
We Built Minds Before We Understood Them
Every generation of technology has a moment where building stops being the hardest part.
We're the research company reverse engineering with interpretability to design, build, and understand intelligence.
Subscribe:.rss.atom.json.md.m3u.pls
Live Last read · last published · next check

Every generation of technology has a moment where building stops being the hardest part.

Activation Steering Can Reduce Bias, But Only If You're Honest About Its Limits

Dictionary learning over transformer activations: what the trainer sees, how feature quality scales with data, and how a trained SAE plugs into inspection, steering, and benchmarks.

No gradient-descent loop runs. Four analytical passes: SAE baseline, gradient landscape, LiSSA influence scoring, and NTK-linearised weight prediction

https://pypi.org/project/aquin

Vivly.in × Aquin Labs Experiment Study

Open-Source Repo

Geometry inspection, retrieval evaluation, fine-tuning monitoring, and embedding diff across checkpoints. Load any sentence-transformers compatible encoder and get the full picture of embedding space.

Aquin supports the full transformer family: dense LLMs and Mixture-of-Experts models. Every tool in the platform is architecture-aware from the moment you load a model.

Adversarial risk detection across the model checkpoint and the boundary between model versions.

Real-time signal detection, behavioral before/after comparison, and per-layer feature diffs: everything the loss curve doesn't show you.

Seven tools that answer two questions: how did the model produce this output, and is the output actually correct?

Three behavioral evals that go beyond accuracy measuring whether a model answers consistently, what it quietly avoids, and where its knowledge runs out.

How Aquin scores and validates SAE features, and how the Benchmark Builder lets you measure model capability without leaving your inspection session.