Definition. An "ALM" is a trained system that generalizes beyond NLA-style reconstruction introduced in this Anthropic's article.
ALMs satisfy ALL of:
(a) Multi-task read side (AV): takes activations from a target LLM and handles at least three distinct natural-language task families - e.g., open-ended explanation, targeted question-answering, inferring context/user properties, predicting the outcome of an intervention without running it - where the system was deliberately trained or fine-tuned for this task.
(b) General write side (AR): maps free-form natural-language descriptions to activation-space objects (steering vectors, probes, or patch targets) as an open-ended interface. Deriving steering vectors by editing an NLA explanation and diffing AR reconstructions does NOT count.
(c) Public introduction: described in a paper, preprint, technical report, or official blog post, with either released artifacts (weights/API) or a reproducible training recipe, by any organization.
Explicitly NOT sufficient: NLAs and successors optimized purely for reconstruction; QA activation oracles alone (read-only, single task family); HyperSteer-style text-to-steering models alone (write-only); an ensemble of separate single-purpose tools not presented as a unified system.
Resolution: YES if such a system is introduced on or before Dec 31, 2027; judged by the creator in January 2028, with ambiguous cases argued in comments before resolving. NO otherwise.
* The name doesn't matter. "ALM," or any successor term, resolves YES if (a) - (c) hold.
** I won't bet on this market.
1,000