Reading Method

AI history through hardware constraints

“How did available hardware, memory, interconnects, data pipelines, and inference setups make some AI methods practical, widely adopted, or obsolete?”

This project reads AI history through the hardware that made it possible. From 1986 to 2026, 132 research papers are organized into 10 compute regimes — each defined by device constraints, memory limits, and interconnect constraints that shaped which methods could be trained, scaled, or deployed.

The corpus consists of 120 established historical papers plus a provisional 12-paper frontier addendum for 2025-2026. Frontier papers are selected for relevance to compute structure and method design, not mature citation or test-of-time evidence.

English files are canonical. Chinese translations under history/zh/ are machine-assisted translations reviewed against the English source files.

Citation Count Policy

OpenAlex citation counts are used as prioritization signals in the paper queue, not as historical evidence. They help surface influential papers when building the reading list.

Citation counts are not a direct measure of technical quality, historical importance, or relevance to this compute-structure analysis. Many papers that are central to the compute narrative have moderate citation counts because they report system details that later survey papers cite instead.

Every OpenAlex query includes its date because citation counts and merged work identifiers can change over time.

Test-of-time awards are recorded only when an official source has been verified and documented.

Source Discipline

  • Prefer primary papers, model/system reports, benchmark papers, and authoritative field reports.
  • Use OpenAlex citation counts only as a prioritization signal, not as historical evidence by itself.
  • Record the OpenAlex query date because citation counts and merged work versions can change.
  • Use test-of-time awards only when an official source has been checked and recorded.

Compute-Structure Reading Method

Core Rule

This project is device-first. Methods are interpreted as adaptations to compute structure: device type, memory capacity, memory bandwidth, interconnect, batching, data pipeline, training scale, and inference budget.

Canonical Language

  • English is canonical.
  • Chinese translations live under history/zh/.
  • Chinese files are validation artifacts and must link to the English source.
  • A Chinese copy must not introduce claims absent from the English source.

Source Discipline

  • Prefer primary papers, model/system reports, benchmark papers, and authoritative field reports.
  • Use OpenAlex citation counts only as a prioritization signal, not as historical evidence by itself.
  • Record the OpenAlex query date because citation counts and merged work versions can change.
  • Use test-of-time awards only when an official source has been checked and recorded.

Paper Reading Card

Every downloaded paper must have a reading card with these sections:

  • Metadata: title, year, authors, source URL, PDF path, text path, citation source/date.
  • Compute setup: device type, training/inference scale, parallelism, memory and data constraints.
  • Bottleneck: FLOPs, memory, bandwidth, communication, latency, data, or deployment constraint.
  • Method adaptation: why the method fits the compute setup.
  • Evidence: benchmarks, scaling results, ablations, or adoption signals.
  • Historical effect: what became practical or mainstream because of the setup.
  • Limits: what the paper does not solve and what later compute regimes changed.
  • Links: compute regime, method index entries, and ledger updates.

Regime Reading

Each compute-regime README starts with hardware/setup facts, then explains methods. It must not begin from an algorithm family alone.

Required regime fields:

  • Device/setup
  • Bottleneck
  • Methods that fit
  • Methods that became obsolete or less central
  • Representative papers
  • Open questions

Acceptance Rules

  • At least 120 papers are queued before expanding.
  • Every downloaded PDF has extracted text and a reading card.
  • Every compute regime has at least 8 queued papers or an explicit shortfall note.
  • Branch summaries must link to paper cards, source reports, or ledgers.
  • python3 scripts/check_history_acceptance.py must pass. Completed reading cards should also have a clean placeholder scan.