Neolab · confidential · stealth

Tiling tree — the lab space.

Systematically exploring what the lab could technically do. Method (Boyden / Wang / Khera / Marblestone): recursively split a solution set into MECE subsets by physical / measurable properties. Avoid retroactive splits (organizing around existing fields), catch-all "other" buckets, vague words, and category errors (mixed split dimensions). Redo trees as tech evolves.

Problem tiled: ways to reduce a frontier model's misalignment (make it verifiably safer) that a new lab could own. Evals/measurement are a different tree — mixing them is a category error.

Tree 1 — the naive split (what not to do)

Split by field-names: pretraining · RLHF/post-training · interpretability · evals · oversight · governance. Pitfalls: retroactive (only re-surfaces known ideas), category error (evals = measurement, governance = distribution — different dimensions), catch-all ("oversight/other" dead-ends). Discard.

Tree 2 — physical split: where do the bits change?

MECE + exhaustive: an intervention edits the weights, or edits runtime activations, or leaves the model untouched.

AWeights change — parameters edited. Sub-split by source of the update: A1 behavioural signal (RLHF/DPO/SFT) · A2 internal signal — a probe defines the objective, train the deception direction to zero ← the wedge · A3 structural edit (ablate a direction, prune experts, delete off-task knowledge → "judge AGI").
BActivations change at runtime — params fixed. B1 fixed steering vector · B2 probe-gated (detect-then-block live) · B3 representation surgery.
CNothing inside changes — only the environment. C1 inputs (prompt/constitution/filter) · C2 scaffold & execution (deterministic wrappers, replace-model-with-software) · C3 access & compute (who runs which weights, where — the per-country institution).

Tree 3 — the split that cuts across the fields

A2 and B2 share one dependency: both need a validated internal readout (a probe). So re-tile on the orthogonal physical axes — must you read internals? × do you change the model or the environment? (a Zwicky box):

	Change the model	Change the environment
Black-box behaviour only	RLHF · DPO · SFT A1 crowded — everyone	prompts · filters · scaffolds · access C1–C3 crowded — the startup swarm
White-box read internals	probe → train-away A2 · structural prune A3 ← the empty tile · Neolab's home	probe-gated runtime block B2 · interp audit

Read-off

The empty tile — Neolab's technical home

White-box × change-the-model, at scale, on open weights. Use a validated probe to drive weight edits (A2 train-away + A3 structural prune). Almost nobody does this on the newest open-weight models — it needs a datacenter + the will. Measurement already exists (FAR's probes); the optimization half is open. That gap is the lab.

The weird lever (Grigory)

Press the empty technical tile (A2+A3) and the governance tile (C3, per-country access) together. Each alone has players. Nobody presses both. The combination is the underexplored idea — a lab that produces realigned open weights and coordinates their access via a per-country institution.

Pitfalls avoided: no catch-all bucket; split by physical loci (weights/activations/environment) not field-names; measurement kept in a separate tree; one consistent dimension per split.

Second tree — the institution's legitimacy (the moat)

Tile sources of legitimate coordinating power by the property who is bound, and by what:

forceenforcement / treaty with teeth — needs states, slow, not ours to start.
ownownership / economics — you control the asset (compute, the realignment pipeline). Lucid's verifiable-compute model.
partparticipation / representation — a seat for every country. Cheap, and hugely underrated (Grigory) — nobody actually does it.
techtechnical indispensability — the room matters because the work in it is the best.

Read-off: Neolab's legitimacy = participation × technical-indispensability (cheap + compounding), optionally backed by ownership of the pipeline/compute. Avoid force as the primary. This is the co-authorable moat FAR won't build.

How to use this

Run the empty-tile experiment first: probe → train-away on an open-weight model with FAR's probes. Does the optimization half work + generalize? It gates everything — and the Juniper 5-month fork.
Draft the participation tile: advisory board = best AI person from each of the 20 most populous countries.
Keep evals as a separate measurement tree — don't contaminate the intervention tree.
Re-run the trees when open-weight capability / probe tooling shifts (the method says trees expire).