$cat ~/posts/historia-da-inteligencia-artificial

bilingual · en / pt

← /blog

artificial intelligence didn't begin with ChatGPT

AI and machine learning#history#learning

In the post "you're not an AI expert", I promised to tell the history of AI from the beginning, in the same format as the cryptography post. That one already covered Turing. What remained was to follow the story through neural networks, language models and the agents we use today.

The frustration was the same: "AI" had become a synonym for chatbot, and knowing how to use the tool had started to pass for knowing the subject. But opening ChatGPT doesn't explain how a machine learns, why it needs so much data, or how it can write an answer that's convincing and wrong.

I'm still studying this on my own. I'm neither a computing historian nor an AI researcher; the sources are linked throughout the text. There are many people and paths in this history, and I'll explain each idea before relying on its acronym. You don't need programming or mathematics to follow along.

The first part follows the field's precursors: how to represent knowledge, execute procedures and, later, learn from examples. Dartmouth marks the formation of a shared agenda, but those questions had already emerged along different paths. We'll follow both the methods and the people, institutions and resources that made them possible.

You can read in order or jump to learning from examples, language models and agents, or institutions, resources and impacts. The "further reading" blocks stay expanded; the "skip further reading" link takes you to the next passage. At the end, the map, chronology and glossary are there for reference. Beside the controls, the experiments explain what they actually calculate and what they only illustrate.

The question running through this post is: what does the machine do, how does it do it, and what does that still not demonstrate? It applies to a seventeenth-century calculator, a neural network and an agent in 2026.

before AI: making reasoning something we can manipulate

the shape of an idea

Before computers existed, people were already trying to describe the structure of reasoning. If that structure could be repeated, perhaps reasoning wouldn't depend on having a good idea every time.

Consider these two statements: “every mammal is an animal” and “every dolphin is a mammal.” They let us conclude that every dolphin is an animal. You don't need to know every dolphin in the ocean to get there. The conclusion depends on the relationship between categories, not on fieldwork carried out at that moment.

Now replace the names: “everything in category A is in category B” and “everything in category C is in category A.” We can still conclude that C is in B. The content is gone, and a form of reasoning remains. Logic studies this kind of structure: which conclusions follow from which premises, and under what conditions.

Aristotle studied these forms of inference long before AI existed as a research field. Calling his work artificial intelligence would be anachronistic, but the question was already there: what lets us move from one statement to another? If we describe the form precisely, perhaps a mechanical operation can carry out part of it.

But a conclusion can follow correctly from false premises. If I say that all birds live on Mars and that a macaw is a bird, the rule delivers an absurd conclusion without making an error in the transition between statements. The validity of reasoning and the truth about the world are different questions. A machine that applies rules flawlessly doesn't thereby gain magical access to correct facts.

An expert system can apply a correct rule to a wrong fact. Later, we'll encounter a similar problem in language-model answers and agent actions.

changing the names doesn't change the relationship. formalizing means making explicit the structure that allows us to move from one statement to another.

what if arguing could become calculating?

In the seventeenth century, Gottfried Wilhelm Leibniz imagined something ambitious: a symbolic language that could represent ideas precisely, and a calculus that could operate on them. Instead of getting lost in words that meant different things to each person, two people could clarify the representation and calculate its consequences.

That ambition had two halves. The first was representing knowledge. The symbol for “animal,” for example, would need a sufficiently definite meaning. The second was establishing permissible operations on the symbols. A collection of labels doesn't reason by itself; it needs relationships and rules.

Leibniz also worked on mechanical calculators. His project for universal reasoning remained an ambition, far from a modern computer. He wanted to extend calculation beyond arithmetic. Perhaps parts of reasoning could be treated as calculation. The account of his work in the Stanford Encyclopedia of Philosophy places that ambition in its intellectual context.

In the nineteenth century, George Boole gave logical operations an algebraic form. As a first approximation, imagine two conditions: “the door is closed” and “the alarm is on.” Each can be true or false. AND requires both; OR accepts at least one; NOT reverses a condition.

This is familiar even if you've never programmed. To qualify for a promotion, you might need to be a registered customer and spend more than a certain amount. For an alarm to sound, the door opening or a sensor detecting motion might be enough. Logic makes these combinations explicit. Boole gave those operations a systematic mathematical treatment. The historical reference on Boole also distinguishes his work from later formulations of Boolean algebra.

Gottlob Frege extended formal logic's reach with his Begriffsschrift, published in 1879. The difference that matters here is moving from simple combinations to a language that can discuss objects, properties, relationships and quantities: “someone exists,” “for every object,” “this person is that person's mother.” A more expressive language represents more things, but also demands more work from whoever operates on it. The reference on Frege documents that contribution.

The sequence so far is this: recognize a form of inference, write it in symbols and define the permissible operations. We can then ask whether a procedure could do part of the work.

It also exposes a problem that has never left AI: representing the world takes work. “The glass is on the table” sounds simple. But what counts as a glass? Which table? Is it on top, resting on the edge, inside a box that's on the table? Good representations simplify a task. Bad ones hide precisely the detail that matters.

the machine that didn't always have to do the same thing

Machines built to imitate actions also existed long before anyone called them AI. The automaton of Diana on a stag, from around 1620, could move across a table. Gears produced a prepared sequence; the appearance of action doesn't demonstrate learning. That tradition helps distinguish three things we'll keep encountering: automating, imitating and learning.

Another tradition ran in parallel: building machines that performed work. Pascal's and Leibniz's mechanical calculators automated arithmetic, but automating an operation differs from having a machine whose behavior can be described by a program.

A music box is a useful comparison. It has a mechanism and a sequence that determines what it plays. Replacing the sequence changes the music without requiring you to reinvent the gears. The machine stays the same; the instructions change.

In the 1830s, Charles Babbage designed the Analytical Engine. The design distinguished storage, operations and instruction control; punched cards would help direct the work. These are conceptual relatives of the memory, processor and program in today's computers. But Babbage didn't build the complete machine. The distinction between a design and a working machine matters here. The Computer History Museum helps separate them.

In her notes on the Analytical Engine, published in 1843, Ada Lovelace recognized that the design could work with things other than numbers. If something could be represented by symbols, a symbol-manipulating machine might work with more than numerical quantities. Music is one example associated with the notes: relationships between sounds could, given suitable representations, enter that world of operations.

To work with music, the machine would need a representation of sounds and operations on it; a musician's experience is a different matter. Lovelace also discussed the limits of what the machine could originate on its own. Following instructions and producing a result that surprises their author were already questions before the electronic computer.

For example, a number can represent a musical note, a color or a letter. “Change this number” means different things depending on the representation. If symbol 65 represents the letter A, adding one may point to B. If it represents a pixel's intensity, adding one brightens part of an image. The mechanism operates on numbers; their meaning depends on the system in which they're used.

This helps explain why a computer can handle images, text, audio and money. It doesn't need a different physical gear for every subject. It needs a way to represent the subject and a suitable procedure. That generality will be decisive for AI: ideas about language, games, logic and learning can be tested on the same class of machine.

But which problems can a machine solve by following instructions? Having a program doesn't yet answer that.

what a machine can do

recipe, proof and limit

A procedure is a sequence of steps that doesn't depend on inspiration at every attempt. The recipe needs to say enough for someone to follow it. “Make good dough” isn't a recipe; “mix these quantities for this long” begins to be one.

In mathematics, there's a stricter version of this question: is there a procedure that always terminates and correctly decides a class of problems? Solving a few examples isn't enough. A rule that works when I choose favorable cases isn't a general solution.

David Hilbert was central to a program seeking secure formal foundations for mathematics. Among its questions was the possibility of general procedures for deciding certain logical problems. If all mathematical reasoning could be transformed into symbols and rules, perhaps deciding what follows from those symbols could be automated.

In 1931, Kurt Gödel showed limits of sufficiently expressive formal systems. The conditions are very specific: the results don't say that "nothing can be proved" or that "every opinion is equally valid." The incompleteness theorems show that certain consistent, effectively axiomatized systems cannot prove every arithmetical truth expressible within them, and establish a limit on proving their own consistency under precise assumptions. The Stanford Encyclopedia of Philosophy's explanation spells out those conditions.

Even without following the proof, we can understand the consequence: powerful rules aren't guaranteed to settle every possible question about themselves. The history of formalization encountered internal limits, and faster electronics didn't make those limits disappear.

In the 1930s, Alonzo Church and Alan Turing gave rigorous descriptions of computation. Church worked with lambda calculus, a mathematical language of functions. Turing described an abstract machine with a tape, symbols, states and movement rules. The formalisms proved equivalent. The claim that they capture every effective procedure is the Church–Turing thesis, not a theorem about an informal definition of computation.

The Turing machine was a mathematical model for studying procedures, much simpler than a physical computer. Imagine a tape divided into cells and a head that reads one cell at a time. Depending on the symbol and its state, it writes something, moves and changes state. It's simple enough to analyze and general enough to represent complex computations.

an illustrative configuration: symbol and state determine the operation. the tape isn't a human memory; it's part of a formal model of computation.

Turing proved that some problems have no general decider of this kind. One familiar result is the halting problem: no single algorithm can correctly determine, for every program and input, whether that program will terminate or run forever. A faster processor doesn't solve this limit.

You can still determine whether a particular program terminates and analyze entire classes of programs. What fails is the promise of one method that correctly decides every case.

from logic to electricity, from electricity to feedback

Claude Shannon established a fundamental connection between logical operations and switching circuits. A condition that's true or false in algebra can correspond to an electrical configuration. Combinations of circuits can perform operations that, on paper, were symbolic relationships. His thesis, with a title page signed in 1937, documents that bridge.

Certain physical behaviors can implement formal operations. This lets us build chains of operations in hardware. Shannon was one contributor to a development involving many people and architectures.

Shannon also formalized a mathematical theory of communication in 1948. Here, information has a technical meaning involving uncertainty and message transmission. Measuring information doesn't measure wisdom, intention or understanding. A file can carry many bits and say little anyone cares about; that difference matters when, decades later, we discuss models processing enormous volumes of data.

Under the name cybernetics, Norbert Wiener organized the study of control, communication and feedback in machines and organisms, presented in his 1948 book. “Feedback” may sound like something from a meeting, but here it's a simple mechanical idea: observe the result of your action and use that observation to adjust the next action.

A thermostat is one example. It measures temperature, compares it with the target and controls heating. Is the room too hot? The system reduces or turns off the heat. Too cold? It acts again. Its response depends on the observed state of the environment. That's different from simply running a heater for two hours and hoping.

The thermostat adjusts its action according to what it observes. This cycle is an idea from cybernetics that runs through industrial control, robotics and learning, without requiring us to consider the thermostat intelligent. Later, agents will also read results and adjust plans, using different technology to organize a similar cycle.

feedback returns information about the result. the cycle allows course correction; by itself, it implies neither learning nor understanding.

neurons as a model, not a copy of the brain

In 1943, Warren McCulloch and Walter Pitts published work connecting networks of idealized neurons with logic. A neuron in that model was a mathematical simplification: receive signals and produce an output under defined conditions. Networks of those units could perform logical operations.

Connecting simple units could also produce computational behavior. There was no need to write each operation as a high-level rule. That didn't make the model a faithful reproduction of a brain. Biological neurons have dynamics and contexts that a threshold unit omits. That's where the inspiration for the name comes from. McCulloch and Pitts's original paper lets us examine those idealizations.

In 1949, Donald Hebb proposed a theory of changes in connections between neurons associated with experience: a connection's strength could change with activity. The familiar phrase “neurons that fire together wire together” is a later simplification. A copy of Hebb's book is preserved by the Max Planck Institute.

Think of a network of paths. If every connection's influence stays fixed forever, behavior changes only when someone changes the design. If those influences can adjust to examples or results, behavior itself can be modified through learning. We still need rules that make those adjustments work, but the possibility is there.

There are already lineages that will later be treated as competitors: operating on explicit symbols, reacting through control cycles, learning relationships from data and adapting networks of connected units. They weren't all born at one conference. They converged in an environment that now also had digital computers on which ideas could be tested.

One such experiment took physical form in 1951: SNARC, built by Marvin Minsky and Dean Edmonds. Circuits represented connections, and a reinforcement mechanism modified recently used paths. Minsky's own biography and his thesis document the work. It was an experimental machine: networks had physical implementations before Dartmouth, still far from reproducing a brain.

Turing changes the question

In 1950, Turing published Computing Machinery and Intelligence. To reframe the question of “thinking,” he began with a game involving a man, a woman and an interrogator, then asked what would happen if a machine took one of the roles. The best-known interpretation asks whether the interrogator could distinguish a machine from a person through textual answers.

We can observe the answers and describe the experiment even while disagreeing about what thinking means. The test that came to bear his name offers a way to discuss the question, not a definitive certificate of intelligence, consciousness or understanding.

A machine can be convincing because of the evaluator's limitations, conversational strategy, a narrow context or real language capability. And a system can solve an important scientific problem without conversing like a person. Being indistinguishable in a conversation is therefore evidence about a situation, not a universal measure of everything we call intelligence.

The paper also discusses learning. Turing didn't imagine only hand-writing every answer for an adult machine; he considered building something that could be educated. That brings the question of intelligence closer to the question of change through experience. The 1950 text is worth more than the one-sentence summary that usually circulates.

The connection with cryptography comes through this interest in procedures, constraints and inference. At Bletchley Park, the work against Enigma was collective: it built on contributions from the Polish mathematicians Marian Rejewski, Jerzy Różycki and Henryk Zygalski, as well as Turing and other British researchers. Exploiting constraints to rule out configurations is an example of structured search. This reveals a continuity of questions, but doesn't automatically turn cryptanalysis into machine learning.

AI will include methods that search for solutions by following rules and methods that adjust parameters from examples. Sometimes they work together. Representation, search and computation are meeting points between these areas, built by many people.

a field gets a name, and ambition grows

Dartmouth didn't start from nothing

On August 31, 1955, John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon signed a proposal for research the following summer at Dartmouth. The phrase artificial intelligence appears in the document. In 1956, the meeting helped consolidate a name and scientific identity for questions that already had a history.

The document began with an ambitious conjecture: aspects of learning and intelligence could be described precisely enough for a machine to simulate them. The proposal included language, neural networks, abstraction and machine improvement. The original preserved in McCarthy's archive lets us see that optimism without relying on a promotional retrospective.

Naming a field has practical consequences. It helps bring researchers together, request funding, open laboratories, create courses and discuss an agenda. Work that previously seemed to belong to different traditions gains a banner under which it can organize.

That also creates disputes. Which work belongs under that banner? Which approach deserves money? Are a neuron-inspired network and a theorem-proving program trying to explain intelligence in the same way? The name unites a field but doesn't settle its internal differences.

The meeting helped organize the field, but the approaches continued to diverge. The participants didn't leave Dartmouth with a plan for the next seventy years.

  1. 1936Church and Turing7 yearsformalizing procedures and investigating their limits.
  2. 1943McCulloch and Pitts5 yearsnetworks of idealized neurons connected to logic.
  3. 1948Shannon and Wiener2 yearsinformation, communication, control and feedback.
  4. 1950Turing5 yearsa behavioral experiment beyond the debate about thinking.
  5. 1955the proposal1 yearartificial intelligence appears as a research program.
  6. 1956Dartmouth2 yearsthe meeting consolidates an identity for the discipline.
  7. 1958perceptronRosenblatt publishes a learning model based on adjusting weights.
before the first great wave: computation, networks and the question of behavior already existed when the field got its name.

the program that searched for a proof

Allen Newell, Herbert Simon and Cliff Shaw developed the Logic Theorist in the mid-1950s. The program worked on theorem proving. In simple terms, it received a goal and tried to find a sequence of operations connecting available statements to the desired conclusion. The 1956 report preserved by Carnegie Mellon documents this work.

A proof can be thought of as a path. You start from accepted premises, apply permissible operations and need to reach your destination. Beyond knowing the operations, you need to decide which to try, in what order, and when to abandon a path that's going badly.

If you've ever tried a jigsaw puzzle, you know the difference. Knowing that compatible pieces can fit doesn't tell you which of hundreds to choose next. A useful clue reduces attempts. In computing, a heuristic is a practical guide to a more promising search, without necessarily guaranteeing that a choice is perfect.

The Logic Theorist produced proofs by searching paths between representations: the operations were given; finding a useful sequence was the program's job.

Newell and Simon continued this work with the General Problem Solver. The ambition was to make the solution mechanism less specific: represent states, goals and differences, choosing actions that reduced the distance between the current and desired state. “General” in the name should be read as a research objective, not the ability to solve any everyday problem.

Imagine making coffee: the initial state contains cold water, ground coffee and a kettle; the goal is brewed coffee. Heating the water makes it possible to move on to brewing. But if the power is out, heating fails and the plan needs another condition or alternative, such as an available stove. This is a teaching example, not a historical GPS execution: explicit states and actions work in a prepared problem; discovering everything that matters in everyday life greatly increases the work of representation.

The early results impressed people because they achieved something previously impossible. The work needed to move beyond problems prepared for a demonstration still had to be measured.

the program that changed with experience

Another direction became clearer with Frank Rosenblatt's perceptron, described in his 1958 paper. Instead of representing an entire chain of human rules, the model combined inputs using adjustable weights and produced a decision.

A weight determines how much an input influences the result. Imagine assessing whether fruit is ripe. Color, firmness and smell can contribute differently. With features on the same scale, a weight larger in magnitude gives smell more influence in the numerical combination; its sign can support or oppose the output. The model has neither a nose nor a concept of fruit; it receives numbers someone chose or measured to represent the features.

The training process changes those weights using examples. If the desired output was “ripe” and the model answered something else, a procedure can adjust the values to reduce later errors. This is a different way to program: a person chooses the structure, examples and learning rule, but doesn't write every final decision directly.

During training, the parameters, adjustable values such as weights, change. During inference, the model executes its operations on an input with the available parameters to produce a prediction or answer. Here, “inference” names that execution, without requiring logical reasoning.

Changing the input can also change the answer without changing weights. This distinction will be fundamental to prompts: examples in a conversation guide that execution, without thereby incorporating a permanent skill into the model.

The perceptron had important limits, which we'll examine in an experiment later. We can already compare the two programs: the Logic Theorist searched for consequences and paths using explicit representations; the perceptron adjusted a decision function from examples. These are two ways of working on AI problems, both older than chatbots.

The models used in chatbots, known as large language models (LLMs), are neural networks with many layers. Networks are a family of machine learning, which is part of AI. The diagram places this piece within the whole field; layers and language come later.

this view locates methods. vision, speech and robotics are problem areas that can use several of these pieces; they don't belong to just one box.

why the early results seemed to promise more

A computer that proves a theorem can seem close to doing science by itself. A classifier that learns from images can seem close to recognizing any object. The early results encouraged expectations on that scale.

But a carefully prepared photograph and a poorly lit object in the street demand different things from a classifier. The same goes for conversing with expected phrases versus following a story full of implicit references. In theorem proving, the premises are already defined; in a social problem, deciding which premises represent the situation may be precisely the hard part.

Every result needs to be read within its conditions. Which inputs were used? Which parts of the task had the representation already solved? Who selected the examples? How much human work surrounded it? Those questions tell us how much of the task the program solved.

intelligence written as rules

representing a world and searching for a path

To manipulate symbols, researchers also needed suitable programming languages. John McCarthy began implementing Lisp in 1958; the formal paper on its recursive functions appeared in 1960. Lisp made lists and expressions easier to work with: a statement, a set of rules or part of a program could be represented as structures another program could examine. McCarthy's account distinguishes development from publication.

Imagine storing “Ana is Bia's mother” as a list with three roles, rather than just a loose sentence. The program can locate the relationship, substitute a name or combine the structure with another. Lisp helped build and transform these representations. Knowledge still had to be put into them. That usefulness helps explain its place in symbolic AI laboratories and, later, the Lisp-machine market.

The approach known as symbolic AI represents things explicitly: objects, relationships, conditions, actions and rules. “A is on top of B” can be a fact; “if A is on top of B, moving B first requires removing A” can be a rule. The program operates on these representations.

The rules make part of the reasoning visible. If the system says it needs to remove a box before pulling out a shelf, we can trace the path to that decision. The facts may still be wrong, but we can examine the operations.

Imagine three stacked boxes. You want to take the bottom one to another room, but can move only one box at a time and can't move through the others. There's an initial state, a goal and permissible actions. The solution involves a sequence: remove the top box, remove the middle one, move the bottom one. The program needs to find that sequence in the space of possibilities.

A state is a relevant description of a situation. An action transforms one state into another. A plan is a sequence of actions to reach a desired condition. A machine can use these concepts without wanting to carry a box or knowing what being in a room feels like.

Finding ways to achieve a goal and organizing actions were tasks associated with reasoning. Programs could perform parts of them, using mechanisms that could differ from ours.

The difficulty appears when possibilities multiply. If each state allows ten actions, two steps can open a hundred paths; three, a thousand. Ten steps can already produce an enormous number of combinations. Not every path is different or needs exploring, but the example shows why “just try the options” quickly stops being practical.

A heuristic can guide the search. On a map, straight-line distance to the destination suggests which city to visit first. It doesn't replace the actual road, but helps avoid very bad attempts. In planning, other approximate measures can play the same role. This reduces the number of paths the program needs to examine.

even with only two alternatives per decision, the number of paths doubles at each level. the diagram shows four levels; the problem grows far beyond the page.

ELIZA: when conversation makes it seem someone is there

In 1966, Joseph Weizenbaum published a paper on ELIZA. The system transformed textual inputs using keywords and decomposition and recomposition rules. A familiar script imitated certain patterns of therapeutic conversation, often returning questions related to words in the input.

A sentence containing “my mother” could trigger an answer about family. A sentence with an expected structure could be rearranged into a question. This wasn't a neural network trained on books or a model with a general understanding of the situation being described.

The paper by Weizenbaum matters both for the mechanism and for the reaction it helps us discuss. People readily attribute intention and understanding to linguistic behavior. A seemingly caring answer can give the impression that the system understood someone's experience, even when it applied a very narrow transformation.

The tendency to see someone behind the conversation was already there. Today's language models use a different mechanism but encounter the same human reaction.

The version below searches for literal fragments and shows the rule it matched. It's a small adaptation so you can see the whole mechanism, not a reproduction of ELIZA.

simplified execution · local rules

rule-based reply

what makes you say you are sad?

matched rule

contains “sad” → return a question about sadness without analyzing negation

a pattern-based conversation inspired by ELIZA. compare ‘I am sad’ with ‘I am not sad’: the same fragment can trigger the same answer even when the sentence says the opposite.

The answer may seem relevant because you supplied a word the script recognizes. But it can fail because that word appeared in a negation, quotation or unexpected context. The program doesn't need to identify all of that to produce a grammatical sentence.

There's also a human part to how this works. The person fills in gaps: interprets a vague question as interest, imagines continuity and attributes intention. The interface contributes to that reading. The same operation, presented as a text transformation in a technical panel, may seem much less human than it does in a conversation window.

SHRDLU and the advantage of a small world

Terry Winograd developed SHRDLU in the late 1960s and early 1970s. The system handled language in a blocks world: objects with properties and relationships that could be represented explicitly. You could request operations on those objects and ask questions within that context.

Imagine asking “put the red block on the blue one.” In a blocks environment, the system can have an object list, know which are red and blue, represent spatial relationships and choose an action. A later sentence can use “it,” and context helps determine which block is meant.

SHRDLU integrated language analysis, contextual memory and actions on the environment. Winograd's own account also explains the limit: the program worked in a universe carefully prepared for its mechanisms.

If I leave the blocks and say “tidy the house for a visitor,” the meaning opens up. Which visitor? What counts as tidy? Is the box on the floor in the way, or is it where someone keeps their medication? The common-sense knowledge a person uses without noticing can be much larger than the explicit list of objects and rules.

The closed blocks world was part of the solution. Outside it, expanding the list of nouns wasn't enough to give the program the knowledge it needed.

Not everyone accepted that more rules would be enough to extend those results. In a critique published by RAND in 1965, Hubert Dreyfus questioned the assumptions behind that ambition. Knowing a situation also involves tacit skills, context and bodily experience: recognizing irony isn't simply finding the word “irony.” The critique opened a philosophical dispute about what formal representations capture. It isn't a mathematical proof that machines could never learn, but it prevents treating a blocks world as a complete sample of human life.

Shakey brings the problem onto the floor

At SRI, Shakey brought perception, representation, planning and action together in a mobile robot, developed between 1966 and 1972. The situation wasn't just a symbolic description inside a program. The robot had to deal with a physical environment and do something in it. SRI's institutional archive preserves this history.

Moving an object involves more than choosing a sentence: locate the object, estimate the space, plan a sequence and execute movement. Each stage can have its own errors. If the representation says a path is clear and a camera misses an obstacle, the plan can be logical and still fail on the floor.

In a robot, each instruction depends on the body and environment: sensors, position, response time and motion. An action correct in the plan can fail when those conditions differ from what the program assumed.

This setting produced STRIPS, by Richard Fikes and Nils Nilsson, published in 1971. An action has preconditions, what must be true to apply it, and effects, facts it adds to or removes from the state. To move a box, the robot may need to be beside it; afterward, “the box is at the origin” no longer holds and “the box is at the destination” does. The original paper formalizes this way of searching for plans.

an example inspired by STRIPS: conditions permit an action, and its effects change the state. a plan must chain actions whose conditions are satisfied.

Logic programming opened another direction. Prolog, developed in 1972 by Alain Colmerauer, Philippe Roussel and colleagues in Marseille, with theoretical exchanges involving Robert Kowalski and researchers in Edinburgh, lets you write facts and relationships and ask the system to seek a proof for a query. The creators' account and Kowalski's history preserve the collaboration between building the language in Marseille and investigating proof procedures in Edinburgh. These are participant accounts, with their own perspectives.

In a family example, supply “Ana is Bia's mother” and “Bia is Caio's mother,” plus a rule about grandmothers. The system can search for a chain answering whether Ana is Caio's grandmother. A person describes relationships and a query; the mechanism seeks combinations that satisfy it. This tradition will connect directly with Japan's investment in knowledge processing in the 1980s.

These demonstrations worked under well-defined conditions. ELIZA depended on conversation patterns, SHRDLU on a blocks world and Shakey on the robot's perception and control. Extending the results to less prepared situations would require more work than the promises suggested.

promises, experts and winters

what cooled down, exactly?

“AI winter” is a useful phrase as long as it doesn't become a worldwide season with exact start and end dates. It usually names contractions in expectations, funding and markets. Fields and countries didn't all cool down together, nor did they stop researching.

In machine translation, early demonstrations and the Cold War context encouraged interest and investment. But translation involves ambiguity, context, idioms and subject knowledge. A demonstration with carefully chosen sentences doesn't measure performance on varied documents.

One concrete demonstration was Georgetown–IBM, on January 7, 1954: an IBM 701 translated Russian sentences into English using a controlled vocabulary and rules. IBM's record and the 1955 technical account describe its conditions. The system translated that set of sentences; how it would handle varied texts remained to be investigated.

Consider “he saw the bank.” It could mean a financial institution or the side of a river. “The bank was beside the river” helps one interpretation; “the bank approved the loan” helps another. Translating individual words and rearranging structures can work in simple cases and break when the choice depends on a larger story.

The 1966 ALPAC report evaluated machine translation in the context of United States investment. It had important effects on priorities and resources. That doesn't mean machine translation ended there: computational linguistics, supporting tools and other research continued.

In the United Kingdom, James Lighthill's 1973 report criticized AI's difficulties, including scalability. The assessment was associated with British retrenchment, but was challenged at the time. McCarthy's response shows disagreements over interpretation and criteria, not a unanimous verdict that intelligent machines were impossible.

When a field depends on funding, what counts as success also depends on institutions. Demonstrating a principle, delivering a useful tool and justifying military investment are different goals. Changed priorities can reduce resources despite scientific results; results falling short of promises can make such changes more likely.

Expectations grew, results fell short of promises and institutions reconsidered resources. This cycle happened at different times and in different countries, combining technical difficulties with funding choices.

the limit of a straight line

A famous part of this history is Perceptrons, by Marvin Minsky and Seymour Papert, published in 1969. The book studied limits of classes of perceptrons. Its best-known result in popular accounts involves linear separation.

Imagine red and blue points on paper. If you can draw a straight line leaving every red point on one side and every blue point on the other, the classification is linearly separable. If the colors alternate at the corners of a square, no straight line does the job.

That's the structure of XOR: two inputs, with output 1 when exactly one is on. “Neither on” and “both on” belong to one category; “only the first” and “only the second” to another. A single linear separator cannot correctly divide all four combinations.

The result applies to that linear separator, not every neural network. An architecture with intermediate units and nonlinear operations can build a different separation. To be useful, it also needs an effective training method.

trying harder with the line won't fix it: the arrangement requires a different boundary. the next section's experiment compares the two architectures.

The book's role in that retrenchment is a historical question, not a consequence of its theorem. Mikel Olazaran, in 1996, investigated how interpretations of the controversy related to symbolic AI's authority and resource allocation. Explaining the period also requires considering hardware, training and continuing research, without assigning one cause to the entire field.

knowledge of one field, rather than intelligence for everything

Expert systems narrowed the task's ambition: capture knowledge of one field and use it for specific problems, rather than start with a general solver.

DENDRAL worked on interpreting chemical structures. MYCIN investigated rules involving infections and treatments. XCON was used to configure Digital Equipment Corporation computer systems. “Expert” came from the scope and knowledge base; it didn't mean the system mastered every subject.

How does this work? Separate a collection of facts from a collection of rules. The facts describe the current case. Rules explain how certain conditions allow further statements. An inference mechanism selects and chains rules, sometimes handling degrees of certainty.

In an invented maintenance example, “the machine won't turn on” is a fact. A rule might say that if there's no electrical power, investigate the connection before dismantling the motor. Another rule may use that investigation's result. The system builds a path guided by recorded knowledge.

A well-organized base uses domain knowledge to avoid bad paths; adding rules alone doesn't solve this.

The book Rule-Based Expert Systems, by Buchanan and Shortliffe, documents MYCIN's architecture and debates. It's important to distinguish experimental studies from routine clinical deployment. Evaluating a program in research doesn't mean it became an autonomous doctor in hospitals.

who keeps the world up to date?

Expert systems become difficult when knowledge changes or one rule interacts badly with another. A single rule may seem clear; thousands can have dependencies, exceptions and conflicts.

Capturing a human expert's experience isn't simply asking “which rule do you use?” Many people recognize situations through practice without being able to describe a complete list. Extracting, formalizing, verifying and updating knowledge takes work.

Imagine a store with rules for recommending compatible parts. One manufacturer changes a connector, another launches a similar model, a product is discontinued and commercial names change. The base needs to keep up. Otherwise, the mechanism keeps correctly applying information that no longer matches the catalog.

That maintenance adds to infrastructure costs. During the commercial wave of the 1980s, specialized Lisp platforms and machines had their place. Changes in the computer market, costs and alternative systems' performance altered that business. A technology can lose commercial viability without all the science behind it being false.

Japan's Fifth Generation Computer Systems project, launched in 1982, invested in knowledge processing, logic programming and parallelism. The Information Processing Society of Japan's record describes eleven years and approximately 54 billion yen, through fiscal year 1992. This was a scientific policy with equipment and research, not just a product promise. Its ending doesn't erase its contributions or prove logic programming useless. ICOT ran the program with MITI support; the ICOT/AIST archive preserves its research and infrastructure. The dominant commercial transformation imagined in the initial enthusiasm didn't materialize.

In the late 1980s and early 1990s, commercial expectations and maintenance costs contributed to another contraction associated with “winter.” Meanwhile, speech recognition, statistics, control, optimization and networks continued advancing. These methods grew while expert systems occupied the commercial spotlight.

learning from examples

who writes the rule?

One question has kept returning: who puts knowledge into the machine? In a rule-based system, people write much of the structure linking conditions and consequences. In machine learning, many behaviors are estimated from data.

Someone still chooses a representation, model family, objective, data and adjustment procedure. The difference is that the result needn't be a collection of individually written decisions. The procedure seeks parameters that work well on the available examples.

Let's use a message filter. Writing “if it contains promotion, mark it as spam” is a manual rule. It may catch a scam and also a legitimate message from a store you follow. A learned approach receives categorized messages and searches for useful regularities: word combinations, sender patterns or other features.

The labels are the answers supplied during training: spam or legitimate message in this example. The model predicts, compares with the labels and is adjusted to improve according to a chosen measure. This is supervised learning: input examples come with a desired answer.

A label isn't a truth that appeared from nowhere. Someone classified the messages, with criteria, doubts and possible errors. If people disagree on what counts as spam, the machine inherits a version of that dispute. Data represents the world under particular conditions.

Other methods seek structure without receiving a category for every example, such as groups of similar items. This usually falls under unsupervised learning. In other methods, the data's own structure supplies the target: hide part of a text and try to reconstruct it, for example. We call many of these approaches self-supervised.

These are ways to formulate a learning problem, and a task can use more than one. The categories don't measure how intelligent a system is.

probability becomes part of the solution

Not every problem admits a deterministic answer based on complete facts. A sensor is noisy, a word has several interpretations and data covers only some possible situations. Statistics offers ways to estimate regularities without pretending to absolute certainty.

An idea associated with Bayes is updating a belief in light of evidence. Think of a coin that might be biased. Before observing tosses, you have hypotheses about it. The results change how compatible each hypothesis seems with what happened. Updating isn't “believing whatever I want”: it involves relating hypotheses, evidence and assumptions.

Judea Pearl and other researchers developed tools for organizing probabilistic relationships. A Bayesian network represents dependencies between variables through a graph. The diagram helps organize how an observation changes estimates about other conditions. Pearl's laboratory bibliography documents that lineage.

Statistical dependence doesn't establish cause: two values can vary together because a third influences both. Inferring causality requires additional assumptions; drawing an arrow doesn't provide them.

Hidden Markov models played an important role in speech recognition. “Hidden” means the system doesn't directly observe the state it wants to infer. It observes signals produced by a sequence of states and seeks a compatible explanation.

Imagine hearing someone behind a wall. You receive an imperfect signal, not a ready-made transcript. Certain sounds and transitions are more likely than others. A model can combine those probabilities to choose more plausible sequences. Lawrence Rabiner's 1989 tutorial systematized this family's problems and methods.

These pieces give uncertainty different roles: Bayes updates beliefs, networks organize dependencies and Markov models estimate sequences of states from signals. In speech or the ambiguous “bank” sentence, a system can compare alternatives without requiring complete facts.

trees, margins and combinations of models

In language, the change also involved grammars, sentence-structure analysis and probabilistic methods. A parser seeks relationships such as subject, verb and object; a language model can estimate the probability of a continuation. N-gram models do this by observing short consecutive sequences. Brown and colleagues' statistical translation work in 1990 sought regularities in paired translations and combined probabilities with search. It didn't eliminate linguistic structure: it changed how part of it was estimated, rather than relying exclusively on expert-written rules.

Not all machine learning involves a neural network. A decision tree organizes tests in sequence: a test's answer leads down a branch until a decision is reached. Unlike a hand-written tree, an algorithm can choose the tests from examples.

In a message filter, a tree can split data by one feature and then another, seeking better separation of categories. Quinlan's work on tree induction shows how a structure resembling rules can be learned, rather than supplied by an expert.

A tree fitted too closely to a few examples may memorize details. Combining trees built in diverse ways can reduce certain weaknesses. Random Forests, associated with Leo Breiman's work, use trees and random choices to combine predictions. Breiman's original page presents the method.

Methods called support vector machines, or SVMs, seek separation with a good margin between categories. Return to the points on paper. If several lines separate the data correctly, a line squeezed against the points may be more sensitive to small changes. The margin helps choose a boundary with distance from the examples most important in defining it.

With techniques called kernels, certain relationships can be treated as separations in another representation. Representation changes the kind of boundary a model can find, although the switch doesn't make every problem simple. Cortes and Vapnik's 1995 paper is a milestone in this work.

These methods remain relevant because many problems involve tables, moderate data volumes and cost constraints. A system predicting a quantity from a customer record doesn't necessarily need a network with billions of parameters. The right technique depends on the data, objective and conditions of use.

training doesn't prove it works outside training

Training and evaluating on the same examples creates a problem. A model can become excellent at reproducing that set and poor on new cases. A child who memorizes the order of answers on a worksheet hasn't necessarily learned the operation.

That's why separating training and test data matters. Training participates in adjustment. Testing seeks to measure behavior on examples not used directly to adjust the model. Other sets can help choose configurations during development; the caution is not to turn the final test into more training material through repeated attempts.

The term overfitting means fitting training details too closely, harming generalization. Training loss can fall without improving real use: on which data is the graph measuring performance?

There's also data leakage. If information available only after a decision enters training or evaluation, the result becomes artificially easy. Predicting tomorrow with data that exists only the day after tomorrow is a misleading demonstration, even if the algorithm performs every calculation correctly.

testing seeks to measure use outside the adjustment material. it needs to represent the real problem; merely making two piles doesn't fix a poor selection of data.

how a network adjusts its own weights

During training, a network tries, measures the difference from the target and adjusts its internal numbers. It combines inputs, weights and nonlinear operations; intermediate units feed other units. Let's name the parts of that cycle.

Training starts by defining a measure of error, the loss function. It compares the prediction with the target and assigns a cost. The training procedure seeks parameters that reduce that cost on the available examples.

But a network can have many weights. How do we know which to change, in which direction and by how much? The gradient describes the loss's sensitivity to small parameter changes. Backward propagation, or backpropagation, calculates those sensitivities through the layers using the relationships between operations.

The 1986 paper by Rumelhart, Hinton and Williams was central to popularizing learned representations in intermediate layers. It wasn't the absolute invention of all reverse differentiation: there were predecessors. What gained momentum was a practical way to train networks that could represent more than a simple separator.

Imagine walking downhill without seeing the whole landscape. The local slope indicates a downward direction. Small steps can reduce elevation but don't guarantee finding the landscape's lowest point. In training, the “landscape” is error as a function of parameters, and the slopes are calculated mathematically.

In the experiment below, choose AND or XOR and compare one neuron with a network containing four intermediate units. The network receives all four examples, calculates predictions and adjusts weights. Training is real, performed here in the browser, with the same initialization whenever you reset.

simplified execution · local training

x1x2-0.990.32y

rounds: 0 · mean loss: 0.8664

inputsexpectedprobability of 1decision
0 / 0031.5%0 ✓
0 / 1138.8%0 ✗
1 / 0114.5%0 ✗
1 / 1019.0%0 ✓
mean loss · 0.8700 rounds0 rounds

output weights: -0.99 / 0.32 · output bias: -0.78

Mean loss is binary cross-entropy. Training uses all four examples in the table. Matching them does not measure generalization to other problems.

start with XOR and one neuron; train a few times. then choose the hidden layer and compare. all four cases are used in training here: the experiment demonstrates representation and adjustment, not an evaluation of generalization.

AND asks for “both on”; XOR asks for “exactly one on.” One neuron can learn a separation for AND. With XOR, trying harder doesn't remove the architectural limit. The hidden layer allows other combinations, and adjustment can reduce the error in this example.

The prediction is a probability; the table applies a threshold to decide the category. Different probabilities can produce the same decision, so counting correct answers and measuring loss answer different questions.

images, sequences and rewards call for different structures

A convolutional neural network, or CNN, exploits local relationships in images. A small filter moves across regions, using the same weights in different places. A learned pattern doesn't have to be relearned separately at every image coordinate. This gives the architecture a useful assumption about how images are organized.

The 1998 work by LeCun and colleagues documents networks and document-recognition applications before AlexNet. The path from pixels to a decision passes through learned representations; the architecture helps exploit spatial structure.

An important predecessor is the Neocognitron, introduced by Kunihiko Fukushima in 1980. Layers combined local-feature detection with tolerance to shifts: recognizing a stroke even when it appears a little farther left or right. Fukushima's paper helps explain the CNN lineage. It didn't use the backpropagation training of later convolutional networks; architectural inspiration and adjustment rules need to be distinguished.

In sequences, order matters. “The dog bit the man” and “the man bit the dog” contain the same words but different relationships. Recurrent networks maintain a state as they traverse inputs. One training challenge is preserving relevant information over many steps.

LSTMs, introduced by Hochreiter and Schmidhuber in 1997, regulate what enters, remains in and leaves an internal state. That's the information flow “memory” describes here.

In reinforcement learning, a system learns to choose actions from consequences and reward signals. A game-playing agent might receive a favorable result for winning and need to discover actions that increase that result's likelihood. The reward may arrive long after the choice; assigning value to intermediate actions is part of the difficulty.

Richard Sutton developed temporal-difference methods, and Watkins and Dayan's Q-learning became a reference. These methods learn value estimates and guide decisions. They don't presuppose natural language or depend on a user having a conversation.

In temporal-difference learning, systematized by Sutton in 1988, a prediction can be updated by comparing it with a recent reward and a new estimate of what will happen next. Imagine revising your arrival-time estimate after completing part of a journey, without waiting to arrive before learning anything. In games, this helps assign value to intermediate situations.

Q-learning learns action values in states and can update a target strategy using experience collected by another strategy; this distinction is called off-policy. Exploring an alternative doesn't require considering it the best future action. The 1992 convergence proof depends on assumptions about exploration, update rates and the environment. It doesn't promise that any agent will learn any task simply by receiving rewards.

Rewards can also be badly designed. A system can maximize a recorded number without achieving the objective a person intended. Learning to earn points isn't learning justice, prudence or human intention. Decades later, when human preferences and verifiers are used to adjust language models, this difficulty will reappear at another scale.

What to retain from these families: the problem type, model structure and learning signal are different choices.

QuestionExamples from this section
Which regularity should we seek?trees learn splits; SVMs seek margins; probabilistic models represent uncertainty.
How should inputs be organized?CNNs exploit image neighborhoods; recurrent networks and LSTMs traverse sequences.
Where does the adjustment signal come from?labels in supervised learning, data structure in self-supervision, consequences in reinforcement learning.

These choices can be combined. More data and computation would allow some of them to be tested at another scale.

data and computers change the scale

a chess victory without a chess player's brain

In 1997, Deep Blue defeated Garry Kasparov in a six-game match. The victory carried enormous symbolic weight: a computer beat one of history's greatest chess players in an activity used to exemplify human intelligence.

The score was 3.5 to 2.5. According to IBM, the system could examine about 200 million positions per second. That number measures search on that hardware. It doesn't convert to a player's "thoughts per second."

The system combined search, position evaluation, domain knowledge and specialized hardware. It wasn't a language model and didn't learn chess like later neural systems. IBM's preserved history documents the result and its engineering.

Think of a player looking a few moves ahead. Every action opens possible responses from an opponent. The combinations grow, and evaluating everything through the end of the game is impractical. A system can explore part of the tree, discard possibilities and estimate which positions are more favorable. Domain knowledge helps decide what to investigate and how to evaluate what it finds.

The victory shows that performance can be achieved through mechanisms different from a person's. We needn't attribute the human experience of enjoying chess, feeling pressure or having conscious intuition to recognize that the program played extraordinarily well.

It also shows the limit of a narrow task. Winning chess doesn't automatically confer the ability to drive, translate or understand conversation. The board has explicit rules, a representable state and a relatively clear objective. The world outside has very different conditions.

Reinforcement learning already offered another path in games. Gerald Tesauro's TD-Gammon used networks and temporal-difference learning to play backgammon, exploring game outcomes. His 1995 account documents the work. Search and learning continued developing and would later be combined in other systems.

why does a graphics card help train a network?

In 2006, Hinton, Osindero and Teh presented a deep-network training method with a layerwise pretraining strategy. The paper on deep belief networks helped revive interest in these networks. "Deep" refers to multiple layers of learned operations.

Training many layers was difficult for reasons of optimization and cost. An expressive architecture isn't very useful if the adjustment procedure can't find a good solution or requires unfeasible resources. Training methods and available resources were changing those conditions.

Some of those conditions came from GPUs, graphics processing units. Drawing a scene involves many similar operations on many elements. That work favors parallel processing. Neural networks also perform enormous numbers of multiplications and additions on groups of numbers.

A matrix arranges values in rows and columns. Such a table lets us collect relationships and calculate many combinations at once. When a layer combines several inputs with several weights, matrix operations provide an efficient language and implementation.

A GPU can execute many of these calculations at high throughput. The gain depends on memory, data layout, workload size, data movement and implementation. Adding powerful hardware to a program that spends its time waiting for data doesn't solve the whole problem.

hardware favors a kind of work. neural architectures became more practical when their operations could exploit this parallelism at scale.

A practical bridge was CUDA, announced by NVIDIA in 2006, with tools made available in 2007: researchers could use GPUs for computation beyond graphics. In a 2009 study, Raina, Madhavan and Ng reported gains of up to 55 times over a dual-core CPU. A network configuration with roughly 100 million parameters went from weeks to about one day. The number belongs to that experiment; the milestone is making larger experiments practical, not promising the same gain for every program.

ImageNet: data is infrastructure

A network needs examples, and evaluating progress requires comparable conditions. ImageNet, introduced in 2009 by Jia Deng and colleagues, organized a large image database by category. Fei-Fei Li played a central leadership role in that effort.

Building the dataset took work. Collecting, organizing and annotating images required design and human labor. People defined categories, selected examples and produced labels. The ImageNet paper documents its construction.

The 2009 presentation described 3.2 million images, 5,247 categories organized through WordNet and twelve major branches of that hierarchy. Some verification used Mechanical Turk workers. These figures describe that version, not every later version's size or each competition's subset.

A dataset is an organized collection of data used in a study. A benchmark is an evaluation setup that allows results to be compared. They may be related but aren't the same thing. One dataset can serve many experiments; an evaluation chooses data, tasks, measures and conditions.

Having a common benchmark changed vision research. Teams could compare methods on a shared task. This encourages measurable progress but can narrow incentives: improving that test's score doesn't necessarily solve every difficulty of vision in the real world.

In 2012, AlexNet, by Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton, demonstrated a striking advantage in the image-classification competition. The combination involved CNNs, GPUs, lots of data, and training and regularization choices. None of these ingredients can be erased from the explanation.

The original paper reports a top-5 error of 15.3% versus 26.2% in the 2012 competition. “Top-5” accepts the correct category among five suggestions. This isn't the error rate for any possible image, or the same measure as requiring one correct answer.

The 15.3% belongs to a combination of seven networks submitted to the competition. The network described in the paper had about 60 million parameters; on the ILSVRC-2010 test, the PDF reports 17.0% top-5 error. Distinguishing edition, set and model combination matters even for such a famous historical result.

With deep learning, a network also learns the representations used to solve the task. In many earlier approaches, people designed features manually: edge patterns, shapes, frequencies or combinations that seemed relevant. Deep networks began learning many of those transformations along with the final decision.

People still choose the dataset, objective, architecture and training. Some work shifts: instead of manually designing every useful feature, part of the representation can emerge from adjustment over examples.

words become positions, sequences become memory problems

In language, embeddings are vector representations: lists of learned numbers representing words, tokens or other objects. Word2Vec, associated with work by Mikolov and colleagues in 2013, made the usefulness of those representations especially visible.

Imagine a city map with many dimensions and coordinates learned from usage. Elements used in related contexts can acquire useful relationships in that space. The original work investigates efficient word representations. We needn't interpret every coordinate as “how happy this word is” or “how human it is”; the value lies in the whole representation and the operations it permits.

Vector analogies aren't universal semantic laws. The map reflects regularities in data, including inequalities and problematic associations.

In 2014, seq2seq models demonstrated a neural way to map one sequence to another, such as an input sentence to its translation. The paper by Sutskever, Vinyals and Le used LSTMs. One part encoded the input; another produced the output.

Compressing a long sentence's information into a fixed representation creates a bottleneck. You can grasp the difficulty without mathematics: read a paragraph, close the book and try translating everything using only what you retained in a very small summary.

Bahdanau, Cho and Bengio proposed an attention mechanism so the translation process could consult representations of input parts while producing output. The work was made available in 2014, before the Transformer. The translator learned to combine information from different positions, an operation we'll follow in the next section.

generating, going deeper and learning to play

Generative adversarial networks, abbreviated GANs, proposed by Goodfellow and colleagues in 2014, set up a game between two models. A generator produces examples; a discriminator tries to distinguish generated examples from the real dataset. Their objectives oppose each other, and training can improve generation.

You can compare it to a counterfeiter and an inspector: one tries to produce something convincing, the other to identify its origin. In the program, these are two functions adjusted through optimization. The GAN paper describes the construction, which also faces difficulties with stability and result diversity.

ResNets introduced residual connections: paths that combine a representation with a learned transformation of it. This helped train very deep networks. He and colleagues' work, released in 2015 and presented at CVPR in 2016, made clear that depth needed an architecture supporting information flow and adjustment.

In reinforcement learning, DQN combined deep networks with value learning to play Atari from pixels. The initial 2013 work preceded the 2015 Nature paper. A network could learn useful representations for acting without first receiving a manual list of every relevant visual feature.

AlphaGo combined policy networks, value networks and search. A policy guides actions; a value estimate helps assess positions; search explores alternatives. The January 2016 paper documents the victory over Fan Hui, and the match against Lee Sedol took place that March. The original paper makes the combination visible, not a victory of “networks alone” over everything else.

Sizes and conditions tell another part of the story: ResNet presented 152-layer networks; the 2015 DQN paper evaluated 49 games, training a separate agent for each game with the same general recipe. AlphaGo beat Fan Hui 5–0 in October 2015, before publication, and Lee Sedol 4–1 in March 2016. These weren't a single model that could play any game without training, or demonstrations outside their evaluation rules.

Transformers and reusable models

first, language has to become input

A computer doesn't receive a word as a ready-made idea. It receives a representation. In language models, text is usually split into tokens, units that may correspond to words, word fragments, punctuation or other sequences, depending on the tokenizer.

“Token” here means something different from the certificate token discussed in the cryptography post. It's a unit of text processing. The same word appears in several technical contexts; learning the specific concept prevents confusing unrelated functions.

Why split text this way? A vocabulary containing every possible word and variation would be enormous and still encounter new forms. Using only very small units makes sequences longer. Methods that learn frequent combinations offer a compromise: common sequences can become reusable units, while others are represented in pieces.

The byte-pair encoding algorithm, BPE, offers one way to build those combinations. In the experiment below, the algorithm starts with the text's bytes and merges repeating adjacent pairs. You choose how many merges to use and can reconstruct the text from the resulting tokens.

simplified execution · local merges

tokens (33)

Bytes that do not yet form a UTF-8 character appear in hexadecimal (0x…). ␣ represents a space.

the␣cat␣saw␣the␣rat␣and␣the␣cat␣saw␣the␣house
rules (oldest first)
  1. t + h → th
  2. th + e → the
  3. the + ␣ → the␣
decoded
the cat saw the rat and the cat saw the house

✓ decode(encode(text)) == text

a small BPE trained on the text you type. merge frequent pairs and watch the units change. production models normally use a vocabulary already trained on much larger datasets.

This experiment teaches an operation but doesn't represent every commercial model's tokenizer. Vocabulary choice, character handling and training data matter. Portuguese, code, numbers and other text types can occupy different token counts for the same apparent amount of information.

After splitting, each token is associated with a numerical representation. Layers operate on those representations. These are learned values, used in layer calculations and transformed according to context. There's no dictionary definition attached to each token.

This also explains why small text edits can change behavior. Token boundaries can differ, positions change and relationships between representations are recalculated. A surface that looks almost identical to a person can produce a different input for the system.

attention is an operation, not a mental state

Return to the sentence: “the cat slept because it was tired.” To a person, “it” can refer to the cat. For a model, a useful question is how to let the representation at “it” combine information from other positions in the sequence.

In attention, scores calculated from representations are converted to weights, and those weights guide a combination of values. One position can use more information from a second and less from a third. In self-attention, the same sequence supplies the positions being queried.

“Attention” names this mathematical operation, not subjective concentration. Its weight map alone doesn't explain the network's entire behavior.

The visual below uses chosen weights to teach the mixture, without running a Transformer. It considers a reading of the complete sentence. In next-token models, masks restrict which positions can be consulted; a position cannot use a future it still needs to predict.

illustration · authored weights

choose the position that queries the others

1. separate positions

  1. theposition 1
  2. catposition 2
  3. sleptposition 3
  4. becauseposition 4
  5. itposition 5
  6. wasposition 6
  7. tiredposition 7

each unit receives a representation, and position helps distinguish order. whole words are teaching units here; a real tokenizer may split them.

weights invented for this explanation; not measurements from a Transformer

choose ‘it’ and observe the share assigned to ‘cat’. the weights are illustrative, not measurements or explanations of a real model.

what changed in 2017

The paper Attention Is All You Need, by Vaswani and colleagues, introduced the Transformer in 2017. The architecture used attention as its central mechanism in place of the recurrence organizing many earlier models. The original paper investigated translation tasks.

A recurrent network has a dependency between steps: the next step's state depends on the previous one. That makes many calculations difficult to perform simultaneously. The Transformer made it possible to process many positions in parallel during training, better exploiting the hardware becoming central to the field.

That parallelism applies during training. In autoregressive generation, each next token depends on context, including earlier tokens: the output remains sequential.

Order also needs to enter the representation, because mixing positions without order information would lose distinctions between sentences. The original Transformer uses positional information. Its architecture includes other operations, such as layers transforming representations and residual connections. The paper's title shouldn't be read as a literal inventory containing attention and nothing else.

Multiple attention heads allow different projections and combinations in parallel. One analogy is examining the same sequence through several views. But we shouldn't label a head “the head that understands pronouns” without investigating the model: distributed representations are more complicated than those labels.

Attention also has a cost. In the traditional dense version, combining every position with every other makes the work grow approximately with the square of sequence length. Doubling context doesn't necessarily cost only twice as much in that operation. Later architectures and techniques seek to reduce bottlenecks, but the limit explains why long context involves difficult engineering.

three ways to organize the same family

The Transformer opened several model families. One way to follow them is to distinguish the roles of input and output.

An encoder builds representations of an input. BERT, introduced by Devlin and colleagues in 2018 and published in 2019, made pretraining representations with context from both sides especially effective. One important task was reconstructing masked parts of text. The BERT paper documents the approach and adaptation to later tasks.

An autoregressive decoder produces sequences by successively predicting the next token, with access to the permitted context. The GPT lineage explored this direction from 2018, as described in OpenAI's initial work. To produce “the cat slept,” the system operates on context, selects a continuation and uses the result to continue.

An encoder-decoder architecture represents an input and generates an output conditioned on it. T5, by Raffel and colleagues, organized many tasks as “text-to-text”: a textual input is transformed into another textual output. The T5 work systematizes comparisons of data, objectives and transfer.

These are architectural categories; a product can combine several with tools. Its commercial name doesn't reveal that composition.

a broad base, adaptable later

Pretraining is a broad learning phase before later adaptations. In a large language model, the task of predicting or reconstructing text can be built from the corpus itself. A person doesn't need to write the ideal answer to every possible future question.

Fine-tuning, a stage of further training, adapts a model to a task, domain or behavior. Here, parameters change. Putting instructions and examples in context changes the input with parameters fixed, although the resulting behavior may change substantially.

GPT-3, described by Brown and colleagues in 2020, made this distinction especially visible. The largest model in the study had 175 billion parameters and could perform several tasks from instructions and examples in its context. The GPT-3 paper documents conditions, results and limits.

Showing two translations and requesting a third lets the model use their pattern in that execution: this is in-context learning, with fixed weights, as in the earlier distinction between input and training.

Interest in scaling laws grew with empirical regularities between data quantity, model size, computation and loss. Kaplan and colleagues' 2020 work documents these relationships. They're useful experimental relationships for planning resources, not guarantees of equal gains in every quality measure.

In 2021, Stanford researchers proposed foundation model for bases trained on broad data and adaptable to many tasks. Bommasani and colleagues' report discusses opportunities and risks: reusing a base can spread capabilities and also propagate its failures across applications.

beyond text, beyond conversation

CLIP, by Radford and colleagues, learned relationships between images and textual descriptions. The 2021 paper shows how representations of the two modalities can be aligned. This offers another bridge: language can help organize vision tasks without being the system's only input.

Training used 400 million image–text pairs. The contrastive objective brings matching representations closer relative to alternatives: a description should match its image better than another. This enabled evaluation of zero-shot transfer, without retraining for that specific task, using textual category descriptions. Zero-shot doesn't mean the model never encountered related concepts during pretraining.

Multimodality means working with more than one information type, such as text, images, audio or video, learning relationships according to architecture and training.

AlphaFold offers a path different from general-purpose conversation. It used neural methods for protein-structure prediction, with striking results in scientific evaluation. Jumper and colleagues' 2021 paper documents the approach and experiments.

Predicting a structure helps with important scientific problems. It still differs from automatically knowing every biological function of a protein or dispensing with experiments. This caution preserves the achievement's scale without making it an explanation of all biology.

These applications already extended beyond language: relating images and descriptions, adapting a base to different tasks and predicting protein structures. In 2022, the conversational interface would make some of this research much easier to use.

So far, three choices help locate the models: BERT represents an input, GPT generates continuations and T5 transforms an input into another sequence. Pretraining produces a base; fine-tuning changes its parameters; examples in context guide an execution with fixed parameters. Multimodality expands input and output types. The family map gathers these relationships for reference.

AI enters everyday life

from predicting text to answering a person

Predicting the next token can teach useful language regularities, but that objective doesn't directly equal answering a question well. An internet text can continue a sentence with sarcasm, fiction, error or quotation. An assistant needs more specific behavior.

The process of instruction tuning uses examples of requests and desired responses to guide that behavior. Asking “explain photosynthesis to a child” differs from merely requesting the continuation of a textbook passage. The model needs to respond to the request, and training can emphasize that pattern.

InstructGPT, described by Ouyang and colleagues in 2022, combined human demonstrations, response comparisons and subsequent adjustment. In one stage, people wrote examples. In another, they compared alternatives, and those preferences helped train a reward model. The language model was then optimized to produce better-rated answers.

This process is a form of reinforcement learning from human feedback, or RLHF. The InstructGPT paper documents the stages. What changed was the objective: predicting text was complemented by signals about desired behavior in an interaction.

In the study, a tuned 1.3-billion-parameter model was preferred to 175-billion-parameter GPT-3 on the evaluated distribution. This teaches an important distinction: size, prediction quality and usefulness to a user aren't the same measure. A larger model may be less prepared to follow the kind of request it received.

A confident, brief or agreeable answer can please people while being wrong. Human evaluation depends on instructions, knowledge and criteria, so approval and factual truth can diverge.

That's why “alignment” shouldn't become a word that ends the discussion. It can mean following instructions, respecting policies, avoiding certain behaviors or serving social values. These objectives can conflict and require continuous evaluation. RLHF addresses parts of the problem; it doesn't settle the whole relationship between a machine and human interests.

Constitutional AI, associated with Anthropic's research, investigated another combination: written principles guide criticism and revision, and model-generated feedback can participate in training. The original research describes a procedure; “constitution” names the principles used, not a guarantee that a system has ethics of its own or perfectly applies any value.

the interface changed the technology's reach

Human feedback also has predecessors: 2019 experiments with GPT-2 already used preferences to adjust behaviors such as summarizing. In 2022, training and public access would bring these techniques to many more people.

ChatGPT launched on November 30, 2022. The original announcement described its relationship to GPT-3.5 models and human-feedback training. The conversation window made it much easier to try a technology that had previously reached many people through demonstrations, papers or less visible integrations.

You could write a question, watch an answer, request a revision and continue in the same thread. This reduces the distance between curiosity and using the system. No programming language was needed to try a summary, explanation or draft.

A new interface can be an enormous historical change even when it combines earlier techniques. A product's importance isn't measured only by whether it invented an architecture. Distribution, access, integration and user experience change who can do what.

Conversation context also lets a later request refer to an earlier one. “Make it shorter” means little in isolation and a lot when the earlier answer is available. This can look like personal memory, but we need to distinguish history presented to the model from permanent changes to its parameters.

In 2023 and 2024, products from several organizations expanded public use of language, code and multimodal models. GPT-4, Claude, Bard/Gemini, Llama and Mistral illustrate paths with different access and distribution regimes. These models entered everyday tools through different forms of access.

That's when chatbots came to occupy so much of the conversation about AI. That was the frustration behind the previous post: using one of these tools began to be confused with knowing the whole field.

A few dates help separate model from interface. GPT-4 appeared on March 14, 2023; its report omitted details of architecture, hardware and training data. Claude 2 arrived in July, Mistral 7B in September and Mixtral 8x7B in December. Google introduced the Gemini family in December 2023, before renaming the Bard interface to Gemini in February 2024. Product name, model family, license and availability are different pieces of information.

hallucination: the sentence fits, the fact may be missing

A language model can produce a fluent answer that doesn't match facts or available sources. The field calls many such cases hallucinations. The term is a metaphor for system output, not a human diagnosis.

Consider asking for a paper's name and reference. The model knows patterns of titles, authors, journals and dates. It may produce a combination resembling a bibliography that doesn't correspond to an existing document. The form can be convincing precisely because training taught regularities of that form.

The model may have learned useful information and still produce an incorrect reference. Fluency doesn't check whether a paper exists; that check needs to enter the answer's evaluation.

Getting a calculation wrong, citing a nonexistent source, omitting a condition and following an improper instruction are different failures. The general category helps discuss a problem, but evaluating a system requires separating failure types. One percentage can hide very different behaviors.

Nor is “don't hallucinate” enough. The instruction is part of the input, but doesn't create evidence or guarantee the model recognizes its gaps. More reliable systems combine training, suitable data, tools, evaluation and workflow design.

consulting documents changes the input

A high score can hide benchmark contamination: if evaluation questions or answers were in training material, the test no longer measures only performance on new examples. Format, tools, attempt counts and inference budgets also need to be comparable. A familiar exam doesn't demonstrate general intelligence, and scaling doesn't guarantee eliminating hallucinations.

The same caution applies to so-called emergent abilities. A skill seems to appear suddenly when a score jumps beyond a certain size. But an “entirely right or entirely wrong” metric can hide gradual improvements in individual steps. A 2023 study showed this effect in specific evaluations. Measurement can create the appearance of a leap. The study doesn't justify dismissing every new capability as illusory.

One way to bring in current or specialized information is to retrieve documents and place them in context. RAG, retrieval-augmented generation, combines a retrieval stage with generation conditioned on the retrieved material. Lewis and colleagues' 2020 paper is a reference for this approach.

Imagine asking for a museum's opening hours. The model may have seen outdated information during training. An application can consult the current page, select a passage and supply it with the question. The answer now has a source available in that execution.

The lookup adds information to context, like the earlier examples, without updating weights through that operation. The document needs to be available in the execution where it'll be used.

RAG also doesn't guarantee that the application found the correct document or interpreted it well. Search may select another museum branch, a passage may omit a holiday and generation may confuse conditions. Retrieval and the answer need evaluation, along with the source's quality.

weights, context and external documents play different roles. saying a system learned or remembered can hide which mechanism was used.

learning to remove noise to generate images

The popularization of image generation has its own technical history. Diffusion models learn operations related to a process that adds noise to data and a generative path that progressively removes noise.

During training, we can start with an image and add known levels of noise. The model receives perturbed versions and learns an estimation task associated with that process. During generation, we start from noise and apply learned transformations to construct a sample.

The 2020 work by Ho, Jain and Abbeel is a milestone in this family. Generation produces a new sample using learned regularities, conditioned or not on information such as text. It doesn't reveal a photograph that was already hidden in the noise.

The visual is schematic: it shows a drawing of a house becoming visible as a programmed noise layer is reduced. It illustrates a path but doesn't execute a diffusion model's mathematical procedure or generate a new image from a prompt.

illustration · fixed drawing

1. noise: in a real generative process, the sample starts from a noise distribution. this visual uses a reproducible pattern.

go back: add noise · The drawing becomes clearer in six stages. The final image is programmed: no model generates images here.

six illustrative stages, with a fixed house and reproducible noise. in a real model, transitions are learned and generation constructs samples; this drawing only helps follow the idea.

In latent diffusion, models perform part of this work in a compressed representation rather than always operating directly on every pixel. Rombach and colleagues' paper, made available in 2021 and published in 2022, explains this combination of efficiency and conditioning.

DALL-E 2, Stable Diffusion and Midjourney helped make text-to-image generation a public activity in 2022. They brought together generative methods, language representations, data, hardware and interfaces. A beautiful image, like a fluent sentence, needs its own evaluation: fidelity to the request, diversity, rights, biases and possible reproduction of training examples.

Visual generation also has different paths. The original DALL-E, from January 2021, used an autoregressive Transformer, not a diffusion model. Diffusion became especially visible to the public in 2022, including through Stable Diffusion. Generative video added the need to maintain relationships over time: producing a convincing frame doesn't guarantee an object keeps its position and identity in subsequent frames.

adapting and reducing cost

As large models became reusable bases, demand grew to adapt them without repeating all of training. LoRA, proposed by Hu and colleagues in 2021, learns lower-dimensional updates to certain operations. The original paper explains how to reduce the set of adjusted parameters.

Quantization represents values with lower numerical precision. This can reduce storage and memory requirements, with consequences depending on the model and technique. QLoRA, described by Dettmers and colleagues in 2023, combines quantization and low-rank adaptation to make certain adjustments more accessible. These are cost reductions at specific stages; training still requires resources.

Distillation trains one system to reproduce another's behavior or signals. It can produce a smaller model retaining some utility, though not automatically every capability. It's a way to transfer behavior under an objective and a collection of examples.

Adapting a model and reducing its memory requirements helps when the question becomes "how much does putting this into an application cost?" Cost and latency also depend on engineering around the network.

reasoning, tools and agents

spending computation while solving

If you entered here, it's worth revisiting training and inference and document retrieval. The question now is: how much work does a system do after receiving a problem and before delivering an answer?

There are several ways to spend more computation at this stage. Producing candidates, comparing alternatives, revising an attempt, searching and checking are examples. The goal is to improve a solution using a larger processing budget during execution.

Requesting intermediate steps in a prompt was already a research direction before 2024. The later change involves combining that possibility with training and resource-allocation strategies. Test-time compute names this axis of computation during problem solving.

The 2024 work by Snell and colleagues studied tradeoffs between model size and processing during problem solving under specific tasks and conditions. Gains depended on the task and the distribution of processing. The comparison opened another option beyond increasing model size, without establishing universal superiority.

In certain comparisons with equivalent operation budgets, inference strategies outperformed simply using a model 14 times larger. Problem difficulty and processing allocation changed the result. The number describes an experimental comparison, not a substitution rule between any pair of models.

OpenAI o1, introduced in September 2024, made this regime especially visible in products. The company described reinforcement training and gains from greater computation during problem solving. The announcement is a source for the launch and reported results; performance comparisons need to preserve their conditions and origin.

DeepSeek-R1, introduced in January 2025, broadened access to models and distillations associated with reinforced reasoning. The technical paper distinguishes R1-Zero, R1 and distilled models. These variants shouldn't be treated as the same procedure under interchangeable names.

“Reasoning” here names capabilities and strategies investigated on problems. Longer steps can contain errors, and a generated explanation doesn't necessarily reveal the internal process leading to the result.

verification changes the game

In reinforcement learning with verifiable rewards, sometimes abbreviated RLVR, an answer can receive a reward for passing tests or reaching a verifiable result. Verification reduces dependence on someone comparing every solution, but the test must represent the task. FastCuRL, from 2025, investigated this approach with a 1.5-billion-parameter model and eight GPUs, adjusting example difficulty and solution length. The example shows a smaller scale than the largest industrial training runs, under the evaluated task's conditions.

If the system generates code, a tool can run tests. If it proposes a move, an environment can measure the result. If it answers a problem with a verifiable answer, there's a clearer signal for comparing attempts.

This makes training and candidate selection easier. A verifier provides information the answer's appearance alone doesn't. An elegant program failing a test needs correction; a less elegant alternative may satisfy the measured condition.

But a verifier checks only what was defined. Incomplete tests can accept a buggy program. A numerical answer may be right through an invalid path, depending on the task. A solution can pass a test and fail when requirements change.

For open-ended topics, such as explaining a historical event or advising an organization, quality is harder to condense into an automatic reward. Sources can disagree, conditions can be missing and values can conflict. Progress in mathematics cannot simply be declared transferable to every form of human judgment.

a tool provides access beyond the weights

A model can produce a structured request for software to perform an action. That action may query a database, run a calculation, open a document or operate another application. The external program executes and returns a result, which enters the context.

When an application consults a calculator and returns the calculation to the model, that's tool use. To evaluate the system, we need to know which result came from the tool and how the model used it, rather than attribute the entire calculation to the network.

An agent, in the sense used by many current systems, puts these operations into a cycle: observe, choose an action, execute a tool, read the result and decide how to continue. “Agent” has a history before LLMs; the term wasn't invented for a chatbot calling a tool.

Let's follow an example: find two museums' opening hours and create a sourced table. The simulation below shows both a path obtaining the information and another with an incomplete result. It doesn't consult real museums, execute tools or contain information you can use to plan a visit.

flow simulation · fictional results

  1. 1observe
  2. 2plan
  3. 3retrieve
  4. 4read results
  5. 5verify
  6. 6answer

observed result

fictional request: create a table of opening hours for Museum A and Museum B

the system receives the task and conditions: two museums, opening hours and verifiable sources.

local simulation: no tool executes and no external information is fetched

an agent is a system in a loop. enable the incomplete result to see why verification must change the plan instead of merely making an answer appear certain.

A tool also has permissions. Knowing how to write an action request doesn't automatically grant access to the world. The application defines which tools exist, which data can be seen and which operations are allowed.

Anthropic announced the Model Context Protocol, MCP, in November 2024 as a protocol for connecting AI applications to sources and tools. The original announcement documents the proposal. A common protocol simplifies connections; it doesn't itself provide memory, competence, safety or unrestricted authorization.

context is part of the engineering

The model receives input organized by the application: instructions, history, documents, tool results and task state. Context engineering describes selecting and organizing that material.

If important information was left out, the model can make a poor decision even if it could use that information when provided. If context mixes a trusted instruction with external text trying to command actions, the application needs to distinguish their roles.

A prompt injection exploits that boundary: text found in a document or page tries to be treated as an instruction to the system. The problem involves integration and control as well as model behavior. A document an agent reads shouldn't inherit the authority of whoever requested the task.

Errors can also accumulate in long tasks. A search returns the wrong page, a summary loses a condition and an action uses that summary. The final result can look well organized while carrying a failure from three steps earlier.

A large context window says how many tokens can fit, not that they're all used equally well. In Lost in the Middle, from 2023, some models had more difficulty when relevant information was in the middle of long texts. Imagine handing over a whole folder and asking for a date hidden on page 87: fitting in the folder and finding the date are different things. The result needs reassessment for each model, version and task, not a universal condemnation of long context.

the frontier in October 2026

At the October 10, 2026 cutoff, variants prioritizing capability, speed and cost coexist with multimodal systems and products integrating tools. Access comes through conversational interfaces, application programming interfaces (APIs, used by other software) and distributed weights, under different conditions. The announcements and access table retains examples and sources; it's a dated snapshot, not a ranking.

Distributed weights, available code, published data and a reproducible training recipe are different dimensions of openness. The specific artifact's license also matters. “You can download it” doesn't settle all these questions.

The OSI Open Source AI Definition specifies freedoms to use, study, modify and share, and requirements for components. Open weights conveys something narrower. Weights can be distributed with restrictions or without sufficient information to study and modify the system under that definition.

Smaller models and local execution coexist with large-scale services. Quantization, distillation and specialization can reduce requirements. In mixture-of-experts, or MoE, architectures, different parts of the model activate during processing; total parameters and parameters used in an operation needn't coincide. “Experts” here names architectural components, not human professionals with defined disciplines.

Gemma and Ministral, along with Qwen variants and R1 distillations, exemplify this direction. On-device execution can reduce latency, enable use without a constant connection and keep certain data local. Those benefits depend on the application and its other components: downloading a model doesn't guarantee a private end-to-end workflow or compatibility with every phone.

In programming, the system around a model is often called a harness. It supplies an editor, terminal, tests, execution limits and context assembly. An agent can edit, test, observe a failure and revise. Grok Build opened components of this system in July 2026, including tools, MCP and subagents; that isn't publishing model weights. Subagents are processes receiving subtasks, unlike internal “experts” in an MoE architecture. Coordinating processes adds communication and completion criteria, not an automatic guarantee of correctness.

The field's current state combines networks, search, representation, external memory, tools, verifiers and infrastructure. Some pieces are recent; others are decades old. Systems can do more because the pieces are connected, but can also fail at their boundaries. The connections need evaluation too.

the history beyond models

We've reached current systems by following methods. Now let's examine the other axis announced at the beginning: applications, institutions, infrastructure and impacts — where systems were already used, who builds them and what sustains them. This work was present in earlier periods, even when the narrative followed an algorithm or demonstration.

the AI you already used without calling it AI

You probably used other forms of AI before opening a chatbot: receiving a movie recommendation, using speech recognition or searching for something in an image.

Recommendation systems organize films, products, music and posts. One familiar approach seeks regularities in people's and items' preferences: those who chose similar things may be interested in other related things. Others learn representations and predictions from interactions.

This includes collaborative filtering and matrix factorization. Imagine a table whose rows are people and columns are films, with most ratings missing. Factorization learns smaller representations of people and films that help estimate missing values. It needn't consciously know why someone likes a film to find useful regularities; nor can it treat an estimate as a guaranteed preference.

The chosen objective influences behavior. Maximizing clicks isn't the same task as increasing diversity, lasting satisfaction or information quality. A system can improve its recorded metric while degrading something it didn't measure.

There's also a cycle: the system selects what a person sees, the person reacts to the selection and the reaction becomes new data. The data records responses to options the system showed, not every option the person could have chosen. This cycle also occurs outside social networks.

Recognition and synthesis of speech have their own histories, from statistical models to neural and multimodal systems. Transcribing audio, producing a voice and interpreting a spoken request are distinct tasks, though an interface may combine them. An underrepresented accent or background noise can change part of the workflow before the question even reaches the language model.

In older systems, acoustic and language models were separate components. Later neural approaches expanded end-to-end training, mapping audio directly to text and also audio to audio. This reduces some divisions in the workflow; it doesn't remove the need to evaluate transcription, content, voice and conditions of use.

In computer vision, locating objects, segmenting an image, classifying a scene and relating images to text are also different problems. Improving a classification measure doesn't automatically solve perception in traffic, industrial inspection or scientific imaging. We need to know which task was evaluated.

Beyond CNNs, the Vision Transformer, released in 2020, divided images into patches and treated their representations as a sequence. Hand-designed descriptors, CNNs and Transformers seek useful representations through different paths to solve vision problems.

In robotics, the tradition of control and feedback remains essential. A vision-language-action model seeks to connect perception, instruction and physical behavior. But moving an arm requires sensors, localization, control and contact with objects. Error tolerance depends on the situation, and a plausible sentence doesn't replace a physically safe action.

In the 1980s, Rodney Brooks explored layers of reactive behavior: a robot could respond to obstacles without first building a complete symbolic description of everything. Another path probabilistically estimated where a robot was and what its environment looked like. Work by Thrun, Fox and Burgard linked simultaneous localization and mapping, usually abbreviated SLAM. A sensor is uncertain; observing and updating estimates can be more useful than pretending to have a perfectly known map. These approaches coexist with planning and control.

VLA, from vision-language-action, names the connection between vision, language and action. A model such as Qwen2.5-Omni, from March 2025, combines text, images, audio and video without thereby controlling a robot. Qwen3, from April, emphasized direct-answer and deliberative-processing modes. “Multimodal” alone doesn't tell us which inputs, outputs and actions a model supports.

In science and healthcare, models can help analyze images, predict structures, organize documentation and support research. Required evidence varies with the use. A research result, regulatory authorization and demonstrated benefit in clinical conditions are different categories; none can silently stand in for another.

AlphaFold 3, introduced in May 2024, extended the problem to interactions between proteins, DNA, RNA and other molecules. Genomics and cellular models also seek regularities in large biological datasets. Predicting a structure or relationship is a scientific contribution; demonstrating that a treatment improves patients' lives requires another evidentiary path.

The AI Index 2026 recorded 258 FDA authorizations for AI-enabled devices in 2025 through September. In a different population, a study published in JAMA in 2025 examined 903 devices authorized through August 2024: 505 had publicly described clinical studies, and 12 of those 505, or 2.4%, included randomized trials. That denominator matters. It doesn't mean 2.4% of the 258 devices from 2025, or that missing public descriptions prove the absence of all clinical evidence.

In 2024, the Nobel Prize in Physics recognized work by John Hopfield and Geoffrey Hinton related to neural networks, and part of the Chemistry prize recognized Demis Hassabis and John Jumper for protein-structure prediction, alongside David Baker's prize for computational protein design. The Physics and Chemistry announcements document that recognition. It encompassed neural networks and scientific applications far beyond chatbots.

a history of countries and institutions

United States companies occupy much of the recent conversation, but we also need to consider research in Europe, Japan, the Soviet Union and China.

Europe appears in the precursors of logic and computing, continues through language, logic programming, statistics and robotics, and remains present in laboratories and companies. Japan's fifth-generation project shows a different industrial and scientific policy, investing in knowledge and parallel computing in ways that influenced international discussions.

The Soviet Union had its own trajectory in computing and cybernetics. Cybernetics' position changed through political and institutional disputes, from a contested subject to a language of modernization in certain periods. Slava Gerovitch's historical study helps explain the relationship between science and political discourse. This history can't simply be translated into being ahead or behind in a single race.

China is central to today's ecosystem, with universities, companies, robotics and model families such as Qwen and DeepSeek. These lines of work show that ideas, skills and infrastructure aren't restricted to one country. Comparing countries requires distinguishing scientific output, industrial capacity, adoption, patents, hardware access and available models.

Patents, released models and investment measure different things. A promised investment, for example, may still be far from delivered equipment. None of these counts alone summarizes the quality of a country's research.

There's also circulation: researchers study and work in different countries, papers cross institutions and implementations reuse published knowledge. Countries and companies are useful units for discussing resources and policy, but aren't closed boxes where an idea belongs forever.

Two measures convey China's scale without conflating activities: the AI Index 2026 attributes 74.2% of 2024 AI patent grants in its database to China; the International Federation of Robotics records 54% of new industrial-robot installations in 2024, around 295,000. Patents and installations measure different activities, and neither summarizes all of a country's research quality.

Brazil didn't enter this history yesterday

Brazil's AI community has a history preceding the generative wave. The SBIA memorial published by the Brazilian Computer Society, SBC, records a first symposium in Porto Alegre in November 1984 and the building of a research community in subsequent decades.

Universities and groups worked on knowledge representation, neural networks, optimization, language, vision and intelligent systems. Institutions associated with this history include UFRGS, USP, Unicamp, UFRJ/COPPE, UFPE, PUC-Rio, ITA and others. The diversity of work exceeds what a short list can represent.

BRACIS resulted from bringing artificial-intelligence and neural-network communities together. The 2025 organizers explain the merger of SBIA and SBRN. The conference brings together communities researching AI and neural networks long before today's assistants.

An important intermediate record is the first National Workshop on Neural Networks and first Neural Networks School, in 1992. BRACIS 2025 identifies itself as the 35th edition in these communities' continuity. This research calendar is much older than the commercial chatbot release calendar.

  1. 1984first recorded symposium26 yearsPorto Alegre: organization of the community documented in the SBIA memorial.
  2. 2010SBIA memorial10 yearsthe community records its own history, people and institutions.
  3. 2020C4AI4 yearsa research center begins with challenges in Portuguese, science and society.
  4. 2024PBIA 2024–20281 yeara plan for education, infrastructure and applications; targets still require execution and evaluation.
  5. 2025BRACISthe conference records continuity and convergence between scientific communities.
a selection of Brazilian records, not a list of all research. gaps between milestones don't represent decades without work.

C4AI, established by USP, IBM and FAPESP, began operating in 2020, addressing Portuguese language, health, environment and work. USP's announcement and FAPESP's page document its formation. It's one example, not the only center or an exclusive national birthplace.

Brazilian Portuguese matters as a research problem. Evaluation in English doesn't establish the same result in Portuguese, and formal written Portuguese doesn't cover every accent, register and context of use. Data and criteria need to represent those who will use the system.

The Brazilian Artificial Intelligence Plan 2024–2028 addresses infrastructure, education and applications. It's a planning document: targets and planned investment must be distinguished from implementation and measured results.

Brazilian records show many institutions and contributions but are less systematic for some periods. Choosing a single "founder of Brazilian AI" would require historical investigation these documents don't support.

the collective work hidden behind a surname

Names help us follow ideas but can distort the story if they hide teams and less visible work. A network trained on millions of examples depends on whoever created, organized, annotated and evaluated them. A product depends on researchers, engineers, operators and people producing preference and safety signals.

Women appear in computing's history from Ada Lovelace onward. Karen Spärck Jones made fundamental contributions to information retrieval and language. Her work on the importance of terms in documents helps explain why a rare word in a collection may convey more than one appearing almost everywhere. The 1972 paper and Cambridge archive preserve that history.

Fei-Fei Li appears in ImageNet's construction; Kristina Toutanova coauthored BERT. In Brazil, the SBC memorial records researchers such as Rosa Maria Viccari and Maria Carolina Monard. Telling these stories through one researcher's name would erase some of the people who did the work.

In impact evaluation, Joy Buolamwini, Timnit Gebru, Margaret Mitchell, Inioluwa Deborah Raji and other researchers helped make questions about disparities, documentation and responsibility explicit. Scientific work includes choosing which errors to seek, not merely lowering an existing average error.

Gender Shades, by Buolamwini and Gebru, examined intersectional disparities in commercial gender-classification systems using images. The 2018 paper has a specific task, sample and conditions. Its result applies to the systems and conditions studied. The question exposes an evaluation problem: a good average can hide much worse performance for certain groups.

Model cards, proposed by Mitchell and colleagues, organize information about a model's purpose, evaluation and limitations. The original work shows that documentation is also part of evaluating technology. An isolated score doesn't tell us for whom, under which conditions and with which risks a system was tested.

Rediet Abebe and colleagues add another question: what can computing do within processes of social change? Roles for Computing in Social Change examines roles and limits of technical contributions. A more accurate model doesn't by itself resolve power distribution, access or institutional decisions. A technical metric may omit those conditions.

the datacenter is part of the model you use

“An algorithm improved” is an incomplete explanation when scale depends on semiconductors, memory, networks and energy. Distributed training needs to move data and coordinate many machines. Fast processing doesn't help if machines spend their time waiting for information or memory can't hold the workload.

GPUs, TPUs and other accelerators offer different computing capabilities. High-bandwidth memory helps feed those operations. Interconnects help machines exchange values. Software must organize the whole system. Bottlenecks change with architecture and task size.

Training cost also differs from usage cost. Expensive training can be amortized across many applications. Cheaper inference can increase use. At the same time, competition at the frontier may demand even larger training runs and infrastructure.

This helps explain why published knowledge and the ability to perform an experiment aren't distributed in the same way. Reading an architecture doesn't give a university the equipment, energy and teams to repeat any industrial training run. Open publication helps ideas circulate, but infrastructure imposes material conditions.

In its 2025 report, the International Energy Agency estimated that all datacenters consumed around 415 TWh of electricity in 2024. Its central scenario projected about 945 TWh for 2030. The energy-demand chapter explains assumptions and scenarios. These figures aren't exclusive to AI, and the 2030 value is a projection from that report.

The 415 TWh represented about 1.5% of worldwide electricity. A relatively small global share can coexist with substantial local pressure when equipment concentrates in one region.

all datacenters, according to the IEA in 2025. the 2030 bar is a scenario, not an observed result or a measurement exclusive to AI.

Efficiency per task and total consumption can also move in different directions. If each query uses less energy but query volume grows substantially, the total can rise. A promise of savings needs to specify its comparison.

Locally, large datacenter concentrations can strain electrical grids and other resources even when their global share seems smaller. The "cloud" depends on physical equipment, and its location changes who deals with those costs.

bias, privacy and authorship enter the same history

Bias can arise in data selection, categories, objectives and usage context. An imbalanced set can underrepresent a population. A poorly chosen category can compress different situations into one label. Average accuracy can conceal disparities.

Addressing this takes more than filtering an offensive word. We need to know the task, the groups and conditions represented, and which errors have consequences. Evaluation criteria are also human choices.

Privacy involves collection, access, retention and the possibility of inferring information. A model can memorize certain training examples, and an application can expose documents through poorly configured permissions. These are different mechanisms; calling everything “the model leaked” can hide where a fix is needed.

Copyright involves distinct questions: how material was used in training, whether output reproduces protected expression, who can claim authorship and which license conditions apply. The discussion varies by jurisdiction and still requires case-specific analysis. U.S. Copyright Office reports distinguish, for example, authorship of generated material from training questions. This documentation isn't a single rule for every country.

Human labor doesn't disappear because an interface generates on its own. Building datasets, annotating examples, comparing answers, moderating content and investigating failures help construct the system. That work often remains outside the interface we use.

These decisions influence the system itself: what enters the data, the training objective, who can access tools, how errors are evaluated and who receives benefits.

employment: tasks, jobs and institutions

An occupation comprises many tasks. Automating some can change a routine without immediately eliminating the job. It can also enable team reorganization, increased production or a shift toward checking and correcting outputs.

That's why exposure to AI isn't the same measure as predicted unemployment. A study can identify tasks potentially achievable by a technology; actual outcomes depend on cost, reliability, adoption, organization and economic decisions. The ILO's 2025 study examines occupational exposure and transformation potential.

The AI Index 2026 reports that 88% of respondents in a 2025 survey said their organizations used AI in some function, and 79% reported regular generative-AI use in at least one business function. These are self-reported survey answers, not a census of every company worldwide. Agent use at scale, per function, was much less widespread in most cases. The report, section 4.3, spells out those conditions.

Solving a task, organizational adoption and distributing benefits are different outcomes requiring their own measures.

In the same survey, 32% of respondents expected workforce reductions the following year because of AI, while 43% expected stability. Self-reported expectations aren't counts of actual layoffs. Confusing task exposure, managerial expectations and actual employment changes manufactures certainty the data doesn't provide.

There's also a question of human learning. If someone stops practicing tasks that taught them to check results, they may become more dependent on the system precisely when they need to judge a failure. This isn't inevitable; it's a choice about organizing use, education and review.

Productivity gains also need context. Generative AI at Work, by Brynjolfsson, Li and Raymond, studied customer-support assistance and found heterogeneous benefits, larger for less experienced workers in that setting. Task, organization and experience level change the result. This provides evidence for a concrete application; it doesn't by itself measure the whole economy's productivity or every profession's future.

intelligence, understanding and consciousness aren't one test

Success on tasks reopens old philosophical questions. What counts as intelligence? Performance, learning, adaptation, understanding, general capability? An unspecified definition leaves two people arguing over words while thinking of different criteria.

John Searle formulated the Chinese Room argument in 1980. Imagine a person following rules to manipulate symbols in a language they don't understand. The resulting answer can look suitable from outside. For Searle, this challenges the conclusion that correct symbol manipulation suffices for understanding. The original paper also includes discussion and objections.

In Searle's distinction, weak AI can be a tool for studying cognition; strong AI holds that the appropriate program constitutes understanding. “Strong” here doesn't simply mean a high score or more parameters. Nor is it a precise synonym for AGI: general capability and understanding are different questions.

The argument is philosophical and remains disputed; it wasn't an experiment settling the question. Critiques concern the whole system, representation and its relationship to the world. Traditions that don't fully agree continue to debate the difference between behavior and experience.

LLMs reopen the dispute because they learn representations and demonstrate capabilities beyond fixed scripts. Those capabilities need investigation. Still, an answer saying “I feel” doesn't establish subjective experience: the system was trained to produce language in which people say such things.

A 2023 interdisciplinary analysis of AI consciousness examined indicators derived from consciousness theories. Its conclusion about the systems studied then shouldn't become universal proof about every future model. The question requires criteria, theory and evidence, not an impression created by conversation.

AGI, artificial general intelligence, also depends on a definition and a way to evaluate it. An isolated benchmark, laboratory promise or impressive task doesn't resolve the whole dispute. We need to know the required scope, learning conditions, transfer and reliability.

For governance, initiatives such as the NIST AI Risk Management Framework and the UNESCO recommendation make evaluation and responsibility an ongoing practice. The problem exceeds labeling a technology good or bad in the abstract: it involves purposes, institutions, evidence and consequences.

History demonstrates a sequence of capabilities built, tested, expanded and sometimes overestimated. It doesn't demonstrate that consciousness, general intelligence or automatic replacement of all human activity is an inevitable destination.

The same holds for systemic risks: concentrated power, surveillance, disinformation and institutional dependence aren't reducible to a chatbot getting a sentence wrong. Adversarial testing, or red teaming, deliberately seeks failures; safety evaluations need to track capabilities and uses. The December 2025 Foundation Model Transparency Index recorded a mean of 41 points, versus 58 in the previous edition. Criteria and participants changed, so the comparison needs that caveat. Publishing more benchmarks isn't the same as transparency about data, risks and operations.

what changed from one end to the other

Our path began with forms of reasoning that could be described. Symbols and procedures made those forms representable; computers made it possible to execute them and test programs.

Search and rules worked on prepared tasks, but knowledge and possible alternatives grew quickly beyond them. Learning from examples shifted some of that work. Deep networks began learning representations too, and data, hardware and software enabled training at another scale.

The Transformer helped exploit parallel processing. Pretraining made a base reusable, and instruction and preference tuning changed its behavior as an assistant. With retrieval, tools and verifiers, the application began consulting information and executing operations beyond weights. The conversation interface brought this to many more people.

the pieces remain available and can work together. a current system can combine representation, search, learning, tools and verification.

Throughout these changes, results kept depending on the task, representation and conditions. ELIZA seemed to understand through conversation. SHRDLU operated in a narrow world. Expert systems faced maintenance. Networks needed data and adjustment methods. Language models and agents still need evaluation and well-organized surrounding systems.

A promise of general capability needs to show how it escaped those specific conditions. Succeeding at one task still doesn't give a system authority over every other task.

A token, a weight and a plan now occupy different places in this history. You can revisit the examples and play with the experiments: which values change when the network trains? Where did the agent's answer come from? Which rule made ELIZA ask about family?

That's the step I wanted to take after the previous post. Using the tool can be useful, but studying the subject requires opening these questions and examining what happens inside.

for later reference

The narrative ends above. This section can wait until a question arises: the family map locates methods; the chronology supplies dates; 2026 announcements describe access conditions; the glossary explains terms; common shortcuts check interpretations; and further reading offers depth.

a chronology of changes

Dates of development, demonstration, paper and launch can differ. One milestone here doesn't necessarily carry the same weight as another. Historical links are in the corresponding explanations and readings; recent releases have their own links.

Date or periodMilestone and significance
AntiquityAristotle systematizes forms of syllogistic inference, a conceptual ancestor of formal logic.
17th centuryEuropean mechanical automata embody the idea of machines imitating complex actions.
1640sPascal builds mechanical calculators, part of the long history of arithmetic automation.
1670sLeibniz develops a calculating machine and articulates ambitions for universal symbolic calculation.
1700sLeibniz's calculus ratiocinator remains a historical reference for mechanizing reason.
1830sBabbage designs the programmable Analytical Engine.
1843Ada Lovelace publishes notes on the Analytical Engine and discusses symbolic manipulation beyond mere numbers.
1847–1854George Boole develops the algebra of logic.
1879Frege publishes Begriffsschrift, a milestone in modern logic.
1900Hilbert formulates problems helping structure mathematics' formalist program.
1920sHilbert's program intensifies the search for formalization and decision procedures.
1931Gödel publishes incompleteness theorems, revealing limits of formal systems.
1936Church formalizes computability through lambda calculus.
1936Turing introduces abstract machines and characterizes computability.
1937Equivalences between formalisms support what became known as the Church–Turing thesis; the thesis about effective procedures is distinct from those proofs.
late 1930sShannon connects Boolean algebra and switching circuits.
1943McCulloch and Pitts publish formal neurons capable of logical operations.
1948Shannon publishes a mathematical theory of communication.
1948Wiener publishes Cybernetics, connecting feedback, control and communication.
1949Hebb publishes a theory of learning through associative synaptic modification.
1950Turing publishes Computing Machinery and Intelligence and the imitation game.
1951Minsky and Dean Edmonds build SNARC, a network experiment in hardware.
1954The Georgetown–IBM demonstration draws attention to machine translation.
1955Newell, Simon and Shaw develop Logic Theorist.
Aug. 31, 1955McCarthy, Minsky, Rochester and Shannon submit the Dartmouth Summer Research Project proposal.
summer 1956The Dartmouth meeting consolidates “Artificial Intelligence” as a research identity.
1956Logic Theorist is demonstrated as an automated proof system.
1957Rosenblatt develops the perceptron.
1958McCarthy begins LISP, which will become central to symbolic AI.
1958–1959Perceptrons receive substantial public and military attention.
late 1950sNewell and Simon advance general problem solvers.
1960sLaboratories at MIT, Stanford, Carnegie Mellon and Edinburgh expand symbolic AI.
1960sDENDRAL begins exploring systems based on scientific knowledge.
1965–1966Weizenbaum develops and publishes ELIZA.
1966The ALPAC report criticizes the results of available machine translation.
1966The Shakey project begins at SRI.
1969Minsky and Papert publish Perceptrons.
1970–1971Winograd develops SHRDLU in the blocks world.
early 1970sSTRIPS formalizes planning actions through preconditions and effects.
1972Early development of MYCIN at Stanford.
1972–1973The Lighthill report criticizes British AI's scalability.
1970sFunding declines in some AI research in the US and UK: the first “AI winter.”
1970sProlog consolidates logic programming through French and European work by Colmerauer and Roussel and influences from Kowalski.
late 1970sExpert systems attract renewed commercial interest.
1980Fukushima introduces the Neocognitron, an important conceptual precursor to modern CNNs.
1980XCON begins demonstrating expert systems' commercial value.
1982Japan launches Fifth Generation Computer Systems.
1980sThe Lisp-machine market grows with expert systems.
1986Rumelhart, Hinton and Williams popularize backpropagation for distributed representations.
1987–1990The Lisp-machine market and commercial expectations for expert systems collapse.
1988Sutton systematizes temporal-difference learning.
1988Judea Pearl's work consolidates Bayesian networks as a framework for probabilistic reasoning.
1989Rabiner publishes a classic HMM tutorial.
late 1980sA second “AI winter” affects companies and funding in parts of the field.
early 1990sStatistical methods gain strength in speech, NLP and pattern recognition.
1992Q-learning receives a formal convergence proof under specific conditions.
1992Brazil holds its first National Workshop on Neural Networks and first Neural Networks School.
1992FGCS ends its eleven-year cycle in fiscal year 1992.
1992–1995TD-Gammon demonstrates self-play learning at an extraordinary level.
1995Cortes and Vapnik publish SVMs in Machine Learning.
1997Hochreiter and Schmidhuber publish LSTM.
May 1997Deep Blue defeats Kasparov 3.5–2.5.
1998LeCun and colleagues publish an influential account of gradient-based learning/CNNs for documents.
2001Breiman publishes Random Forests.
early 2000sDigital data, the internet and clusters expand machine learning's material base.
2006Hinton, Osindero and Teh publish fast training for deep belief nets.
2006–2007CUDA and general GPU programming broaden researchers' access to parallel computation.
2009ImageNet is introduced with millions of hierarchically organized images.
2009Raina, Madhavan and Ng demonstrate major GPU speedups in deep learning.
2012AlexNet dominates the ImageNet competition, catalyzing deep learning in vision.
2013Word2Vec popularizes efficient neural word embeddings.
2014Sutskever, Vinyals and Le present LSTM-based seq2seq.
2014Bahdanau, Cho and Bengio introduce learned attention in neural translation.
2014Goodfellow and colleagues introduce GANs.
2015DeepMind publishes DQN in Nature.
2015ResNet wins vision competitions using residual connections.
Jan. 2016The AlphaGo paper is published; the system had already defeated Fan Hui.
Mar. 2016AlphaGo beats Lee Sedol 4–1.
2017Vaswani and colleagues publish Attention Is All You Need.
June 2018OpenAI introduces Transformer-based generative pretraining.
2018BERT is released, later published at NAACL 2019.
2019T5 proposes a unified text-to-text framework.
2019OpenAI experiments already use human feedback to tune GPT‑2.
Jan. 2020Kaplan and colleagues publish scaling laws for language models.
May 2020GPT‑3 demonstrates strong few-shot/in-context performance at large scale.
2020C4AI USP–FAPESP–IBM begins operating in Brazil.
2020AlphaFold2 achieves a historic result at CASP14; detailed publication follows in 2021.
Jan. 2021OpenAI introduces DALL‑E.
2021CLIP demonstrates contrastive image–text pretraining on 400 million pairs.
July–Aug. 2021AlphaFold2 is published in Nature.
Aug. 2021Stanford publishes the report consolidating “foundation model.”
2022InstructGPT demonstrates the strength of instruction tuning + RLHF.
2022Diffusion models and text-to-image systems become a public phenomenon.
2022Stable Diffusion accelerates the visual-generation ecosystem with accessible weights.
Nov. 30, 2022ChatGPT launches as a research preview.
early 2023Anthropic expands access to Claude.
Mar. 14, 2023OpenAI releases GPT‑4.
July 2023Anthropic releases Claude 2.
Sept. 2023Mistral releases Mistral 7B under Apache 2.0.
Dec. 2023Google introduces Gemini and integrates it into Bard.
Dec. 2023Mixtral 8x7B reinforces sparse mixture-of-experts use.
Feb. 8, 2024Google renames Bard to Gemini.
2024Generative video and advanced multimodality broaden the generative race.
2024AlphaFold 3 extends biomolecular modeling to broader interactions.
Aug. 2024Work on optimal test-time compute scaling quantifies new inference/model tradeoffs.
Sept. 12, 2024OpenAI introduces o1-preview, making “reasoning models” a commercial category.
late 2024MCP emerges as an open protocol connecting models to data and tools.
Jan. 20, 2025DeepSeek releases R1, broadening access to reasoning models.
Mar. 2025Qwen2.5-Omni exemplifies integrated text, image, audio and video input.
Apr. 2025Meta releases Llama 4 Scout and Maverick, multimodal MoE models with accessible weights.
Apr. 2025Qwen3 introduces a hybrid line with thinking and direct-answer modes.
2025Reinforcement learning for reasoning spreads rapidly in open research.
2025Coding agents become an important commercial and research category.
Feb. 2026Anthropic releases Sonnet 4.6, emphasizing coding, computer use and agents.
Mar. 2026OpenAI release notes record GPT‑5.4 Thinking and a reorganization of the 5.x line.
Apr. 2026Meta introduces Muse Spark on its website and app; the API was in private preview. Don't confuse each interface's availability.
May 2026Google announces the Gemini 3.5 family and Managed Agents infrastructure.
July 21, 2026Google introduces Gemini 3.6 Flash, distinguishing model iteration from agent infrastructure.
July 2026xAI opens Grok Build harness components for coding agents.
July 2026xAI makes Grok 4.5 available in a coding integration; quality claims remain self-reported.
July 2026OpenAI announces the GPT‑5.6 family in retrieved official documentation.
Aug.–Sept. 2026Google continues the Gemini 3.7/3.8 cycle.
Sept. 22, 2026Anthropic documentation records Claude Opus 5.5.
Oct. 6, 2026Mistral announces Large 4 in public API preview; weights are promised for after this article's cutoff.
Oct. 7, 2026Anthropic announces Haiku 5.5.
Oct. 8, 2026Google announces a Gemini agent integrated for enterprise work.
Sept.–Oct. 2026The GPT-6 documentation records expanded access. A September 29 update confirms earlier announcements without by itself establishing the first launch date.
Oct. 10, 2026The frontier features reasoning, multimodality, agents, tools and competition between large and small models, without consensus on AGI.

announcements and access at the 2026 cutoff

These are examples announced by October 10, 2026. A manufacturer's page documents the launch and access conditions; capability claims require evaluation. Don't use this list as a ranking.

EcosystemExample documented at the cutoffWhat it helps illustrate
OpenAIGPT-6.1 Sol and GPT-6 access expansionmodels, interfaces and tool use are developed as parts of products.
AnthropicOpus 5.5 and Haiku 5.5variants address different capability, speed and cost constraints.
GoogleGemini 3.5 familylanguage, modalities and action appear in the same systems agenda.
AlibabaQwen ecosystemChinese families include language, image and other modalities.
DeepSeekV4 preview and V4.1-Flash workpapers and previews must be read with their availability conditions.
MistralLarge 4 public preview, October 6the announcement offers API testing and promises weights later; a promise isn't verified availability.
xAIGrok 4.6long-running tasks and agents appear as commercial priorities.
MetaLlama 4, announced in 2025a milestone in multimodality and weight distribution; this row doesn't claim to describe the company's latest release.

a glossary for when a word slips away

WordMeaning used here
AIa field investigating systems capable of tasks associated with intelligence, through several traditions and methods.
Symbolicworking with explicitly represented objects, relationships and rules.
Searchexploring alternatives to find a path or solution.
Heuristicpractical guidance for choosing which alternatives to explore.
State / plana relevant description of a situation / a sequence of actions to reach a goal.
Machine learningadjusting behavior from examples, data or consequences.
Dataset / benchmarkan organized data collection / a task and conditions for comparable evaluation.
Parameter / weighta value adjusted by training / a parameter modulating an operation or connection.
Training / inferencean adjustment stage / executing the model to obtain an output.
Lossa measure of predictions' cost relative to the chosen objective.
Gradient / backpropagationsensitivity of loss to parameters / calculating those sensitivities through operations.
Generalization / overfittingworking on new examples / fitting too closely to training details.
Supervised / self-supervisedlearning from supplied targets / obtaining targets from the data's own structure.
Deep learninglearning with multilayer networks that transform representations.
CNN / RNN / LSTMarchitectures exploiting local structure, recurrence and control of sequence memory, respectively.
Reinforcement learning (RL) / policylearning from reward signals / behavior guiding action selection.
Token / embeddinga discrete unit being processed / a learned numerical representation.
Attention / self-attentiona weighted combination of information / a combination among positions within the same sequence.
Transformera family of architectures with attention as a central mechanism.
Autoregressivegenerating a sequence by conditioning each continuation on permitted context.
Pretraining / fine-tuningbroad initial learning / subsequent training to adapt behavior.
Context / context windowinformation made available to the model during an execution, with limited processing capacity.
In-context learningusing supplied instructions and examples without necessarily changing weights.
LLM / foundation modela large language model / a broad base adaptable to multiple tasks and modalities.
Multimodalworking with more than one information type, such as text, images and audio.
GAN / diffusiongeneration through adversarial objectives / generation associated with learned denoising transitions.
RLHFreinforcement training using signals associated with human preferences.
RAGretrieving external information and including it in answer generation.
LoRA / QLoRAadaptation with lower-dimensional updates / combining that adaptation with quantization.
Quantization / distillationreducing numerical precision / transferring behavior or signals between models.
MoEan architecture selecting subsets of components for certain operations.
Test-time computecomputation invested during solving after receiving a problem.
Chain of thoughtintermediate steps expressed during solving; neither a correctness guarantee nor faithful access to the internal process.
VLAa vision-language-action model linking perception and instructions to actions; more specific than multimodality.
Verifiera component testing a condition or evaluating a candidate under defined criteria.
Tool use / agentexternal software executing operations / a system deciding and acting through observation cycles.
MCPa protocol connecting AI applications, sources and tools.
Hallucinationplausible output that doesn't match facts or relevant evidence.
AGI / consciousnessdistinct questions of general capability and subjective experience, requiring criteria and investigation.

where to continue reading

Sources are linked throughout. To continue, I'd start with papers' abstracts and figures, trying to understand which problem each solved. You don't need to understand all the mathematics on the first pass.

I used a documentary survey of AI history as a starting point and checked events against papers and historical documents. The recent overview runs through October 2026. Manufacturers' comparisons remain identified as their own results, with the conditions and limitations accompanying each announcement.

follow the next posts

Add this RSS feed to your feed reader to follow new posts. No account or email required.

open RSS feed ↗