tszzl: Clinging to Plain-Text CoT Is the Alchemy Age of Observability

On 09-11, researcher tszzl (author of the Loop Transformer architecture) argued that future observers will regard our efforts to preserve plain-text chain-of-thought (CoT) for model observability as an "alchemy age." The better path, he contends, is to understand the model's "alien ontology" from the ground up, achieving observability far superior to reading CoT verbatim. The take sparked discussion among several researchers.

Confirmed

• tszzl posted the view on 09-11; bookwormengr, UltraRareAF, and gandamuml relayed it with consistent framing.
• doodlestein responded by introducing his new project FrankenAlignment: continuing earlier ideas, it uses a small sidecar model to compress a large model's activations, aiming to understand models from the ground up as an alternative to text-CoT observability.
• Ryan Greenblatt of Redwood Research joined the discussion on CoT observability.
• gandamuml likened the debate to the Esperanto movement: "not accurate, but fun."

Why it matters

• The debate touches a core tension in AI safety and alignment: text CoT is the main tool for monitoring model reasoning, yet may lose information about the model's true computation.
• FrankenAlignment represents a concrete technical attempt to build observability on internal representations such as compressed activations, potentially as an alternative to text CoT.

2026-09-11 ~ 2026-09-11 · 6 related posts

• [source] tszzl: plain-text CoT observability will look like an alchemical era — tszzl · 2026-09-11
• Plain-text chain-of-thought will look like an alchemical era of observability — UltraRareAF · 2026-09-11
• Preserving plaintext CoTs will look like an alchemical era of observability, argues loop transformer architect — bookwormengr · 2026-09-11
• [source] FrankenAlignment project replaces text CoTs with compressed activations via sidecar model — doodlestein · 2026-09-11
• [source] Ryan Greenblatt questions whether plain-text CoTs will really become an alchemical era of observability — RyanGreenblatt · 2026-09-11
• Preserving plain-text CoTs will look like 'the alchemical era of observability' — gandamu_ml · 2026-09-11