Enric Boix Adsera

Value of the Information & Strength of the Argument

The talk presents valuable insights into AI interpretability and alignment, with concrete examples and empirical results. The speaker argues for the multi-index representation hypothesis over the linear one, supported by experiments showing improved probe performance. The dynamic adaptation scheme for defenders is innovative and addresses a real problem, with a simple mathematical model and promising empirical results. However, the arguments are based on limited experiments and some informal setups, and the theoretical justification for the adaptation scheme is simplified.

Scientific Rigor, Source Quality, Title Accuracy

The speaker references several papers and sources, including the Alibaba incident, the linear representation hypothesis (Mikolov et al.), and representation engineering. He mentions his own papers and ongoing work. The sources are relevant and credible, but not all are formally cited with URLs. The title is just the speaker’s name, which is typical for seminar recordings and does not mislead. The content is rigorous for a research talk, but some claims are based on preliminary results.

The title is just the speaker's name, so it does not reflect the content, but this is typical for seminar recordings.

7/10

Talk by a researcher presenting ongoing work, with references to papers and empirical results, but limited peer-reviewed validation and some informal experiments.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 00:00Introduction and motivation: Alibaba agent hacking incident.
• 02:00Linear representation hypothesis and concept vectors.
• 05:00Multi-index representation hypothesis and recursive feature machines.
• 08:00Probing results: hallucination and toxicity detection.
• 10:00Steering models by adding concept vectors.
• 12:00Application to honesty in role-playing scenarios.
• 15:00Introduction to multi-agent adversarial games.
• 18:00Cybersecurity example: planting bugs and defender failure.
• 21:00Dynamic adaptation scheme: best-of-n self-distillation.
• 24:00Mathematical model and empirical results.

• Alibaba incident report — Mentioned as motivation for alignment issues.
• Mikolov et al. word embeddings — Cited as origin of linear representation hypothesis.
• Representation engineering paper — Referenced for steering method.

• Anthropic research on monitoring — Mentioned as evidence that agents can evade monitors.
• OpenAI research on monitoring — Mentioned as evidence that agents can evade monitors.