Latest / SleepWise: Learn Softly, Sleep Deeply / Artificial Intelligence Explained Gently For Sleep
Transcript
- 0:00Good evening and welcome back to Sleep Wise.
- 0:04Tonight we begin at the newest edge of an old human wish, the
- 0:10wish to build a thinking tool, then set it down the way you set
- 0:15down. A heavy day outside the world
- 0:19can be loud with certainty, but here, in the soft light of
- 0:24night, we can move gently and reverse from the present back
- 0:30through the quiet centuries that led us here.
- 0:33Picture a modern room after midnight, a laptop lid half
- 0:37closed, a phone face down on a bedside table somewhere far away
- 0:44in a data center cooled by fans and careful engineering.
- 0:49Machines are awake in a different way, not with desire
- 0:54or intention, but with activity. Tiny calculations ripple through
- 1:00chips like the faint hum you barely notice until the house
- 1:05grows still. And in many homes, people type a
- 1:10question into a simple box and words arrive back fluent, calm,
- 1:18sometimes surprising. In practical terms, what we call
- 1:24artificial intelligence to day is not one single invention.
- 1:30It is an umbrella for many methods that help computers
- 1:33recognize patterns and make useful decisions.
- 1:38Sorting photos. Translating languages.
- 1:42Recommending songs. Detecting fraud.
- 1:46Adjusting a thermostat. Steering the robot arm.
- 1:51Over the last decade, these systems have spread because two
- 1:56things became abundant data and the ability to compute on that
- 2:02data quickly. When a computer learns from
- 2:05examples, it does not memorize them the way a person remembers
- 2:10A childhood St. Instead, it adjusts internal
- 2:16numbers so that overtime it becomes better at map as inputs
- 2:21to outputs an image to a label, a sentence to another sentence,
- 2:27a sound to a transcript. You can think of this learning
- 2:32as a long practice session. The system sees an example,
- 2:37makes a guess, measures how wrong it was, and nudges itself
- 2:43slightly toward being less wrong next time.
- 2:47That cycle repeats millions, sometimes billions of times
- 2:53until the model becomes reliable at a task.
- 2:58The models themselves can be many shapes, decision trees,
- 3:03linear models, neural networks. And the training can happen in
- 3:10the cloud on specialized hardware, with enormous batches
- 3:14of examples flowing through like a river.
- 3:18But tonight, we won't stay in the present too long.
- 3:23We'll let it fade behind us like city lights shrinking in a rear
- 3:28view mirror. We'll drift backward through the
- 3:32era of smartphones and cloud services, back through the rise
- 3:38of powerful chips, back through earlier waves of machined
- 3:42learning, and further still to the first dreams of mechanical
- 3:48calculation, the first symbolic logics, the first attempts to
- 3:54measure uncertainty, and the ancient fascination with
- 3:58automata devices that moved as if they had a tiny life inside
- 4:05them. For now, let your breathing find
- 4:09its own pace. Let the idea of these bright
- 4:12machines feel distant and quiet, like a constellation.
- 4:17You don't need the name. We are simply walking backward
- 4:22along the path, step by step, until the night grows older and
- 4:27simpler and the world slows. And so do we.
- 4:33The present begins to soften at the edges, and we step back into
- 4:38the recent decade, when AI stopped feeling like a distant
- 4:43research word and started feeling like a quiet utility.
- 4:49It appears in small places first, the way night air slips
- 4:55in through a cracked window. Your photo app groups similar
- 5:00faces. A spam filter keeps your inbox
- 5:04from filling, A map suggests the quickest turn before you decide.
- 5:12None of this asks for applause. Yet together it changes the
- 5:17texture of ordinary life. Underneath there is usually a
- 5:23trained model, a system of adjustable parameters fitted to
- 5:29examples. In image recognition, the input
- 5:33is a grid of pixel values and the output is a label or a set
- 5:38of probabilities. In speech recognition, the input
- 5:44is a waveform shaped into features and the output is text.
- 5:49In recommendations, the input is a history of views, clicks, and
- 5:54pauses, and the output is an ordered list designed to be
- 5:59useful. The mathematics differs from
- 6:02task to task, but the rhythm is similar.
- 6:06The system is shown an example. It produces an output, and a
- 6:11measure of error is computed by comparing that output to a
- 6:15target. Then the parameters are nudged
- 6:19slightly toward better performance.
- 6:21The adjustments are small, almost unimaginably small, but
- 6:26they add up. After millions of examples, the
- 6:30model begins to capture patterns that are hard to write as
- 6:35explicit rules. It can notice that certain
- 6:40arrangements of edges often form a face, that certain rhythms
- 6:46often form a word, that certain sequences of choices often form
- 6:51a preference. Importantly, these systems do
- 6:56not store a single tidy explanation.
- 7:00They store tendencies encoded as numbers, which you only see when
- 7:07the system makes a prediction. Once deployed, the model meets
- 7:11the real world, and the real world is never as neat as a
- 7:15training set. Lighting shifts.
- 7:19Accents vary. People phrase requests in
- 7:23unexpected ways. Engineers watch performance,
- 7:28collect new examples and retrain because learning systems can
- 7:34drift. As the world drifts in this way,
- 7:38modern AI became less like a one time invention and more like
- 7:44ongoing maintenance, a garden that must be tended with careful
- 7:51attention to reliability and fairness.
- 7:55This is why the last decade felt like an acceleration 2 Resources
- 8:02became abundant at once. First, data.
- 8:06More of life was digitized, and everyday actions left traces
- 8:12that could be gathered into training sets, photos,
- 8:16recordings, transactions, sensor readings, searches, and logs.
- 8:22Second, compute chips grew faster and more specialized, and
- 8:29cloud infrastructure made large training runs possible without
- 8:33owning a private room of servers.
- 8:36And now, as we continue our backward walk, we let the
- 8:41conveniences fade into their hidden machinery.
- 8:47We follow the path from polished apps to the era of smartphones
- 8:52and cloud services, then further into the hardware that made
- 8:57modern learning practical, and further still toward older
- 9:03methods, smaller data sets, and earlier dreams of teaching
- 9:08machines at all. A little farther back, the story
- 9:12takes on the feel of a pocket sized revolution.
- 9:17Before AI became a headline, the smartphone became a habit.
- 9:24It gathered the world into a single object you carried from
- 9:28room to room, from street to cafe, from morning to night.
- 9:34In doing so, it also gathered data, images, motion, location,
- 9:42touch and voice. A phone is a small bundle of
- 9:46sensors, and sensors turned life into numbers.
- 9:51Those numbers did not automatically become training
- 9:55material, but they made it possible.
- 9:59As services moved online, organizations built systems to
- 10:04store and organize immense volumes of information.
- 10:10They built pipelines to clean it, label parts of it, remove
- 10:15duplicates, and handle missing values.
- 10:19They learned to split data into training sets and test sets so
- 10:26that a model could be evaluated honestly, not rewarded for
- 10:31simply repeating what it had already seen.
- 10:35They also learned, often through public debate and internal
- 10:39friction, that privacy and consent are not optional
- 10:44details. Data has owners, contexts and
- 10:49consequences. Connectivity changed the feel of
- 10:53what a device could be. When networks became fast and
- 10:58reliable, a phone could send a request to distant servers and
- 11:04receive a result. In the time it takes to blink,
- 11:08intelligence began to feel like a living layer over the device.
- 11:14Some predictions ran locally for speed and battery life.
- 11:20Others ran in the cloud, where larger machines could do heavier
- 11:25computation. This division shaped modern
- 11:29products, lightweight models on the device, and heavier training
- 11:34and experimentation on centralized infrastructure.
- 11:39The edge mattered because of constraints.
- 11:42Phones have limited battery, limited heat budget, and limited
- 11:47memory. So engineers learn to compress
- 11:51models, prune unnecessary parts, and simplify calculations,
- 11:58making systems small enough to run smoothly without draining a
- 12:03device. They also learn to measure
- 12:06latency, because a model that is accurate but slow can feel worse
- 12:13than a model that is slightly less accurate.
- 12:17But instant app stores and frequent updates created a new
- 12:22rhythm. Products could change weekly,
- 12:25sometimes daily. That encouraged constant
- 12:29evaluation. In the real world, teams watched
- 12:32how often a feature was used, whether it reduced errors,
- 12:37whether it increased satisfaction, and whether it
- 12:41behaved consistently across languages, regions and devices.
- 12:47In factual terms, this period taught the world how to
- 12:52operationalize machine learning, deploy models, monitor them,
- 12:59collect fresh examples, retrain and roll improvements out again.
- 13:06Now we step backward once more. We leave the glass screens and
- 13:13wireless signals behind, and we approach the era when the hidden
- 13:18engines of progress became unmistakable.
- 13:23The racks, the cooling, and the chips built to repeat
- 13:29mathematics at enormous scale. Now the not air cools and we
- 13:35find ourselves closer to the machine rooms.
- 13:39Before modern products could rely on learning systems
- 13:43everywhere, training had to become feasible at scale.
- 13:49One of the quiet enablers was specialized hardware, especially
- 13:54the graphics processing unit, the GPU.
- 13:59It began as a tool for drawing images quickly, shading,
- 14:04lighting, rendering, the illusion of depth.
- 14:07But it turned out to be good at something else too, repeating
- 14:13the same numerical operations many times in parallel without
- 14:19slowing down. Many learning methods, and
- 14:23especially neural networks, rely on long chains of arithmetic
- 14:27that can be expressed as vast collections of simple steps.
- 14:32Multiply, add, compare, repeat. In earlier eras, training larger
- 14:39models on general purpose CPUs could be painfully slow.
- 14:45But GPU's were built to handle thousands of similar
- 14:49computations at once because graphics demanded it.
- 14:55Researchers realized they could borrow that speed for training.
- 15:01A run that once took weeks could be shortened, and experiments
- 15:06that once felt impossible began to fit into human schedules.
- 15:12Around this hardware grew the infrastructure of the modern
- 15:16era. Data centers, warehouse sized
- 15:20rooms filled with racks, cooling systems, backup power, careful
- 15:25networking, and teams who think in terms of airflow and uptime.
- 15:33These places can sound dramatic, but they are also strangely
- 15:37patient. They do not hurry.
- 15:40They run jobs, store results, run the next jobs.
- 15:46They are the modern equivalent of mills, turning raw material
- 15:51into something shaped. Here, the raw material is
- 15:55information and the turning is computation.
- 16:00Hardware progress is not only about speed.
- 16:05Memory bandwidth matters because models and data must be moved
- 16:10efficiently. Interconnects matter because
- 16:14training can be split across multiple machines.
- 16:18Reliability matters because long runs are costly to interrupt.
- 16:24And energy matters because computation becomes heat, and
- 16:29heat must be managed. The story of AI is in part a
- 16:36story of engineering limits being pushed gently outward,
- 16:42year by year, as we drift backward.
- 16:45You can let the words chips and compute become softer, like
- 16:52distant weather. The details matter, but they do
- 16:57not ask you to stay alert. We are simply noticing the
- 17:01engines beneath the surface and then letting them fade as we
- 17:06step into an earlier time, when data sets were smaller, training
- 17:11was slower, and progress depended heavily on how
- 17:15carefully it was measured. So we step back again toward the
- 17:20era of benchmarks and shared yardsticks, where the field
- 17:25learned to agree quietly on what it meant to improve.
- 17:30We arrive in a period when the field learned to trust its
- 17:35measurements. Progress in machine learning is
- 17:39not only about building models. It is also about deciding how to
- 17:44judge them. In earlier decades, researchers
- 17:48often worked on small data sets, sometimes collected by hand, and
- 17:54compared results in scattered ways.
- 17:57Overtime, shared benchmarks became common, standardized sets
- 18:03of examples that many teams could use so that improvements
- 18:07were visible, repeatable, and comparable.
- 18:11A benchmark might be a collection of images with
- 18:15labels, recorded speech with transcripts, or text paired with
- 18:21answers. The point is not that a
- 18:25benchmark represents all of reality, it never does, but that
- 18:30it becomes a lighthouse. If you can improve performance
- 18:36on the same set under similar conditions, you can tell whether
- 18:41a new method is truly better or merely different.
- 18:46This practice also made the field more honest.
- 18:50It reduced the temptation to claim progress without evidence,
- 18:56and it encouraged careful evaluation across multiple
- 19:01tasks. Along with benchmarks came the
- 19:04quieter work of data preparation.
- 19:07Data is rarely ready as it is. It arrives messy duplicates,
- 19:14errors, strange formats, missing labels and biases embedded in
- 19:20what was collected and what was ignored.
- 19:24Cleaning and curating data sets became a craft.
- 19:28So did deciding how to split data into training, validation
- 19:33and test sets, and how to avoid leakage where hints from the
- 19:38test set sneak into training and make results look better than
- 19:44they truly are? Metrics became part of the
- 19:47shared language. Accuracy, precision, recall,
- 19:53error rates, and other measures help teams describe performance
- 19:58clearly. Some tasks needed new metrics
- 20:02and some needed human evaluation especially.
- 20:07When outputs were subjective or when correctness was hard to
- 20:11define, but the impulse was the same, create a common map of
- 20:17progress. Tonight you can imagine these
- 20:20data sets as quiet piles of photographs, boxes of
- 20:24recordings, and stacks of printed pages.
- 20:29Not as stories you must read, but as raw material.
- 20:34The work of building benchmarks is not glamorous, yet it shapes
- 20:39what gets built, because what gets measured gets optimized.
- 20:46And as we step backward again, those piles become smaller, the
- 20:52stacks thinner, and the experiments more dependent on
- 20:56handcrafted features and classical methods.
- 21:01We are approaching the era when machine learning often meant
- 21:05carefully chosen inputs, simpler models, and a different kind of
- 21:11patience. Now we step into the time when
- 21:16machine learning often meant a narrower set of tools and a
- 21:22different balance between human design and automated learning.
- 21:28Before deep neural networks dominated many applications,
- 21:33practitioners frequently relied on models like logistic
- 21:38regression, decision trees, random forests, and support
- 21:45vector machines. These methods could be powerful,
- 21:49but they usually depended on features, carefully crafted
- 21:54representations of the input. In a vision task, features might
- 22:00describe edges, corners, textures, or color histograms.
- 22:05In speech, features might capture frequency patterns over
- 22:10time. In text, features might include
- 22:14counts of words or short phrases.
- 22:18Creating these representations required domain knowledge and
- 22:23experimentation. The model then learned on top of
- 22:27those features, adjusting its parameters to separate
- 22:31categories or predict outcomes. In this era, a large portion of
- 22:38the intelligence was in the pipeline, the preprocessing, the
- 22:42feature engineering, the careful selection of signals that
- 22:46mattered. These approaches were not crude.
- 22:50They were often elegant and statistically grounded.
- 22:55They taught the field how to think in terms of
- 22:59generalization. A model should perform well not
- 23:04only on the data it is seen, but also on new data drawn from the
- 23:09same underlying process. They encouraged regularization
- 23:15methods to avoid over fitting by limiting complexity.
- 23:19They encouraged cross validation, testing models
- 23:23across different splits of data to estimate robustness.
- 23:29And they encouraged humility because small changes in the
- 23:34data could change results. Computing constraints shaped the
- 23:40style of work. Training had to be efficient,
- 23:45and memory had to be respected. That led to an emphasis on
- 23:51convex optimization in some settings, problems where a
- 23:55single best solution could be found reliably.
- 24:00It also led to practical heuristics and to a culture of
- 24:05careful baselines. Compare your new idea to
- 24:09something simple and strong before celebrating.
- 24:12As we drift backward, we can feel the field in this period as
- 24:17a workshop of careful craft. Less spectacle, more tuning,
- 24:24less scale, more thought about what to include and what to
- 24:29ignore. And yet, beneath the surface,
- 24:33older neural ideas never vanished.
- 24:36They waited quietly for the conditions that would let them
- 24:41return. So we step back again toward the
- 24:46years when neural networks began to re emerge, then further to
- 24:52the techniques that made them trainable, and to the earlier
- 24:57hopes and disappointments that shaped the field's memory.
- 25:02We move backward into the early 2000 tens when a familiar
- 25:07approach, neural networks, began to feel newly capable.
- 25:13Neural networks had existed for decades, sometimes celebrated,
- 25:18sometimes dismissed, often emphasized in theory and limited
- 25:24in practice. But as larger data sets became
- 25:28available and faster hardware became common, deeper networks,
- 25:34networks with many layers started to perform strikingly
- 25:39well on tasks like image recognition and speech.
- 25:45In simple terms, a neural network learns by adjusting
- 25:50weights, numbers that determine how strongly 1 unit influences
- 25:55another during training, the network processes.
- 26:00An example produces an output, compares it to a target, and
- 26:05then uses an error signal to nudge many weights in small
- 26:10increments. The key is that this error
- 26:14signal can be propagated backward through the layers so
- 26:19that each weight is updated in the direction that reduces
- 26:25error. It is a method of learning by
- 26:28tiny corrections repeated countless times, like smoothing
- 26:35a stone by steady water. This period also sharpened
- 26:41practical techniques, better ways to initialize weights,
- 26:46better activation functions, better ways to prevent
- 26:51overfitting, and better optimization methods that made
- 26:57training stable. Researchers learned to use large
- 27:01batches of data to normalize inputs, to manage exploding or
- 27:07vanishing gradients, and to select architectures suited to
- 27:13the structure of the problem. The results felt like a hinge.
- 27:21Systems that once seemed brittle began to feel practical.
- 27:26At the same time, costs and constraints became part of the
- 27:30conversation. Larger models demanded more
- 27:34energy and more compute. Access was uneven, and data
- 27:41reflected the world as it is uneven, imperfect, sometimes
- 27:46unfair. So models could inherit those
- 27:50patterns unless care was taken. The field began to broaden its
- 27:56view, not only asking does it work, but also for whom and
- 28:02under what conditions. Tonight we do not need to solve
- 28:07those questions. We only need to notice the shape
- 28:12of the era, an old idea returning with new strength,
- 28:18powered by compute, guided by benchmarks, and made practical
- 28:25by patient iteration. From here, the path continues
- 28:30backward into the core training technique itself, how learning
- 28:36is turned into mathematics, and into earlier decades.
- 28:41When optimism cooled into long winters, we step back again
- 28:48toward the engine of training and the history that gathered
- 28:53around it. Now we step back to the
- 28:56mechanism that quietly underlies much of modern learning,
- 29:02optimization. It is a simple word for a deep
- 29:05habit, adjusting something until it fits.
- 29:11In machine learning, the something is a set of
- 29:15parameters, and the fit is measured by a loss function, a
- 29:21numerical score that represents how far a model's outputs are
- 29:25from the targets we want. Training is the process of
- 29:29reducing that loss. One of the central ideas is
- 29:33gradient descent. Imagine you're on a landscape at
- 29:38night, and you cannot see the whole terrain, but you can sense
- 29:43which direction slopes downward from where you stand.
- 29:47You take a small step down, then another, then another.
- 29:53Each step is tiny because large steps might overshoot overtime.
- 30:00You descend toward a valley, toward a set of parameters that
- 30:05produces better predictions. This is not guaranteed to find
- 30:10the absolute best solution in every case, but it is often good
- 30:15enough and remarkably effective at scale.
- 30:20For neural networks, the practical question was how to
- 30:24compute gradients efficiently when there are many layers.
- 30:30The answer is a method that computes the contribution of
- 30:34each parameter to the loss by working backward through the
- 30:39network, reusing intermediate computations.
- 30:44This makes training feasible even when there are millions of
- 30:48parameters. Without this efficiency, deep
- 30:52networks would be too slow to train and too expensive to
- 30:58experiment with. Optimization also requires
- 31:01choices, learning rate, batch size, momentum, regularization.
- 31:08Each one shapes how training behaves.
- 31:12Too fast and the model may bounce around without settling.
- 31:17Too slow and training may take impractically long.
- 31:23Too flexible and the model may memorize quirks of the data.
- 31:28Too constrained and it may never learn enough.
- 31:33Much of machine learning practice is the calm, repetitive
- 31:38tuning of these settings, guided by validation, results and
- 31:43experience. As we drift backward, this world
- 31:47of gradients and losses begins to thin into earlier eras.
- 31:53We will soon reach the decades when the field leaned heavily on
- 31:57hand crafted features and simpler models, and then further
- 32:03back into a time of bold early promises and sharp limitations,
- 32:10when neural ideas were introduced, challenged, and
- 32:14sometimes set aside. So we step back again, toward
- 32:19those earlier hopes, and toward the long, quiet winters that
- 32:24taught researchers patience. We step backward into the
- 32:30periods often called AI winters, when enthusiasm cooled and
- 32:37funding became harder to find. These winters were not times
- 32:43when nothing happened. They were times when progress
- 32:47was slower and expectations were corrected.
- 32:52Early systems had promised a great deal, but they struggled
- 32:57in the wild. They required too much hand
- 33:00tuning. They failed when the environment
- 33:04changed. They could not easily scale to
- 33:08the full messiness of human language or perception.
- 33:12One reason was that early approaches often leaned on
- 33:15explicit rules. If you can write down the rules
- 33:19of a domain, a computer can follow them quickly.
- 33:24But the world is full of exceptions, edge cases, and
- 33:28tacit knowledge. Rules multiply, then collide.
- 33:34Maintaining them becomes its own problem.
- 33:38Another reason was compute. Many methods that looked good on
- 33:43paper or on small toy problems became infeasible when data grew
- 33:48and tasks became realistic. A third reason was data itself.
- 33:54Many domains simply did not have large, well labeled data sets
- 33:59available. In these quieter years, a
- 34:02different kind of work continued.
- 34:06Researchers refined statistical methods.
- 34:10They developed better evaluation practices.
- 34:14They built data sets where they could.
- 34:18They explored learning in constrained settings, they
- 34:23improved algorithms for search and planning in specific
- 34:28domains, and they learned the value of bass lines and
- 34:33reproducibility. The field matured even when it
- 34:38was not in fashion. It is also in these eras that
- 34:42you can feel the rhythm of science more broadly.
- 34:45A new idea arrives. Bright and confident.
- 34:50Reality pushes back. The idea is reshaped or set
- 34:56aside or combined with other ideas.
- 35:00Sometimes it returns later, when the conditions are right.
- 35:05This is not failure in the dramatic sense.
- 35:07It is endurance, the slow tightening of a craft as we
- 35:13drift backward. We can imagine the laboratories
- 35:17of that time, fluorescent lights, whiteboards, modest
- 35:22machines, small data sets and long patience.
- 35:28The world outside might have moved on to other obsessions,
- 35:32but inside the work continued measuring, testing, revising,
- 35:39and further back still, we find the early neural inspirations
- 35:44that sparked both optimism and disappointment, perceptrons,
- 35:51simple learning rules, and the first attempts to mimic the
- 35:56brain with mathematics. So we step back again toward the
- 36:01beginning of learning machines and toward the early decades of
- 36:06computation that made such dreams imaginable.
- 36:12Now we are closer to the beginning of the learning
- 36:17machine dream. In the mid 20th century,
- 36:22researchers began proposing systems that could adjust
- 36:26themselves based on experience. One early family of ideas was
- 36:32the perceptron, a simple model that takes inputs, multiplies
- 36:38them by weights, sums them, and produces an output based on
- 36:44whether the sum crosses a threshold.
- 36:47It is modest by modern standards, yet it carries an
- 36:51important promise. The weights can be learned from
- 36:56examples. If a system can change itself in
- 37:00response to data, then intelligence might be something
- 37:04you can train rather than something you must hand code.
- 37:10That was the hope, and perceptrons could learn some
- 37:14kinds of patterns, especially patterns that can be separated
- 37:19cleanly in a mathematical sense. But they also had clear
- 37:24limitations. Some seemingly simple patterns
- 37:28cannot be learned by a single layer perception.
- 37:33This led to criticism, disillusionment, and a shift of
- 37:39attention toward other approaches.
- 37:42Still, the perceptron era mattered deeply.
- 37:48It introduced the language of weights and learning rules.
- 37:54It encouraged thinking about generalization and noise.
- 37:59It raised the question of representation.
- 38:02What form should inputs take so that a learning system can use
- 38:07them? And it forced researchers to
- 38:11confront the gap between appealing metaphors of neurons
- 38:17and the hard requirements of mathematics and data.
- 38:21Around the same time, the world of computation was expanding
- 38:25rapidly. Early electronic computers were
- 38:28being built and improved storage became more reliable.
- 38:33Programming languages emerged. The notion of a stored program,
- 38:39data and instructions living in memory made computers flexible
- 38:45in a way that earlier calculating devices were not.
- 38:51Problems could be expressed as procedures, and procedures could
- 38:56be changed without rebuilding the machine.
- 39:00As we drift backward from this point, we begin to leave AI as a
- 39:06named field and return to the broader foundations, computation
- 39:12itself, information as a measurable quantity, and the
- 39:17mathematics of uncertainty. We will pass through war time
- 39:24and post war engineering through the careful articulation of what
- 39:29a universal machine means and into the earlier centuries where
- 39:36logic and calculation were first formalized.
- 39:41For now, let the earliest learning machines feel like
- 39:46small lanterns, not spotlights. They did not solve everything,
- 39:53but they lit a path, and we continue stepping backward
- 39:59toward the deeper roots of the idea that thought can be
- 40:02mechanized. We stepped back into the era
- 40:06when computation itself was being clarified, not as a single
- 40:12device, but as a concept. Early electronic machines were
- 40:18powerful. But the deeper breakthrough was
- 40:21the realization that many different tasks could be
- 40:25expressed in the same abstract language.
- 40:29Symbols manipulated by rules. A computation is a procedure.
- 40:37A procedure can be written down, and if it can be written down
- 40:42precisely, a machine can carry it out.
- 40:47This is where the idea of a universal machine becomes so
- 40:52important. Rather than building a separate
- 40:56device for every task, you could build 1 machine capable of
- 41:02reading a description of any task and performing it.
- 41:06In practice this meant storing instructions in memory alongside
- 41:12data so the machine could be reconfigured by software.
- 41:19This is one reason computers became general purpose tools.
- 41:24They were no longer just calculators, they were engines
- 41:28for procedures. Alongside computation grew
- 41:33information theory, a way to quantify messages.
- 41:39If you treat communication as signals that can be noisy, then
- 41:44you can ask how much information is contained in a message, how
- 41:49efficiently it can be compressed, and how reliably it
- 41:55can be transmitted. This is not AI in the modern
- 42:00sense, but it becomes part of the foundation.
- 42:05Compression and prediction are closely linked because to
- 42:09compress well, you must capture patterns.
- 42:13To capture patterns is, in a sense, to predict, and beneath
- 42:18both computation and information lies probability, the
- 42:23mathematics of uncertainty. Long before computers, thinkers
- 42:30developed ways to reason about chance and incomplete knowledge.
- 42:35They built tools for estimating what is likely and for updating
- 42:40beliefs when new evidence arrives.
- 42:44These tools, when combined with computers, became statistical
- 42:50learning systems that do not claim certainty but offer well
- 42:56calibrated guesses. As we drift backward, the modern
- 43:01machinery becomes quieter. We are leaving behind racks of
- 43:07servers and entering older rooms, chalkboards, mechanical
- 43:11calculators and the early formal languages of logic.
- 43:17Soon we will reach the 19th century's dreams of programmable
- 43:23machines, then earlier centuries where logic was turned into
- 43:28algebra and uncertainty was given numbers for to night.
- 43:33Notice how gentle this chain is. Each step is a small
- 43:40clarification. What is a procedure?
- 43:44What is a message? What is uncertainty?
- 43:49The world slows and the questions become simpler.
- 43:54And in that simplicity, there is a kind of rest.
- 43:59We continue backward toward the older crafts that prepared the
- 44:05ground for everything that followed.
- 44:08We step back to a time when the question of intelligence felt
- 44:14less like a product feature and more like a thought you could
- 44:20hold in your hand, turning it slowly, seeing new edges in the
- 44:26lamp light. In the middle of the 20th
- 44:29century, as electronics and mathematics began to meet, Alan
- 44:35Turing asked a simple, unsettling question.
- 44:40If a machine can carry out any well defined procedure, how
- 44:46would we decide whether it is thinking?
- 44:50He proposed an imitation game as a practical lens, because we
- 44:55only ever judge a mind from the outside by the words and actions
- 45:00it offers us. Behind that question sit
- 45:04something even more foundational.
- 45:08The idea of a universal machine. Instead of building a separate
- 45:14device for every task, you could build 1 machine that reads
- 45:19instructions and follows them whatever those instructions
- 45:23describe. In practical terms, that means a
- 45:28program can be stored, copied, edited and run again.
- 45:34The machine becomes a general performer of procedures, and the
- 45:39procedures become an invisible architecture that can change
- 45:43without changing the physical hardware.
- 45:47This is the quiet pivot from a calculator to a computer.
- 45:53There is a sleepy clarity in that a computer is a device for
- 45:59symbol manipulation. It moves marks according to
- 46:03rules. Yet that simplicity is powerful
- 46:07because so much of life can be expressed as rules operating on
- 46:12symbols, arithmetic, accounting, navigation, tables, scheduling,
- 46:19and code breaking. Later, those same machines would
- 46:23be used to tune learning systems from examples.
- 46:27But the first step is simpler, to show that doing can be
- 46:34formalized, that a method can be written down South precisely,
- 46:40that it runs without interpretation.
- 46:43Still, the night invites A softer distinction.
- 46:47Repeating a procedure is not the same as understanding it.
- 46:52Turing did not claim that machines already possessed an
- 46:56inner life. He asked us to be honest about
- 47:00what we mean by mind when we judge it.
- 47:04If a system produces the same outward behavior as a thinking
- 47:09person in a particular setting, where do we draw the line?
- 47:15Is intelligence an inner glow or a pattern of responses shaped by
- 47:22context, memory and learning? As we drift backward, picture
- 47:28paper stacked in neat piles, the slow scratch of chalk, and a
- 47:35world realizing that procedures can be written, stored and
- 47:41carried out at speed, We are still in the age of circuits and
- 47:46urgency, and that is where our steps carry us next.
- 47:51Softly, without hurry, we move back into the war time and post
- 47:57war years when computation was pulled forward by necessity.
- 48:02The story is often told with grand language, but in the quiet
- 48:07truth of it, much of this era was about logistics and
- 48:12patience. Messages that needed to be
- 48:15decoded, trajectories that needed to be calculated,
- 48:20supplies that needed to be routed, patterns that needed to
- 48:25be detected in noise. Human beings did what they could
- 48:30by hand, with tables and teams, but speed mattered and error
- 48:35mattered, and so machines were built to help.
- 48:40Early electronic computers were large, hot, and temperamental.
- 48:46They filled rooms, demanded careful maintenance, and could
- 48:51fail in ways that felt almost biological, sudden, stubborn,
- 48:57unpredictable. Yet they offered something
- 49:01precious Repeatability. Once a procedure was set up
- 49:06correctly, the machine could perform it over and over without
- 49:12growing tired, without losing its place, without drifting into
- 49:18distraction. This repeatability became a kind
- 49:23of trust. A key practical shift was the
- 49:27stored program idea, instructions and data living
- 49:33together in memory. When you can store a program,
- 49:37you can change it without rewiring a machine.
- 49:41Software becomes a new kind of craft.
- 49:45It also becomes a new kind of vulnerability because bugs are
- 49:52not mechanical, they are conceptual.
- 49:56A single misplaced step can send an entire calculation quietly
- 50:01off course. This pushed early programmers
- 50:05toward discipline, checking, testing, documenting, building
- 50:12confidence through careful repetition.
- 50:15And beneath the engineering ran a philosophical undercurrent.
- 50:21If a machine can follow a procedure, Florida State, then
- 50:25the boundary between calculation and reasoning begins to blur,
- 50:32not because a machine feels, but because many tasks we once
- 50:38thought required intelligence can be broken into steps.
- 50:43The more steps we can specify, the more we can delegate.
- 50:49This does not reduce human thought to mere procedure, but
- 50:53it does reveal that procedure can go surprisingly far.
- 50:58As we drift backward, we feel this era like a corridor between
- 51:04worlds. On one side, modern computing,
- 51:08flexible and general. On the other, older traditions
- 51:13of measurement and table making, where humans built knowledge
- 51:18into lists, charts, and hand computed reference books.
- 51:23We'll keep stepping away from humming rooms of electronics
- 51:28toward the quieter idea that information itself can be
- 51:32counted, and that messages can be treated like measurable
- 51:37things. Now we step back into a moment
- 51:42when information became something you could measure with
- 51:47the same calm certainty as weight or distance.
- 51:53Claude Shannon, working in the middle of the 20th century,
- 51:58showed that messages can be treated mathematically.
- 52:03A message in this view is not defined by its meaning, but by
- 52:10its structure, by the set of possible signals, and how
- 52:14surprising a particular signal is.
- 52:18The more surprising, the more information it carries.
- 52:23This is a strange idea at 1st, and then it becomes beautifully
- 52:29simple. With that shift came a new way
- 52:32to think about noise. Every channel, wire, radio wave,
- 52:37paper, air adds distortions. Information theory asks how much
- 52:44can you compress a message without losing it, and how much
- 52:49redundancy do you need to add so it survives the noise.
- 52:54It is a theory of limits and possibilities, and in a quiet
- 52:59way it connects to learning. Because compression and
- 53:05prediction are close cousins. If you can predict what comes
- 53:11next, you do not need to transmit it in full.
- 53:16If you cannot predict, you must send more detail.
- 53:21This doesn't mean that meaning disappears, it means that before
- 53:27meaning there is pattern. A language has regularities.
- 53:33A melody has regularities, A stream of numbers has
- 53:40regularities. Once you see inflammation as
- 53:44pattern plus surprise, you begin to notice that many problems are
- 53:50really about capturing structure, finding what is
- 53:54stable beneath what is noisy. The practical results were
- 54:00immense. Communication systems improved,
- 54:04data storage improved, codes and checksums improved, and a
- 54:09conceptual bridge was built between the world of signals and
- 54:14the world of computation. Bits became a common currency,
- 54:20and with a common currency, different domains could speak to
- 54:25one another. A message could be an image, a
- 54:29sentence, a measurement, a command.
- 54:33Everything becomes encodable. To night.
- 54:38We let this become a gentle thought, the world as a stream
- 54:44of signals and the human mind as something that has always tried
- 54:49to find the signal inside the noise.
- 54:53We are walking backward. So soon we will leave the
- 54:58electronics behind and move toward the older mathematics
- 55:03that taught people to live with uncertainty, probability,
- 55:08statistics, and the careful art of inference long before data
- 55:14meant the modern kind of data at all.
- 55:17We step back into the older habit of making sense of
- 55:22uncertainty. Before computers, before
- 55:26circuits, before stored programs, people still had to
- 55:32decide with incomplete information.
- 55:36Merchants weighed risk. Ship captains weighed weather.
- 55:42Governments counted populations imperfectly.
- 55:46Physicians made judgments from partial symptoms.
- 55:50The world was never fully known, so mathematics learned to make
- 55:55room for maybe. Statistics, in its broadest
- 56:00sense, is a language for describing variation.
- 56:05Not just the average, but the spread, not just the typical,
- 56:10but the outliers. Over centuries, people develop
- 56:15tools for summarizing data and drawing conclusions from
- 56:20samples, learning that a small set of observations can hint at
- 56:26a larger truth but can also mislead.
- 56:31The discipline grew alongside practical needs, astronomy,
- 56:36navigation, insurance, public health and economics.
- 56:42When you measure repeatedly, you see error.
- 56:46When you see error, you want a way to account for it.
- 56:50One important idea is estimation.
- 56:54If you have noisy measurements, what is the best guess of the
- 57:00underlying value? Another is confidence.
- 57:04How sure can you be, given the number of observations in their
- 57:10variability? These ideas were refined by
- 57:14generations of mathematicians and scientists, and later became
- 57:19essential to machine learning, because learning is, at heart,
- 57:25the act of fitting a model to data that is never perfect.
- 57:30And there is a humbling lesson here.
- 57:33The more carefully you measure, the more you notice uncertainty
- 57:39not as a failure, but as a feature of reality.
- 57:44The world is not a clean equation.
- 57:47It is an ongoing process with randomness, hidden causes, and
- 57:52limits to observation. Statistics teaches you to build
- 57:57models that are not rigid claims but structured guesses, useful,
- 58:03revisable, and aware of their own fragility as we drift
- 58:08backward. Imagine candlelight over a desk,
- 58:12a hand drawing columns of figures, a careful mind checking
- 58:18sums twice. Long before machines could
- 58:21learn, humans were already practicing a kind of learning,
- 58:26updating their beliefs when evidence changed, refining
- 58:30methods, improving predictions. In a moment, we'll step into the
- 58:37age when measurement expanded, when States and institutions
- 58:42began to collect numbers at scale, creating the early raw
- 58:47material from modern data thinking.
- 58:50Now we step back into an age of counting.
- 58:55Not the fast counting of computers, but the slow
- 58:58institutional counting of societies as they grew more
- 59:03complex. Censuses, tax records, shipping
- 59:08logs, mortality tables, weather Diaries.
- 59:11These were early data systems created for practical reasons,
- 59:16governance, trade, planning and survival.
- 59:21Each record was a small act of translation, turning lived
- 59:26reality into symbols on paper. This era matters because it
- 59:32changed what people believed was possible.
- 59:36When you keep records across years, patterns emerge.
- 59:40Prices rise and fall. Birth rates shift.
- 59:45Diseases move through cities. Harvests correlate with
- 59:50rainfall. Trade routes respond to war and
- 59:54policy. The patterns are never perfect,
- 59:59but they are visible enough to guide decisions in a quiet way.
- 1:00:05Large scale record keeping becomes a tool for prediction.
- 1:00:11It also shapes the culture of measurement.
- 1:00:14Standard units are adopted. Instruments improve.
- 1:00:19Time keeping becomes more precise, maps become more
- 1:00:25accurate. The world becomes increasingly
- 1:00:29legible to itself, not in a complete sense, but in a way
- 1:00:35that supports coordination. And coordination is one of the
- 1:00:40hidden roots of later computation.
- 1:00:43If many people agree on formats and standards, information can
- 1:00:49travel farther and be reused. You can feel tonight how gentle
- 1:00:56this is. A Ledger is not a machine, but
- 1:01:00it is a memory. A table is not an algorithm, but
- 1:01:06it is a method. A standardized form is not
- 1:01:10intelligence, but it is a kind of shared structure.
- 1:01:15These are the quiet scaffolds that let later systems scale.
- 1:01:20When computers arrive, they will inherit not only mathematical
- 1:01:26ideas but administrative habits. Categorizing, indexing,
- 1:01:33summarizing, revising. As we continue backward, we
- 1:01:38approach a time when probability itself was being formalized,
- 1:01:45when chance stopped being merely a feeling and became something
- 1:01:49you could write down and reason about.
- 1:01:53That shift begins with games, with wagers, with the simple
- 1:01:59human desire to know what is likely.
- 1:02:02Even when the future refuses to be certain, we step back into
- 1:02:08the early mathematics of chance, when probability began to take a
- 1:02:14recognizable shape. The setting is surprisingly
- 1:02:19ordinary. Games, wagers, questions of
- 1:02:24fairness, and the restless curiosity of people who wanted
- 1:02:29to understand what random might mean.
- 1:02:34When you roll dice or draw cards, you cannot predict the
- 1:02:38next outcome, but you can predict the balance of outcomes.
- 1:02:43Over time, that balance, this steadying tendency, became a
- 1:02:49kind of comfort for the mind that likes order.
- 1:02:54Probability gives you a way to speak about the future without
- 1:02:58pretending to control it. It does not say what will
- 1:03:02happen, but what tends to happen and how strongly.
- 1:03:08It teaches you to separate possibility from likelihood.
- 1:03:14It also teaches you to update when new evidence arrives.
- 1:03:19The best view of the world should change.
- 1:03:22This is not indecision, it is discipline.
- 1:03:27A belief that never updates is not stable.
- 1:03:31It is frozen even before later refinements.
- 1:03:36This early era created tools that would echo forward.
- 1:03:40Expected value, for example, is a way of weighing outcomes by
- 1:03:45their probabilities, turning uncertainty into a single useful
- 1:03:52number for decision making. The law of large numbers is the
- 1:03:57quiet reassurance that over many trials, averages become stable.
- 1:04:05These ideas travel far beyond games.
- 1:04:08They shape insurance, finance, science, and eventually the
- 1:04:13mathematical foundations of learning systems that choose
- 1:04:18parameters to minimize expected error.
- 1:04:22And there is a deep link between probability and prediction.
- 1:04:28To predict well is to assign reasonable probabilities.
- 1:04:33To learn from data is to adjust those probabilities as patterns
- 1:04:38become clearer. Modern machine learning often
- 1:04:42looks sophisticated, but beneath it is this ancient bargain.
- 1:04:48We accept uncertainty, and we build methods for navigating it.
- 1:04:53As we drift backward, we can imagine the gentle clatter of
- 1:04:58dice in a wooden cup, the scratch of ink as someone writes
- 1:05:04down outcomes, the patience of repeating an experiment again
- 1:05:09and again. The night invites the same
- 1:05:13patients. Your mind does not need
- 1:05:17certainty right now. It only needs a soft path to
- 1:05:22follow. And that path leads further back
- 1:05:27into the roots of logic, the older dream that reasoning
- 1:05:33itself can be written in symbols.
- 1:05:37Now we step back into the long tradition of formal reasoning,
- 1:05:43where thinkers try to describe logic as a structure independent
- 1:05:48of any particular person. Long before binary circuits and
- 1:05:53programming languages, there were arguments, syllogisms, and
- 1:05:59the careful separation of valid reasoning from persuasion.
- 1:06:04The setting could be a shaded colonnade, a monastery library,
- 1:06:11a small classroom with a slate board, places where ideas were
- 1:06:17refined through repetition. Logic at its simplest, is the
- 1:06:24study of what follows from what. If certain premises are true,
- 1:06:31what conclusions must also be true?
- 1:06:35The power of logic is that it aims to be impersonal.
- 1:06:41It is not about who speaks, but about structure.
- 1:06:45This aspiration would later become essential for
- 1:06:49computation. Because a computer does not know
- 1:06:54what a sentence means in the human sense, it can only follow
- 1:07:01formal rules. Logic is one of the oldest
- 1:07:06attempts to make rules for thought explicit.
- 1:07:11Over centuries, logical systems were debated, extended, and
- 1:07:17sometimes tangled in philosophical disputes.
- 1:07:21But the enduring contribution was the idea that reasoning can
- 1:07:27be represented. You can write down a form.
- 1:07:31You can check whether an argument matches the form.
- 1:07:36You can separate the content from the structure.
- 1:07:40That separation is a quiet kind of abstraction, and abstraction
- 1:07:46is the air computation breathes. This doesn't mean human thinking
- 1:07:52is nothing but logic. Humans also imagine, feel,
- 1:07:57intuit, and improvise. Yet logic provides a skeleton, a
- 1:08:04way to test consistency, to avoid contradictions, to make a
- 1:08:11chain of reasoning stable enough to build upon.
- 1:08:15Science leans on it, law leans on it, mathematics leans on it,
- 1:08:22and later computer science will lean on it too, even when the
- 1:08:28systems being built are probabilistic rather than
- 1:08:33strictly logical. As we drift backward, logic
- 1:08:37begins to look less like a modern discipline and more like
- 1:08:42a craft of language. Careful definitions, careful
- 1:08:47distinctions, careful steps. It is slow work.
- 1:08:54It belongs to quiet rooms, and from here our path leads to even
- 1:09:02older aids for thinking, tools for calculation, tables and
- 1:09:10devices that help people extend the reach of their minds beyond
- 1:09:16what memory alone could hold. We step back into the world of
- 1:09:22calculation before electricity, when the mind reached for tools
- 1:09:27the way a hand reaches for a railing in the dark.
- 1:09:32Long before programmable computers, there were abacuses,
- 1:09:36counting boards, and written algorithms, procedures carried
- 1:09:42in words. There were logarithm tables that
- 1:09:46turned multiplication into addition, and navigation tables
- 1:09:51that helped ships find their way across open water.
- 1:09:56These were not machines that thought, but they were
- 1:10:01extensions of thought, quiet prosthetics for memory and
- 1:10:06arithmetic. The practical idea here is
- 1:10:09simple. If a task is repeated often
- 1:10:13enough, you can codify it. You can create a method that
- 1:10:18anyone can follow step by step and get the same answer.
- 1:10:24This is the beginning of an algorithmic mindset.
- 1:10:28It does not require circuits. It requires clarity, and clarity
- 1:10:35can be written on paper in many places.
- 1:10:39Calculation was not an abstract hobby.
- 1:10:43It was survival. Merchants needed reliable
- 1:10:47arithmetic. Engineers needed measurements.
- 1:10:52Astronomers needed predictions of planetary positions.
- 1:10:58Governments needed accounts. These needs shaped a culture
- 1:11:03that valued accuracy and repeatability, and that valued
- 1:11:08the creation of tools, mental and physical, that reduced the
- 1:11:12chance of error. You can imagine the texture of
- 1:11:16this era, the grain of wood beneath an Abacus, the smell of
- 1:11:22ink, the thinness of paper worn by repeated handling.
- 1:11:28People built tables the way later engineers would build
- 1:11:32software libraries, shared resources that make future work
- 1:11:37easier. And when a table contains enough
- 1:11:40of the world, a kind of compression happens.
- 1:11:45Complex reality becomes a page you can consult.
- 1:11:51As we drift backward, calculation becomes slower still
- 1:11:57and instruments become more mechanical.
- 1:12:01Clocks, gears, escapements, devices built to keep steady
- 1:12:06time and steady motion. Precision engineering begins to
- 1:12:12grow. It is a quiet prelude to the age
- 1:12:16when someone will dream not only of calculating but of building a
- 1:12:21machine that can follow sequences of instructions,
- 1:12:26programmable machinery waiting just ahead of us on the backward
- 1:12:31road. Now we step back into the craft
- 1:12:36of precision, where gears and springs taught people that
- 1:12:40regularity can be engineered. Time keeping is one of the
- 1:12:46oldest demands for dependable mechanism.
- 1:12:50A good clock is a promise. The world can be measured
- 1:12:55consistently, and once you can build consistent motion, you can
- 1:13:01begin to build repeatable operations.
- 1:13:05Turns, clicks, steps, carries. The physical world becomes a
- 1:13:11place where procedures can live. This is not yet computation in
- 1:13:17the modern sense, but it is the soil that computation grows
- 1:13:22from. Mechanical devices can embody
- 1:13:26rules. A gear ratio is a rule.
- 1:13:31An escapement is a rule. A Cam that lifts a lever at a
- 1:13:37particular point in a rotation is a rule.
- 1:13:41In the quiet language of machinery, cause and effect are
- 1:13:47made dependable, and dependability is what makes long
- 1:13:52chains of operations possible without constant human
- 1:13:57correction. Alongside clocks came other
- 1:14:00instruments, balances, sextants, compasses, measuring chains,
- 1:14:08surveying tools. Each one is a way of turning a
- 1:14:13messy world into a readable quantity.
- 1:14:18Each one makes it easier to coordinate across distance and
- 1:14:23time. When you can trust measurements,
- 1:14:27you can trust plans. When you can trust plans, you
- 1:14:32can build larger systems. Modern computing inherits this
- 1:14:36sensibility. Accuracy matters, repeatability
- 1:14:41matters, and error must be accounted for.
- 1:14:46There is also something psychological here.
- 1:14:50A well made mechanism invites a certain kind of wonder.
- 1:14:56It moves with purpose, yet without intention.
- 1:15:00It seems alive though it is not. That feeling of lifelike motion
- 1:15:07emerging from parts is an old human fascination, and it will
- 1:15:14soon lead us into the era of ambitious mechanical calculation
- 1:15:19and early programmability, where brass and steel begin to imitate
- 1:15:25not only motion but method. For now, let the image be
- 1:15:30simple. A quiet workshop, the faint
- 1:15:34sound of a ticking clock, metal cooling after being shaped, a
- 1:15:40steady rhythm that does not rush.
- 1:15:44We are moving backward and the world is becoming older and
- 1:15:48slower, but also clearer in its foundations.
- 1:15:54In the next stretch, we will approach the 19th century's
- 1:15:58dream of programmable machinery and the careful steps that made
- 1:16:04it imaginable. We've arrived at a century that
- 1:16:09loved gears the way our own century loves code.
- 1:16:13In the 1800s, factories were learning rhythm, railways were
- 1:16:18learning schedules, and nations were learning to count
- 1:16:22themselves with ledgers and forms.
- 1:16:25In that same atmosphere of precision, Charles Babbage
- 1:16:30imagined machines that could do more than calculate a single
- 1:16:35table. He wanted machinery that could
- 1:16:39follow a sequence of operations reliably, repeatably, and with
- 1:16:46fewer human mistakes. His early vision, the Difference
- 1:16:50Engine, was aimed at producing mathematical tables, those long
- 1:16:56columns used for navigation and engineering, without the slip of
- 1:17:01a tired hand. But his larger dream was the
- 1:17:06analytical engine, a general purpose calculating machine that
- 1:17:12could be directed by instructions.
- 1:17:14The striking part is that the idea of programmability appears
- 1:17:21here in Metal and motion. Rather than rewiring the machine
- 1:17:26for each task, you could feed it different instructions and it
- 1:17:32would perform different procedures.
- 1:17:35That is a familiar idea now, but then it was a quiet leap Babbage
- 1:17:41drew on the industrial technologies of his day.
- 1:17:47Looms used punched cards to control patterns in cloth,
- 1:17:52turning design into a sequence of holes and absences.
- 1:17:58That same notion, encoding a set of steps externally, then
- 1:18:03letting a machine execute them, became one of the inspirations
- 1:18:09for early programmable devices. The machine's gears would carry
- 1:18:15numbers, but the program would be carried by the cards, an
- 1:18:20early separation between hardware and instructions.
- 1:18:24Ada Lovelace, working with Babbage's concepts, offered an
- 1:18:30insight that echoes forward. She understood that if a machine
- 1:18:36can manipulate symbols according to rules, then it may not be
- 1:18:41limited to arithmetic alone. Numbers can represent other
- 1:18:47things, notes in music, letters in a language, steps in a
- 1:18:53process. In modern terms, this is the
- 1:18:57idea that computation is not bound to one domain.
- 1:19:02It is a general method applied to encoded representations.
- 1:19:08It's important to keep the scale gentle.
- 1:19:12These machines were not fully built in the way Babbage hoped,
- 1:19:17and they did not transform society in his lifetime.
- 1:19:22But the concept they introduced is foundational, a machine that
- 1:19:27can be instructed, a system that can execute a method described
- 1:19:34in a form it can read. Tonight you can imagine brass
- 1:19:39turning slowly, card stacked like thin leaves, and a mind
- 1:19:45realizing that procedure itself can be mechanized.
- 1:19:50As we continue backward will meet an even more abstract
- 1:19:55shift. Logic becoming algebra,
- 1:20:00reasoning becoming something you can write as symbols that behave
- 1:20:06predictably. The metal will fade and the idea
- 1:20:13will remain. And in a sleep wise kind of way,
- 1:20:17there is comfort in how modest this beginning is.
- 1:20:20No flashing screens, no rush, only careful drawings and
- 1:20:25patient thought. The future arrives here as a
- 1:20:29blueprint, not a spectacle. You can let that be reassuring.
- 1:20:34Progress is often quiet at 1st, and the night is good at holding
- 1:20:40quiet things. The workshop air clears and we
- 1:20:45find ourselves in a different kind of construction site.
- 1:20:49Not gears on a table, but symbols on a page.
- 1:20:55In the 19th century, George Boule pursued a strange, elegant
- 1:21:00ambition to treat logic as a form of mathematics.
- 1:21:06Instead of arguing in sentences, you could write reasoning in a
- 1:21:12symbolic language and manipulate it with consistent rules.
- 1:21:19The way you manipulate numbers. At first it can sound dry, true
- 1:21:26and false and or not, but the leap is profound.
- 1:21:34If logical statements can be represented in a formal system,
- 1:21:39then reasoning becomes something you can calculate.
- 1:21:45Bool's algebra takes propositions and gives them
- 1:21:49operations. You can combine statements,
- 1:21:52negate them, simplify them, and check whether one follows from
- 1:21:58another. It is a way of turning the
- 1:22:01structure of thought into a kind of machinery, except the
- 1:22:06machinery is abstract. This is one of the deepest roots
- 1:22:12of later computing. Electronic circuits can be built
- 1:22:16so that switches represent states on and off, and
- 1:22:22combinations of switches implement logical operations.
- 1:22:28Bool did not design modern computers, but his framework
- 1:22:33made it natural to think of logic as something that can be
- 1:22:37embodied in a system of operations.
- 1:22:42The distance from a page of symbols to a network of gates is
- 1:22:48shorter than it appears because both are about reliable
- 1:22:53transformations. There is also a more human
- 1:22:57lesson here. Formal logic does not capture
- 1:23:02everything we do when we think, but it offers a standard of
- 1:23:08clarity. It separates persuasion from
- 1:23:13validity. It lets you see when a
- 1:23:15conclusion is supported by premises, and when it is merely
- 1:23:20surrounded by confident language.
- 1:23:24In science and engineering, that clarity becomes a kind of safety
- 1:23:30rail. It keeps complex reasoning from
- 1:23:34collapsing under its own weight in the broader backward story.
- 1:23:39This is where method becomes portable.
- 1:23:43A logical structure can be copied, taught, checked, and
- 1:23:48reused. It does not depend on one
- 1:23:52person's memory or intuition. And once methods are portable,
- 1:23:57they can be layered, one method supporting another, until a
- 1:24:03whole architecture of reasoning emerges.
- 1:24:07Tonight, notice how soothing that portability can feel.
- 1:24:12Your mind does not need to hold everything at once.
- 1:24:17It can rest on structures. It can let one thought support
- 1:24:21the next without strain. And as we keep moving backward,
- 1:24:27the symbols will loosen from strict true and false and enter
- 1:24:32the softer world of uncertainty, where beliefs can be updated and
- 1:24:39knowledge is expressed not as certainty but as probability.
- 1:24:45Imagine a quiet desk lamp pooling light over a page of
- 1:24:49tidy notation, the ink still slightly raised where it has
- 1:24:54dried. In that small circle of light,
- 1:24:59an idea is settling into place that some parts of reasoning can
- 1:25:06be made dependable. Like a well built clock, the
- 1:25:11world outside can be ambiguous, but inside the symbol system
- 1:25:16each step is crisp, and that crispness will later help
- 1:25:21machines act consistently. Before we reach the oldest myths
- 1:25:27and mechanisms, we pass through a quiet doorway that many modern
- 1:25:34fields share. The mathematics of updating your
- 1:25:39mind. Thomas Bayes is often associated
- 1:25:43with a simple, powerful idea. When new evidence arrives, A
- 1:25:49rational belief should change in a specific way.
- 1:25:55The details can be written as a formula, but the spirit is
- 1:25:59gentle. Start with what you currently
- 1:26:03believe, then revise it in proportion to how strongly the
- 1:26:07new evidence supports or contradicts it.
- 1:26:10Bayesian thinking did not appear fully formed in one moment.
- 1:26:16It sits within a broader history of probability, statistics and
- 1:26:22inference that grew from astronomy, navigation, games of
- 1:26:29chance, insurance, and scientific measurement.
- 1:26:34Over time, people realized that uncertainty is not merely
- 1:26:39ignorance, it is a measurable feature of decision making.
- 1:26:45If you can assign probabilities, you can compare choices.
- 1:26:50If you can update probabilities, you can learn.
- 1:26:55This is foundational to many approaches in machine learning,
- 1:27:00especially those that treat models as hypotheses and data.
- 1:27:07As evidenced, even when modern systems are trained with
- 1:27:11different techniques, the conceptual link remains.
- 1:27:16Learning is, at heart, the improvement of expectations.
- 1:27:22In the presence of uncertainty. You do not need perfect
- 1:27:27knowledge. You need a disciplined way to
- 1:27:31move from less informed guesses to better informed ones.
- 1:27:36Probability also teaches humility.
- 1:27:41A high probability outcome is not guaranteed.
- 1:27:46A low probability outcome is not impossible in a world full of
- 1:27:53randomness and hidden causes. Probabilistic models offer a
- 1:27:58kind of calm realism. They let you say this is likely
- 1:28:06without pretending to be omniscient.
- 1:28:10They also invite better questions.
- 1:28:13What evidence would change my mind and how much?
- 1:28:19Historically, these ideas matured as people gathered more
- 1:28:23measurements. Astronomers compared
- 1:28:26observations to predicted orbits.
- 1:28:30Surveyors dealt with error. Physicians compared treatments.
- 1:28:36Governments counted populations and discovered drift.
- 1:28:41Each domain needed tools to separate signal from noise, and
- 1:28:47probability supplied a language for that separation.
- 1:28:53As we move backward tonight, let that language soften into
- 1:28:57something almost soothing. Updating is a kind of release.
- 1:29:04You do not have to cling to an earlier belief.
- 1:29:08If the world shows you something new, you can adjust, you can
- 1:29:14loosen, you can let your mind become lighter.
- 1:29:19And with that lightness we can drift further back, away from
- 1:29:24equations and into the older human fascination with lifelike
- 1:29:29machines, automata that moved, sang and startled people into
- 1:29:36wonder. In practical terms, a prior is
- 1:29:40simply a starting point, and a posterior is the new view.
- 1:29:45After learning tonight, you can treat your own thoughts the same
- 1:29:51way. Begin where you are, then gently
- 1:29:55update toward rest, like a tide correcting its line along the
- 1:30:00shore. Each small revision brings you
- 1:30:04closer to stillness. Long before anyone spoke of
- 1:30:08algorithms, there were machines made to astonish automata.
- 1:30:16Devices that imitate life through clever mechanisms appear
- 1:30:21again and again across history. In ancient Greece, engineers
- 1:30:28like Hero of Alexandria described pneumatic and
- 1:30:35mechanical contraptions, doors that opened seemingly by
- 1:30:40themselves, figures that moved in temples, little theatrical
- 1:30:46scenes powered by weights, water and hidden tubes.
- 1:30:51Later centuries brought clockwork birds, mechanical
- 1:30:55musicians and intricate dolls that could write or play
- 1:30:59instruments, their motions guided by cams and gears.
- 1:31:05The point of an automaton was not productivity, it was wonder.
- 1:31:11It asked a quiet question without speaking it aloud.
- 1:31:17If movement can be produced by parts, what else might be
- 1:31:22produced by parts? People knew these creations were
- 1:31:26not alive and still. The mind tends to lean toward
- 1:31:31life when it sees coordinated motion.
- 1:31:35A bird that flutters, a head that turns, a hand that lifts.
- 1:31:44The illusion is thin, and yet it is emotionally potent.
- 1:31:50Technically, automata are systems of stored behavior.
- 1:31:55A Cam profile encodes emotion over time.
- 1:31:59A gear train transforms one rhythm into another.
- 1:32:04A spring stores energy, then releases it in a controlled way.
- 1:32:10In modern language, you might call this a program, except the
- 1:32:15program is carved into metal. The behaviour repeats because
- 1:32:22the structure repeats. There is no learning here, no
- 1:32:27adaptation, only design. But the idea that behaviour can
- 1:32:32be encoded is a crucial ancestor of later computation.
- 1:32:38Automata also influenced culture.
- 1:32:42They appeared in courts and exhibitions, symbols of mastery
- 1:32:47over matter. They lived in stories as well.
- 1:32:51Artificial servants, enchanted statues, crafted companions.
- 1:32:58Even when the details were fantastical, the fascination was
- 1:33:04real. Humans have long imagined minds
- 1:33:08and bodies as things that might be assembled.
- 1:33:12Tonight. We can hold that fascination
- 1:33:16gently without turning it into drama.
- 1:33:21Imagine a quiet hall after visitors have gone home.
- 1:33:26A mechanical bird sits still on its perch.
- 1:33:31A clockwork musician is silent. The gears are at rest and the
- 1:33:37illusion sleeps. In that stillness, we can sense
- 1:33:42why these devices mattered. They trained the imagination to
- 1:33:47see motion as a result of method.
- 1:33:50And as we drift backward one last time, we leave the
- 1:33:55workshops and exhibition halls behind.
- 1:33:59We return to the oldest technology of all, shared
- 1:34:03stories passed from voice to voice, compressing experience
- 1:34:09into language so that a community can remember what one
- 1:34:15person alone cannot. In the 18th century, makers like
- 1:34:21the Jacques Draws family built celebrated automata, the writer,
- 1:34:27the draftsman, the musician, showing how far meticulous
- 1:34:32mechanism could go when paired with patience.
- 1:34:36The lesson was not that metal had become human, but that
- 1:34:41sequence and timing could be captured, preserved, and
- 1:34:46replayed with exquisite fidelity.
- 1:34:50At the far end of our backward walk.
- 1:34:53The world grows older than engines and older than ink.
- 1:34:58Here, intelligence is not a machine at all, but a habit
- 1:35:05practiced by people who had to read the environment carefully
- 1:35:10before formal mathematics. There were patterns.
- 1:35:14The stars rose and fell in familiar roots.
- 1:35:18The moon waxed and thinned. Seasons returned, not perfectly
- 1:35:24but reliably enough to guide planting and migration.
- 1:35:28Wind changed with geography. Tides followed rhythms.
- 1:35:32Animals moved with weather. To notice these regularities was
- 1:35:39to survive in this earliest layer.
- 1:35:43Prediction lived inside the body.
- 1:35:46You learned the feel of approaching rain in the air's
- 1:35:50pressure. You learned the smell of water
- 1:35:53near by. You learned which clouds meant a
- 1:35:57storm and which meant nothing at all.
- 1:36:02Then, slowly, communities learn to share these observations.
- 1:36:07One person's experience became a group's memory.
- 1:36:12That is where language begins to behave like a tool.
- 1:36:17It carries more than emotion. It carries compressed knowledge.
- 1:36:22A story can hold a map. A myth can hold a warning.
- 1:36:27A proverb can hold a rule of thumb.
- 1:36:32Oral tradition is not random entertainment.
- 1:36:35It is a storage system tuned for minds that remember through
- 1:36:41rhythm and repetition. In that sense, storytelling is
- 1:36:46an ancient form of compression. It removes unnecessary detail
- 1:36:52and keeps what tends to matter, the pattern beneath the noise.
- 1:36:59This is one reason the human mind is so at home in narrative.
- 1:37:04We'd learned long ago to store the world in sequences of words.
- 1:37:11When you look back from modern AI to this older place, a
- 1:37:15continuity appears. Machines today learn from data
- 1:37:22by adjusting internal parameters humans long ago learned from
- 1:37:28experience by adjusting expectations.
- 1:37:33Both are forms of fitting ways of becoming better at
- 1:37:39anticipating what comes next. The difference is not only power
- 1:37:45or speed, but meaning. Human learning is tangled with
- 1:37:51desire, responsibility, and care.
- 1:37:56It changes who we are, not just what we predict.
- 1:38:00Tonight, we can let that be a calming thought.
- 1:38:05The newest tools are built on the oldest materials, pattern,
- 1:38:10language, memory, and the patient urge to make sense of an
- 1:38:15uncertain world. And if the modern world
- 1:38:19sometimes feels too bright with its inventions, you can remember
- 1:38:24how deep the roots really are, and how quietly they began.
- 1:38:30Somewhere long ago. A small group sat near firelight
- 1:38:35and listened as a voice traced the shape of a river, the path
- 1:38:40of a herd, the safest crossing after rain.
- 1:38:45The listener did not download knowledge.
- 1:38:49They absorbed it slowly with trust and carried it forward.
- 1:38:54Tonight, you can do the same with calmness.
- 1:38:58Let the timeline fade. Let the gears stop turning.
- 1:39:02Let the cymbals loosen and fall like soft snow.
- 1:39:08Your mind doesn't have to solve anything.
- 1:39:11It only has to rest, held by the same gentle rhythms that held
- 1:39:16our ancestors. The turning sky, the returning
- 1:39:21seasons, the slow breath in the dark.
- 1:39:25Good night.