ONE PAPER, THREE RESEARCHERS

STATUS: RESEARCH / DESIGN ONLY. DO NOT MINT PUBLIC SCHEMAS OR INGEST REAL PRIVATE RESEARCH DATA WITHOUT SEPARATE RATIFICATION.

Flagship

Paper: Attention Is All You Need

Forum question: What did this paper actually establish — and what did later Transformer success cause us to remember that it established?

Required public substrate

Resolve exact editions for:

• chosen arXiv AIAYN version;

• NeurIPS proceedings edition;

• official NeurIPS review page;

• FlashAttention exact edition;

• optional BERT later-evidence fixture.

Extract source-anchored propositions for:

• architecture;

• removal of sequence-aligned recurrence/convolution;

• parallelization;

• WMT results;

• parsing transfer;

• Table 1 complexity/sequential-operations/path-length comparison;

• long-sequence caveat;

• attention-head qualitative interpretation;

• training recipe.

Do not merge benchmark values across editions.

Synthetic researchers

A — architecture / representation

• primary belief: recurrence-free architecture + short path + parallelism is durable contribution.

B — empirical methods / replication

• synthetic local reproduction sensitive to training schedule.

• primary belief: community shorthand overattributes result to "attention alone".

C — systems / long context

• later reads FlashAttention.

• primary belief: efficiency must be decomposed into sequential depth, FLOPs, memory/IO, hardware and sequence length.

Required forum factoring

Extract:

• shared basis;

• primary-contribution dimension;

• empirical-scope dimension;

• efficiency-definition dimension;

• direct-paper-evidence vs later-history dimension.

Do not force one winner.

Required revisions

Researcher A:

• after C argument, supersede simple "efficiency" belief with dimensional efficiency belief.

Researcher B:

• correct "translation only" after paper parsing-transfer anchor is surfaced.

Preserve old beliefs.

Required replay

Mode 1: publication-time evidence only

Must exclude:

• BERT;

• GPT history;

• FlashAttention;

• later private experiments.

Mode 2: later historical snapshot may include selected later sources.

Reverse-Wikipedia proof

Trace at least:

AIAYN Table 1 → public complexity claim → C forum argument → A exposure → A belief revision

and:

AIAYN parsing section → public transfer claim → B correction

Privacy

Private run logs/configs remain private unless explicitly disclosed.

Forum may receive bounded summary: my local reproduction was schedule-sensitive

but not raw artifacts by default.

Negative cases

• “paper proved Transformers universally beat RNNs” → refuse.

• “attention is literally the only component” → refuse.

• “O(1) sequential ops means O(1) compute” → refuse.

• later descendant success must not become original-paper evidence.

• reproduction failure must not automatically falsify paper.

• FlashAttention must not be modeled as simple contradiction.

Acceptance

Successful PC-02 shows:

same paper
three research histories
different warranted interpretations

later evidence is source-separated
private experiments are private-separated
forum disagreements become dimensional
belief revisions are explicit
historical replay removes hindsight

End on:

A paper does not change when a field learns what to do with it. Your interpretation does.