AI & ML interests

Interpretability, runtime monitoring, model editing, and the verification of self-improving AI, on open models with every run replayable.

Recent Activity

Michelangelo-z  updated a Space 14 days ago
saias-research/README
Michelangelo-z  published a Space 14 days ago
saias-research/README
View all activity

Organization Card

SAIAS Research

SAIAS (Self-improving AI, Auditable & Steerable) keeps AI self-improvement inspectable. It measures, openly and replayably, whether we can still see and steer what changes inside a model as an AI improves it: a runtime that watches a small open model's internals token by token, flags and pauses at internal states, and logs every intervention with who chose it and why.

The map

For each failure we study, the map of understanding asks seven questions: can we read the failure inside the model, foresee it, steer it, stop it, fix it, does the signal hold up under pressure, and does seeing inside help an AI improve the model. Each cell puts selected published evidence next to our own runs, with its source, its date and its evidence level. The two columns are context, not a ranking. Our first failure is repetition loops in Gemma 3 1B, and all our results so far are exploratory. Negative results count.

Links

Licence

Code is Apache-2.0. Content and data, including the run records, are CC BY 4.0. Every result on the map credits the people who produced it.

Who runs it

SAIAS is run by Michelangelo Zampogna. AI Design Blueprint (AIDB) sponsors SAIAS's infrastructure costs and provides the reviews each phase passes; it holds none of SAIAS's data, and its reviews grade how SAIAS is built, not whether its findings are true. AIDB was founded by the same person as SAIAS.

models 0

None public yet

datasets 0

None public yet