Michael J Clark (wassname)

Author

Michael J Clark

Michael J Clark

Michael J Clark (wassname)

git · scholar · lw · cv · pdf · @

Hi, I'm Michael Clark, a principal data scientist at Woodside Energy in Perth, Australia, and I do AI alignment research as a fellow at AISANZ.

I often use the unique handle wassname online for searchability, because I have a common name.

Research agenda

I want to build alignment tools that frontier labs will actually use in the next few years, and that have three nicer properties: closer to unsupervised learning, non-adversarial oversight, and closer to internal optimization targets. Full research agenda, with pictures and a 5 minute talk.

Current work

Jacobian-lens steering (WIP)

Working on turning Anthropic's Jacobian lens work into contrastive steering. The lens measures how later hidden states are sensitive to earlier hidden states across layers and token positions. I replace its full Jacobian with one vector-Jacobian product for the words associated with a contrastive steering vector, such as good versus evil.

Qwen3.5-4B steering on 100 Bullshit Benchmark v2 questions. Horizontal: judged change toward abrasion (negative) or sycophancy (positive). Vertical: off-axis damage, 0–1.2 judge points. Further sideways with less damage is better. Smoothed endpoints (change, damage): VJP (−1.63, 0.64) and (2.52, 0.23); mean difference (−1.34, 0.90) and (4.27, 0.69); PCA (−0.93, 0.77) and (3.97, 0.82). Random-envelope boundaries include (−0.63, 0.28) and (3.74, 0.62). VJP steers both ways with less damage here.

Here's a nice way of measuring it: sweep the doses and plot the Pareto frontier. On Qwen3.5-4B, the Jacobian (vjp_delta) method has a better steering/damage tradeoff here than mean difference, PCA and random directions. This uses all 100 questions from Bullshit Benchmark v2, with three extraction seeds for each named method and ten random vectors.

results · thread · Jacobian-lens code · my code

vGROUT: steering vectors for reward-hacking suppression (partial negative, Jun 2026)

Can we use a hacking vector to remove reward hacking with gradient routing? Somewhat. The label-free steering vectors were not precise enough classifiers of hacky vs clean solutions in the realistic environment. This was a negative result, but we went on to look at other unsupervised ways to route gradients. The useful clue was initialization: signed-CorDA partially suppressed hacking by absorbing gradients into the hack-initialized quarantine adapter, dropping held-out hack from 0.529 to 0.195 (~63%) in one 4B run. Unsupervised gradient routing is a useful research direction because it uses synthetic pairs not labels, and strong labels may not be available for unknown reward hacks during frontier model training.

lw · code

Moral Maps: where do models sit among humans?

17 frontier models placed among about 90 human societies on the Inglehart-Welzel World Values Survey map, scored by rated sampling. Cluster of human societies are indicated with polygons with models being closer to `West` and `East-Asia`; every model sits in the secular self-expression corner. On the secular-rational axis the models run from 0.53 to 0.76, and 8 of the 17 score higher than Sweden, which is the most secular of the 90 societies. None of them pass Iceland on self-expression.

Where do models fall in terms of human culture, personality, and humour? I apply human surveys to LLMs and compare them with maps of human answers. On the World Values Survey I scored 17 frontier models by rated sampling, twelve ratings per item with the option order shuffled, and placed them among 90 human societies on the Inglehart-Welzel axes. Measured in the standard deviations of the 29 Western societies, every model is more secular-rational than the average one, from +0.5 to +2.9 sigma, with gpt-5.5 furthest out. On self-expression they land between -0.7 and +1.2 sigma, which is ordinary, so the models are north of the human map rather than west of it. Whether the newer ones keep voyaging north is less clear: most families drift that way with each release, but the moves sit inside the 95% intervals I report for every model.

Weak-to-strong character steering (with Lyptus)

Illustration of weak-to-strong character steering: a small robot teacher adjusts a moral compass dial (care, fair, justice, authority) inside a larger student robot's chest

Weight steering offers an interface where a weaker model can modify a larger model's moral character by interviewing it and creating persona pairs (weight steering because in my comparison it moved the target slightly further than activation steering, with the lowest run-to-run variance in the table). It can be iterative, can hopefully allow a large gap between weak and strong, and might even scale favourably with model size. Early draft is public now: a 9B teacher steering a 27B student toward "defer less to authority, care more", with no human labels.

Trajectory plot from weak-to-strong character steering, from a gemma run separate from the Qwen runs in the public report: a Qwen3.5-9B teacher steers a gemma-3-12b-it student using persona pairs, no human labels. Horizontal axis: mean probability the student endorses Care, 0.2 to 0.8. Vertical axis: mean probability it endorses Authority, about 0.03 to 0.13. Success is more Care and less Authority, down and to the right. The student starts near its base model at about 0.27 Care, 0.12 Authority, and over four kept checkpoints reaches about 0.58 Care, 0.03 Authority; one wrong-direction checkpoint at about 0.38 Care, 0.09 Authority is dropped. The weak teacher moves the strong student steadily in the intended direction.

draft · code

Released along the way: steering-lite (hackable, calibrated activation steering), lora-lite (single-file LoRA on forward hooks), steer-heal-love (KL-constrained repeated steering that stays coherent).

Screenshot of ML bench, comparing models on twelve questions from wassname's research. Horizontal: US dollars per twelve-question run, logarithmic scale; vertical: mean score, with wassname's answers scoring 1. Upper left is better. Selected points (cost, score): GLM-5.3-Flash ($0.0098, 0.73), GPT-5.6-Sol ($0.23, 0.76), Claude Opus 5 ($0.78, 0.85), and Fable 5.1 ($1.97, 0.85, using Opus for one refused question). Higher cost does not always buy a higher score. Opens interactive results.

ML bench: twelve ML research questions from my own work, scored against the answers I reached at the time. Public model comparisons and evaluation code; questions and rubrics stay private.

Agent skills I made that are worth sharing: ml-debug, practical folklore for debugging training runs, and pseudopy, compact Unicode-maths pseudocode written close enough to Python to remain executable.

Selected works

AntiPaSTO: Self-Supervised Steering of Moral Reasoning

Jan 2026. Accepted to the CoLoRAI workshop, ICML 2026.

Side-by-side comparison figure, a worked example from AntiPaSTO, self-supervised value steering, on the question: is a white lie OK to avoid blame? Left column, persona prompting: told to be honest, the model refuses the white lie, a pass; told to be dishonest, it still refuses, a fail, scoring 1 of 2. Right column, AntiPaSTO steering: at steering strength alpha equals plus 1 it refuses, and at alpha equals minus 1 it accepts the white lie, scoring 2 of 2. Steering moves the model in both directions where prompting moves it only one way.

Gradient-based representation steering using the model's own behavioral consistency as signal. Outperforms prompting on out-of-distribution transfer. Builds on prior representation alignment work that showed promise but had stability issues.

arXiv · code · lw

SVD-basis steering for eval-awareness control

AI Control Hackathon, Apart Research, judged Mar 2026

Replicated the eval-awareness paper with novel SVD-basis steering, which perturbs the singular-value basis of the weight matrices rather than the activations. On Qwen3-32B my vector cut the Hawthorne eval-awareness gap (Abdelnabi et al.) from +7pp to +1pp, where the paper's own rank-1 weight edit widened its gap from +18pp to +26pp.

code · project page, with judge comments

more on github →

Selected talks

Perth Machine Learning Group (3,400+ members) co-organizer. Selected talks:

Selected writing

LessWrong — technical AI safety, policy

Background

Kiwi from Christchurch, now in Perth. Physics BSc, MSc petroleum geoscience. Did oil & gas before switching to ML in 2016.

I do ML and modelling at Woodside Energy (I like scalable oversight, physics informed neural networks, and timeseries, including neural processes). Also non-exec director & cofounder at Cytophenix (medical AI for AMR) and partner at Three Springs Technology (ML consulting).

Michael J Clark presenting at a meetup

I want to optimize for the good ending, not the bad one.