SEA SALT
LABSFollow the lab
RESEARCH PROGRAM / 2026

RESEARCHATLAS

Sea Salt Labs investigates how neural networks form, preserve, and expose learned computation.

Current work is organized into four distinct research programs.
04 PROGRAMS01 CONFIRMATORY STUDY2,800 TRANSFORMERS
PROGRAM INDEX / CURRENT

Current programs

Each program defines a separate research objective, the work underway, and the questions that remain unresolved.

01ACTIVE / PRIMARY

Mechanistic interpretability

Recover the computations that actually cause a trained network's behavior, then test whether those computations survive intervention and substitution.

NOW TESTING

Separating shared causal organization from implementation-specific coordinates across independently trained Transformers.

OPEN QUESTIONS
  1. Can a causal circuit be replaced and still produce the same behavior?
  2. Do disjointly trained models converge to equivalent computations?
  3. Can an extractor recover structure rather than correlation?
02ACTIVE / EXPLORATORY

Learning dynamics

Track when useful information forms during training and whether later learning preserves, reorganizes, or destroys earlier capabilities.

NOW TESTING

Measuring information emergence, residual-stream contribution, compensation, and retention across checkpoints and layers.

OPEN QUESTIONS
  1. At what layer does new task information first become recoverable?
  2. Do later blocks compensate when an earlier block is removed?
  3. Does a new abstraction accumulate or overwrite prior computation?
03EXPLORATORY

Abstraction and generalization

Distinguish memorized input-output associations from procedures that remain useful under new names, inputs, and runtime definitions.

NOW TESTING

Formal-language curricula that compose primitives into unseen operations without relying on English instruction or familiar task labels.

OPEN QUESTIONS
  1. Can a model execute a new operation defined only at runtime?
  2. Does learning a primitive make a composed procedure easier to acquire?
  3. What survives held-out operator names and input pairs?
04BUILDING / VALIDATING

Experimental systems

Build the datasets, runners, controls, and audit trails required to distinguish a scientific result from a measurement artifact.

NOW TESTING

Locked complete-domain datasets, resumable population training, failed-run ledgers, intervention suites, and reproducible releases.

OPEN QUESTIONS
  1. Which controls expose extractor and architecture bias?
  2. Can dataset overlap be removed without destroying task identifiability?
  3. Can every result be regenerated from one frozen manifest?
RESEARCH DISPATCH

Follow the work while the answer is still uncertain.

research@seasaltlabs.com