Paper Authors Respond to Alignment Controversy
aran_nayebi · x · 2026-07-15
This post responds to criticism of a paper, primarily clarifying the scope of its conclusions:
- The authors state that the exact-case theorems (Thm 1 and Thm 3) are meant to illustrate concepts more clearly through idealized limit cases, not to claim that reality strictly meets these strong assumptions.
- They also provided asymptotic (Thm 2) and soft (Thm 4) versions, which they believe better reflect practical scenarios.
- A key claim is that usedness/minimality itself does not equate to a disguise for "strong alignment"; they require tasks to genuinely activate nonlinearities within the network, referred to here as task-visible nonlinear signatures.
- Furthermore, the authors emphasize that the zippering phenomenon they observed is also covered in the paper, so the results cannot simply be dismissed as "representation metrics don't matter."
Related event: Debate Sparks Over Metric Dependence in AI Research(4 posts)→
More from Research
- Jacob Tsimerman interview frames LLMs as a turning point for mathematical discovery — stevenstrogatz · 2026-07-21
- New survey bridges continual learning and parameter-efficient fine-tuning — v_lomonaco · 2026-07-21
- Daniel Hanchen’s 2-hour workshop covers open models, reward hacking and RL — danielhanchen · 2026-07-21
- Codex’s claimed proof of a math problem turns into a “CEO of math” meme — builderjaydub · 2026-07-21
- Tau Ceti launches as an AI-formalized mathematics library for Lean — wellecks · 2026-07-21
- Krea2 users find a 4-step Raw plus 4-step Turbo workflow that preserves quality — PropagandaOfTheDude · 2026-07-21