Blog

Curated alignment, interpretability, and world models content

Highlighted Works

From Aurelius

Proba: predict what your fine-tuning data will do, before you train

The announcement of Proba, an interpretability platform for open-weight models: read your fine-tuning data through the model's internals and predict what it will teach, before you commit the run. Launching in August.

From Aurelius

Aurelius Competition Live on Apex: Activation Steering

Aurelius launches an activation steering competition on Apex, Bittensor subnet 1: live now, with miners incentivized to steer model behavior from the inside.

Alignment

An Off Switch for Dual Use Knowledge in AI Models

AE Studio, an Aurelius advisor, with Anthropic: GRAM trains dual use knowledge into removable modules, so one model can ship with sensitive capabilities switched on for trusted users and off everywhere else.

Alignment

Logits as a New Monitor for Evaluation Awareness

Goodfire reads a model's logits to estimate how likely it is to say it knows it is being tested, catching eval awareness with 10 to 100 times fewer rollouts than judging its words.

Interpretability

A Global Workspace in Language Models

Anthropic finds an internal workspace inside Claude, the J-space, where the model thinks about concepts without writing them down and deliberate reasoning takes shape.

Interpretability

Natural Language Autoencoders

Anthropic trains Claude to translate its own activations into plain text and back: a verbalizer describes what a thought contains, a reconstructor rebuilds the thought from the description alone.

World Models

A Functional Taxonomy of World Models

The follow up to From Words to Worlds: world model is an overloaded term, so Fei-Fei Li splits it into renderers, simulators, and planners, and argues the three are converging into a single foundation model.

World Models

Google DeepMind

Genie 3: A New Frontier for World Models

DeepMind's general purpose world model: type a prompt and walk through a live, navigable world it generates in real time, holding consistent for minutes at a stretch.

“If it is not right do not do it; if it is not true do not say it.”

Marcus Aurelius, Meditations Book XII