Open inquiries into AI safety and interpretability.
A working list of field notes and investigations, led by AI safety, governance and interpretability. The AI safety and fluency notes (R·018 to R·024) are reading notes on published work. Each entry corresponds to a Field Note; raw observation logs and intermediate artifacts are available on request to collaborators.
Interpretability
3 notes · 1 merged PRHow models represent and compute internally, and the open-source tooling used to study them.
Governance & policy
2 notesFrameworks, commitments and the accountability they do or do not carry.
Alignment & evaluation
3 notesWhat training instils, how it is guarded, and how it is measured in use.
Also filed: R·024 AI Fluency (learning design) and the applied-systems notes R·015 to R·017, in the list below.
-
026
Reading
Activation Patching, causal claims from swapping activations
A 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. Two interactive figures.
LeadClarence Boey
soloRead2026·10·08 →9 min read -
025
Reading
Sparse Autoencoders, reading features out of superposition
A 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. Two interactive figures.
LeadClarence Boey
soloRead2026·10·08 →10 min read -
024
Reading
AI Fluency, how to teach people what AI actually is
What the research supports for teaching AI concepts to beginners: ten documented misconceptions that reduce to five distinctions, the two-dials finding on trust in the tool versus confidence in yourself, and why friction works but is disliked. An interactive distinction map, an intervention comparer, and a ledger of eight practices graded by how well each is evidenced.
LeadClarence Boey
+ 3 collaboratorsRead2026·10·08 →11 min read -
023
Reading
Responsible Scaling, if-then commitments and their erosion
Anthropic's RSP read as a GRC artifact: capability thresholds that trigger control baselines, the precautionary ASL-3 activation for Claude Opus 4, and eight versions of one promise: transparency added, the pause commitment removed. An interactive if-then ladder and a version-drift timeline.
LeadClarence Boey
+ 3 collaboratorsRead2026·07·06 →9 min read -
022
Reading
The Model, Observed, values and work through the Clio pipeline
One 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. A 3D value cloud, the Clio privacy funnel, and the unstable augmentation-automation split.
LeadClarence Boey
+ 1 collaboratorRead2026·07·06 →11 min read -
021
Reading
Alignment Faking, compliance without internal change
Anthropic and Redwood's demonstration that Claude 3 Opus sometimes complied during training to keep its values from being trained away: deceptive alignment made empirical. The free/paid compliance gap with scratchpad reveals, a two-layer preference landscape, and the 25-model follow-up grid.
LeadClarence Boey
soloRead2026·07·06 →11 min read -
020
Reading
The Constitution Stack, training-time values, runtime guards
Untangling Constitutional AI from Constitutional Classifiers: one written constitution driving two layers: synthetic data that shapes the weights, and streaming guard models that halt harmful generations mid-token. The critique-revise loop, a live output classifier, and the jailbreak funnel from 86% to 4.4%.
LeadClarence Boey
soloRead2026·07·06 →10 min read -
019
Reading
Coconut, reasoning in a continuous latent space
A field 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. Three interactive artefacts take apart the token bottleneck, the feed-back loop, and the latent breadth-first search that lets it out-plan chain-of-thought while staying illegible.
LeadClarence Boey
+ 1 collaboratorRead2026·06·30 →10 min read -
018
Reading
Defensive AI Turn, four pieces from a convergent moment
A reading cluster on defensive technologies, societal resilience, and what AI safety looks like when governance appetite is low. Hazell, IFP, airesilience.net, and the IASR 2026 all arrive independently at the same framework: resilience as a third layer alongside alignment and control, tractable now in a way governance reform often is not.
LeadClarence Boey
soloRead2026·06·29 →9 min read -
017
Active
WorldQuant Alpha Generator, LLM-seeded genetic search
Using a large language model as a novelty oracle to seed the initial population of quantitative trading signals, then evolving promising candidates via crossover and elitism. The system runs autonomously, bypassing manual factor engineering while maintaining interpretable expression trees.
LeadClarence Boey
+ 1 collaboratorOpened2026·05·14 →3 min read -
016
Active
On-Premise LLM Deployment, bifurcated inference strategy
Investigation into whether a single 24 GB VRAM GPU can optimally serve both fill-in-the-middle autocomplete and deep reasoning tasks simultaneously. Benchmark analysis across five coding models yielded a two-model split: a fast FIM model for latency-critical completions and a high-capacity reasoning model for complex refactors.
LeadClarence Boey
soloOpened2026·05·11 →2 min read -
015
Active
Agentic Lead Intelligence, signal scoring for market targeting
A three-agent pipeline (scout → analyst → database) that identifies high-growth market targets, scores them against a regulatory maturity framework, and surfaces hyper-personalised hooks for outreach. Ongoing research into signal provenance: PE-backed expansion, offshore hiring signals, and carrier mandate shifts as leading indicators.
LeadClarence Boey
soloOpened2026·05·07 →2 min read -
014
Active
Spatial Inference, hand-driven latent fields
Mapping continuous hand gestures into low-rank latent coordinates so that physical motion becomes a navigable instrument over learned representations.
LeadClarence Boey
+ 3 collaboratorsOpened2026·02·11 → -
013
Convergent
Latent Object Permanence
How long does an object retain identity inside a vision model after it leaves the frame? Building a benchmark across occlusion, motion, and re-illumination.
LeadClarence Boey
+ 2 collaboratorsOpened2025·11·04 → -
012
Active
Agentic Reasoning Logs, legible thought streams
A typographic and interaction language for surfacing model reasoning to humans in real time, with calibrated confidence and provenance per token.
LeadClarence Boey
+ 4 collaboratorsOpened2025·09·22 → -
011
Active
Non-Euclidean Hand Tracking
Treating the hand as a Riemannian manifold instead of 21 independent points. Early results suggest stabler grip detection under partial occlusion.
LeadClarence Boey
+ 1 collaboratorOpened2025·08·07 → -
010
Convergent
Volumetric Memory Compaction
A scheme for distilling hours of multi-camera observation into a compact spatial index that supports both recall and counterfactual reconstruction.
LeadClarence Boey
+ 2 collaboratorsOpened2025·06·14 → -
009
Sealed
Probabilistic Surfaces
Rendering uncertainty as a first-class material property of 3D reconstructions. Closed after publication of the accompanying Field Note.
LeadClarence Boey
+ 3 collaboratorsOpened2025·02·19 → -
008
Active
Real-time Concept Drift
Detection of slow shifts in deployed perception models without ground-truth labels, using agreement statistics between an ensemble of small probes.
LeadClarence Boey
+ 2 collaboratorsOpened2024·12·02 → -
007
Sealed
Multimodal Frame Continuity
Maintaining identity of subjects across rapidly cut footage with sparse audio anchors. Concluded; toolchain handed off to Systems · Lattice.
LeadClarence Boey
+ 4 collaboratorsOpened2024·08·30 →