Michael J Clark (wassname) | CV

Michael J Clark

Perth, Australia · name@wassname.org · GitHub · Scholar · ORCID · LinkedIn

Summary: ML engineer (8+ yrs) with peer-reviewed publications. Currently principal data scientist at Woodside Energy, Non-Executive Director at Cytophenix. Research interests: representation engineering, model evaluation, steering methods that work when output-level alignment fails.

AI Safety Research

vGROUT: Steering Vectors for Reward-Hacking Suppression (Jun 2026, partial negative) | LessWrong | GitHub

  • Tested whether label-free hacking vectors can route reward-hacking gradients during GRPO
  • Negative result: the vectors were not precise enough classifiers of hacky vs clean solutions in the realistic environment
  • Useful clue: signed-CorDA initialization partially absorbed hacking gradients into a quarantine adapter in one 4B run; mechanism evidence, not a deployable operating point

Weak-to-Strong Character Steering (2026, in progress, with Lyptus) | Draft | GitHub

  • Weight steering as an interface where a weaker model can modify a larger model’s moral character by interviewing it and creating persona pairs (early results)
  • Early public draft: a 9B teacher steering a 27B student toward “defer less to authority, care more”, with no human labels
  • Weight steering because it beats activation steering by my measures; might scale favourably with model size

Open-source steering tools (2026) | tinymfv · steering-lite · lora-lite · steer-heal-love · cwsteer · persona-steering-template-library

  • Fast moral-preference eval, calibrated activation steering, single-file LoRA on forward hooks, KL-constrained repeated steering, contrastive weight steering, and persona-pair templates

AntiPaSTO: Self-Supervised Steering of Moral Reasoning (Jan 2026) | arXiv:2601.07473 | GitHub | LessWrong

  • Gradient-based representation steering using model’s own behavioral consistency as signal
  • Works without preference labels; outperforms prompting on out-of-distribution moral dilemmas
  • Builds on prior representation alignment work that showed promise but suffered from instability

S-space Steering for Eval-Awareness Control (Mar 2026) | GitHub | Project page

  • Replicated eval-awareness paper with novel singular-value-basis (S-space) steering
  • Hawthorne gap on Qwen3-32B reduced to almost zero (1% vs prior work’s 26%)
  • Judged submission, AI Control Hackathon, Apart Research, Mar 2026

Publications

Ibrahim M, Clark M, Castelnau W. “Improving operational efficiency through condition-based monitoring and IoT technologies.” Australian Energy Producers Journal, 65(2), 2025. DOI: 10.1071/EP24092

Scott NJA, Butler AP, Butler AP, Berg KB, Butler PH, Carr JM, Cook NJ, Clark MJ, Anderson NG. “Pilot study to confirm that ovine fat and liver can be distinguished by spectroscopic tissue response on a MARS-CT.” Endocrine Journal, 57, S421-S422, 2010.

Zeller H, Dufreneix S, Clark M, Butler PH, Butler APH, Cook N, Tlustos L. “Charge sharing between pixels in the spectral Medipix2 x-ray detector.” IEEE IVCNZ, 363-366, 2009.

Berg KB, Carr JM, Clark MJ, Cook NJ, Anderson NG, Scott NJ, Butler AP, Butler PH, Butler AP. “Pilot Study to Confirm that Fat and Liver can be Distinguished by Spectroscopic Tissue Response on a MARS-CT.” AIP Conference Proceedings, 1151(1), 106-109, 2009.

Zainon R, Butler APH, Cook NJ, Butzer JS, Schleich N, De Ruiter N, Tlustos L, Clark MJ, Heinz R, Butler PH. “Construction and Operation of the MARS-CT Scanner.” University of Canterbury, 2009.

Industry Experience

Woodside Energy — ML SME & Technical Lead | 2023-present

  • Principal data scientist for major Australian energy company
  • Published “Improving operational efficiency through condition-based monitoring and IoT technologies.” (see above)

Cytophenix — Non-Executive Director | 2023-present

  • Founding board member of medical AI spinout (Perkins Institute, UWA)
  • AI-powered antimicrobial susceptibility testing; TGA/FDA regulatory pathway
  • Awarded $1M in CUREator grants (Federal Medical Research Future Fund); raised $1.3M pre-seed (Nov 2025)

Three Springs Technology — Director, Partner | 2019-present

  • ML consulting for mining and energy; 15+ projects from research to deployment
  • Developed open-source deep learning curriculum for E&P major
  • Landgate SPUR grant (2016) for satellite-based water leak detection

ThinkCDS — Technical Director | 2016-2019

  • Founded ML consulting firm (merged with Three Springs 2019)
  • Point cloud ML, satellite imagery, reinforcement learning for mining

Schlumberger / OMV — Geophysicist | 2011-2017

  • Seismic data loading, geophysics workflows; transitioned to ML 2016

Education

MSc Petroleum Geoscience — Victoria University of Wellington | 2013-2014
Thesis: The Neogene seismic stratigraphy and uplift history of the Western Chatham Rise

BSc Physics (1st Class Honours) — University of Canterbury | 2006-2009

Skills

Languages: Python, PyTorch, Transformers, einops, Pandas, NumPy

Infrastructure: Docker, Kubernetes, AWS, Git

Methods: Representation engineering, activation steering, model evaluation, interpretability, experiment design, time-series forecasting

Community

Perth Machine Learning Group (3,400+ members) co-organizer since 2018. Active contributor on LessWrong and GitHub.

Selected talks:

  • AntiPaSTO: Self-Supervised Value Steering (Jan 2026) — interpretability research
  • AI Governance: Risk and Regulation (May 2023) — WA Data Science Week panel
  • Experiments with GPT-2 Chatbots (Aug 2019) — early LLM exploration
  • Transformer Network Architecture (Jun 2019) — attention mechanisms, BERT/GPT
  • Deep RL for Bucketwheel Excavator Control (Oct 2018) — industrial RL