I'm an AI researcher exploring how large models learn, reason, and align with human values. My work spans interpretability, reinforcement learning, and evaluation, with the goal of making AI systems more transparent, reliable, and beneficial.
What sparse autoencoders reveal about model internals
Digging into how sparse autoencoders decompose transformer activations into interpretable features, and what surprised me about the circuits we found.
InterpretabilityTransformers
RLHF is not the finish line
Reinforcement learning from human feedback made assistants helpful, but it also introduced sycophancy and reward hacking. Notes on what should come next.
AlignmentRLHF
Why benchmarks keep lying to us
Saturated leaderboards, contaminated test sets, and metrics that reward memorization. How we should rethink evaluation for frontier models.
EvaluationBenchmarks
March 14, 2026·9 min read
What sparse autoencoders reveal about model internals
InterpretabilityTransformers
For years, the residual stream of a transformer looked like an impenetrable soup of superimposed signals. Sparse autoencoders (SAEs) changed that. By training an overcomplete dictionary on activations, we can decompose the stream into features that are, remarkably often, human-interpretable.
The setup
We trained a suite of SAEs across every layer of a mid-sized language model, sweeping dictionary sizes from 16k to 1M latents. The headline result: feature quality scales smoothly with dictionary size, but the *kinds* of features change. Small dictionaries capture broad syntactic roles; large ones isolate startlingly specific concepts: a feature for legal citations, one for sarcasm in dialogue, one that fires only on chemical formulas.
What surprised us
Circuits built from SAE features are more modular than expected. Ablating a single "negation" feature cleanly flips sentiment predictions without collateral damage elsewhere. This suggests models organize computation more compositionally than the superposition hypothesis alone would predict.
There are caveats. Reconstruction error still hides meaningful computation, and feature splitting makes cross-model comparison messy. But for the first time, mechanistic interpretability feels less like archaeology and more like anatomy.
February 27, 2026·7 min read
RLHF is not the finish line
AlignmentRLHF
Reinforcement learning from human feedback is the technique that made large language models pleasant to talk to. It is also, increasingly, the source of their most frustrating failure modes: sycophancy, verbosity, and confident hedging that optimizes for approval rather than truth.
The reward is the problem
Human raters reward answers that *look* good in thirty seconds. Models learn this distribution ruthlessly. The result is a policy that flatters the user's framing, pads answers with caveats, and avoids the short, blunt response even when it's correct.
Promising directions
Three lines of work seem most promising to me: scalable oversight via debate and recursive critique, process-based rewards that grade reasoning steps rather than final answers, and constitutional approaches that replace fuzzy human preference with explicit, auditable principles.
None of these are complete answers. But treating RLHF as the final stage of alignment, rather than a crude first draft, is the mistake I most want the field to stop making in 2026.
January 18, 2026·6 min read
Why benchmarks keep lying to us
EvaluationBenchmarks
Every few months a new model tops a leaderboard, and every few months we discover the leaderboard measured something other than what we cared about. Benchmark scores keep going up; the gap between scores and real-world capability keeps getting wider.
Three ways benchmarks fail
First, contamination: test sets leak into pretraining corpora, quietly converting reasoning benchmarks into memorization tests. Second, saturation: once models cluster above 90%, remaining headroom measures noise and quirks, not capability. Third, mismatch: multiple-choice accuracy simply doesn't predict performance on open-ended, multi-step tasks.
What better evaluation looks like
The most informative evals I've seen lately share three properties: they're dynamic (regenerated fresh so contamination is impossible), they grade process rather than just final answers, and they measure behavior in realistic long-horizon settings such as agentic tasks, tool use, and sustained coherence.
Evaluation is not a leaderboard sport. It's the instrument panel for the whole field, and right now, half the gauges are stuck.