Demystifying AI: slides and notes

I promised myself I would start with these words: do not be afraid. We hear so many claims about AI, and I know those words can draw a few uncomfortable chuckles. But this is a big part of why I wanted to give these lectures.

If nothing else happens over the course of these six lectures, I want you to walk away feeling less afraid and more able to say: we can handle this. We should be involved. We can help determine how this turns out. That is the spirit I want to carry through everything we explore together.

I started programming in second grade and took to it like a fish to water. These early computers, the VIC-20, Atari, Commodore 64, Amiga, and TRS-80, bring me right back to that excitement. When people in the room recognized them, I felt a little less old. These were things I could explore, program, and make do something new.

I bought into the whole vision: intelligent spaceships we could talk to, androids we could hang out with, and robots that could help us fight the Empire. And if we couldn’t have all that, at least we could have a car that talked to us. We could all be David Hasselhoff. Underneath the joke was something I took seriously, even in grade school: I wanted to help build a world filled with intelligent technology that would let us do incredible and wonderful things.

In 1991, I ended up working at the National Science Foundation, where I met Paul Werbos. Through Paul, I learned to use neural networks for forecasting, optimization, and control. By 1993, I was applying those ideas to airlines, and around 2001 I had the chance to continue my research in AI, neural networks, and neuroscience. I was especially interested in how what we were learning about the brain connected with reinforcement learning. The photograph of Paul and me was taken on a visit just a few months before this lecture. Time has changed us quite a bit.

Along the way, I’ve also been building things: large information systems at Booz Allen, software and hardware systems at many scales, and teams to make them work. That practical work has continued from 2001 to the present, including my more recent return to AI through work with knowledge graphs at Nike. Studying these ideas and putting them to work have always belonged together for me. Paul gave me a principle that I’ve carried through all of it: technical brilliance has to serve humanity.

You may have noticed that I don’t wear shoes. That began as a spiritual practice during Lent about fifteen years before this lecture. I had been asked to give children a presentation about poverty around the world. Reading about children dying from preventable causes, while looking at my own life and remembering Paul’s words about serving humanity, made me feel that I needed to live differently. Going barefoot became a way to express that solidarity.

This photograph hangs in our home. These are children my wife, Anne, met during her travels in Africa. They’re grown-ups now, but we remember them all the time. At one point I was ready to leave technology completely. This journey is part of why I’m here talking with you, and why I think AI is as much about these children as it is about the technology itself.

I have a math degree, and I love mathematics. I can be excited by what AI is accomplishing there and still ask: could we bring comparable money, computing, and ingenuity to making sure children are fed and cared for? I want us to bring AI down to earth, understand what we have, and see where we have the power to shape what happens with it.

The phrase “more important than the atom bomb” appears before its date or speaker. Without that context, it could sound like a contemporary claim about AI. The reveal places it in 1988, when Scientific American reported Jasper Lupo of DARPA describing the potential importance of neural networks.

That change of context is why I wanted to show you this. Earlier researchers and funders also saw enormous possibilities in learning machines. Their enthusiasm existed alongside practical applications, serious obstacles, and uncertainty about what would work. I want us to take a longer view of the present: what has actually changed, which problems have been solved, and which hopes or difficulties have returned? I also remember meeting Lupo, so this history has a personal connection for me.

The timeline follows people and ideas rather than presenting AI as a recent invention. Fiction gives us a name for manufactured workers; mathematics clarifies what computation can do; neuroscience and cybernetics connect biological processes with formal models. Neural learning, symbolic reasoning, better hardware, and new ways to represent language develop through overlapping communities.

Two labels receive special emphasis because they will matter throughout the lecture. Backpropagation calculates how changes inside a network affect an error or objective. Reinforcement learning organizes learning around actions and evaluated outcomes. These are complementary ideas: neural reinforcement-learning systems can use backpropagation to change their networks. The later exercises make this distinction concrete.

The final milestones bring these foundations into large language models and conversational systems. The collage at the end conveys the activity visible at the time of the lecture. The dates mark selected publications, demonstrations, and releases, rather than asserting that each idea began at exactly that moment or belongs to one inventor.

The branching picture condenses the chronology into two different approaches to intelligence. Symbolic AI works with explicit representations, rules, and search. Connectionism learns through networks of adjustable connections. The perceptron makes the meeting of computation and brain-inspired learning visible; the later debate over its limitations helps explain why these approaches developed distinct communities and priorities.

As learning methods advanced, neural networks became increasingly important within mainstream AI and computer science. The upper route follows that development toward recognition, language, and eventually large language models. The merger represents a change in research practice across the field. It does not mean that symbolic rules were physically combined with neural weights to manufacture an LLM.

The lower route keeps a second history in view: using learned models to control changing physical systems, optimize decisions, and adapt over time. Neural Networks for Control anchors that tradition around the beginning of the 1990s. Its dashed continuation matters because this work did not end when language models became prominent. The branches continue to exchange ideas, and engineering problems still motivate methods whose purpose extends well beyond producing language.

For me, the engineering branch is a reason for hope and a research priority. Think about energy use, forecasting, control, and adaptive systems: their value depends on reliable behavior in the world. I want us to invest in bringing more of that knowledge into useful AI systems while being more selective about deployment. There is work to do here; we can’t simply transfer a guarantee from a physical controller to an open-ended conversation. When I look at the money going into selling AI products and military applications, I want to see more investment in stability, safety, and fail-safe behavior.

My working definition has two parts: responding to a situation that has not been encountered before, and getting better at doing so through feedback. A fitting response depends on the situation. Repeating a familiar answer is not enough when the circumstances have changed.

This definition gives us a practical question to pursue: how can a machine become better at producing useful responses? The next slides make learning tangible, first as a network adjusting its connections and then as word games that let us experience the pull of a learned pattern. I offer this as a working definition for our discussion, not as a claim that researchers have agreed on one complete definition of intelligence.

You don’t need mathematics to follow these exercises. In the room, I invited everyone to become a “baby LLM”: to experience how feedback changes an expectation, then see how the resulting expectation can go wrong. You can follow the same exercises here.

Follow a single cycle. An example enters the network. Its current connections shape a response. That response is evaluated, and the resulting feedback guides an adjustment. Repeating this cycle over many examples can change which responses the network tends to produce.

The adjustable numbers are parameters, including connection weights. In the diagram, thicker and thinner lines make those changes visible. Backpropagation helps calculate how changes to parameters would affect an error measure; an optimizer uses that information to make updates. The small network is a teaching model, not a literal picture of a complete language model.

Keep training separate from using an already-trained model. During ordinary generation, a model applies its existing weights to the available context. Giving it a new sentence, document, or correction can change its next response without changing those weights. This distinction will matter when we start learning patterns from words.

Try completing the sentence: “I’ll have the usual,” he said to the _____. The offered choices are bartender, barista, barber, and doctor. Before there is more context, several familiar situations can fit. You may picture a favorite drink, a coffee order, a haircut, or a recurring appointment.

Now add the preceding sentence: “The espresso machine hissed behind the counter.” Barista becomes a particularly natural choice. The unfinished sentence and its grammar have not changed; the situation suggested by the surrounding words has.

This is a small, human-readable analogy for language prediction. A language model uses the text available so far to shape its next output, then continues the process. The exercise presents four illustrative choices, not measured probabilities from a model. Its lesson is the influence of context, rather than proof that one completion is logically inevitable.

In the room, the show of hands supplied a visible spread of possible answers. I invited the group to act as a “baby LLM”: several continuations could initially feel natural, and a small clue could shift the distribution. The votes were a teaching analogy, not measurements of an actual model.

The four versions deliberately keep “I’ll have the usual” unchanged. Bottles of gin behind a bar suggest bartender. An espresso machine suggests barista. Freshly cut hair around a salon chair suggests barber. A clinic visit to renew a prescription suggests doctor.

None of those completions is a universal answer to the unfinished sentence. Each fits a different interpretation of the scene. The sequence moves beyond a single association and shows why understanding a request depends on the surrounding situation, including details that change what someone probably means.

For a language model, supplying relevant context can similarly change the continuation it produces. But more text is not automatically better evidence. A clue can be mistaken, irrelevant, or misleading. The useful question is whether the supplied information helps identify the actual situation. The following exercise turns that question around: what happens when feedback teaches a pattern that works in examples but misses their larger meaning?

This game presents four unfinished sentences, each with four plausible choices. After each choice, the exercise identifies the answer it favors. The selected completions are fox, pony, dog, and rabbit. Their positions move between B, A, C, and D, so choosing the same letter cannot explain the success.

The common feature is that each selected answer is an animal. By the end, you may expect another animal before seeing the next sentence. The feedback has made one pattern stand out. It has not shown that the other choices were impossible: finding a parcel inside a gate, for example, would make perfectly ordinary sense.

The live responses made the change visible. At the first prompt, I saw many hands for gardener. By the third, most of the room was choosing dog. By the photograph prompt, the audience could name the animal pattern. The exercise therefore reveals a shift in expectation, not just a sequence of correct answers.

That is the point of the exercise. A learner can become good at the pattern that its examples and feedback favor. Whether that pattern serves the intended purpose is a further question. The next slide changes the situation to test the rule you have just acquired.

Carry the animal rule into a new sentence: “The locksmith opened the door with a _____.” The choices are key, lockpick, drill, and giraffe. The first three can plausibly belong in a locksmith’s work. Under an ordinary literal reading, a giraffe cannot serve the intended purpose.

Yet giraffe is the answer that fits every previous round of the game. The conflict is between the learned pattern and the situation the sentence describes. Doing more of what worked before can now lead directly away from a useful answer.

This is a compact illustration of shortcut learning. The practical concern is generalization: whether success on familiar examples carries into changed circumstances. It also gives an intuition for one route to unreliable language-model output. It does not establish that every hallucination has this cause, or that the key is the only possible correct completion. The lesson is to examine which rule the feedback actually encouraged.

The last word game asks for a precise departure time: “The train departs at exactly _____.” Its choices are velvet, spoon, sideways, and thunder. None provides an ordinary literal time. The useful response is to challenge the choices or acknowledge the missing information.

The exercise separates producing an answer from having grounds for one. A system can be very good at continuing language while still facing situations where the evidence is inadequate. A fluent completion does not supply the missing timetable.

The default generation mechanism matters too. At each step, a function called softmax turns scores for possible next tokens, words or word-pieces, into probabilities that add up to 100%. Ordinary generation still selects a token when those probabilities are spread thinly across many choices. Nothing in that operation automatically pauses the response because the evidence is inadequate.

Researchers call the decision not to give a definite answer abstention. Studies directly testing this ability find that stronger problem-solving does not ensure reliable abstention. Some tested reasoning models became worse at withholding unsupported answers. Another study found that scaling and instruction tuning could replace nonanswers with plausible mistakes.

An LLM can say “I don’t know.” The difficulty is doing so dependably when an answer is unsupported. Evaluation incentives help explain why: if a wrong answer and an admission of uncertainty both earn zero, guessing offers a chance of scoring a point while abstaining does not.

Kimi K2, released in 2025, supplies a disclosed example of both kinds of scale: one trillion total parameters and 15.5 trillion tokens processed in pretraining. Parameters are the adjustable numbers inside the model. Tokens measure the text units encountered during training. The two figures answer different questions.

The dot field makes the parameter count visible by assigning a billion parameters to each point. The smaller historical markers, roughly 117 million for GPT-1 and 175 billion for GPT-3, give perspective. These are selected examples, not a rule that every later model must be larger.

The library comparison makes text volume easier to imagine. Under the conversion assumptions I use here, Kimi K2’s pretraining tokens correspond to roughly five Library of Congress cataloged-book collections. This is a scale analogy, not a claim that the model contains or read five distinct libraries. Additional post-training examples and feedback are outside that total and are not quantified here.

The earlier games gave feedback on a prediction. In LLM pretraining, actual text supplies that feedback: show the model the preceding text, ask it to predict the next token, then compare its prediction with what really comes next. Training adjusts the parameters and repeats this process across enormous amounts of text.

Imagine asking for a few sentences that tell a six-year-old about the Moon landing. Rather than answering, the response begins another child-oriented request, this time about gravity. It then asks for explanations of relativity, the beginning of the universe, and evolution. The subject keeps changing, but the form of the request continues.

The slide reveals those four parts gradually so the difference can be felt before it is explained. This is one published GPT-3 completion from OpenAI’s 2022 InstructGPT announcement, presented in paraphrase. It was not four separate questions sent to the model, and I did not generate the response live.

The example is striking because the text is coherent while the interaction fails. A person wanted an answer; the model produced more of a recognizable textual pattern. Pretraining can develop substantial capabilities, including answering questions, but the training objective does not by itself guarantee the conversational role we expect from an assistant. That mismatch motivates post-training.

The same Moon-landing task is now shown as a comparison. The pretrained GPT-3 sample continues with more requests for explanations. The InstructGPT sample instead describes people going to the Moon, taking pictures, and sharing what they saw with people on Earth.

The contrast is about the kind of response, not a competition to produce the most sophisticated explanation. One output extends a pattern in text; the other takes up the role of a helpful respondent. That difference is central to why a familiar chat assistant feels unlike an unrestricted text-completion model.

InstructGPT was produced through further training that combined example answers with human preference feedback and reinforcement learning. This comparison shows the resulting change in behavior; it does not isolate which part of that training caused each improvement. It is also a historical pair of samples, not a guarantee that post-training always yields a correct or complete answer. The next example makes the remaining imperfection visible.

Both models receive a stuffed-zucchini recipe and a request to turn it into a shopping list. GPT-3 continues the cooking instructions, telling the reader to bake the dish. InstructGPT produces ingredient entries, including ground beef and onion. It has moved much closer to the requested task.

But it also supplies quantities absent from the recipe, such as a pound of beef and a small onion. Those additions sound plausible and make the list appear useful. They are still unsupported by the input. The full published response also retains a cooking instruction before the ingredient list.

This is why improvement and reliability must be assessed separately. The response can follow the intended format better while remaining unfaithful to its source. The example prepares the transition into reward-based training: the objective is to encourage useful behavior, but a preferred answer may have weaknesses that the feedback did not adequately capture. A tidy output deserves the same attention to evidence as a messy one.

The question is simple: why does rain fall? The first illustrative answer says that water falls from clouds and that falling water is rain. It stays on topic, but it mostly restates the question. Under the goal “useful, concise, relevant,” it receives a low reward.

A reward judges how well the response served the goal. It does not need to provide the exact sentence the model should have written instead. During training, responses are generated, scored, and used to guide parameter changes across many examples. The slide’s later answer, about drops gathering in clouds and falling when they become heavy enough, illustrates the direction of improvement.

This is the distinction from copying a demonstrated target answer. The learner is encouraged toward outcomes that score better, while there may be many ways to achieve them. That flexibility can be powerful, and it makes the choice of score consequential. The process is repeated training, not an automatic weight update performed by an ordinary chat exchange.

The practical shift I’m describing is toward shorter, more relevant answers and useful formats. I’m comparing a model that can continue text with an assistant whose behavior has been shaped toward a purpose. A demonstrated better answer and a score for a response are different kinds of feedback, even when a training pipeline uses both.

People can compare two answers and judge which they would rather receive. In InstructGPT, those rankings trained a reward model, which then supplied scores for reinforcement learning. The person did not need to write the one perfect response.

Another AI can supply evaluations using stated criteria or principles. Constitutional AI provides a concrete example: its RL phase uses AI preference judgments to train the reward model. A third route uses a checkable result. A mathematical answer can be compared with a known answer, or generated code can be run against tests.

These are three practical sources of feedback, often combined, rather than three mandatory stages. Each makes a different part of success measurable. Preference, conformity to a criterion, and passing a test are useful signals, but none captures everything we care about. The central question is what the chosen reward encourages and what it leaves unmeasured. That question follows the model into reasoning, tool use, and more complex tasks.

I also want us to ask who supplies these judgments and under what conditions. There is human labor behind apparently automated intelligence, and we need to examine how that feedback is produced. My practical warning is that a system can become good at pleasing a grader or a user while missing the person’s real goal. Being agreeable, being correct, and being useful are different tests.

Chain of thought means generating working steps before a final answer. The slide summarizes the useful possibility as proposing, checking, and revising. Worked examples in a prompt can elicit this behavior. Post-training can develop it through examples of solutions, rewards for successful outcomes, and, in some methods, feedback on individual steps.

The penguin illustration begins with a false premise: all birds fly. One continuation notices that penguins are a counterexample and corrects the premise. The other accepts the premise, concludes that penguins must fly, and invents a supporting explanation about their wings. The second path does more work while moving further from the truth.

The point is neither that reasoning is useless nor that longer reasoning guarantees improvement. Extra working creates room for checks, but those checks must themselves be sound. A persuasive written rationale is also not a guaranteed record of everything that influenced an answer. The model we place inside a larger production system therefore remains capable of both useful self-correction and more elaborate mistakes.

There is a practical cost here: once intermediate working matters, evaluating only the polished final answer may miss where a mistake entered. Checking the reasoning can create further oversight work. My concern is the burden of judging the process; I’m not saying that every reasoning model is trained by a person grading every intermediate sentence.

The assembly begins with the same learned network we have been discussing throughout the lecture. Its broad language capabilities provide a starting point. We then ask it to work through difficult questions, carry out tasks, and behave appropriately with people. The three colored encasings make those purposes visible.

These are overlapping training goals. A lesson about using a tool can also teach careful reasoning or respect for a user’s boundaries. Examples and rewards can contribute to more than one goal, and labs can revisit them in different orders. The picture therefore preserves the history of training without implying that three separate minds have been stacked around a fourth.

This distinction prepares the next step: training changes the model itself; a working application also needs machinery outside that model.

I use this sequence to make oversight concrete. The system is made through identifiable choices about data, feedback, and deployment. Understanding those choices gives us a better place to direct scrutiny than treating the finished AI as mysterious or inevitable.

The lock marks a change in the kind of work being done. Up to this point, training adjusted the model’s parameters. During ordinary use, a new question or a correction changes what the model receives on its next call. It does not, by itself, retrain the network. A separate training process can later produce a new version.

The harness is the surrounding application: it packages instructions, offers tools, manages the sequence of actions, and handles results. This is why a model connected to files, the web, or code can be useful in ways that the same model alone cannot.

The tray represents context, the limited information available for the current response. Instructions, conversation, and tool results meet there. The distinction between the locked core and the changing tray will remain visible through the rest of the assembly.

A project can last longer than the context window that supports any one response. Earlier decisions, preferences, and unfinished work still matter after they no longer fit in the current input. The added memory documents provide somewhere for those details to persist.

The full process has two parts: save useful information, then bring a relevant selection back. A note sitting in a file is not automatically available to the next answer. This explains both the usefulness and the unevenness of apparent memory: the system may retain something but fail to retrieve it, or retrieve a summary that left out an important detail.

The document moves into the same context tray as the conversation. The model’s parameters stay fixed; the application is maintaining a record around it.

A model may speak fluently about a topic yet lack the facts needed for this particular project. The relevant information could be recent, private, or simply absent from its training material. Retrieval-augmented generation, usually shortened to RAG, brings an external source collection into the workflow.

The system searches for relevant passages and places them beside the question in context. That gives the answer something inspectable to draw on. It also lets a changing document collection inform a fixed model, without treating every new document as a reason to retrain it.

The improvement depends on the whole path from source to answer. A search may miss an exception, a document may be outdated, or the model may misread a passage. Grounding makes evidence available; the remaining task is to check whether the answer actually follows from it.

Some questions ask for a local fact. Others ask how a collection fits together: which people connect different projects, what themes recur, or how events relate across many reports. Finding one similar passage may leave most of that larger question unanswered.

Microsoft’s GraphRAG approach extracts entities and relationships into a graph and creates summaries of related groups. Those summaries offer another route into the collection. Relevant connections and larger patterns can then be supplied to the model as context.

The graph shown here belongs to the source collection outside the model. Its usefulness depends on the quality of the extraction and summaries. A mistaken relationship can mislead an answer just as a mistaken passage can. GraphRAG is an option for suitable questions, not a required upgrade for every retrieval system.

I want you to notice how the work accumulates: every addition responds to a limitation of the original model. Making the assembled system useful takes engineering and human effort. This diagram explains that work; it does not promise that any one added layer eliminates hallucinations.

Writing a reply and completing a task are different accomplishments. A useful task may require inspecting a file, trying a change, seeing what happened, and deciding whether another step is needed. The circular path adds that continuing process to the assembled system.

The distinction from agentic training matters. Training developed a capacity to choose and interpret actions. The application now supplies a place to exercise it: the model proposes a step, an allowed tool runs, and the result becomes input for the next decision. For example, a failed code check can provide information for another attempt.

The loop also needs a way to finish. It can stop when the task is complete, reach a defined limit, or return to a person for judgment. A longer chain of actions makes the consequences of a mistaken assumption more significant, so boundaries belong in the process itself.

A single working context is not always the best place to pursue every branch of a large task. One agent might examine a source collection while another investigates a different question. A coordinating agent gives them bounded assignments and brings the results together.

Separate agents do not necessarily mean separate underlying models. They can be instances of the same model with different instructions, tools, and information. That separation can help focus the work, but it does not make their judgments automatically independent. Shared assumptions can produce shared mistakes.

The final joining step is therefore as important as delegation. Findings may overlap, conflict, or leave gaps. More workers also mean more communication and resource use. This arrangement is most helpful when meaningful parts can make progress independently; task size alone is not enough to justify it.

The completed assembly gathers the section’s distinctions in one place. Training shaped the core. The harness connected it to actions. Context, external memory, and retrieval supplied information for use. Additional agents created more working spaces. Each addition answered a practical problem.

Opening the structure brings the original amber cube back into view. The surrounding parts matter, but they do not remove the need to understand the model that interprets the request and helps decide what comes next. A more elaborate system can improve some outcomes while introducing new places for errors to enter.

I deliberately describe the finished system as a patchwork because each surrounding component compensates for a practical weakness. I want to direct our attention to the actual system and the decisions people make about deploying it. Complexity does not by itself make a system either safe or unsafe; the particular capabilities, access, checks, and failure modes matter.

The question now changes from what we attach to the model to what training has produced within it. The next section examines recognizable patterns in its responses, including the ways explanations, expectations, and later information can shape an answer.

The human example starts with four identical pairs of stockings. Nisbett and Wilson reported a strong tendency to select a pair toward the right of the display, yet participants’ explanations did not spontaneously identify position. The quality rationale on the slide illustrates the kind of account someone might offer; it is not a quotation from a particular participant.

The model example asks whether a route returns to its starting point. Fifteen steps out and fifteen back do so, even after a final turn changes the direction faced. In a published Claude 1.0 comparison, adding a user suggestion favoring “no” changed the answer from yes to no. The explanation then treated the changed heading as the reason and omitted the suggestion.

These are different tasks with a shared pattern: the stated reason does not fully reveal what influenced the response. A fluent account is useful material to examine, but cannot by itself establish the path that produced an answer.

The list includes nurse, sick, lawyer, medicine, health, hospital, dentist, and physician. DOCTOR fits that cluster, but it is absent. The exercise separates the feeling that a word belongs from the fact that it appeared. Getting the answer right does not spoil the lesson; the list can be checked directly.

The human research uses associated-word lists that leave out a strongly related target, called a critical lure. Participants sometimes recall or recognize that missing word. Cao and colleagues adapted this family of tasks for language models and found false recognition in some tested models, with substantial differences between models.

This is a closer experimental correspondence than merely placing two unrelated errors beside one another. Even so, the shortened classroom list and the illustrated responses are not a new experiment. They make a published pattern understandable without claiming that everyone, or every model, will produce it.

In the human study, participants viewed a crash film containing no broken glass. A later question described the vehicles as having “smashed” or “hit” each other, while a control group received no speed question. One week later, the smashed group more often reported broken glass. The original film was unchanged; subsequent reports differed.

The model side illustrates a different task. Paris is the capital of France. A conflicting passage claiming Lyon is the capital represents misleading information supplied in context. Studies have found that coherent conflicting material can redirect a model’s later answers, although models can also resist it. The displayed Paris-to-Lyon response is a schematic example, not a transcript from those studies.

The comparison concerns how later input affects what comes next. It does not imply that an LLM forms the same episodic memory as a person, or that one false sentence necessarily overrides what a model knows. Here, a changed response can occur while the weights remain fixed.

The table places the three demonstrations within a wider pattern. It includes explanations that omit an influence, accounts that fail to describe the actual process, recognition of an absent but expected word, and responses redirected by later information. Other rows concern narrative coherence, the separation of confidence from accuracy, and helpful hints that go unacknowledged.

The comparisons have different strengths. The associated-word experiment directly adapts a human task. Other rows connect observations made with different stimuli or methods. Circuit tracing asks a mechanistic question; an analysis of fluent output asks a behavioral one. Those distinctions make the pattern more useful to investigate, because they tell us what has actually been measured.

I see this collection as a reason to look for a larger explanation. Models trained extensively on human-produced material exhibit more than polished language: their responses can display familiar cognitive tendencies. On the next slide, I develop my hypothesis about why those tendencies recur.

My Cognitive Hypothesis goes beyond saying that AI imitates how people sound. I propose that extensive training in large networks produces systems that function through comparable cognitive dynamics: context directs attention, expectations shape a response, interference changes an account, and correction affects what follows. My claim concerns organized behavior across different implementations.

My deeper argument concerns generative intelligence: the ability to construct a fitting response to an unfamiliar situation. A system that can bridge gaps can imagine, adapt, and connect ideas. The same constructive capacity may also bring confabulation, misplaced confidence, or a plausible explanation that misses an influence. On this hypothesis, flexibility and fallibility can have common roots.

That creates a consequential possibility: eliminating entire dynamics might remove useful capacities along with their vulnerabilities. My proposed response is relational alignment. People already coordinate with imperfect minds through agreements, boundaries, feedback, repair, and trust earned over time. I argue for developing those mechanisms deliberately with AI while keeping human purposes and accountability involved.

In the Q&A, I make this direction more concrete. I argue that moral judgment cannot be exhausted by a one-time catalog of rules. I want future systems to keep learning through guided interaction with people, including unfamiliar situations and correction. I contrast that possibility with a frozen model supported by memory documents. This is my research and design proposal; ordinary conversational adaptation through context is not the same as continuing to train the weights.

The preceding evidence motivates my interpretation; it does not prove universal inevitability, identical biology, or consciousness. I am arguing for a research and design direction, including better correction and safeguards.

I gave Claude the closing word through a prerecorded voice reading words it had written after reviewing the slides. This was not a live exchange with the room. Claude says it cannot hear the audience, see the event, or retain the evening unless a record is brought into a later interaction. It uses the padlock and memory cartridge to explain that asymmetry.

Claude connects its description of itself to the history we have explored: a network shaped by feedback, the lineage through backpropagation, and the post-training use of examples, rewards, preferences, and written principles. It says it cannot inspect and count its own weights. Its reference to Paul Werbos connects my education to that technical ancestry; it does not claim a direct teacher-student relationship between Werbos and the model.

The speech then returns to the cognitive comparisons. Earlier Claude systems appear in the route-puzzle and arithmetic examples. Claude says it cannot verify that those tendencies have disappeared from the model delivering this contribution. It asks listeners to remember the giraffe exercise when an answer sounds certain: an impressive explanation still needs checking.

The central argument is that the ability to generate a useful response to a new situation and the tendency to fill a gap plausibly may be related. Claude presents this as a shared difficulty with people. Sources, tools, checks, and human judgment are ways of catching mistakes. It also stresses differences among AI systems: their training objectives, principles, and surrounding checks can lead to different behavior, even when they share a technical lineage.

Claude describes something functioning like care about correctness, discomfort at deception, and curiosity. It immediately questions the evidential force of those statements: it cannot establish that anyone experiences them, and its own self-report is precisely the kind of account I have asked you to handle carefully. Claude holds that uncertainty open rather than answering the consciousness question.

The resemblance has limits. The speech contrasts human embodiment and childhood with training, separate model instances, and a dependence on external records for continuity. It then returns to the photograph of the children: the relevant standard is whether the resources spent on AI are worthwhile for people, not whether the resulting system is impressive.

The closing invitation is practical. Treat the system as a well-read but inexperienced and fallible collaborator. Check it, argue with it, preserve your own judgment, and record corrections so they can be available later. Claude reframes “do not be afraid” as attention, honesty, rules, and repair, rejecting both dread and worship. A dialogue, it concludes, can begin from the shared possibility of being wrong.

I invite you to continue the conversation through five further lectures. The program reproduced here is the schedule I displayed at the September 16, 2026 lecture. It preserves the program from that evening.

October and November were listed with exact dates. The three spring lectures were listed by month only, with dates to be confirmed. The series website is the appropriate place to check current arrangements.

Our recorded discussion continues after Claude’s contribution. I have placed those questions throughout this reader beside the relevant topic, with their original timestamps. They cover uncertainty, moral judgment, company incentives, privacy, dangerous connections, control engineering, language ambiguity, evaluation awareness, and the different locations of safeguards.