On: Claude discovers a novel enzyme system with CRISPR-like repeats

[oofbey]

In the research literature this is generally called coconut after the first paper published on the idea. It has advantages and disadvantages. I agree it’s more efficient. But the lack of observability and transferability are real downsides that block its adoption.

[iririririr → oofbey]

the point is it may be a feature. most providers don't want clients reading those and almost nobody provide thinking text anymore.

[nomel → iririririr]

> almost nobody provide thinking text anymore.

Isn't the goal to be able to "debug" and identify alignment issues?

[cubefox → nomel]

Exactly. This is how the Huggingface incident was reconstructed. But GPT-6 uses partially Neuralese, and its monitorability has dropped sharply according to benchmarks.

[goolz → iririririr]

Another reason to switch off a central provider to open models as quickly as possible.