What is Artificial Intelligence?

• The dream: building machines that can think, learn, and act.
• AI: making computers do things that–when done by humans–require intelligence.
• Two traditions:
  • Symbolic AI (logic, reasoning, rules)
  • Connectionist AI (neural networks, learning from data)

AI Conceptual Graphic

Defining “Intelligence” in AI

• Human intelligence includes:
  • Perception (seeing, hearing, sensing).
  • Reasoning (logic, planning, problem-solving).
  • Learning (improving from experience).
  • Action (making decisions in the world).
• AI systems often focus on narrow slices of intelligence.

Symbolic vs. Connectionist AI

• Symbolic AI (1950s–1980s):
  • Uses rules, logic, and expert knowledge.
  • Strength: clear reasoning and explanations.
  • Weakness: brittle, hard to adapt.
• Connectionist AI (1980s–today):
  • Uses neural networks and statistics.
  • Strength: adapts well to messy data.
  • Weakness: hard to explain decisions.

Early Artificial Intelligence (Pre-1950s)

• Philosophical roots: Descartes, Leibniz, Hobbes → thinking as computation.
• 1800s: Babbage & Lovelace imagine mechanical computation.
• 1940s: Alan Turing formalizes the concept of a universal machine.
• Turing Test (1950): “Can machines think?”

Alan Turing

Turing and Computability

• Alan Turing (1912–1954):
  • Formalized computability with the Turing Machine.
  • Proposed the Imitation Game (Turing Test).
• His question reframed AI research: not “can machines think?” but “can they act as if they think?”
• Foundation of modern computer science + AI.

The Turing Test

What is the Turing Test?

• Proposed by Alan Turing in 1950 (“Computing Machinery and Intelligence”)
• Seeks to answer: “Can machines think?”
• Uses an imitation game instead of defining “thinking” directly

How It Works

• Participants: A human judge, a human, and a machine
• Interaction: Text-only conversation (no voice/appearance)
• Goal: If the judge cannot tell which is the machine, the machine “passes”

Significance

• Foundational idea in artificial intelligence
• Shifts focus to observable behavior
• Sparked debates on intelligence, language, and consciousness

Criticism

• Tests mimicry, not real understanding

• Chinese Room Argument (John Searle): passing is not equal to comprehension

  • The Chinese Room Argument is a thought experiment against the Turing Test

  • Imagine a person in a room with a rulebook for manipulating Chinese symbols

  • They receive Chinese input, use the rules, and return correct Chinese output

  • To outsiders, it looks like they understand Chinese

  • But inside, the person doesn’t understand the language–just symbol manipulation

  • The Chinese Room Argument suggests that passing the Turing Test is not equivalent to real understanding or consciousness

• Modern AI (like ChatGPT) challenges the test’s adequacy

• Play the turing test online against a human or LLM

The Birth of AI (1950s–1970s)

• 1956: Dartmouth Conference – the official birth of AI.
• Optimism: “Machines will rival human intelligence within a generation.”
• Achievements:
  • Early logic programs (Newell & Simon’s Logic Theorist).
  • Game-playing AI (chess, checkers).
  • Natural language attempts (ELIZA, SHRDLU).

Dartmouth AI Proposal (1955)

Early Success Stories

• Logic Theorist (1956): proved 38 of 52 theorems from Principia Mathematica.
• General Problem Solver (1959): aimed for universal reasoning.
• Game-playing programs: checkers programs learned strategies better than some humans.
• Set the tone: AI could match or exceed humans in narrow tasks.

Limits of Early AI

• Overestimated progress:
  • Language understanding far harder than expected.
  • Commonsense reasoning remained elusive.
• Hardware limited:
  • Very small memory (kilobytes!).
  • Slow processors.
• Early optimism gave way to frustration.

Image from Dalle-3

The First AI Winter (1970s)

• Bold promises → disappointing results.
• Computers were too slow, memory too small.
• Governments cut funding.
• Overhype + underdelivery = AI winter.

Why the First Winter Happened

• Natural language understanding proved intractable.
• Commonsense reasoning: trivial for humans, impossible for machines.
• Cost of computing was high; funding dwindled.
• Media hype made failures more visible.

Consequences of the Winter

• Research funding shifted to other areas.
• AI became an “academic backwater.”
• Yet… important foundations were laid:
  • Search algorithms.
  • Knowledge representation.
  • Early neural network concepts.

Image from Dalle-3

Thawing from the First Winter (1980s)

• Japan’s “Fifth Generation Computer Project.”

  • Timeline: 1982–1992
  • Goal: Leap beyond conventional computing -> AI-driven, knowledge-based systems
  • Massively parallel architectures (Parallel Inference Machines)
  • Logic programming (Prolog) as foundation
  • Natural language processing, expert systems, theorem proving
  • Advanced parallel computing research
  • New logic programming tools
  • Trained a generation of AI researchers
  • Fell short of ambitious goals
  • Outpaced by conventional computing advances
  • Limited commercial impact, but lasting research influence

• Renewed funding & optimism.

• Expert Systems emerge: rule-based decision-making (e.g., medical diagnosis).

Backward Chaining (Expert Systems)

Expert Systems Explained

• Mimic human experts by encoding if-then rules.
• Example: MYCIN (1970s) for diagnosing blood infections.
• Advantages: explainable reasoning.
• Limitations: hard to maintain, rules explode in number.

The Fifth Generation Project

• Japan (1982–1992): aimed for computers with human-like reasoning.
• Focus on logic programming (Prolog).
• Sparked global competition (U.S., Europe).
• Ultimately fell short, but drove research and funding.

Image from Dalle-3

The Second AI Winter (1990s)

• Expert systems proved expensive & brittle.
• Failed to scale → loss of confidence.
• Another crash in funding and interest.

Why Expert Systems Failed

• Knowledge engineering bottleneck:
  • Extracting rules from experts was slow.
  • Systems could not adapt to new situations.
• Maintenance nightmare: rules conflicted.
• Costs outweighed benefits.

Shift in Focus During the 1990s

• Rise of statistical approaches: probability, statistics, data-driven learning.
• AI blended with computer science fields:
  • Databases.
  • Information retrieval.
  • Robotics.
• AI rebranded as “machine learning” in many contexts.

Thawing Again (Late 1990s–2000s)

• New hope: statistical methods & machine learning.
• Internet explosion = big data.
• Cheaper, faster hardware.
• IBM’s Deep Blue defeats Kasparov (1997).

IBM Deep Blue

Deep Blue’s Victory

• IBM Deep Blue vs. Garry Kasparov (1997).
• Specialized hardware + search + heuristics.
• Landmark: machine beats world chess champion.
• Symbolized AI as practical and competitive.

Big Data Revolution

• Internet users generated huge amounts of data.
• AI shifted from hand-coded rules to pattern-finding in data.
• Statistical learning → recommendation systems, spam detection.
• More data → better performance.

AI in the Early 2000s

• Search engines revolutionize knowledge access.
• Speech recognition improves.
• AI used in spam filters, recommendation systems, and everyday apps.

Early Google Logo (1998)

Search and Recommendation

• Google’s PageRank transformed search (late 1990s).
• Amazon pioneered personalized recommendations.
• AI became embedded in daily life–often invisibly.

Speech and Language in the 2000s

• Hidden Markov Models (HMMs) drove progress.
• Systems like Dragon NaturallySpeaking became usable.
• Foundation for voice assistants in the 2010s.

Deep Learning (2010s)

• Inspired by the brain → neural networks with many layers.
• Requires massive data + GPUs.
• Achievements:
  • ImageNet breakthrough (2012).
  • Near-human speech recognition.

Neural Network

The ImageNet Moment

• 2012: AlexNet by Hinton’s team.
• Deep CNNs cut image classification errors nearly in half.
• Catalyzed explosion of deep learning research.

The Digit Predictor

• One simple predictor that we can easily train with a neural network is a digit predictor. Let’s see one in action.

Deep Learning Applications

• Vision: face recognition, medical imaging.
• Language: machine translation, speech recognition.
• Games: AlphaGo (2016) beat Go champion Lee Sedol.
• Became state-of-the-art in many fields.

Self-Driving Cars

• Early DARPA challenges (2004–2005).
• Google’s self-driving car (Waymo).
• Symbol of AI entering the physical world.

Waymo Self-Driving Car

DARPA Grand Challenges

• 2004: most teams failed; none finished.
• 2005: Stanford’s “Stanley” won by completing the desert course.
• Sparked commercial and academic self-driving research.

From DARPA to Waymo

• Google project launched in 2009.
• Progressed from retrofitted Priuses to dedicated self-driving cars.
• Challenges: safety, regulation, ethics.
• Still an active frontier today.

Transformers and GPT

• 2017: Transformers revolutionize AI.
• GPT (Generative Pretrained Transformer): AI that can write, converse, create.
• A leap in natural language understanding.

Attention (Transformers)

Why Transformers Matter

• Key innovation: attention mechanism.
• Handles sequences efficiently.
• Enables training on massive text corpora.
• Foundation of modern NLP breakthroughs.

GPT and Generative AI

• GPT-2 (2019), GPT-3 (2020), GPT-4 (2023).
• Can generate essays, code, stories, poetry.
• Raises questions about creativity and authorship.

LLMs in the 2020s

• Chatbots like ChatGPT become mainstream.
• AI used in art, code, medicine, education.
• Raises questions: ethics, bias, jobs, creativity.

ChatGPT Logo

Ethical Challenges of LLMs

• Biases in training data → biased outputs.
• Risks of misinformation and “hallucination.”
• Intellectual property questions: who owns AI-generated text?
• Need for transparency and accountability.

AI in Society Today

• Ubiquitous in:
  • Search and recommendation.
  • Healthcare and drug discovery.
  • Creative tools (art, music, writing).
• Balancing opportunity and risk is key.

The Future of AI

• Superintelligence? Hype or possibility?
• Collaboration between humans and AI.
• Key questions:
  • How do we align AI with human values?
  • How do we regulate and govern AI?
  • What role will YOU play in shaping AI’s future?

Future AI

Possible Futures

• Optimistic: AI accelerates science, cures diseases, solves climate change.
• Pessimistic: job loss, surveillance, misuse in war.
• Balanced: humans + AI collaborate, but with regulation.

Your Role in the Future

• Every user, policymaker, engineer contributes to AI’s trajectory.
• Think critically:
  • Where should AI be trusted?
  • Where should humans remain in control?
• The future is not predetermined–it’s shaped by choices now.

Discussion

• What excites you most about AI?
• What concerns you the most?
• Where do you see yourself in this AI-powered world?