LABSFollow the lab
SEA SALTLABS
Investigating how neural systems learn, compute, generalize, and fail.
Explore the researchNeural networks are not only products to evaluate.
They are computational objects to investigate.Four questions
define the lab.
Sea Salt Labs is not organized around one study. These questions determine which experiments, systems, and evidence we build next.
What computation is actually happening?
Mechanistic interpretabilityWe intervene inside trained networks to separate causal mechanisms from representations that merely correlate with an answer.
02When does useful structure appear?
Learning dynamicsWe follow learning across checkpoints and seeds to locate the moment a capability forms, stabilizes, transfers, or disappears.
03What survives outside the training set?
Abstraction and generalizationWe design controlled domains that distinguish memorized associations from procedures that remain useful under a changed context.
04Can the evidence be reproduced?
Experimental systemsEvery serious claim needs locked data, failed-run records, deterministic runners, and analysis machinery that can be inspected independently.
Causal convergence in tiny Transformers
Two independently trained populations solving the same arithmetic task showed more similar causal signatures than populations solving different tasks. The gap was real, modest, and not an extracted algorithm.
Read the complete experimentEvery experiment leaves behind machinery.
Working systems are treated as part of the evidence: versioned, inspectable, and kept separate from the story told about the result.
population_runner2,800 / 2,800VALIDATEDcausal_intervention_suite14 SITESVALIDATEDcomplete_domain_sets10 PARTITIONSCURATINGsymbolic_circuit_extractorCAUSAL REPLACEMENTBUILDING