Categorizing AI from a Computer Science Perspective

Timeframe: 1950’s–1980’s

Core idea

Hand‑crafted rules and logic. Systems do exactly what experts encode: if‑then rules, symbolic reasoning, decision trees, early planners.

Classic tech/achievements

• Early chess engines (Turing‑style algorithms, pre‑learning engines, even Deep Blue’s core search + handcrafted eval).
• Expert systems like MYCIN and DENDRAL.
• Classic game “AI”: Pac‑Man ghosts, early FPS bots, RTS scripts, finite‑state machines, behavior trees.
• Limits: No learning; brittle outside predefined situations; cannot improve from data.

Timeframe: 1990’s–2000’s

Core idea

Statistical learning on large datasets to predict or rank

Tech Examples

• Web searches, credit scoring, fraud detection, ad click‑through prediction.
• Classic Google Search core ranking and query understanding: heavy ML on logs and click data.
• Big‑data stacks: Hadoop/HDFS/MapReduce/YARN powering data lakes, offline feature generation, large‑scale ETL.
• “Big data” connection: The term big data became mainstream in this era, cheap storage + distributed compute + huge logs = feeding predictive models.

Limits: Great at prediction, weak at perception or open ended content generation; mostly works on structured or engineered features.

Timeframe: 2010’s

Core idea

Deep learning for perception, recognizing patterns in unstructured data: images, audio, video, sensor streams.

Tech Examples

• Image classification, object detection, speech recognition, medical imaging diagnostics. Enables applications like driverless cars and general “understand what I’m seeing” capabilities.
• Systems can recognize objects and patterns in unstructured data like images, audio, and video (distinguishing cats vs dogs, pedestrians vs road, cancerous vs benign lesions).

Limits: Perceives well but doesn’t inherently plan, act, or create still largely task‑specific.

Timeframe: Late 2010’s – 2020’s

Core idea

• Models that can generate new content by learning high‑dimensional distributions: text, images, code, audio, video, designs.
• Systems can generate new artifacts, text, code, images, video, designed by learning to “fill in the blank” in the media its working on.

Tech Examples

• Large language models, diffusion models, and other generative architectures.
• Modern “AI answers” in search: Gemini/LLM‑powered AI Mode summarizing and synthesizing results on top of classic retrieval.

Impact: Enables creative workflows, code generation, and AI designed artifacts.

Limits: Still often ungrounded, can hallucinate, can lack robust physical/causal understanding without embodiment.

Timeframe: 2010’s – 2020’s+

Core idea

Intelligence connected to bodies operating in the physical world, robots that sense, plan, and act in messy, dynamic environments.

Tech Examples

• Self‑driving cars as embodied systems: perception (Wave 3) + prediction (Wave 2) + generative planning (Wave 4) + low‑level control on real hardware.
• Manipulation robots, legged robots, drones, warehouse and logistics robots.

Characteristics: Must handle dynamics, energy limits, safety, and irreversible physical errors; mistakes are expensive and sometimes dangerous.

Limits: Hardware is hard: actuators, power, robustness and scaling beyond well structured environments is slow.

Timeframe: 2020’s+

Core idea

Systems with explicit self‑models that can imagine themselves in the future, reason over those imagined futures, and adapt, taking on “self‑awareness” and AGI (Artificial General Intelligence).

Tech Examples (early/prototypes)

• Self‑modeling robots that infer their own body plan from sensor data and learn to walk by simulating themselves.
• “Machine scientists” that infer physical laws from data and autonomously generate and test hypotheses.
• This is often associated with Artificial General Intelligence (AGI), which excites and alarms many observers.

Dr. Lipson expects this to be achievable and likely sooner than many think, seeing consciousness as an engineering problem rather than mysticism.

Open questions: Safety, governance, interpretability, and how to “steer” such systems rather than fear them.