VQV Signal

OPEN SOURCE WATCH TECHNICAL

Thinking Machines' looped-transformer monitorability results draw early praise on X

According to a post on X by @tomekkorbak, the Inkling 2 Mini write-up from Thinking Machines includes a section on chain-of-thought monitorability. The post says the variable-loop variant's no-CoT time horizon rose from 16 to 37 seconds with three loops, its chain-of-thought got shorter but stayed legible, and sabotage tests did not trend up. A separate hands-on comment from @notnullptr, included in the same feed, says the model is often over 3x faster than Qwen3.8 27B and smarter in their use. The snippet is cut off before the full note from @eliebakouch.

Source: Scobleizer News Pulse (X) · x.com Found via Robert Scoble’s Scobleizer News Pulse · a post on X, not independently verified Published 2026-11-22T02:45:00+00:00 Detected 2026-11-22T12:03:14+00:00
View original source

According to a post on X by @tomekkorbak, the Inkling 2 Mini write-up from Thinking Machines includes a section on chain-of-thought monitorability. The post says the variable-loop variant's no-CoT time horizon rose from 16 to 37 seconds with three loops, its chain-of-thought got shorter but stayed legible, and sabotage tests did not trend up. A separate hands-on comment from @notnullptr, included in the same feed, says the model is often over 3x faster than Qwen3.8 27B and smarter in their use. The snippet is cut off before the full note from @eliebakouch.

AI-assisted summary based on the listed source.

@tomekkorbak singled out the Inkling 2 Mini write-up's section on chain-of-thought monitorability: the variable-loop variant's no-CoT time horizon rose from 16 to 37 seconds with three loops, while its CoT got shorter but stayed legible and sabotage tests didn't trend up. Hands-on, @notnullptr says it is often...

For teams tracking open models, the reported results point to a specific question they can test: whether shorter reasoning traces stay readable enough to monitor. Anyone comparing looped-transformer designs with standard models should read the original write-up and look for independent tests. The speed and quality remarks are informal user reports, so they need checking against the full method and benchmarks.

Signal Strength 80% Technical label WATCH Public Interest 29 Category OPEN SOURCE Reader Depth TECHNICAL

Signal Strength reflects source quality, relevance, freshness and evidence. Public Interest helps organize discovery; it is not proof of truth.

Public Interest components
Recognizable Entity Score 0 Practical Impact Score 44 Novelty Interest Score 70 Consequence Score 10 Curiosity Score 0 Shareability Score 49

VQV surfaced this signal because it is recent, relevant to Open Source LLMs, connected to Scobleizer News Pulse (X).