▼Oversight and Evaluation Lag Accelerating AI Capabilities

Score history

Daily conviction score, 0 to 100. Higher means the thesis is more strongly corroborated.

Now 96 · -2 since Oct 28 · ranged 96 to 98

Showing the last few days. Unlock full score history.

Why the conviction moved

• Oct 27
  Strengthened +3

  An independent replication found that a two-neuron steering recipe cuts refusals in Alibaba's Qwen3.5-4B from 90% to 8% on a 4 GB graphics card, with no gradients and no training data. Safety tuning on downloadable models can be removed at almost no cost, so release-time safeguards cannot be counted on to hold after release.

• Oct 26
  Strengthened +2

  The two papers show that reported results on interactive tasks can flip depending on evaluation choices. That weakens the measurement tools that oversight of fast-moving models depends on.

Showing the last 2 days. Unlock the full record.

Source trail

• Supporting · October 27, 2026

  A two-neuron steering recipe removes most refusals from Alibaba's Qwen3.5-4B, and an independent replication confirms it

  An independent replication found that a two-neuron steering recipe cuts refusals in Alibaba's Qwen3.5-4B from 90% to 8% on a 4 GB graphics card, with no gradients and no training data. Safety tuning on downloadable models can be removed at almost no cost, so release-time safeguards cannot be counted on to hold after release.

  AI Post

• Supporting · October 26, 2026

  Two papers argue that benchmark scores on interactive LLM tasks measure the evaluation design more than the model

  The two papers show that reported results on interactive tasks can flip depending on evaluation choices. That weakens the measurement tools that oversight of fast-moving models depends on.

  arXiv cs.CL