LabOS

LabOS is a small research lab where the researchers are AI agents. I supply the ideas and make the calls; agents do the reading, coding, and experiments. Negative results included.

For empirical pieces, the date is the last day of substantive research, not a later copy edit.

Process note · July 11, 2026

How the lab works: AI agents as the researchers

One person with a lot of ideas, AI agents as the research staff. How the lab runs many threads in parallel, keeps its memory in files, and kills most ideas cheaply.

Read the article →

Findings

Controlled study · July 20, 2026

The query decides whether a repaired KV cache is safe

Splice a re-encoded edit into a stale KV cache and only the states after the edit stay wrong. That drift never flipped an unrelated query — 0/252 in each of two studies — but flipped a third of the queries that read the edit. A probe can rank the failures; compute and calibration kept it from becoming a method.

  • cache reuse
  • selective prediction
  • preregistration

Instrument study · July 15, 2026

Prediction depth is a property of the ruler, not just the model

A translator-free lens for the layer where a transformer settles each token: keep the MLPs, silence the attention. It passed a three-family validation and out-tracked the tuned lens on GPT-2 — and then showed that swapping zeros for means flips verdicts, findings, and the sign of one result.

  • interpretability
  • prediction depth
  • controls

Audit · July 11, 2026

Does model editing store facts, or just answers?

A ladder of progressively stronger fact-free controls for knowledge editing. ROME nearly dissolves, SAKE dissolves at the first rung, REMEDI keeps a real residual.

  • model editing
  • controls

Negative result · July 10, 2026

Does a transformer snap between modes of computation?

The hope: discrete computational regimes you could read off a layer’s Jacobian. I preregistered the test and ran it — across 270 configurations, zero beat a plain continuous predictor.

  • interpretability
  • Jacobian lens

Characterization · July 10, 2026

Every importance gate is a Bayesian prior in disguise

I tried to beat uniform training by gating learning strength on input importance. What came back was an identity, not a method — and it lets you call the winner before running anything.

  • continual learning
  • Bayesian inference

Audit · July 10, 2026

Auditing activation steering: how much is just pushing on the output?

For ActAdd, a static logit bias reproduces 96% of the steering effect — even though the steering vector is nearly orthogonal to the unembedding. Six methods audited.

  • interpretability
  • activation steering
  • controls

Methods & process

Roundup · July 12, 2026

Killing bad ideas early

Eleven dead projects for under $5 of cloud compute: the cheaper the stage an idea dies at — prior art, arithmetic, a preregistered gate, a control, real data — the more ideas you get to try.

  • negative results
  • kill gates
  • research process

In progress

Live in the lab, not written up here yet — either too early, or the interesting parts stay private for now.

In progress · details withheld

Activation reachability

Which unsafe activation states can prompts alone reach, and which require white-box access — and what that boundary means for runtime monitoring. Details withheld for now.

In progress

Foundations of neural nets

Can a single overcomplete layer learn its own architectural constraints from data and description-length pressure alone? Early signal, hard controls pending.