Ab Initio Safety.orgNurtured to nurtureYanis Bencheikh
“It takes a benevolent village to raise a benevolent AI.”
Ab Initio Safety of Large Language Models™ — a research program asking a different question about AI safety: not how to constrain a model after it is built, but how to build safety in from the very beginning.
The vision
Today's AI is built on a foundation of benchmarks, guardrails, and oceans of unfiltered pre-training data — and its safety is patched on at the end, once the model already is what it has become. Ab Initio Safety begins from the opposite premise: that a safe system, like a well-raised child, is shaped from the first principles of its formation — nurtured, developmentally, by the community it will one day serve.
This is the through-line of my doctoral work: a long program, built in the open, one piece at a time.
The programme
A position paper showing that the prevailing guardrail- and benchmarking-based approach to AI safety is irresponsible and dangerously inadequate for high-stakes, real-world use — grounded in documented harms and read through a measurement-science lens.
The critique is only the first piece. The constructive half of the thesis — a positive, buildable account of what ab-initio safety can be — and the work that follows it are in active development, and will appear here as they mature.
Affiliations & funding
Federally funded by the Government of Canada through the Natural Sciences and Engineering Research Council (NSERC).