SAE-based mechanistic interpretability will be superseded by manifold-based analysis for understanding semantic concepts within 24 months.

{
  "doc": "/Users/antonborzov/Documents/Research.nosync/notes/2026-05-14_phil-trans-A-goodfire-aboutblank-impact.md",
  "claim": "SAE-based mechanistic interpretability will be superseded by manifold-based analysis for understanding semantic concepts within 24 months.",
  "horizon": "24 months",
  "section": "A-Tier — strongly load-bearing on specific vectors",
  "category": "technical",
  "confidence": null,
  "depends_on": [
    "Manifold steering (Wurgaft et al) adoption in mechanistic interpretability"
  ],
  "extracted_at": 1778899659,
  "extracted_by": "haiku-4-5",
  "status_check_log": [],
  "evidence_threshold": "PROPOSED: Geiger's critique that 'SAE features shatter manifolds into unrelated pieces' becomes mainstream consensus; major interpretability teams (Anthropic, DeepMind, Redwood) publish work prioritizing manifold-based over SAE-based feature analysis."
}