Research
Notes and investigations in AI safety, governance and interpretability. Most are reading notes on published work; some are my own experiments and open-source contributions. Each links to its full version, with interactive figures.
In progress
Work without a published note yet.
- OngoingSparse autoencoder accuracy and the Johnson-Lindenstrauss lemma
Benchmarking how faithfully SAEs recover features, using random-projection distance bounds as the yardstick. Builds on contributions to SAEBench and SAELens.
- OngoingAI diffusion in Singapore, and its effects on what people believe
With a classmate: the downstream economic effects of AI diffusion in Singapore, with the scope widening to epistemic effects across ASEAN.
Interpretability
How models represent and compute internally, and the open-source tooling used to study them.
4 notes
- 026Activation patching, causal claims from swapping activationsA reading note on turning "this component seems important" into a causal claim: clean and corrupted runs, why denoising and noising answer different questions (OR versus AND circuits), how causal tracing localises facts without telling you where to edit, and the ablation and metric choices that change the answer.
- 025Sparse autoencoders, reading features out of superpositionA reading note on how SAEs decompose a model's activations into a dictionary of features: the superposition geometry that makes them necessary, the four ways the dictionary goes wrong (absorption, hedging, splitting, shrinkage), and the sanity checks SAE benchmarks have since had to survive.
- PRA JumpReLU loader fix for SAELens, so SAEs run as the SAEs they areA merged contribution to SAELens: loading JumpReLU sparse autoencoders trained with dictionary_learning either failed outright or, before v6.23, silently ran them as plain ReLU SAEs. The fix, the evidence, and the tests that pin it.
- 019Coconut, reasoning in a continuous latent spaceA note on Hao et al.'s Coconut: feeding a language model's last hidden state straight back as its next input embedding, so reasoning happens in continuous vectors instead of tokens.
Governance and policy
Frameworks, commitments and the accountability they do or do not carry.
3 notes
- 028AI incident reporting, learning from the near missA reading note on nine days of Singapore policy, from the primary sources: the Prime Minister's proposed international pilot for reporting AI incidents, a study of how domestic reporting should cover near misses that cause no harm, and the rule for when government AI agents need a human to approve.
- 023Responsible scaling, if-then commitments and their erosionAnthropic's Responsible Scaling Policy read as a governance, risk and compliance artifact: capability thresholds that trigger control baselines, the precautionary ASL-3 activation for Claude Opus 4, and eight versions of one promise, with transparency added and the pause commitment removed.
- 018The defensive AI turn, four pieces from a convergent momentA reading cluster on defensive technologies, societal resilience, and what AI safety looks like when governance appetite is low.
Alignment and evaluation
What training instils, how it is guarded, and how it is measured in use.
4 notes
- 027Net influence, finding colluding agents from their moves aloneMy own experiments from the Apart Research collusion sprint: a published collusion detector run on 295 games between Llama and Mistral agents in Rock-Paper-Scissors and a double auction.
- 022The model, observed, values and work through the Clio pipelineOne privacy-preserving instrument pointed twice. Values in the Wild maps 3,307 values expressed across 700,000 real conversations; the Economic Index maps usage onto occupational tasks.
- 021Alignment faking, compliance without internal changeAnthropic and Redwood's demonstration that Claude 3 Opus sometimes complied during training to keep its values from being trained away: deceptive alignment made empirical.
- 020The constitution stack, training-time values, runtime guardsUntangling Constitutional AI from Constitutional Classifiers: one written constitution driving two layers. Synthetic data shapes the weights, and streaming guard models halt harmful generations mid-token.
AI fluency
How people learn what AI actually is, and how to teach it well.
1 note