Lecture 18: NLP, Linguistics, and Philosophy

What this lecture covers

This is the final lecture of the course, and instead of introducing new technical methods it steps back to reflect on the field as a whole. Christopher Manning reviews the major ideas the course built up, from word vectors through Transformers to pre-training and post-training, then discusses open problems such as generalization, interpretability, and multilingual coverage. The lecture then moves into more philosophical territory: the century-old debate between symbolic and neural ("cybernetics") approaches to AI, the role of language in human thought, and competing theories of linguistic meaning. It closes with a discussion of societal risks from AI, including job disruption, concentration of power, and disinformation.

After watching, you should be able to summarize the arc of ideas covered across the course, describe several concrete limitations of current large language models, explain the historical distinction between symbolic AI and neural/cybernetic approaches, and describe two competing theories of what "meaning" is for a word.

Key ideas

• Distributional semantics: the idea that a word's meaning can be captured by the contexts it appears in ("you shall know a word by the company it keeps") underlies word vectors and everything built on them, including Transformers.
• Generalization vs. memorization: large pre-trained models often succeed by having seen almost every pattern before; on tasks requiring learning from very limited examples, older architectures like LSTMs can generalize better than Transformers.
• Benchmark contamination: because pre-training data now includes much of the public web, models may be memorizing benchmark answers rather than solving problems fairly, which undermines claimed performance gains.
• The physical symbol system hypothesis: Newell and Simon's claim that a symbol-manipulating system is necessary and sufficient for general intelligence, the founding assumption of classical, symbolic AI.
• Cybernetics: the rival, earlier tradition (Norbert Wiener, Frank Rosenblatt) focused on control, communication, and learning systems, which neural networks are a continuation of, not the symbolic AI tradition.
• Denotational vs. distributional/use theories of meaning: classical linguistic semantics ties a word's meaning to what it refers to in the world; an alternative view (traced to Wittgenstein and J.R. Firth) ties meaning to how a word is used and connected to other things.
• Present-day AI harms vs. existential risk: the lecture argues that concentrated economic power, disinformation, bias, and hallucination are more pressing concerns than speculative existential risk from superintelligence.

Walkthrough

Reviewing the course's arc of ideas (1:05)

The lecture opens with a recap of the course's progression: word vectors, feedforward networks, sequence models (RNNs, LSTMs), the Transformer, and the modern pre-training-then-post-training pipeline. Manning highlights distributional semantics as the thread connecting all of these, from early word vectors to how Transformers are trained today, and notes the empirical pattern that performance has scaled roughly linearly with data, compute, and model size, even though this does not mean every NLP problem is solved.

Where models still struggle: generalization and interpretability (4:12)

Manning raises open problems, starting with the question of whether large models are really learning to generalize or mostly memorizing patterns they have seen before, describing them as closer to "a talking encyclopedia" than a generally intelligent reasoner. He points to research showing that on data generated by a finite automaton, LSTMs can generalize from limited exposure better than Transformers, which need far more data to learn the same pattern, in contrast to how quickly humans can learn a new skill from a single demonstration. He also discusses interpretability research, citing an early example from Andrej Karpathy's work on LSTM neurons tracking line length, as part of ongoing efforts (now including mechanistic interpretability work on Transformers) to understand what is happening inside these largely black-box models.

Multilingual gaps and benchmark contamination (8:16)

The lecture discusses unequal progress across languages: GPT-4 improves performance over GPT-3.5 for many languages, including some large ones like Punjabi, Marathi, and Telugu, but the vast majority of the world's languages have far too little written data to build comparable models, and many are primarily oral and at risk of extinction. Manning then raises benchmark contamination as a growing concern, using an example where a model solved all pre-2021 coding-benchmark problems but none of the recent ones, suggesting memorization rather than genuine problem-solving. He briefly touches on domain-specific risks in biomedical and legal NLP, citing a study from Stanford's RegLab finding a roughly one-in-six hallucination rate in legal AI answers, and notes that NLP systems still carry cultural and social biases.

How good are today's language models, really? (16:31)

Manning shares a live example of asking GPT-4o to write a sonnet about Transformer architecture where every line begins with "T," noting that despite years of NLP experience, the result still seems close to "magic." He then cites a study of Boston Consulting Group consultants: those using GPT-4 completed 12% more tasks, worked 25% faster, and produced work judged 40% higher quality, with the biggest gains for lower-skilled workers. In contrast, he cites a study finding GPT-4's creative fiction was rated three to ten times worse than professional New Yorker writers, and quotes a skeptical Financial Times piece questioning whether generative AI investment is outrunning its actual capabilities.

Symbolic AI versus the neural, cybernetic tradition (25:42)

Manning traces the history of two competing AI traditions: symbolic AI, founded by John McCarthy and formalized by Newell and Simon's "physical symbol system hypothesis" (that manipulating symbols is necessary and sufficient for general intelligence), versus cybernetics, the earlier tradition of Norbert Wiener and Frank Rosenblatt's perceptron, concerned with control, communication, and learning. He argues that while human language is unmistakably a symbolic system, that does not mean the brain (or an NLP system) must itself be a symbol-manipulating processor; neural networks, in the cybernetics lineage, can plausibly model language processing without being symbolic in that sense. He suggests linguistic concepts like compositionality and reference remain useful for framing questions about what NLP systems should be able to do, even if the underlying computation is not symbolic.

Language, thought, and theories of meaning (39:00)

Drawing on Wilhelm von Humboldt, Daniel Dennett, and the classic distinction between symbolic and use-based theories of meaning, Manning explores the role language plays in human cognition, including Dennett's four grades of intelligence (Darwinian, Skinnerian, Popperian, and Gregorian), with human language proposed as the paradigm example of a "Gregorian" thinking tool. He then contrasts denotational semantics, the traditional view that a word's meaning is what it refers to in the world (associated with Alfred Tarski and formalized for natural language by Richard Montague), with a use theory of meaning drawn from Wittgenstein and J.R. Firth, where meaning comes from how a word connects to other words and experiences. Manning illustrates this with the word "shehnai," showing how meaning can be built up gradually through pictures, descriptions, or textual context, arguing that meaning is gradient rather than binary.

AI's future risks: jobs, power, and disinformation (1:03:22)

The lecture closes with a discussion of societal risks. Manning compares historical fears of automation-driven unemployment (citing 1928 and 1961 news articles) to today's concerns, arguing that job loss fears have repeatedly not materialized as predicted, but that concentration of wealth and power among a small number of technology companies is a more credible present-day risk. He expresses skepticism about existential-risk narratives around superintelligence, citing critics like François Chollet and Joelle Pineau, and argues that framing focused on existential risk can distract from concrete, current harms: bias, disinformation, hallucination, copyright violations, and surveillance. He closes by quoting Carl Sagan's warning in "The Demon-Haunted World" about a future where technological power concentrates in few hands while the public loses the ability to question it, framing education as a safeguard against that outcome.

Before you watch

• This lecture assumes familiarity with the course's full technical arc: word vectors, RNNs and LSTMs, Transformers, and pre-training/post-training, so it is best watched after completing the earlier lectures.
• No new mathematical content is introduced; the lecture is discursive and benefits from having some grounding in the course's discussion of evaluation and reasoning from earlier weeks.

Check your understanding

1. Why might a Transformer generalize worse than an LSTM on a task with very limited training examples?
2. What does benchmark contamination mean, and why does it undermine claims about model performance?
3. What is the difference between the "physical symbol system hypothesis" and the cybernetics tradition that neural networks descend from?
4. How does a use theory of meaning differ from denotational semantics, using the "shehnai" example?
5. According to the lecture, why does the speaker consider concentration of power a more pressing risk than existential risk from superintelligent AI?