Trace

Salmaan Barday

individual

OtherTechnical AI safety research

Received
$650
from 1 funders
Grants
2
Funders
1
Years active
2026

Grants over time

$0$200$400$600$8002026

Biggest funders

BlueDot Impact$650100%2

Grants received

DateFunderViaAmountCauseSourcePurpose
Jul 2026BlueDot Impact$500OtherbluedotExperiment on AI deception. I find that reading when a model is lying is 'easy', but stopping it is much harder and only works at certain layers. Detecting a problem isn't the same as fixing it.
May 2026BlueDot Impact$150EvalsbluedotTesting whether LLM agents develop compressed or private communication under repeated interaction, and building evals to distinguish benign compression from concerning non- interpretable drift.