9. Introduction to Gen AI

Summary

The lecture introduces generative AI as a powerful tool for research and everyday work—not a technology that independently replaces human expertise. AI can help researchers identify promising directions, explore ideas more quickly, and support experiments, but people still need to validate results and carry out practical work. The speaker cautions against treating AI as either a magic solution or a substitute for researchers.

AI’s history and changing approaches

• AI predates ChatGPT by many decades. In 1950, Alan Turing asked whether machines can think and proposed a behavioral test: whether a person could distinguish a machine from a human through interaction.
• The term artificial intelligence was coined at the 1956 Dartmouth conference.
• Early examples included a checkers program that learned through repeated play—an early form of reinforcement learning—and ELIZA, a conversational program.
• AI has gone through repeated cycles of high expectations, disappointing results, and reduced investment, often called AI winters. Neural-network limitations, scaling problems, and brittle systems contributed to earlier downturns. Later progress in deep learning helped revive the field.
• AI is not a single technology. The lecture surveys several families:
  • Rule-based or expert systems: Apply rules derived from human expertise.
  • Search: Explore possible outcomes, as in game-playing systems.
  • Logic and reasoning approaches: Combine rules and search.
  • Deep learning: Learns patterns from large datasets, with applications in vision, speech, and language.
  • Recommendation systems: Rank or suggest content, products, or connections.
  • Reinforcement learning: Learns through actions and feedback.
  • Hybrid systems: Combine methods, as illustrated by the lecture’s discussion of AlphaFold.
• Many valuable AI applications predate generative AI. Researchers should choose a method suited to the problem rather than assume an LLM is always the right tool.

How the lecture explains language models

The speaker uses next-token prediction to give an intuitive account of large language models (LLMs):

1. Given a prompt such as “I watched a…”, a model assigns different probabilities to possible next tokens. “Movie” or “game” may be more likely than an unrelated word.
2. A more specific prompt—such as “In the new theater, I watched a…”—changes the likely continuations. More context often narrows the possibilities and can improve relevance.
3. The model generates text by repeatedly selecting a next token, adding it to the sequence, and predicting again. This is called autoregressive generation.
4. Because the model learns patterns from large collections of text, it can produce grammatically plausible sentences and reflect patterns that resemble world knowledge or arithmetic. The lecture emphasizes that this does not necessarily mean the model understands those things as a person would.

The speaker introduces four LLM terms:

• Vocabulary or token size: The set of tokens the model can use. Tokens may be words, subwords, or other text units.
• Context size: How many tokens from the prompt or preceding conversation the model can take into account.
• Parameter count: The learned values used to represent patterns in the data.
• Embedding-vector size: The dimensionality of a vector representation used to encode information.

The lecture explains the model as learning a very large set of conditional probabilities: which token is likely given a context. Since storing every possible context and its probabilities explicitly would be impractical, a neural network compresses patterns into vector representations. This is a conceptual explanation, not a claim that the model stores a literal table. Similar contexts may be represented in overlapping ways, which can contribute to errors and variability.

Alignment and retrieval

• Alignment adjusts model behavior toward preferred outputs, often using human feedback or AI-generated preferences. In the lecture’s example, feedback can increase the likelihood of a correct answer and reduce the likelihood of incorrect ones.
• Alignment can also introduce or amplify bias if preferences or data are flawed, or if the model is over-adjusted.
• Retrieval-augmented generation (RAG) adds retrieved information—such as material from a web search or internal repository—to the prompt. The added context can make an answer more focused and improve the chance of a correct response, though it does not guarantee one.

Limitations and responsible use

• Hallucinations: When a model lacks strong evidence for a specific prompt, it may assemble plausible-sounding details that are false. The lecture’s examples include fabricated statistics and a nonexistent economics paper described with convincing details.
• Hallucination reflects a tension: models need to generalize beyond exact examples to be useful, but that same capacity can produce unsupported answers. More context, retrieval, and verification can reduce the risk, but not eliminate it.
• Misinformation: A model may repeat misleading information from a source it treats as credible. An answer can therefore be wrong because its supporting information is misleading, not only because the model invented it.
• Bias and representation: Training data and alignment choices can produce biased outputs or narrow cultural portrayals. The lecture highlights examples involving race and generated images of Indian people, homes, and women, and argues that models need more culturally relevant data and evaluation.
• Unequal performance: The speaker cites research suggesting that chatbots may give less accurate answers to users who communicate in less standard English or are otherwise underrepresented.
• Sensitive uses: The lecture warns against relying on general-purpose chatbots for mental-health support or personal and medical advice without appropriate expertise or supervision. Models may affirm a user’s framing rather than provide sound guidance.
• Deepfakes and copyright: Generative tools can create convincing images or imitations. Users should consider the potential for misuse and avoid inadvertently reproducing copyrighted material.
• Human verification remains essential: Use domain knowledge, advisers, or independent research to check outputs. AI can narrow the search and reduce effort, but its answers should not be accepted as automatically trustworthy.
• Research and productivity: Generative AI can contribute to research and help people work faster, but humans still need to assess, combine, and validate its outputs. The speaker’s central recommendation is to learn to use AI effectively while retaining expert oversight.

Speakers and sources mentioned

• Speaker: An unnamed lecturer presenting the introduction.
• Historical figures and examples discussed: Alan Turing; John McCarthy; Marvin Minsky; Claude Shannon; Nathaniel Rochester; Arthur Samuel; Geoffrey Hinton; John Hopfield; Garry Kasparov; ELIZA; and IBM’s Watson/Jeopardy system.
• Organizations, systems, and AI examples mentioned: Dartmouth conference; DeepMind; OpenAI; ChatGPT/GPT; AlphaGo; AlphaFold; Stable Diffusion; Midjourney; Google Gemini; Siri; Alexa; Google Home; Google Maps; and Google Search.
• Other sources or examples referenced: Sam Altman’s ChatGPT launch announcement; a historical newspaper headline from 1958; Claude Shannon’s 1948 work; a Nature special issue on AI and scientific discovery; Stanford studies concerning chatbot advice and accuracy; World Bank statistics used for comparison; and a Reddit prank post about putting glue on pizza. The lecture also refers to a mathematician’s report of using Claude and Gemini on a proof, but the name is unclear in the subtitles.