- IV.01Turing: computation and the 1950 intelligence question
What could a participant in Turing’s imitation game judge from the written exchange alone?
Turing's 1950 paper reframed a broad question about thinking through an imagined interaction and a set of objections. Its continuing value is the demand to specify the conditions under which a machine is evaluated. The earlier theory of computation addresses another question: which procedures can be formalized. Keeping these questions distinct makes a historical account more useful for evaluating present systems.
- IV.02McCulloch–Pitts models and early artificial neurons
Which logical operations can a network of threshold units express?
McCulloch and Pitts studied networks of idealized all-or-none units using logical relations. Their model makes a historical bridge between nervous activity and formal computation visible, while its simplifying assumptions remain part of the account. The first useful exercise is to see how a small network implements a logical function and where timing, inhibition, or a different unit rule changes its behavior.
- IV.03Dartmouth and the formation of AI
What did the Dartmouth proposal ask its participants to accomplish over one summer?
The Dartmouth proposal, dated 1955 for the following summer's project, documents an agenda assembled by McCarthy, Minsky, Rochester, and Shannon. It is especially useful as a primary record of what the authors thought could be studied together. The historical question is how those ambitions were translated into particular research programs, institutions, demonstrations, and measures of progress.
- IV.04Symbolic AI and expert systems
When an expert changes a rule, which conclusions of the system change with it?
Symbolic AI represents knowledge and inference through explicit structures such as rules, predicates, and search spaces. The MYCIN authors' documentation makes the approach concrete by connecting a knowledge base to explanation and evaluation. The reusable lesson is that a rule's apparent clarity does not remove the need to check its premises, conflicts, scope, and behavior when information is missing.
- IV.05AI winters and changing research priorities
Which promises, budgets and technical limits shaped the withdrawal of research funding?
The phrase AI winter compresses several distinct episodes of changing expectations, evaluation, and funding. The 1966 ALPAC report supplies one specific primary record concerning machine translation and linguistic research. A useful history compares what systems were expected to do with what evaluators measured and what institutions subsequently supported. Different episodes need their own evidence rather than a single repeated story of enthusiasm and disappointment.
- IV.06Backpropagation and neural learning
How does an error at the output determine the adjustment to a weight several layers earlier?
Backpropagation computes how a network's loss changes with its parameters by applying the chain rule through its operations. Rumelhart, Hinton, and Williams's 1986 paper gives a historical example of learning internal representations with this procedure. The practical investigation separates a correct gradient calculation from the larger questions of optimization, generalization, and what the learned representation captures.
- IV.07Bayesian AI
Which predictions change most when a different prior is fitted to the same evidence?
Bayesian AI makes uncertainty about parameters, models, or predictions explicit through probability distributions. MacKay's work on interpolation connects this framework to regularization and model comparison. The important questions are what is uncertain, what prior assumptions are encoded, and whether the resulting predictions remain calibrated when evaluated on new observations.
- IV.08Kernel methods
When does a similarity function let a classifier recover a boundary missed in the original coordinates?
Kernel methods allow a learning procedure to use similarities corresponding to a richer feature space without always constructing that space explicitly. Cortes and Vapnik's support-vector work gives a concrete historical setting for studying margins, errors, and generalization. The useful comparison holds the data split fixed while varying the representation and the control on model flexibility.
- IV.09Deep learning and the 2012 hinge
How much of the change in image recognition can be attributed to training data, accelerators and architecture?
The 2012 ImageNet paper connected a deep convolutional architecture with a large supervised image benchmark and a practical training system. Its historical significance is clearest when architecture, data, optimization, and computational resources are examined together. This topic asks which changes actually improved a measured task and which lessons survive when the images or deployment conditions change.
- IV.10Transformers and attention
What information can a token receive through an attention head, and at what computational cost?
The Transformer organizes sequence processing around attention and associated feed-forward operations. Vaswani and colleagues' original paper supplies the architecture and evaluation context. A useful visual explanation follows how token representations become queries, keys, and values, then shows the weighted combination produced by attention. The weights describe an internal computation; explaining a final prediction requires additional analysis.
- IV.11Scaling laws, Kaplan, and Chinchilla
At a fixed training budget, how does allocating computation between examples and parameters affect loss?
Scaling-law studies fit relationships between model performance and resources over measured ranges. Kaplan and colleagues and the later Chinchilla study provide distinct experiments and compute-allocation conclusions. Their comparison is useful because the resource budget is not a single number: parameters, training tokens, optimization, and evaluation distributions all matter. An empirical curve should retain the conditions under which it was fitted.
- IV.12Large models and agents
Which actions depend on a model’s output, and which depend on permissions granted by its surrounding software?
An agentic system connects a model's outputs to actions in an environment and to observations of the results. ReAct provides a documented approach to interleaving task-oriented reasoning and external actions. The practical research question is whether this loop completes a task reliably under defined permissions, tool failures, and incomplete information. A fluent plan and a successfully observed result are separate evaluation events.
- IV.13Quantum machine learning
Once data preparation and readout are counted, how does the quantum method compare with a classical baseline?
Quantum machine learning must be evaluated through a complete computational task, including data access and output. The HHL linear-systems algorithm is a useful starting point because its formal speedup depends on specific assumptions. Translating a quantum subroutine into a learning advantage requires checking state preparation, conditioning, measurement costs, and the strongest relevant classical comparison.
- IV.14“Nigel compression” identity resolution and Dynamic Markov Compression
Which early publications match the description of dynamic Markov compression, and what do their algorithms and bibliographies identify?
The label Nigel compression remains an identity-resolution question in this atlas. Cormack's author bibliography provides a documented route to Dynamic Markov Compression, which should be studied under its actual authorship and publication record. The immediate task is to establish whether the informal label refers to that work, a different algorithm, a person's implementation, or a transcription error before joining their histories.
- IV.15Compression-based classification and contemporary reuse
How does the compression-based classifier reported in 2006 perform on a new corpus under the original evaluation conditions?
Compression-based classification uses statistical regularities learned from data to help distinguish categories. Bratko and colleagues studied statistical compression models in spam filtering, where the evaluation includes a changing and adversarial text environment. The useful lesson is to specify exactly how a model becomes a class score, which data trained it, and how errors are priced in the intended use.
- IV.16Efficient learning and inference
What happens to accuracy, memory use and response time when a model is quantized?
Efficient inference and learning involve concrete tradeoffs among memory, latency, energy, accuracy, and engineering complexity. GPTQ offers one studied example of post-training weight quantization. The relevant question is whether a change improves the intended workload on the actual hardware, including loading and transfer costs, while preserving the behavior that matters for that workload.