Two Ways to Classify AI
AI systems get categorized along two largely independent axes — confusing them is a common beginner mistake, so this chapter treats them separately before connecting them at the end.
Two Independent Classifications
─────────────────────────────────────────
BY CAPABILITY: how broadly/deeply can it perform? (Sections 2-5)
Narrow → General → Superintelligent
BY APPROACH: HOW is it built? (Section 6)
Symbolic (rules/logic) vs Sub-symbolic (learning-based)
─────────────────────────────────────────
2. By Capability: Narrow, General, Super
The Capability Spectrum
─────────────────────────────────────────
ANI AGI ASI
(Narrow) (General) (Super)
────────────────────────────────────────────────────
Excels at ONE Matches human-level Exceeds human
specific task performance ACROSS intelligence
ANY intellectual task across EVERY
domain
EXISTS TODAY DOES NOT EXIST YET SPECULATIVE
─────────────────────────────────────────
This spectrum is the single most important classification to internalize early — it directly shapes realistic expectations about what today's AI (including advanced LLMs) can and cannot do.
3. Artificial Narrow Intelligence (ANI) — Where We Are Today
Every AI system that exists today, without exception, is narrow. Even the most impressive modern systems are highly capable within a bounded domain, not generally intelligent across all domains.
# A narrow AI's defining trait: extremely good at ONE thing, unable to
# do something UNRELATED, even if it seems "simple" to a human
# A chess engine — vastly superhuman at chess...
def chess_engine_move(board_state):
return "best_move_for_this_position"
# ...but it cannot, without being an ENTIRELY different system,
# hold a conversation, recognize a face, or drive a car.
# Even a modern LLM, despite being broadly capable across many
# LANGUAGE tasks, cannot natively see, physically act, or learn
# genuinely new skills the way a general intelligence would.| Example | Domain | Limitation |
|---|---|---|
| Chess/Go engines | Game-playing | Cannot do anything outside the game |
| Spam filters | Text classification | Cannot understand the email's actual content beyond spam/not-spam |
| Recommendation engines | Predicting preferences | Cannot explain WHY beyond statistical pattern |
| Large language models | Text generation/understanding | No persistent memory across sessions by default, no physical embodiment, can produce fluent but factually wrong output |
Even a system as broadly capable as a modern LLM is still classified as ANI — it's extraordinarily general within language tasks, but that's still a bounded domain, not "any intellectual task a human can do" (it can't, for instance, physically repair a car, or learn a brand-new skill the way a human apprentice does through real-world practice).
4. Artificial General Intelligence (AGI) — The Goal, Not Yet Reached
AGI refers to a hypothetical system with human-level cognitive ability across the full range of intellectual tasks — able to reason, learn, and adapt in genuinely novel domains without being specifically engineered for each one.
What AGI Would Require (Not Yet Achieved)
─────────────────────────────────────────
- TRANSFER learning across wildly different domains without
being retrained from scratch for each one
- Genuine common-sense reasoning about the everyday physical
and social world (an area where even top LLMs still fail
in surprising, inconsistent ways)
- The ability to set and pursue its OWN novel goals, not just
optimize a goal a human specified in advance
─────────────────────────────────────────
Why this matters for a beginner to understand clearly: media coverage and marketing sometimes blur the line between today's impressive narrow systems and AGI — this curriculum draws that line explicitly and consistently, since conflating the two leads to both unrealistic fear and unrealistic hype.
5. Artificial Superintelligence (ASI) — Speculative
ASI refers to a hypothetical intelligence that would exceed human capability across every domain — a concept discussed extensively in AI safety research (Module 6) but which remains firmly speculative, with no current technical path agreed upon by researchers.
Where ASI Fits in This Curriculum
─────────────────────────────────────────
ASI is NOT a technique you'll learn to build — it's a
discussed possible FUTURE state, relevant mainly to Module 6's
coverage of AI safety and alignment, where questions about
long-term risk and governance are explored.
─────────────────────────────────────────
6. By Approach: Symbolic vs Sub-Symbolic AI
Independent of capability level, AI systems are also built using fundamentally different techniques — this axis directly maps onto how this curriculum itself is organized.
Symbolic AI Sub-Symbolic AI
───────────────────────────── ─────────────────────────────
Represents knowledge EXPLICITLY, Learns patterns implicitly
using rules, logic, and symbols from DATA, without explicit
a human can read directly hand-coded rules
"If patient has fever AND rash, "Here are 10,000 labeled
THEN suspect measles" X-rays — learn to recognize
pneumonia yourself"
Module 2 (search), Module 3 The ML, DL, and NLP/LLM
(logic, knowledge representation, notes — neural networks,
planning), Module 4 (probabilistic gradient descent, learned
reasoning) — all primarily representations
SYMBOLIC techniques
───────────────────────────── ─────────────────────────────
# Symbolic approach — an explicit, human-readable RULE
def diagnose_symbolic(has_fever, has_rash):
if has_fever and has_rash:
return "possible measles"
return "unclear"
# Sub-symbolic approach (conceptual sketch) — a model LEARNS
# this mapping from thousands of labeled examples, and its
# internal "reasoning" is not a simple readable rule at all,
# but millions of learned numerical weights (full depth in
# the Deep Learning notes)Neither approach is strictly "better" — symbolic AI is transparent and precise but brittle outside its explicitly coded rules (echoing Chapter 2's expert systems lesson); sub-symbolic AI generalizes well from data but is often a "black box," motivating Module 6's coverage of explainable AI.
7. Why Both Classifications Matter for This Curriculum
Mapping This Curriculum onto Both Axes
─────────────────────────────────────────
Capability axis: Every technique in this ENTIRE curriculum
(including deep learning and LLMs, covered
in separate notes) produces ANI systems —
none of it builds AGI or ASI.
Approach axis: THIS AI Notes curriculum focuses heavily on
SYMBOLIC techniques (search, logic, planning,
probabilistic reasoning) — the Machine
Learning, Deep Learning, and NLP/LLM notes
cover the SUB-SYMBOLIC side in depth.
─────────────────────────────────────────
Understanding this mapping clarifies why this AI Notes curriculum doesn't re-teach neural networks — that's thoroughly covered elsewhere, and this curriculum's unique value is the classical, symbolic side of AI that a modern AI practitioner still needs to understand.
8. Summary & Next Steps
Key Takeaways
- AI is classified along two independent axes: capability (Narrow → General → Super) and approach (Symbolic vs Sub-symbolic).
- Every AI system that exists today — including the most advanced LLMs — is Artificial Narrow Intelligence (ANI); AGI and ASI remain, respectively, an unreached goal and a speculative future concept.
- Symbolic AI represents knowledge explicitly via rules/logic (transparent but brittle); sub-symbolic AI learns patterns from data (generalizes well but often opaque) — neither is universally superior.
- This AI Notes curriculum focuses on symbolic, classical AI techniques; the dedicated ML/DL/NLP notes cover the sub-symbolic, learning-based side in depth.
Concept Check
- Why is even a highly capable modern LLM still classified as Artificial Narrow Intelligence?
- What's the key trade-off between symbolic and sub-symbolic AI approaches?
- Which of this curriculum's upcoming modules will lean primarily symbolic, and which existing notes folders cover the sub-symbolic side?
Next Chapter
→ Chapter 4: The Agent Framework
Jr Codex — 1-on-1 Personalized Coaching | Back to Module Index