Skip to main content
TechGeneTechnology histories
中文

From Symbolic Reasoning to Language Model Agents

Follow rule-based problem solving, learning from experience, deep networks, and language models to understand how computers acquired the ability to generate content and act, while tracing the parallel development of probabilistic reasoning, statistical learning, reinforcement learning, and diffusion-based generation.

How did computers move from solving problems by rules to learning from experience, generating content, and using tools to complete tasks?

Asking a machine to make a move in a board game, recognize a handwritten digit, answer a question using reference material, or even complete a task with tools may seem to call for different abilities. Yet a few questions keep recurring: How should knowledge be represented? How should experience change behavior? How can errors be detected? And how should actions be tested against their results?

Starting with the programmable computer, this story follows these questions through the existing AI technology entries. Each approach has its own conditions for use, and approaches can also be combined; where necessary, the reading sequence looks back at earlier work.

1. Once Rules Could Be Written Down, Search and Knowledge Became the Challenge

Stored-program computers provided a platform for running different algorithms. The Dartmouth proposal of 1955 placed language, neural networks, and self-improvement in a single research program, showing that AI encompassed multiple approaches from the time it acquired its name.

Symbolic methods explicitly represent facts, goals, and rules, then use reasoning or search to solve problems. When there are too many candidates, heuristic search determines the order of exploration; when specialist experience is needed, expert systems bring domain knowledge into the process. DENDRAL's work on chemical structures provides an example, and the related original materials preserve the practical work of combining knowledge with search. When evidence is incomplete, Bayesian networks use probabilities to represent and update judgments.

2. Moving from Writing Rules to Learning Patterns Opened Several Paths

Learning from experience allows systems to improve their performance using data, while requiring us to reconsider the data, objectives, and evaluation methods. Support vector machines address classification through statistical learning; neural networks explore connection structures and parameter learning. The logical model of a neuron from 1943 predates the naming of AI, while the perceptron of 1958 demonstrated an early approach to a trainable classifier.

Multilayer networks need to assign error to parameters across their layers. The 1986 backpropagation paper demonstrated how gradient computation could be combined with the learning of hidden representations. Convolutional networks used the local structure of images, while deep learning further combined multilayer representations, training methods, data, and computing power. Architecture and training together drove progress in capability.

3. Beyond Recognition, Systems Can Learn to Act and Generate

Reinforcement learning focuses on the long-term consequences of actions. It converged with deep networks in methods such as DQN, enabling systems to choose actions from game images; a 2015 study reported representative experiments with this combination. Another branch learned to generate data through diffusion and denoising, with the 2020 DDPM paper demonstrating image synthesis results. These approaches address different problems and can continue to develop independently.

4. Moving from Language Generation to Task Execution Requires More Pieces

The Transformer improved interactions among elements of a sequence and parallelism in training. Large-scale language training subsequently demonstrated broader task adaptation, with the GPT-3 study providing a representative example of few-shot in-context learning. Once models could generate answers, questions remained about how to align behavior with requirements, update reference material, and check results.

Training with human feedback uses demonstrations and preferences to adjust behavior, retrieval-augmented generation introduces material from outside the model's parameters, and ReAct demonstrates an approach in which language models alternate reasoning, action, and reading feedback from the environment. These methods affect training, information access, and execution respectively, and can be combined according to the task; an agent does not have to use all of them.

The 21 steps below cover 20 AI nodes and reuse the stored-program computer as an entry point. The main line, branches, and points of convergence organize the reading of this story; specific technical relationships and their evidence can be checked in the entries and graph. The representative papers associated with the listed years locate stages of development and do not necessarily mark the origin of an entire method.

01
Main line1948–1951

Begin with a Computer That Can Run Different Programs

With instructions and data in memory, a computer can take on different tasks by changing its program. This provides a general platform for executing search, reasoning, and learning. Computing speed and memory capacity constrain program size, but representing knowledge and improving from experience still require new methods.

Stored-program computer

A general-purpose computer architecture that stores both instructions and data in readable and writable memory and executes instruction sequences automatically.

Read technology entryExplore local graph
02
Main lineNamed in the 1955 Dartmouth proposal; the research project took place in 1956

Make Learning, Reasoning, and Language Research Questions

The Dartmouth proposal of 1955 used the name “artificial intelligence” for a research project in 1956, discussing language, neural networks, and self-improvement together. This gave AI a shared scope of research; several of its ideas had appeared earlier, and subsequent development continued along multiple parallel paths.

Artificial intelligence

A set of reusable methods and system capabilities for learning, reasoning, perception, and goal-directed action through computational systems.

Read technology entryExplore local graph
03
BranchTook shape gradually in the 1950s; a 1976 treatment of symbols and search offers a representative synthesis

Represent Knowledge as Symbols a Machine Can Manipulate

One direct approach is to represent facts, rules, and goals explicitly, then let a program derive conclusions step by step. This makes the problem-solving process inspectable while shifting the difficulty to knowledge acquisition: Which facts should be kept? Under what conditions do rules apply? What happens when unstated common sense is needed?

Symbolic artificial intelligence

An approach to artificial intelligence that represents knowledge and solves problems using explicit symbols, rules and reasoning operations.

Read technology entryExplore local graph
04
Main lineEarlier explorations existed; the 1968 A* paper serves as this entry’s verified reference point

Give Search a Direction

Even with explicit rules, there may be too many candidate steps to try individually. Heuristic search uses estimates of the goal or cost to order exploration. The 1968 A* paper is a representative example: suitable estimates can reduce unproductive exploration, while optimality depends on the conditions satisfied by the algorithm and its estimates.

Heuristic search

A method that uses problem-specific estimates to guide the search order and reduce the number of candidate states explored.

Read technology entryExplore local graph
05
ConvergenceKnowledge-based systems such as DENDRAL developed in the 1960s

Bring Domain Expertise into Problem Solving

Expert systems combine knowledge organization with inference mechanisms. In the 1960s, DENDRAL used chemical knowledge to constrain the search for candidate structures, showing how specialist experience could greatly narrow a problem. Knowledge and search work together here, but as a domain changes, adding, correcting, and maintaining knowledge still takes effort.

Expert systems

Systems that organize domain knowledge and reasoning processes to provide analysis or advice within a limited problem area.

Read technology entryExplore local graph
06
BranchDeveloped into a systematic approach in the 1980s; a 1986 paper and a 1988 book are representative works

Use Probabilities to Represent Uncertain Information

Real-world evidence is often incomplete, and judgments are not always simply true or false. Bayesian networks use graphs and conditional probabilities to represent dependencies among variables and update judgments as observations arrive. This probabilistic approach adds tools for handling uncertainty; using an ordinary graph of probabilistic dependencies for causal inference requires additional assumptions.

Bayesian networks

A method that represents dependencies between variables using a directed acyclic graph and conditional probabilities, and updates beliefs after observing evidence.

Read technology entryExplore local graph
07
Main lineSamuel's 1959 paper on learning in a board game is an early representative example, not the sole origin

Let Experience Change a Program’s Performance

When rules are difficult to write out individually, a system can adjust its model using data or experience from interaction. Samuel’s 1959 checkers research is an early example. The questions then change: What experience should be supplied? How should improvement be defined? How should performance be tested in unseen situations? Progress during training needs independent evaluation.

Machine learning

Methods that adjust a model using data or experience from interaction to improve a system's performance on a target task.

Read technology entryExplore local graph
08
BranchThe 1995 support-vector networks paper is an important reference point for the mature method

Find a Classification Boundary Through Statistical Learning

Support vector machines construct a decision boundary by balancing classification margin against training error, and can use kernel methods to represent nonlinear structure. The 1995 support-vector networks paper marks an important development. This is an independent learning approach; feature, kernel, and parameter choices affect performance, and training on large datasets also has costs.

Support vector machines

A supervised learning method that constructs decision boundaries using a margin criterion and can handle nonlinear data through kernel methods.

Read technology entryExplore local graph
09
Main lineThe 1943 McCulloch–Pitts model of a logical neuron is an early reference point

Look Back at Earlier Ideas in Neural Computation

Early logical models of neural networks date to 1943, before the name “artificial intelligence.” They performed computation by connecting simple units; later networks added learnable parameters. How a network should be connected and how its parameters should learn to perform a task are two separate problems to solve.

Artificial neural networks

Interconnected artificial computational units use their connection structure and parameters to represent transformations from inputs to outputs.

Read technology entryExplore local graph
10
BranchRosenblatt's 1958 paper; this entry focuses on the single-layer classification model

Let a Classifier Adjust Its Weights from Errors

The perceptron of 1958 provided an early approach combining neural computation with learning from examples. Its single-layer classification form makes decisions from weighted inputs, then adjusts weights using misclassified examples. It made the learning process concrete while exposing limits in expressive power: a single-layer linear classifier cannot directly handle arbitrary nonlinear classification problems.

Perceptron

An early neural learning model that classifies inputs using a weighted sum and a threshold, and adjusts weights using misclassified examples.

Read technology entryExplore local graph
11
ConvergenceEarlier work on differentiation and training preceded it; a 1986 paper advanced representation learning in multilayer networks

Propagate Output Error Back Through a Multilayer Network

Adding intermediate layers allows more complex transformations, but how should parameters in each layer share responsibility for output error? Backpropagation computes gradients backward through the computation, providing the basis for updates by an optimization method. The representative 1986 paper demonstrated the learning of hidden representations; earlier work on differentiation already existed, and gradients alone do not guarantee a global optimum.

Backpropagation

A method for computing parameter gradients in differentiable networks by applying the chain rule to propagate derivatives from the output toward earlier layers.

Read technology entryExplore local graph
12
Main lineResearch began earlier; the 1998 document-recognition paper serves as this entry’s engineering reference point

Let Networks Use the Local Structure of Images

Similar patterns can appear in different parts of an image. Convolutional networks use local connections and shared parameters to exploit this structure, combined with gradient-based learning. A 1998 paper on document recognition demonstrated this engineering approach. The architecture reduces unnecessary parameters, but effective representations still require suitable data and training.

Convolutional neural networks

Neural networks that use local connections and weight sharing to process spatial or sequential structure.

Read technology entryExplore local graph
13
ConvergenceResearch on training deep models in 2006 and large-scale vision applications in 2012 are important development milestones

Train Multilayer Representations into Useful Capabilities

Deep learning combines multilayer representations, training methods, data, and computing power. Research on training deep networks in 2006 and the ImageNet convolutional network results of 2012 demonstrated important advances in methods and engineering respectively. Several conditions come together here: multilayer networks were not new at this point, and greater scale alone does not ensure reliability.

Deep learning

A machine-learning approach that progressively extracts and combines features through multiple layers of trainable representations.

Read technology entryExplore local graph
14
BranchIts historical roots are earlier; the 1992 Q-learning paper is the representative milestone used in this entry

Learn from the Results of a Sequence of Actions

Some tasks provide no correct answer at each step, only rewards after actions. Reinforcement learning uses these rewards to improve a policy, focusing on long-term cumulative return. The 1992 Q-learning paper is an important example of value learning. How rewards are defined, how costly exploration is, and whether the environment changes all affect the behavior learned.

Reinforcement learning

A method for optimizing long-term cumulative return by interacting with an environment, receiving rewards and improving an action policy.

Read technology entryExplore local graph
15
ConvergenceThe 2015 DQN paper is a representative reference point for decision-making from pixel inputs; research began earlier

Use Deep Representations to Select Actions

Deep reinforcement learning combines representation learning for complex inputs with sequential decision-making. In 2015, DQN used convolutional networks to estimate action values from game images, bringing two approaches together. This expanded the range of inputs systems could handle, while raising issues of training stability and resource requirements; game performance needs to be tested again in new environments.

Deep reinforcement learning

Reinforcement learning methods that use deep networks to represent values, policies, or models of the environment.

Read technology entryExplore local graph
16
Branch2015: the diffusion-based generative approach; 2020: representative DDPM image-generation results

Learn to Generate Along Another Path

Diffusion generative models gradually add noise to data, then learn the reverse process of recovering structure. A 2015 study proposed this approach to generative modeling, and DDPM demonstrated high-quality image synthesis in 2020. It offers another way to generate content, developing alongside language sequence prediction, while leaving issues such as the cost of multistep sampling.

Diffusion generative models

Generative models that recover data structure from noise by learning a stepwise denoising or reverse-diffusion process.

Read technology entryExplore local graph
17
Main lineThe architecture was proposed in Attention Is All You Need in 2017

Let Positions in a Sequence Exchange Information Directly

The Transformer uses attention to organize sequence information, allowing different positions to aggregate one another’s content as the task requires. The original 2017 paper demonstrated the advantages of this architecture and training parallelism on translation tasks. The architecture provided a foundation for later models, but training objectives, data, and scale still determine what a model can learn.

Transformer

A neural network architecture that organizes sequence information processing through attention mechanisms and feed-forward transformations.

Read technology entryExplore local graph
18
Main lineLanguage models and pretraining existed earlier; the 2020 GPT-3 paper is a representative milestone in scaling

Use One Language Model for Many Types of Tasks

Large-scale pretraining enables a language model to handle multiple language tasks through context. The 2020 GPT-3 paper demonstrated representative results with few-shot prompting: a task description or examples can guide the same model to produce appropriate outputs. Language models and pretraining have earlier histories; fluent answers may still contain factual errors.

Large language models

Models trained on large-scale language data that can predict, generate and process language sequences.

Read technology entryExplore local graph
19
Convergence2017: research on rewards learned from preferences; 2022: application to language-model instruction following

Use Human Evaluations to Adjust Model Behavior

Reinforcement learning from human feedback turns preference comparisons into a learning signal, then uses reinforcement learning to adjust behavior. A 2017 study explored learning rewards from human preferences, while language model research in 2022 combined this with training on demonstrations to improve instruction following. Preferences reflect evaluators’ judgments; factual accuracy and task performance still require separate checks.

Reinforcement learning from human feedback

A method that learns a reward signal from human evaluations or preferences, then uses reinforcement learning to adjust model behavior.

Read technology entryExplore local graph
20
BranchThe paper naming RAG appeared in 2020; research combining retrieval and generation predates it

Give Generation External Reference Material

Knowledge in a model’s parameters is difficult to update and hard to trace to specific sources. Retrieval-augmented generation conditions generation on external material, with the 2020 RAG paper as a representative example. It combines information retrieval with language generation; whether retrieval is complete, whether the material is trustworthy, and whether the answer is faithful to it remain three separate questions to check.

Retrieval-augmented generation

A method that introduces content retrieved from external materials during generation to supplement knowledge stored in the model's parameters.

Read technology entryExplore local graph
21
ConvergenceRepresented by the 2022 ReAct preprint and 2023 conference paper; the concept of agents predates them

Connect Language, Tools, and Environmental Feedback

Language model agents place a model within a multistep task process: choose an action, call a tool, read the result, and decide what to do next. The ReAct study of 2022 demonstrated a representative method for alternating reasoning and action. Systems can add retrieval or use models trained with human feedback as needed; neither is required for every agent. The final check is the task result, and errors across successive steps also need to be detected promptly.

Language-model agents

Task-execution systems that combine language models with tool use, environmental feedback, and multistep decision-making.

Read technology entryExplore local graph

These approaches keep offering new solutions to questions about knowledge, experience, representations, objectives, and feedback. Symbolic rules express explicit constraints; statistical and neural methods learn from data; reinforcement learning addresses the consequences of actions; and retrieval and tools connect models to external information and environments.

Each advance also leaves new things to check: whether training data are suitable, whether rewards represent the real objective, whether the reference material supports an answer, and whether a sequence of actions achieves the expected result. Language model agents bring these questions into a single execution process, whose reliability still has to be judged from evidence in specific tasks.

To revisit how computing hardware took shape, continue with From Vacuum Tubes to the Microprocessor. To compare different AI approaches, explore the local graph of artificial intelligence.

Relations in this story

Selected technologyCurrent rangeOutside this rangeDrag to pan · Scroll to zoom
Relation filters
Visible relationsLoaded · 30 relations

References

  1. A Formal Basis for the Heuristic Determination of Minimum Cost Paths

    A Formal Basis for the Heuristic Determination of Minimum Cost Paths

    Authors / institution: Peter E. Hart, Nils J. Nilsson, Bertram Raphael.

    Publication date: 1968.

    Original text / materials: A Formal Basis for the Heuristic Determination of Minimum Cost Paths

    Scope of evidence: Supports heuristic graph search and optimality under specified conditions; 1968 is the date of the A* paper, not the origin of all heuristic search.

    Accessed on: 2026-10-13.

  2. Learning Representations by Back-propagating Errors

    Learning Representations by Back-propagating Errors

    Authors / institution: David E. Rumelhart, Geoffrey E. Hinton, Ronald J. Williams.

    Published: 1986.

    Original text / resources: Learning Representations by Back-propagating Errors

    Scope of evidence: Supports error backpropagation, weight adjustment, and representation learning in hidden units. The 1986 paper is a representative publication, not a basis for claiming that chain-rule differentiation or reverse-mode differentiation began in 1986.

    Accessed: 2026-10-13.

  3. Fusion, Propagation, and Structuring in Belief Networks — Commentary on a Classic Paper Hosted by the Original Author

    Fusion, Propagation, and Structuring in Belief Networks — Commentary on a Classic Paper Hosted by the Original Author

    Authors / institution: Judea Pearl; AIJ Classic Paper Award.

    Publication date: 1986 paper, 2015 commentary.

    Original text / materials: Fusion, Propagation, and Structuring in Belief Networks — Commentary on a Classic Paper Hosted by the Original Author

    Scope of evidence: Supports the use of graphs to represent independence in Bayesian networks and belief propagation in tree-structured or singly connected networks. Inference in arbitrary networks is not described as having linear complexity.

    Accessed on: 2026-10-13.

  4. Gradient-based Learning Applied to Document Recognition

    Gradient-based Learning Applied to Document Recognition

    Authors / institution: Yann LeCun, Léon Bottou, Yoshua Bengio, Patrick Haffner.

    Published: 1998.

    Original text / resources: Gradient-based Learning Applied to Document Recognition

    Scope of evidence: The author's page and the original paper support the combination of convolutional networks, gradient-based learning, and document recognition. The 1998 paper summarizes engineering work; it does not mark the first appearance of CNNs.

    Accessed: 2026-10-13.

  5. A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence

    A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence

    Authors / institution: John McCarthy, Marvin Minsky, Nathaniel Rochester, Claude Shannon.

    Publication date: 1955-08-31.

    Original text / materials: A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence

    Scope of evidence: Supports the name “AI,” the 1956 research project, and the use of automatic computers as a platform for research. The year 1955 is the date of the proposal, not a common birth year for all intelligent technologies.

    Accessed on: 2026-10-13.

  6. Denoising Diffusion Probabilistic Models

    Denoising Diffusion Probabilistic Models

    Authors / institution: Jonathan Ho, Ajay Jain, Pieter Abbeel.

    Published: 2020.

    Original text / resources: Denoising Diffusion Probabilistic Models

    Scope of evidence: Supports an example of high-quality image generation with denoising diffusion probabilistic models. It is not directly extrapolated to arbitrary video or scientific tasks.

    Accessed: 2026-10-13.

  7. A Fast Learning Algorithm for Deep Belief Nets

    A Fast Learning Algorithm for Deep Belief Nets

    Authors / institution: Geoffrey E. Hinton, Simon Osindero, Yee-Whye Teh.

    Publication date: 2006.

    Original text / materials: A Fast Learning Algorithm for Deep Belief Nets

    Scope of evidence: Supports one historical approach to training deep representations. Layer-wise pretraining is not treated as a prerequisite for all deep learning.

    Accessed on: 2026-10-13.

  8. DENDRAL — Original Papers and Software Archive

    DENDRAL — Original Papers and Software Archive

    Authors / institution: Computer History Museum, with original materials by Buchanan, Sutherland, Feigenbaum, and others.

    Published: Includes materials from the 1960s onward.

    Original text / resources: DENDRAL — Original Papers and Software Archive

    Scope of evidence: Supports an early example of knowledge-based systems using domain knowledge to constrain search. DENDRAL is used only as an example, without a separate product node.

    Accessed: 2026-10-13.

  9. Deep Unsupervised Learning using Nonequilibrium Thermodynamics

    Deep Unsupervised Learning using Nonequilibrium Thermodynamics

    Authors / institution: Jascha Sohl-Dickstein et al.

    Publication date: 2015.

    Original text / materials: Deep Unsupervised Learning using Nonequilibrium Thermodynamics

    Scope of evidence: Supports an approach to generative modeling in which forward diffusion destroys the structure of data and a learned reverse process restores it.

    Accessed on: 2026-10-13.

  10. Human-level Control through Deep Reinforcement Learning

    Human-level Control through Deep Reinforcement Learning

    Authors / institution: Volodymyr Mnih et al.

    Published: 2015.

    Original text / resources: Human-level Control through Deep Reinforcement Learning

    Scope of evidence: Supports the use of deep networks to approximate action values and the combination of pixel inputs with reinforcement learning. Results in specific games do not establish general capability in real-world environments.

    Accessed: 2026-10-13.

  11. ImageNet Classification with Deep Convolutional Neural Networks

    ImageNet Classification with Deep Convolutional Neural Networks

    Authors / institution: Alex Krizhevsky, Ilya Sutskever, Geoffrey E. Hinton.

    Publication date: 2012.

    Original text / materials: ImageNet Classification with Deep Convolutional Neural Networks

    Scope of evidence: Supports the engineering combination of large-scale image data, deep convolutional networks, and GPU training. It supports this particular example of the approach, not a claim that deep learning replaces all other methods.

    Accessed on: 2026-10-13.

  12. Language Models are Few-Shot Learners

    Language Models are Few-Shot Learners

    Authors / institution: Tom B. Brown et al.

    Published: 2020.

    Original text / resources: Language Models are Few-Shot Learners

    Scope of evidence: Supports an example of large-scale autoregressive language models and in-context few-shot learning. GPT-3 is a research example rather than a separate commercial product node; this paper is not a basis for dating the start of language modeling to 2020.

    Accessed: 2026-10-13.

  13. What is Artificial Intelligence? — Basic Questions

    What is Artificial Intelligence? — Basic Questions

    Authors / institution: John McCarthy.

    Publication date: Author's online explanation.

    Original text / materials: What is Artificial Intelligence? — Basic Questions

    Scope of evidence: Supports the engineering scope and goals of artificial intelligence. The definition is not treated as evidence that general intelligence has already been achieved.

    Accessed on: 2026-10-13.

  14. A Logical Calculus of the Ideas Immanent in Nervous Activity

    A Logical Calculus of the Ideas Immanent in Nervous Activity

    Authors / institution: Warren S. McCulloch, Walter Pitts.

    Published: 1943.

    Original text / resources: A Logical Calculus of the Ideas Immanent in Nervous Activity

    Scope of evidence: Supports an early logical model of artificial neural networks. The model itself is not equivalent to a modern trainable deep network.

    Accessed: 2026-10-13.

  15. The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain

    The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain

    Authors / institution: Frank Rosenblatt.

    Publication date: 1958.

    Original text / materials: The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain

    Scope of evidence: Verifies the paper's author, year, and the perceptron as an early learning model. The bibliographic page has no abstract; the account of principles also draws on papers on neural networks and backpropagation to avoid extending the limitations of a single layer to all networks.

    Accessed on: 2026-10-13.

  16. Deep Reinforcement Learning from Human Preferences

    Deep Reinforcement Learning from Human Preferences

    Authors / institution: Paul F. Christiano et al.

    Published: 2017.

    Original text / resources: Deep Reinforcement Learning from Human Preferences

    Scope of evidence: Supports learning a reward function from human comparisons of trajectory segments, then using it for reinforcement learning.

    Accessed: 2026-10-13.

  17. Q-learning

    Q-learning

    Authors / institution: Christopher J. C. H. Watkins, Peter Dayan.

    Publication date: 1992.

    Original text / materials: Q-learning

    Scope of evidence: Supports action-value learning from rewards received through interaction. Q-learning is used here only as a representative reference point for reinforcement learning, not to assert that reinforcement learning originated in 1992.

    Accessed on: 2026-10-13.

  18. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks

    Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks

    Authors / institution: Patrick Lewis et al.

    Published: 2020.

    Original text / resources: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks

    Scope of evidence: Supports the combination of a pretrained generative model with external non-parametric memory. The 2020 paper presents the method under this name; it does not mark the beginning of all combinations of retrieval and generation.

    Accessed: 2026-10-13.

  19. ReAct: Synergizing Reasoning and Acting in Language Models

    ReAct: Synergizing Reasoning and Acting in Language Models

    Authors / institution: Shunyu Yao et al.

    Publication date: 2022 preprint, 2023 ICLR.

    Original text / materials: ReAct: Synergizing Reasoning and Acting in Language Models

    Scope of evidence: Supports language models alternating between reasoning and actions and reading environmental feedback. It does not establish that the concept of agents began in 2022 or that reliable general-purpose autonomy has been achieved.

    Accessed on: 2026-10-13.

  20. Training Language Models to Follow Instructions with Human Feedback

    Training Language Models to Follow Instructions with Human Feedback

    Authors / institution: Long Ouyang et al.

    Published: 2022.

    Original text / resources: Training Language Models to Follow Instructions with Human Feedback

    Scope of evidence: Supports a language-model post-training process involving supervised demonstrations, preference rankings, a reward model, and reinforcement learning. Improved preference alignment does not guarantee correctness or safety.

    Accessed: 2026-10-13.

  21. Some Studies in Machine Learning Using the Game of Checkers

    Some Studies in Machine Learning Using the Game of Checkers

    Authors / institution: Arthur L. Samuel.

    Publication date: 1959.

    Original text / materials: Some Studies in Machine Learning Using the Game of Checkers

    Scope of evidence: Supports an example of machine learning in which a game-playing program improves through experience. It does not constitute the invention of all machine-learning methods.

    Accessed on: 2026-10-13.

  22. Support-vector Networks

    Support-vector Networks

    Authors / institution: Corinna Cortes, Vladimir Vapnik.

    Published: 1995.

    Original text / resources: Support-vector Networks

    Scope of evidence: Supports support-vector methods, feature spaces, and the extension to non-separable training data. The 1995 paper acknowledges earlier work on the separable case.

    Accessed: 2026-10-13.

  23. Computer Science as Empirical Inquiry: Symbols and Search

    Computer Science as Empirical Inquiry: Symbols and Search

    Authors / institution: Allen Newell, Herbert A. Simon.

    Publication date: 1976.

    Original text / materials: Computer Science as Empirical Inquiry: Symbols and Search

    Scope of evidence: Supports an approach based on symbolic representation, symbol manipulation, and search. The physical symbol system hypothesis is a theoretical claim and is not recorded as a proven universal theorem.

    Accessed on: 2026-10-13.

  24. Attention Is All You Need

    Attention Is All You Need

    Authors / institution: Ashish Vaswani et al.

    Published: 2017.

    Original text / resources: Attention Is All You Need

    Scope of evidence: Supports the original Transformer's attention architecture and translation experiments. It does not establish the first proposal of the attention mechanism itself.

    Accessed: 2026-10-13.

  25. First Steps: Lectures from the Dawn of Computing

    First Steps: Lectures from the Dawn of Computing

    URL: https://computerhistory.org/blog/first-steps-lectures-from-the-dawn-of-computing/

    Publisher: Computer History Museum

    Accessed: 2026-08-17

  26. 1946: Williams demonstrates CRT storage

    1946: Williams demonstrates CRT storage

    URL: https://www.computerhistory.org/storageengine/williams-demonstrates-crt-storage/

    Publisher: Computer History Museum

    Accessed on: 2026-08-17