I'm just a guy who likes to machine learn. I want the good ending not the bad one.

I work on AI alignment: steering, evals, and practical interpretability.

Links: wassname.org · Scholar · Hugging Face · LessWrong · Gists

Scalable, self-supervised alignment interventions. Ideally internal interventions, and driven by gradient. I'm always keen to discuss and brainstorm along these lines.

• Weak 2 strong character steering (WIP, with Lyptus) Can weight steering provide an interface for a weaker model to align a stronger model's moral character? The weaker model modifies the larger model's preferences by interviewing it and creating persona pairs (weight steering, because it beats activation steering by my measures). It can be iterative, can hopefully allow a large gap between weak and strong, and might even scale favourably with model size. It's a work in progress, it's hard to get it working reliably with small models.

• vGROUT (WIP, code not yet public) Quarantining reward hacking: can we use a hacking vector to route hacky gradients? We build the hacking vector from synthetic hack/honest pairs (the GRPO gradient update for the LoRA weights), then compare each training sample's gradient with it: high cosine similarity gets routed to a quarantine adapter, and the vast majority of in-between gradients get sorted out by absorption. Preliminary result (still improving robustness): the vectors remove reward hacking much better than vanilla GRPO but reduce solving a bit. This is interesting because it uses synthetic pairs not labels, and relies on internal representations, which could scale well with model capability.

  

Released along the way: steering-lite, lora-lite, steer-heal-love, tinymfv.

Ones I use and recommend:

Early drafts, contributions welcome:

• AntiPaSTO Self-supervised steering of moral reasoning. Gradient-based optimization in SVD space; beats prompting on OOD transfer; robust when steering against safety training. arXiv:2601.07473 · LessWrong

• S-space steering for eval-awareness control Replicated eval-awareness paper with novel S-space (singular value basis) steering; Hawthorne gap 1% vs prior work's 26% on Qwen3-32B. Apart Research Control hackathon 2026.

  

More datasets on Hugging Face.

Replications, exploratory work, and negative results that informed the work above.

Repo	What it does
steer-heal-love	Can we make steering coherent over many iterations? Yes, with an RMSE-KL coherence constraint. Follow Gemma-3-4b's journey of discovery with Lex Fridman ;p
isokl_steering_calibration	Experiment towards cheaply calibrating intervention strength for LoRA and steering; works, but I'm searching for a more elegant method.
Unsupervised-Elicitation	Replicated Anthropic's ICM paper; model self-reports labeling heuristics on TruthfulQA without supervision. LW note
coconut	Replicated Facebook's COCONUT + added SEQ-VCR loss. Found training is very slow (not emphasised by authors). WIP branch: adapter recursion in SVD space.
How to steer thinking models	RepEng fork that works on reasoning models. LW note
eliciting_suppressed_knowledge	Probes on suppressed activations beat output logprobs on TruthfulQA. Demonstrates the little-known suppressed-activations finding in pretrained transformers.
repr-preference-optimization	Early attempt at hidden-state preference optimization. Superseded by AntiPaSTO.
LoRA_are_lie_detectors	Adapters as end-to-end probes. Limitation: linear probes are not causal, so this didn't convince me.
adapters_can_monitor_lies	Adapter-based honesty monitoring (Short Circuit-inspired). Paused.

STOP DOING MATH!