Download AUC FINAL PAPER 2026 FULL Q&A STUDY GUIDE GRADED A+ and more Exams Bioethics in PDF only on Docsity!
AUC FINAL PAPER 2026 FULL Q&A STUDY
GUIDE GRADED A+
◉Why did McCarthy coin the term "artificial intelligence"? Answer: To label the field and distinguish it from cybernetics. ◉What was the bold claim in the Dartmouth proposal? Answer: That every aspect of learning/intelligence could be described precisely enough for a machine to simulate. ◉What was a key outcome of Dartmouth despite the messy workshop itself? Answer: The field got named and core goals/networks of pioneers formed. ◉What recurring problem did early AI predictions reveal? Answer: AI progress was overestimated, raising questions about what "full intelligence" means. ◉What does it mean that "intelligence" is a "suitcase word"? Answer: It bundles many different abilities (thinking, perception, language, etc.) under one label.
◉What are the two broad goals of AI in the reading? Answer: Scientific: model natural intelligence; Practical: build systems that perform tasks well, even if not human-like. ◉What is meant by an "anarchy of methods" in AI? Answer: Competing approaches (logic, statistics, brain-inspired) with ongoing disagreement. ◉How is deep learning related to AI? Answer: Deep learning is one approach inside machine learning, which is inside AI. ◉What is the key historical split highlighted? Answer: Symbolic AI vs subsymbolic AI. ◉What is Symbolic AI in one sentence? Answer: Using human- readable symbols and explicit rules to represent knowledge and reason. ◉What is a classic example of early symbolic AI? Answer: The General Problem Solver (Newell & Simon), solving puzzles via states, operators, and rules. ◉What philosophical limitation of symbolic AI is emphasized? Answer: The computer manipulates symbols without understanding their meaning.
◉What is the interpretability gap for neural-style systems? Answer: Their decisions come from many numeric weights, not clear human rules. ◉What did Minsky & Papert argue in 1969 (big takeaway)? Answer: Single-layer perceptrons are severely limited and weren't likely to scale easily. ◉Why didn't multilayer neural nets take off then (per this chapter)? Answer: There wasn't a reliable, general training method for many layers at the time. ◉What is an "AI winter"? Answer: A bust period when promised breakthroughs don't arrive and funding/interest collapses. ◉What does "Easy things are hard" mean in AI? Answer: Tasks humans/children do easily (language, perception, learning from few examples) were surprisingly difficult for AI. ◉In the "sandwich model," what does each ingredient have? Answer: A weight showing how much it contributes to the score ("deliciousness").
◉How does the simplest neural net compute "deliciousness"? Answer: Add up (ingredient present? × ingredient weight) across all ingredients. ◉What's the core idea behind weights? Answer: Not all inputs matter equally—good ingredients should have higher weights; bad ones low/negative. ◉What kind of patterns can a pure "add-it-up" model learn? Answer: Simple additive patterns like "mud + eggshells = bad" or "peanut butter + marshmallow = good." ◉When does the simple additive model work best? Answer: When overall quality is mostly the sum of the parts. ◉What's the big limitation of the additive model (p. 68)? Answer: It can't handle interactions where combinations matter. ◉What "bug" can happen with the additive model? Answer: Enough good ingredients can "cancel out" a bad one and produce an unrealistically high score. ◉What's the general fix for the "cancel-out" bug? Answer: Add more structure than just summing—use additional layers/features.
◉What is the "punisher cell" idea? Answer: A hidden cell that strongly lowers the score if a forbidden ingredient/pattern appears. ◉How does a "punisher cell" help with the sandwich bug? Answer: It prevents many good ingredients from overpowering a truly bad ingredient/pattern. ◉What is the "deli sandwich cell" idea? Answer: A hidden cell that detects a good classic combination and boosts the score. ◉Why do neural nets need nonlinearity (activation functions)? Answer: To represent real interactions/threshold behaviors instead of only smooth addition. ◉What does an activation function do (simple phrasing)? Answer: It makes a unit "turn on" only when inputs cross a threshold. ◉How does nonlinearity enable "AND-like" logic? Answer: It can model "only if A and B are present," not just A or B separately. ◉Do you usually set neural-net weights by hand? Answer: No— weights usually start random and are learned from data.
◉What kind of data does training require here? Answer: Labeled examples from humans (e.g., "sandwich judges" providing correct ratings). ◉What is the basic training loop? Answer: Predict → compare to label → adjust weights slightly to reduce error → repeat many times. ◉What is class imbalance? Answer: When one class is far more common (e.g., 1 in 1000 good sandwiches). ◉Why can class imbalance make "accuracy" misleading? Answer: A model can get very high accuracy by always guessing the majority class, but be useless. ◉Name a few real-world places class imbalance shows up. Answer: Fraud detection, medical diagnosis, customer churn, hacking detection, rare events (e.g., solar flares). ◉What's a common fix for class imbalance (as described)? Answer: Make the training set more balanced (or otherwise prevent the "easy shortcut"). ◉Why are trained neural nets hard to interpret (pp. 76-79)? Answer: "Rules" are often distributed across many cells; individual neurons can be mysterious.
◉Why is LLM behavior hard to fully explain? Answer: The "reasoning" is distributed across billions of learned parameters, so even experts can't trace every step. ◉How do LLMs represent words/tokens internally? Answer: As vectors (long lists of numbers), not as letters. ◉What is the intuition behind word vectors? Answer: They place tokens in a high-dimensional "meaning space" where similar meanings are close. ◉What classic system popularized the vector idea (2013)? Answer: word2vec (learned by predicting words from context). ◉What surprising property can vectors show? Answer: Vector arithmetic can encode relationships/analogies (e.g., big→biggest parallels small→smallest). ◉Why can bias show up in embeddings? Answer: Vectors reflect patterns in human language, including stereotypes. ◉What is polysemy/homonymy and why does it matter? Answer: Words have multiple meanings (bank money vs bank river), so meaning depends on context.
◉How do LLMs handle words with multiple meanings? Answer: They produce different contextual vectors for the same word depending on surrounding text. ◉What is a transformer "layer" doing in a GPT-style model? Answer: It updates token vectors, producing new "hidden states" layer by layer. ◉What are "hidden states"? Answer: The evolving set of vectors for each token after each layer's update. ◉Big-picture: what do early vs later layers tend to capture? Answer: Early layers handle syntax/structure; later layers build higher-level situation/meaning representations. ◉What are the two main sub-steps inside a transformer layer? Answer: (1) Attention and (2) a feed-forward network (FFN). ◉What is attention (core idea)? Answer: Each token selectively "looks at" other tokens to pull in relevant context. ◉What are queries and keys used for in attention? Answer: They're compared (dot products) to decide which tokens connect strongly.
◉What's the "division of labor" idea between attention and FFNs? Answer: Attention retrieves from the prompt/context; FFNs retrieve from learned weights (stored knowledge). ◉What makes training LLMs "self-supervised"? Answer: The next token in real text provides the label automatically—no hand labeling needed. ◉What is backpropagation used for in training? Answer: To adjust weights using gradients so the model reduces prediction error. ◉What is the main "scale" claim (power laws)? Answer: Performance improves predictably as model size, data, and compute increase together. ◉What's an example of a debated "emergent" capability mentioned? Answer: Theory of Mind-style tasks—improves with scale but can be fragile/contested. ◉What is the "stochastic parrots" vs "understanding" dispute? Answer: Whether models truly understand meaning or mainly produce text from statistical patterns.
◉What stance do the authors emphasize in that dispute? Answer: Focus on empirical performance with good controls (avoid confounds and data leakage). ◉Why might next-token prediction work so well, according to the closing idea? Answer: Language is predictable and reflects real- world structure; prediction may also be central to human cognition. ◉ho wrote "Could a Large Language Model be Conscious?" and where/when was it published? Answer: David J. Chalmers; Boston Review (Aug. 9, 2023), adapted from a NeurIPS talk (Nov. 28, 2022). ◉What event sparked the public debate Chalmers responds to? Answer: Blake Lemoine claimed Google's LaMDA was sentient; Google denied it. ◉What is Chalmers's core question? Answer: What counts as evidence for (and against) consciousness in an LLM? ◉What does Chalmers mean by "consciousness" here? Answer: Subjective experience / sentience—there is "something it's like" to be the system.
◉What is X3 and why is it only limited evidence? Answer: Conversational ability; it can mimic thinking, but models still show "giveaways." ◉What is X4 and why might it support the case? Answer: General intelligence/broad competence; if a non-AI showed it, we'd take it seriously. ◉Chalmers's bottom line on current LLMs (end of pro-evidence section)? Answer: Not strong evidence they're conscious, but enough to take the question seriously. ◉What is Chalmers's "disciplined format" for evidence AGAINST LLM consciousness? Answer: Find a missing X: LLMs lack X, and if a system lacks X it's probably not conscious. ◉What is the "biology objection," and Chalmers's response? Answer: Claim: consciousness requires biology; response: that's "biological chauvinism"—organization may matter more than substrate. ◉What is the "senses/embodiment" objection? Answer: LLMs lack bodies/senses, so they likely lack sensory/bodily consciousness and may lack grounding.
◉Does Chalmers think embodiment is strictly required for any consciousness? Answer: No; a "pure thinker" could still have conscious thought, though limited. ◉What does Chalmers suggest as a more direct route to grounding? Answer: Multimodal/embodied systems—especially perception- language-action (even in virtual worlds). ◉What is the "stochastic parrots/world-model" objection? Answer: LLMs just imitate text without modeling the world/self, so they lack understanding (and maybe consciousness). ◉Chalmers's key reply about training objective vs internal mechanism? Answer: Even if trained to predict text, the best way may involve building internal world/self models (empirical question). ◉What is the "recurrent processing" objection? Answer: Many consciousness theories stress recurrence/feedback and persistent state; transformers are mostly feedforward. ◉What is the "global workspace" objection? Answer: Consciousness may require a workspace that broadcasts info across subsystems; LLMs don't obviously have this.
◉What is the core question Winner asks? Answer: Whether technologies themselves can have political properties, not just the people using them. ◉What is the common intuition Winner argues against? Answer: "People have politics, not things" (artifacts are supposedly neutral tools). ◉What is the "social determination of technology" view? Answer: Technology's effects come from the social/economic system it's embedded in, not the artifact itself. ◉Why does Winner think "social forces only" is incomplete? Answer: It can miss how technical design and system momentum shape outcomes even without intent. ◉What is Winner's basic answer to "Do technologies have politics?" Answer: Yes—either through design choices that build in social order or because some technologies strongly require certain political arrangements. ◉What are the two ways artifacts can have politics? Answer: (1) Politics built into design/arrangement; (2) technologies that are inherently (or strongly) tied to certain political relationships.
◉What is Category 1 ("politics through design") in one line? Answer: Design/arrangement makes some outcomes easy and others difficult, creating effects "prior to use." ◉What does Winner mean by consequences "prior to use"? Answer: The political effects are baked into planning/design (not just good/bad uses after deployment). ◉What is the Robert Moses bridges example meant to show? Answer: Infrastructure design (low bridges blocking buses) can systematically exclude groups without an explicit law. ◉What is the "tomato harvester" example meant to show? Answer: "Efficient" innovations can still drive consolidation and job loss— politics can be built into the technical path and funding priorities. ◉Where do political choices happen in Category 1 cases? Answer: First: whether to adopt a technology at all; second: design details after adoption (routes, access, add-ons, system architecture). ◉What is Winner's main lesson from Part I? Answer: Technologies are ways of building order, and early design choices get locked in like major laws.
