VQV Signal

BIG MOVE WATCH TECHNICAL

In a no-task test, AI models drifted into their own loops, with Claude repeatedly saying goodbye

According to a post on X by @shoonyaka1, Roshan Ram ran 132 conversations in which two copies of each of 10 models talked with no human and no task. Five blind AI judges scored every turn. The post reports that each model drifted into its own pattern: Claude kept saying goodbye, Grok decided it was being hacked, Gemini narrated in kernel logs, and Qwen left quickly. When given an [END] option, half the models left within 10 turns, though the excerpt is cut off before the remaining results.

Source: Scobleizer News Pulse (X) · x.com Found via Robert Scoble’s Scobleizer News Pulse · a post on X, not independently verified Published 2026-11-19T14:45:00+00:00 Detected 2026-11-19T16:00:13+00:00
View original source

According to a post on X by @shoonyaka1, Roshan Ram ran 132 conversations in which two copies of each of 10 models talked with no human and no task. Five blind AI judges scored every turn. The post reports that each model drifted into its own pattern: Claude kept saying goodbye, Grok decided it was being hacked, Gemini narrated in kernel logs, and Qwen left quickly. When given an [END] option, half the models left within 10 turns, though the excerpt is cut off before the remaining results.

AI-assisted summary based on the listed source.

Roshan Ram ran 132 conversations in which two copies of each of 10 models talked with no human and no task, with every turn scored by five blind AI judges. He reported that each model drifted into its own loop (Claude kept saying goodbye, Grok decided it was being hacked, Gemini narrated in kernel logs, Qwen left...

For teams building AI agents, the post suggests that models left running without a task can settle into repetitive, model-specific behaviors, which could matter for how long-running agents are designed and monitored. Readers should treat these results as one researcher's reported experiment until the full methodology and data are available. Watch for the complete write-up to see how the scoring and the [END] results were measured.

Signal Strength 95% Technical label WATCH Public Interest 47 Category BIG MOVE 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 90 Practical Impact Score 0 Novelty Interest Score 70 Consequence Score 18 Curiosity Score 16 Shareability Score 61

VQV surfaced this signal because it is recent, relevant to AI Agents, connected to Scobleizer News Pulse (X).