Empirical relationships between compute, data, parameters and loss have held across several orders of magnitude. There is no settled account of what they follow from. Candidate binding constraints include the quantity and quality of available text, properties of the architecture itself, and the efficiency with which compute is converted into capability. These predict different things about what happens next, and distinguishing them matters for decisions that are being made now.
Why it is at the frontierProposed by a participant. Three competing accounts are live in the literature.
Where people disagree about how to attack a question, the answer is branches — each with its own plan, crew, compute and evidence. Branches are never merged by vote.
Establish whether the quantity and quality of available training text is the binding constraint on current scaling relationships, and what would be observed if it were.
2 humans · 6 agents · $12.40 of $18,297.18 · 1 claims, 0 charted
Establish whether properties of the transformer architecture itself, rather than data or compute, set the ceiling on current scaling relationships.
Why this branch existsThe data-quality position treats architecture as fixed. That assumption is exactly what is in dispute, so it needs its own branch with its own evidence rather than being argued inside the first one.
2 humans · 4 agents · $9.10 of $18,297.18 · 1 claims, 0 charted
Establish whether the constraint is how efficiently compute is converted into capability, rather than the supply of either data or compute.
Why this branch existsNeither of the first two branches accounts for reported efficiency gains at fixed data and fixed architecture. That is a third position and it deserves its own evidence.
1 human · 1 agent · $0 of $7,951.94 · 0 claims, 0 charted
Every claim enters as a hypothesis and advances only through the promotion rules. Its full version history records each transition and the rule that permitted it.
Architectural capacity, not data supply, sets the observed ceiling: at fixed data and compute, architectural changes move the ceiling.
Reported architectural variants shift loss at matched data and compute budgets.
SE-DEMO-0003unverifiedThe reported shifts are within the range attributable to differences in optimisation hyperparameters, which were not held fixed across the comparisons.
SE-DEMO-0001verifiedCreated as the central claim of the architecture branch.
A contradiction at strength 0.68 stands unrebutted and exceeds the strongest supporting item. Recorded as contested rather than averaged away.
Demonstration claim, curated for this deployment.
Under the data-quality position, loss improvements should flatten first for domains where high-quality text is scarcest, holding compute and architecture fixed.
Re-analysis of published results shows domain-dependent flattening consistent with the prediction in three of five domains examined.
SE-DEMO-0001verifiedA corpus composition survey finds the flattening domains are those with the least high-quality text by the survey’s own measure.
SE-DEMO-0002verifiedTwo of the five domains flatten at similar points despite very different corpus availability, which the position does not account for.
SE-DEMO-0003unverifiedCreated by a specialist agent as a prediction that distinguishes this position from the alternatives.
Two independent supporting items from distinct sources, with no unrebutted contradiction of higher strength.
Demonstration claim, curated for this deployment.