Control mechanism ↑ · ai artefact
Interpretability first
Mechanistic understanding is a precondition for reliable oversight; behavioural evaluation without interpretability cannot rule out deceptive alignment.
Mechanism
Scale mechanistic interpretability on frontier models so training, evaluation, and governance decisions read internal structure, not just behaviour.
If it succeeds: what binds next
Models are legible. Legibility itself is now an asset and a target, interpretability tools become security-sensitive infrastructure.
A strategy that produces a worse next problem than the one it solved has not done durable work.
Falsification signal
Leading labs cannot produce mechanistic explanations of their own frontier models within two to three years of release.
A strategy held without a falsification signal is not strategy; it is affiliation. Continued support after this signal lands is identity, not bet. See the identity diagnostic.
Self-undermining threshold
overshoot riskWhen safety talent concentrates away from governance and resilience.
If interpretability then fails to scale, the stripped adjacent families had no time to mature as alternatives.
Every strategy has a stable region where it reinforces itself and an unstable region where pursuit defeats it. The threshold between them is usually narrower than advocates acknowledge.
Addresses 1 failure scenario
all scenarios →People on the record
15Profiled figures appear first, with their tier in small caps. Each face links to the person and their full quote record. Tag: interpretability-bet.
expertise mix · 4 profiled
recognition mix
A strategy whose endorsement skews to commentators or external-domain experts is in a different epistemic state from one endorsed mostly by frontier-builders. The mix is read carefully across both axes; see the board for criteria. Counts are over the 4 profiled people on this strategy (11 unprofiled excluded).
Chris Olah
Builds frontier systems · Known across the AI/safety field
Connor Leahy
Deep ML / safety technical · Known across the AI/safety field
Cynthia Rudin
Deep ML / safety technical · Known across the AI/safety field
Neel Nanda
Deep ML / safety technical · Recognised inside subfield
Asma Ghandeharioun
Google DeepMind; 'Patchscopes' for LLM interpretability
David Bau
Northeastern; mechanistic interpretability of LLMs
Fernanda Viégas
Harvard; ex-Google PAIR; data visualization
Jacob Andreas
MIT NLP; language models as belief reports
John Wentworth
Independent alignment researcher; natural abstractions
Lucius Bushnaq
Apollo Research; mech interp
Martin Wattenberg
Harvard; ex-Google PAIR; visualization for ML
Rich Caruana
Microsoft Research; interpretable ML
Roger Grosse
U Toronto; Anthropic; influence functions for LLMs
Trenton Bricken
Anthropic mechanistic interpretability
Tristan Hume
Anthropic mechanistic interpretability
Coordinates
Conflicts, grouped by mechanism
0No strict conflicts catalogued. This strategy pulls a lever that nothing else pulls in the opposite direction.
Complements, grouped by mechanism
5Same-lever reinforce
same lever, same pull, different mechanismBoth strategies pull the same lever in the same direction by different means. They stack: doing both amplifies the pull, at the cost of double-counting in portfolio audits.
Adjacent bet
different levers, loosely coupledDifferent levers, different directions of action. They reinforce only via the general principle that covering more bets dominates covering fewer.
Cross-side bridge
one AI-side, one world-sideOne acts on the model, the other on institutions or culture. The bridge hedges against both artefact-level and substrate-level failure.
Same-lever twins
2Both use the same lever in the same direction. Usually redundant inside a portfolio: each dollar or effort unit only buys one lever pull, even if two strategies are named.
Axis position
Source note: Interpretability first strategy.md