AI Safety

Where I am

Cognitive neuroscience and AI interpretability ask the same question: what is happening inside the black box? Neuroscience has spent decades building methods to answer it, and I bring those methods to AI models. I pursue AI safety through independent research alongside my consultancy, and my aspiration is to become a full-time AI safety researcher.

The shared problem is understanding a system's internal representations as well as its behavioural output. Representational similarity analysis, probe-based decoding and dissociation logic transfer directly to interpretability. Representation engineering (Zou et al., 2023) anchors my framing.

What I am working on

• 01

  A geometrical exploration of alignment faking

  A BlueDot sprint project on GPT-OSS-20B over Anthropic's alignment-faking transcripts. A layer-11 probe reached 0.807 cross-condition AUROC against 0.806 for a TF-IDF unigram and bigram baseline: no clear advantage. My original RSA design could not separate the label from transcript length, so I withdrew it. I am redoing the study with grouped cross-validation, confidence intervals, stronger baselines and a redesigned RSA.

• 02

  A monitoring device for agent swarms

  I am building a monitor that watches a swarm of coding agents and flags coordinated activity for human review. The work is solo and runs on local models only, in two studies. It is my entry to the Apart Research Swarm Alignment Hackathon (6 to 8 November 2026, multi-agent oversight and control).

  Study A, first-line monitoring. Can small local models serve as the first line of monitoring for code written by a real agent swarm? A cascade of a rules baseline, d3 and Swift-1.5 labels each record, and an escalation policy sends uncertain records up the cascade. A stronger external model supplies the reference labels, so results are agreement with that reference.

  Study B, collective monitoring. Records link into a graph, Leiden community detection groups them, and community scores decide what goes to review. Study A's record-by-record cascade is the baseline, compared at fixed review budgets of 1 to 10% of records. The quiet swarm test strips the swarm's obvious markers and reruns both approaches, with a placebo arm to separate marker removal from text damage. Claims cover records and communities of linked records.

• 03

  The Geometry of Sycophancy

  Llama 3.2-3B, base and instruct, studied with RSA, linear probing and temporal generalisation. At late layers, social framing dominates truth in residual-stream geometry (framing ρ ≈ 0.82 against truth ρ ≈ 0.02 at layer 27), while truth stays linearly decodable (AUROC ≈ 0.90). The dissociation between probe and RSA is the point of the design.

Methods I trust

RSA, cross-decoding and activation patching have non-overlapping failure modes, so the combination beats any single probing approach. Method-construct fit matters. I lean towards interpretability and AI control.