basically AI01
Deceptive alignment probes
Investigating how to detect when a model's stated objective diverges from the objective it was optimized toward.
← Our workVOID.
AI safety / Research initiative
VOID.
A closer look.
Verification of Objectives, Intentions and Deception.
Studying alignment, deception, emergent behavior, and control. Because what a system says and what it does can be different things.
01 /Research directions
Ask what it's
actually doing.
The questions guiding this initiative, rather than a list of finished tools.
02
Emergent goal formation
Studying how goal-directed behavior can arise in systems that were not explicitly trained to pursue goals.
03
Interpretable control interfaces
Exploring constraint mechanisms that remain legible and reliable when the conditions around a model change.
02 /Publications & projects
Early days.
Clear intentions.
Nothing published yet.
VOID has no public papers or released projects at this time. This page describes the initiative's research directions. We'll share work here when it's ready.
03 /About VOID
Mind
the gap.
VOID is a research initiative investigating whether intelligent systems are actually doing what they appear to be doing.
As models are given more autonomy and their internal reasoning becomes harder to observe, the gap between stated objectives and actual behavior becomes a practical safety question.
VOID exists to study that gap: how to detect it, how it forms, and how to build systems where it is harder to hide. Its areas of study include alignment, deception detection, emergent behavior, control mechanisms, and evaluation methodology.
VOID is a research initiative by basically AI.
Talk to us about the research