Milestones

How we got from early neural networks to today’s language models and agents: a selective timeline for locating the course’s ideas in time

This page is a selective intellectual timeline, not a history of artificial intelligence. It makes no claim to completeness. Each entry was chosen because it helps place a concept taught in this course in time; many famous systems, and many important lines of work, are left out because the course does not depend on them. Dates are those of the cited publication: an arXiv first posting, a journal issue, or a dated article. Where a paper was revised later, the entry says so.

A note on wording. The history of this field is full of overlapping antecedents, and “first”, “invented”, and “discovered” are rarely safe. This page therefore says what a cited source did and how it described itself, and it uses words such as “became influential” or “helped establish” rather than asserting priority. Where a paper’s own abstract calls its method new, the entry quotes that self-description rather than endorsing it. [S-74]

What the timeline should make clear

Read as a whole, the entries below support seven points that are easy to get wrong from the outside.

  • Neural networks substantially predate language models. Learning machines built from adjustable units were being studied in the 1950s.
  • Backpropagation did not arrive with transformers. The loss-gradient-update loop this course works by hand was influential in neural-network training by the mid-1980s.
  • A transformer is a neural-network architecture, not the beginning of neural networks. It is built from the same units, layers, and training procedure as its predecessors, arranged differently.
  • Attention had important precursors before 2017. A learned, weighted selection over input positions was used in translation systems several years before the Transformer made it the whole architecture.
  • Large pretrained models required more than the architecture. The practice of pretraining once and reusing, the effect of scale, and very large text corpora were separate developments.
  • Post-training and agent systems are later layers. They sit on top of, or around, a pretrained model; neither is a new kind of network.
  • Today’s systems are the result of several interacting lines of work, not one isolated invention. [S-74]

The rail

Neural networks before language models

1958 · The perceptron

What changed. Frank Rosenblatt’s paper in Psychological Review developed the perceptron, described in its title as “a probabilistic model for information storage and organization in the brain”, as a hypothetical learning system. [S-64] The 1986 backpropagation paper listed below names the “perceptron-convergence procedure” as an “earlier, simpler” learning method than its own, which locates the perceptron as an ancestor of the training loop this course teaches. [S-65]

Why it matters here. The basic object of the neural-networks lesson, a unit that forms a weighted sum and passes it through a nonlinearity, is decades older than any language model. Nothing in the perceptron involves language; it is here to fix the starting point.

In the course: Neurons, Layers, and Learning §3–§5.

Source: Rosenblatt, F. (1958), “The perceptron: A probabilistic model for information storage and organization in the brain”, Psychological Review 65(6), 386–408, https://doi.org/10.1037/h0042519. The entry describes the paper by its title and by how the 1986 paper characterizes it; it does not summarize the paper’s own mechanism. [S-64]

1986 · Backpropagation becomes practically influential

What changed. Rumelhart, Hinton, and Williams described in Nature what their abstract calls “a new learning procedure, back-propagation, for networks of neurone-like units”, which “repeatedly adjusts the weights of the connections in the network so as to minimize a measure of the difference between the actual output vector of the net and the desired output vector”, so that hidden units “come to represent important features of the task domain”. [S-65]

Why it matters here. This is the loss, gradient, and update loop the course works through on a single unit, and this influential 1986 formulation predates the Transformer by more than three decades. Two precisions. The course makes no claim about who first derived the gradient computation; the paper is listed for the importance of its formulation, and its “new” is the abstract’s own word. And the 1986 paper uses “back-propagation” more broadly, for the whole learning procedure that adjusts weights, whereas this course uses backpropagation specifically for the gradient computation and reserves the update for the optimizer. [S-14]

In the course: Neurons, Layers, and Learning §8–§10.

Source: Rumelhart, D. E., Hinton, G. E., and Williams, R. J. (1986), “Learning representations by back-propagating errors”, Nature 323, 533–536, https://www.nature.com/articles/323533a0.

1997 · Long short-term memory

What changed. Hochreiter and Schmidhuber addressed a known difficulty in training recurrent networks to use information across long time intervals, which their abstract attributes to “insufficient, decaying error back flow”, by introducing what the abstract calls “a novel, efficient, gradient-based method called ‘Long Short-Term Memory’ (LSTM)”. The paper reviews earlier recurrent-network work at length and builds on an earlier analysis of the problem. [S-66]

Why it matters here. Recurrent networks, and LSTM in particular, were the established approach to sequence modeling that the Transformer paper compared itself against two decades later. [S-21] The sequential dependency the course uses to motivate self-attention is a property of this family: each step’s summary waits for the previous one.

In the course: Why Transformers §1–§2.

Source: Hochreiter, S., and Schmidhuber, J. (1997), “Long Short-Term Memory”, Neural Computation 9(8), 1735–1780, https://doi.org/10.1162/neco.1997.9.8.1735; author-hosted copy at https://www.bioinf.jku.at/publications/older/2604.pdf.

Learning representations of language

2003 · A neural probabilistic language model

What changed. Bengio, Ducharme, Vincent, and Jauvin proposed to “learn simultaneously a distributed representation for each word along with the probability function for word sequences”: each word is mapped to a learned vector through a table that is “a \(|V| \times m\) matrix of free parameters”, and the network’s parameters are tuned “to maximize the log-likelihood of the training data”. [S-29] [S-41]

Why it matters here. Two of the course’s central objects are already here: a learned embedding table, and next-token prediction trained by maximum likelihood. The network is not a transformer and its vocabulary is words rather than subword tokens, and the course does not claim this was the first such model; it is a clear early statement of the idea.

In the course: From Text to Numbers §5; How Pretraining Works §3.

Source: Bengio, Y., Ducharme, R., Vincent, P., and Jauvin, C. (2003), “A Neural Probabilistic Language Model”, Journal of Machine Learning Research 3, 1137–1155, https://www.jmlr.org/papers/volume3/bengio03a/bengio03a.pdf.

2014 · Sequence-to-sequence learning

What changed. Sutskever, Vinyals, and Le presented “a general end-to-end approach to sequence learning” that “uses a multilayered Long Short-Term Memory (LSTM) to map the input sequence to a vector of a fixed dimensionality, and then another deep LSTM to decode the target sequence from the vector”. [S-67]

Why it matters here. This is a recurrent network generating an output sequence one token at a time from an encoded input, the encoder–decoder shape that the Transformer later kept. The single fixed-length vector between encoder and decoder is also the bottleneck that the next entry addressed.

In the course: Why Transformers §1–§3.

Source: Sutskever, I., Vinyals, O., and Le, Q. V. (2014), “Sequence to Sequence Learning with Neural Networks”, arXiv:1409.3215 (v1 2014-09-10; v3 2014-12-14), https://arxiv.org/abs/1409.3215.

2014–2015 · Attention for neural translation

What changed. Bahdanau, Cho, and Bengio conjectured “that the use of a fixed-length vector is a bottleneck in improving the performance of this basic encoder-decoder architecture” and proposed “allowing a model to automatically (soft-)search for parts of a source sentence that are relevant to predicting a target word”. The paper was posted in September 2014 and, per its arXiv listing, accepted at ICLR 2015. [S-68]

Why it matters here. A learned, weighted selection over input positions, the mechanism now generally called attention, was in use several years before the Transformer. The 2017 architecture’s contribution was to build the network from this mechanism rather than to add it to a recurrent one. The abstract itself does not use the word “attention”.

In the course: How Self-Attention Moves Information §1–§2.

Source: Bahdanau, D., Cho, K., and Bengio, Y., “Neural Machine Translation by Jointly Learning to Align and Translate”, arXiv:1409.0473 (v1 2014-09-01; v7 2016-05-19), https://arxiv.org/abs/1409.0473.

2015 · Subword vocabularies

What changed. Sennrich, Haddow, and Birch proposed making a translation model “capable of open-vocabulary translation by encoding rare and unknown words as sequences of subword units”, using “a segmentation based on the byte pair encoding compression algorithm”. [S-27]

Why it matters here. The paper helped establish subword tokenization as a practical way to handle open vocabularies in neural language systems, which is the reason a token in this course is often a word fragment rather than a word. Modern tokenizers use several related subword- and byte-based approaches, with their own rules about spaces and punctuation, so the course treats this as an important step rather than the unique origin of current tokenization.

In the course: Tokens, Embeddings, and the Context Window §2–§4.

Source: Sennrich, R., Haddow, B., and Birch, A., “Neural Machine Translation of Rare Words with Subword Units”, arXiv:1508.07909 (posted August 2015; v5 2016-06-10), https://arxiv.org/abs/1508.07909.

The transformer, pretraining, and scale

2017 · The Transformer

What changed. Vaswani and colleagues stated that the dominant sequence-transduction models of the time were “based on complex recurrent or convolutional neural networks”, with LSTM and gated recurrent networks “firmly established as state of the art”, and proposed an architecture built from attention and feed-forward layers without recurrence, reporting it “superior in quality while being more parallelizable and requiring significantly less time to train” on the two translation tasks tested. [S-21] Removing the sequential dependency of recurrence is given as a central motivation. [S-16] [S-23]

Why it matters here. Transformer architectures became the dominant foundation for many modern large language models, and the course’s why-transformers lesson is built around this paper’s own comparison of path lengths and sequential operations. Two cautions. The Transformer is a neural-network architecture, assembled from the units, layers, and training loop of the earlier entries. And its attention mechanism has the 2014–2015 precursor above; the paper’s defining architectural move was to build sequence transduction around attention and feed-forward computation rather than recurrence or convolution.

In the course: Why Transformers; How Self-Attention Moves Information.

Source: Vaswani, A., et al. (2017), “Attention Is All You Need”, arXiv:1706.03762 (v1 2017-06-12; v7 2023-08-02), https://arxiv.org/abs/1706.03762.

2018–2019 · Pretraining and transfer

What changed. Three reports helped establish transformer pretraining and transfer as a central recipe in natural-language processing. Radford, Narasimhan, Salimans, and Sutskever (2018) demonstrated “that large gains on these tasks can be realized by generative pre-training of a language model on a diverse corpus of unlabeled text, followed by discriminative fine-tuning on each specific task”, using a transformer in a two-stage procedure; the paper names earlier pretrain-then-fine-tune work as its closest antecedent. [S-69] Devlin and colleagues (BERT, 2018) reported that a pretrained bidirectional transformer “can be fine-tuned with just one additional output layer” for a wide range of tasks. [S-24] Radford and colleagues (GPT-2, 2019) reported that language models “begin to learn these tasks without any explicit supervision” when trained on a large corpus of web text. [S-44]

Why it matters here. The three contributions are distinct: generative pretraining followed by task fine-tuning; pretrained bidirectional representations with light task adaptation; and task behavior appearing without explicit task-specific supervision. Together they show that large-scale pretraining, data, compute, and the reuse of pretrained models were additional developments beyond the Transformer architecture itself. GPT-2 is also the source of the course’s tokenizer figures.

In the course: Why Transformers §5; How Pretraining Works §6.

Sources: Radford, A., Narasimhan, K., Salimans, T., and Sutskever, I., “Improving Language Understanding by Generative Pre-Training” (OpenAI, 2018), https://cdn.openai.com/research-covers/language-unsupervised/language_understanding_paper.pdf. Devlin, J., et al., “BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding”, arXiv:1810.04805 (v1 2018-10-11; v2 2019-05-24), https://arxiv.org/abs/1810.04805. Radford, A., et al., “Language Models are Unsupervised Multitask Learners” (OpenAI, 2019), https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf.

2020 · Scale, and behavior from demonstrations in the input

What changed. Kaplan and colleagues reported that language-model loss “scales as a power-law with model size, dataset size, and the amount of compute used for training, with some trends spanning more than seven orders of magnitude”. [S-25] Brown and colleagues reported that GPT-3, “an autoregressive language model with 175 billion parameters”, could be applied “without any gradient updates or fine-tuning, with tasks and few-shot demonstrations specified purely via text interaction”, and that “scaling up language models greatly improves task-agnostic, few-shot performance”. [S-26]

Why it matters here. These results showed that scale, training data, and inference-time demonstrations were important empirical developments beyond the architecture itself. The second result shows behavior being steered by inference-time context alone, with no parameter change: the distinction between parameters and context that runs through the whole course. Both are empirical findings for the systems and setups studied; the course describes in-context behavior as observed task performance, not as learning in the training sense.

In the course: Why Transformers §5; Tokens, Embeddings, and the Context Window §8.

Sources: Kaplan, J., et al., “Scaling Laws for Neural Language Models”, arXiv:2001.08361 (v1 2020-01-23), https://arxiv.org/abs/2001.08361. Brown, T., et al., “Language Models are Few-Shot Learners”, arXiv:2005.14165 (v4 2020-07-22), https://arxiv.org/abs/2005.14165.

2020 · Retrieval-augmented generation

What changed. Lewis and colleagues observed that pretrained models’ “ability to access and precisely manipulate knowledge is still limited” and described models “which combine pre-trained parametric and non-parametric memory for language generation”, the non-parametric memory being an index of retrievable text. [S-33]

Why it matters here. This became an influential formulation of combining pretrained generation with retrieved external information. Modern assistants and agents use a broader range of retrieval and tool-based mechanisms; the course abstracts from the paper’s specific architecture, a learned retriever paired with a generator, to the general idea of bringing external evidence into inference-time context, where it changes the inputs to one call and not the parameters.

In the course: Tokens, Embeddings, and the Context Window §10; From Assistant to Agent §7.

Source: Lewis, P., et al., “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks”, arXiv:2005.11401 (v4 2021-04-12), https://arxiv.org/abs/2005.11401.

Post-training, tools, and agents

2022 · Chain-of-thought prompting

What changed. Wei and colleagues reported that “generating a chain of thought – a series of intermediate reasoning steps – significantly improves the ability of large language models to perform complex reasoning”, elicited by providing “a few chain of thought demonstrations” in the prompt, with the effect appearing in “sufficiently large” models. [S-70]

Why it matters here. It is an early, widely cited demonstration that on some difficult tasks, generating intermediate reasoning tokens provides additional inference-time computation and can improve performance. Note what it is: a prompting technique, applied at inference time, with no parameter change. The 2025 entry below is about training models to produce such traces. The course treats generated reasoning as output, not as a transcript of internal computation.

In the course: Post-training §7; Using Models Well §9.

Source: Wei, J., et al., “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models”, arXiv:2201.11903 (v1 2022-01-28; v6 2023-01-10), https://arxiv.org/abs/2201.11903.

2022 · Instruction following from human feedback

What changed. Ouyang and colleagues fine-tuned a pretrained model first on “labeler demonstrations of the desired model behavior” with supervised learning, then on “rankings of model outputs” using reinforcement learning from human feedback, and reported that outputs of their 1.3-billion-parameter post-trained model were preferred to those of the 175-billion-parameter pretrained model. [S-46] The preference-learning component builds on earlier reinforcement-learning work in which a reward was learned from “human preferences between pairs of trajectory segments” rather than specified by hand. [S-71]

Why it matters here. This is the post-training layer: parameter updates after pretraining that shape a model into an assistant. The course presents it as one provider’s 2022 recipe and as one route through a family of methods; “RLHF” names that route, not post-training as a whole.

In the course: Post-training §3 and §5.

Sources: Ouyang, L., et al., “Training language models to follow instructions with human feedback”, arXiv:2203.02155 (v1 2022-03-04), https://arxiv.org/abs/2203.02155. Christiano, P., et al., “Deep reinforcement learning from human preferences”, arXiv:1706.03741 (v1 2017-06-12; v4 2023-02-17), https://arxiv.org/abs/1706.03741.

2022 · Reasoning traces interleaved with actions

What changed. Yao and colleagues explored “the use of LLMs to generate both reasoning traces and task-specific actions in an interleaved manner”, where actions “allow it to interface with external sources, such as knowledge bases or environments, to gather additional information”, for example through a simple Wikipedia API. [S-72]

Why it matters here. The model, action, and observation loop this course describes, in which the model generates a request, software executes it, and the observation returns to the context for the next step, closely resembles this pattern, and ReAct is an influential early formulation of it in the language-model setting. It is not claimed as the origin of agents; tool use and environment interaction have longer histories the course does not trace.

In the course: From Assistant to Agent §2 and §5.

Source: Yao, S., et al., “ReAct: Synergizing Reasoning and Acting in Language Models”, arXiv:2210.03629 (v1 2022-10-06; v3 2023-03-10, the ICLR camera-ready version), https://arxiv.org/abs/2210.03629.

2023 · Direct preference optimization

What changed. Rafailov and colleagues characterized the reward-model route as “a complex and often unstable procedure” and proposed updating the language model directly from preference pairs with a classification-style loss, “eliminating the need for sampling from the LM during fine-tuning”, reporting results that matched or improved on the reward-model route on the tasks tested. [S-47]

Why it matters here. It shows that preference-based post-training is a family with more than one implementation, which is why the course keeps “RLHF” and “post-training” apart. Both routes remain in use.

In the course: Post-training §5.

Source: Rafailov, R., et al., “Direct Preference Optimization: Your Language Model is Secretly a Reward Model”, arXiv:2305.18290 (v1 2023-05-29; v3 2024-07-29), https://arxiv.org/abs/2305.18290.

2024 · Workflow versus agent

What changed. An engineering article from Anthropic distinguished workflows, “systems where LLMs and tools are orchestrated through predefined code paths”, from agents, “systems where LLMs dynamically direct their own processes and tool usage”, recommended “finding the simplest solution possible”, and described the basic building block as a model augmented with retrieval, tools, and memory. [S-50] As of this course’s access date, the article carries a note that much of the tooling landscape it describes has changed since December 2024. [S-73]

Why it matters here. This is the engineering vocabulary the course adopts for the last step of its story. It is engineering guidance, not a scientific result and not a product launch, and the course states that “agent” terminology is not universally standardized.

In the course: From Assistant to Agent §4.

Source: Anthropic, “Building effective agents”, published 2024-12-19, accessed 2026-09-15, https://www.anthropic.com/engineering/building-effective-agents.

2025 · Reinforcement learning against checkable rewards

What changed. The DeepSeek-R1 report stated that reasoning ability “can be incentivized through pure reinforcement learning (RL), obviating the need for human-labeled reasoning trajectories”, using rule-based rewards on tasks whose answers can be checked; the resulting model, R1-Zero, produced steadily longer reasoning traces but suffered from poor readability and language mixing, and the released R1 was trained through a multi-stage framework combining supervised fine-tuning, rejection sampling, and reinforcement learning. [S-49]

Why it matters here. It provides a detailed, high-profile public case in which reinforcement learning with rule-based, checkable rewards produced increasingly long reasoning traces, and it is the reason the course keeps R1-Zero (reinforcement learning without supervised fine-tuning as a preliminary step) and R1 (a multi-stage recipe with supervised stages and reinforcement learning) apart. It is one lab’s published recipe; other providers have published less, and the course generalizes neither recipe to reasoning-oriented models as a class. Producing reasoning traces was a trained behavior here, where the 2022 prompting entry elicited them at inference time.

In the course: Post-training §7.

Source: DeepSeek-AI, “DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning”, arXiv:2501.12948 (v1 2025-01-22; v2 2026-01-04), https://arxiv.org/abs/2501.12948.

What this timeline leaves out

Several lines of development that mattered are absent because the course does not build on them: the hardware and software that made large training runs feasible; the construction and filtering of training corpora; convolutional networks and the vision work that shaped deep learning in the early 2010s; earlier word-vector methods; alternatives to dense attention; open-weight model families; evaluation methodology; and most of the commercial products through which people meet these systems. Omission here is a statement about the course’s scope, not about importance. [S-74]

For the meaning of any term used above, see the glossary. For what to read next, see resources.