Demystifying AI with Rachel Chung and Nikolas Wolfe
Episode Description (100 words)
In this episode of the Language Innovators podcast, hosts Dr. Linh Phung and Nikolas Wolfe are joined by Carnegie Mellon University professor Dr. Tingting (Rachel) Chung. Together, they demystify artificial intelligence, tracing its historical roots and local ties to Pittsburgh's pioneering work. The panel breaks down complex concepts like machine learning, neural networks, deep learning, and large language models into accessible math and simple Excel operations. They explore real-world applications in education, museum management, and healthcare while critically examining AI risks, human responsibility, and societal impacts.
Take a listen here or Spotify, Apple Podcasts, or any other major podcast platform.
Watch the episode on YouTube and subscribe to the Language Innovators channel for notifications of new episodes.
Episode Summary & Highlights
1. Introduction & Local Connections
Host Introductions: Dr. Linh Phung introduces herself as CEO and co-founder of Eduling, a task-based, AI-integrated platform focusing on developing language learners’ communicative competence. She introduces co-host Nik Wolfe, CTO of Eduling.
Guest Profile: Dr. Tingting Rachel Chung is a Professor of AI Education at Carnegie Mellon University (CMU) at the Heinz College of Information Systems and Public Policy. She holds an MS in information science and two PhDs, one in psychology and one in information science. She is an award-winning educator and author of AI the Magic Box and AI for Business.
Pittsburgh AI Legacy & Personal Ties: Linh and Rachel were former colleagues at Chatham University a decade ago. Nik’s wife studied under Rachel at Chatham, and Nik holds a master's degree from CMU's Language Technologies Institute (LTI). Rachel shares her journey moving from Taiwan to Pittsburgh 31 years ago on a scholarship at the University of Pittsburgh to study psychology, where she learned that early AI was shaped by psychologists. The hosts and guest highlight Pittsburgh's history as a birthplace of AI, emphasizing how cross-institutional collaboration between Pitt and CMU allowed early researchers to learn alongside Turing Award winners and Nobel laureates.
2. Demystifying AI Terminology & Historical Context
The "Three Dialects" of AI: Rachel explains that many terms (e.g., machine learning, neural networks, applied statistics, deep learning) represent the same underlying mathematical concepts, developed by researchers in different fields working in silos before the internet age. Example: Computer scientists used terms like "weights" and "bias," whereas statisticians in adjacent buildings referred to "beta coefficients" and "intercepts." Computer science was a very young field in the 1950s–1970s, so early AI concepts were studied primarily on paper and using algorithms rather than hardware execution.
Rebranding and "Deep Learning": For 20–30 years, researchers who claimed to study "neural networks" faced skepticism and funding rejections because early models didn't yield practical results. Rebranding the field to "deep learning,” notably championed by Nobel laureate Geoffrey Hinton, helped researchers avoid academic stigma and secure funding.
Naming Regrets: Rachel notes that early inventors did not anticipate a billion people using terms like "neural network" in everyday conversations. Nik points out parallel naming conventions in computer science, such as "dynamic programming," which was coined by researchers seeking sexy titles for grant funding rather than descriptive accuracy. Nik and Rachel agree that neural networks are often over-explained with brain analogies, whereas mathematically they closely mirror logistic regression or chained Excel formulas.
3. Defining Intelligence: Humans, Animals, Machines, and the "Eliza Effect"
Psychological Perspectives on Intelligence: Rachel defines intelligence through learning curves, the capacity to observe mistakes, adjust behavior, and reduce errors over time.
Domains of Intelligence: Intelligence spans multiple domains (navigation, physical coordination, and so on). While a Roomba navigating stairs displays spatial intelligence, physical tasks like loading a dishwasher remain major hurdles for robotics.
Language as Signal of Intelligence: Humans heavily weight language as the primary signal of intelligence because no other animals talk or write. The Eliza Effect: Rachel and Nik recall ELIZA, the 1960s chatbot created with ~300 lines of code in what’s known as the Eliza Effect. By merely repeating user inputs back as questions, users anthropomorphized the program. Nik explains that producing natural language serves as a "cheat code" that bypasses human critical filters, making a statistical system appear intelligent.
Analogies to Reality vs. Perception: Rachel explains that while cinematic CGI dinosaurs in the Jurassic Park movie look terrifyingly real, viewers know they are made of numbered pixels while technology like the airplane can be considered “super human” but hasn’t been mistaken as an "artificial bird." Rachel suggests that calling AI "software" or "a computer program" would reduce public confusion.
4. How Large Language Models (LLMs) & Prediction Models Work
Core Principles of Prediction: Large Language Models are statistical prediction models designed to output the most probable next word or token. Other prediction models underpin weather forecasting, stock price modeling, and medical research (e.g., predicting protein folding or cancer drug treatments).
From N-Gram Models to LLMs: Nik traces language modeling from 1940s N-gram models, which calculated basic word probabilities by dividing occurrence counts by total words, to modern LLMs. Modern LLMs extend these statistical counts over longer contexts so that the 100th word generated remains semantically coherent with the first token.
Explaining "Hallucinations": Rachel uses a classroom demonstration: Prompting "How are you ___?" usually yields "doing," "feeling," or "today." But if her target sentence is "How are you Connor?" (her son's name), standard high-probability guesses fail. Hallucination is simply the model guessing high-probability words based on training data when the true target token differs.
Illiterate Computers & Excel Math: Computers cannot read words; text must be converted into numerical vectors. Rachel explains that a single artificial neuron can be constructed using middle-school arithmetic inside Microsoft Excel. By iteratively updating weights based on errors, a model moves from 100% incorrect to 100% accurate in minutes.
5. Technical Mechanics: Objective Functions & Gradient Descent
The Optimization Process: Nik details how machine learning relies on an objective function that quantifies the difference between a model's guess and the correct target. Using gradient descent, the system calculates the slope of the error and tweaks internal weights step-by-step to "climb the hill" toward optimal performance.
Data Sources & Training: LLMs perform exceptionally well on standardized exams (SAT, GRE, ACT) because they have swallowed massive exabytes (zettabytes to be accurate) of scraped web data containing those exact answers and patterns. Nik highlights how queries like "What is a black hole?" often yield output identical to existing educational webpages in the training set.
6. Real-World Applications & Business Use Cases
Human Knowledge vs. Machine Execution: Linh notes that domain complexity (e.g., assessing language fluency or IELTS speaking rubrics) stems from human expertise accumulated over decades. Rachel emphasizes that while human knowledge takes years to acquire, AI can ingest criteria and written knowledge rapidly to perform repetitive analytical tasks.
Offloading Tedium in Business & Academia: Machines complement humans because they never tire of repetitive tasks. Examples that Rachel gives include her students’ work with administrative offices to automate complex, error-prone course scheduling spreadsheets. A former student implemented agentic AI tools at a museum with 300,000 artifacts. Annual re-cataloging that previously required three months of tedious manual work was completed in just two hours, allowing staff to focus on creative curatorial work.
7. Addressing AI Risks, Malicious Use, and Accountability
Tool vs. Weapon: Rachel argues that technology itself is neutral; a standard car can be used as a weapon if driven maliciously. Society manages existential threats (like nuclear capabilities) through regulatory oversight and human governance.
Attributing Blame to Humans, Not Code: Nik criticizes sensationalist narratives claiming AI acts on its own volition. When AI systems generate malware or target military objectives, accountability rests entirely with the human engineers, corporations, and operators deploying the models on powered hardware. Placing blame on the "AI" shifts focus away from bad actors and irresponsible deployment.
8. Educational Resources & Closing Remarks
Featured Books by Dr. Rachel Chung: AI the Magic Box: A beginner-friendly book co-authored with former students that uses a story about children training an AI to find a family puppy to teach neural network math without complex Greek notation. AI for Business: A college textbook used at CMU featuring real-world business scenarios (e.g., Zillow's Zestimate) to guide students through building and computing neural networks by hand.
