AI Post Transformers • Visual Companion

Unified Neural Scaling Laws Across Regimes

A visual map of a scaling-law family that tries to predict what happens when parameter count, data, steps, inference compute, and recipe choices all move together. The page focuses on bends, bottlenecks, hyperbreaks, and forecast failure zones instead of retelling the episode.

arXiv:2605.26248 Caballero • Jaini • Krueger • Rish Posted May 25, 2026 Broken multivariate scaling Transcript IDs: scanning…
Overview

One forecast, several competing forces

The paper’s template is less one smooth curve than a negotiation between helpful scaling, bottleneck terms, and harmful opposition terms that appear when a training regime turns bad.

Surface Explorer

Heatmap slice through a regime-changing surface

This synthetic slice holds total compute roughly fixed and shows why one clean law misses the balanced basin, the hyperbreak ridges, and the replay-pressure wall.

Broken Geometry

From smooth laws to hyperbreak-aware slices

Flip axes and assembly stages to watch a classic smooth fit turn into a broken, opposition-aware surface with separate joint and bottleneck structure.

Axis Slice
Assembly Stage
Forecast Bench

What the extra structure is supposed to buy

Illustrative holdout behavior near regime shifts: lower fit error on curated surfaces, plus better behavior when a withheld budget point lands near a suspected break.

Target
View

All plotted values are realistic mock data shaped to the paper’s claims, not reproduced paper tables.

References

Scaling-law lineage