Technical notes
Experiments, results and reproducible work.
-
November 13, 2026 · Benchmark · Model surgery
Rewire: can an AI improve a model by editing its weights and structure?
A benchmark for agents that make models better, more efficient, more competent and more general, by editing their weights and structure, scored against blind edits of the same size with a sealed scorer. Its first task, a repair, passes every harness check: blind editing fixes half a planted defect, and an expert fixes all of it.
-
November 3, 2026 · Results · Grokking
Results: can a model sleep its way to grokking sooner?
FailNo. The positive control fails: even sleep toward the hindsight circuit (median 2,300 epochs) doesn’t beat the best weight decay (800), so the sleep line stops. A label-free score for the circuit’s neurons stays at chance.
-
November 3, 2026 · Results · Self-supervised learning
Results: can you score a learning rule without labels?
PassYes, with RankMe. Across 108 local learning rules, RankMe tracks linear-probe accuracy with Spearman ρ = 0.58 (95% CI 0.44 to 0.70) and passes the gate. LiDAR falls short, and cross-seed agreement is negatively correlated.
-
November 2, 2026 · Results · Learning rules
Results: is there a reason for a learning rule to develop?
PassAll three checks pass on a contextual bandit: a wrong learning timescale costs 15 points of reward, the analysis detects a built-in developmental effect, and evolution recovers 93% of it. The plain bandit gives development nothing to do.
-
November 2, 2026 · Results · Learning rules
Results: where does backprop clearly win?
No headroomNowhere at laptop scale. Feedback alignment falls up to 23 points behind backprop, but sign-symmetry stays within 2.5 points on three of four networks, so no network clears the 10-point bar.
-
November 2, 2026 · Pre-registration · Overview
Four go/no-go tests, written down before the run
Four cheap overnight tests decide which research directions earn more time. This overview links the four pre-registrations and the protocol they share.
-
November 2, 2026 · Pre-registration · Grokking
Can a model sleep its way to grokking sooner?
Periodic self-distillation toward a model’s own generalizing circuit, tested against tuned weight decay, shams and generic controls across 15 arms and 20 seeds.
-
November 2, 2026 · Pre-registration · Learning rules
Is there a reason for a learning rule to develop?
Three checks before evolving developmental learning rules: whether the task rewards development, whether we could detect it, and whether an optimizer could find it.
-
November 2, 2026 · Pre-registration · Self-supervised learning
Can you score a learning rule without labels?
Do LiDAR, RankMe or cross-seed agreement rank 108 local learning rules the way a linear probe does?
-
November 2, 2026 · Pre-registration · Learning rules
Where does backprop clearly win?
The cheapest CIFAR-10 network where feedback alignment and sign-symmetry fall at least 10 points behind backprop, or none.
-
November 1, 2026 · Updated November 13 · Architectures · Efficiency
Can a language model choose its own layers, token by token?
Per-token paths that loop and skip layers look powerful in hindsight, but random nudges of the same size do as well, and no router fitted to a frozen model finds them. A gate trained with the model helps on benchmarks; our finer per-layer router lost to it, and a diagnostic shows why.
No notes for this project yet.