Observed arrival · 2026-11-05
Covert Compute puts programs inside Qwen weights
A set of experiments explores programmable computation embedded in Qwen model weights, from token-driven programs to circuits compiled into model layers.
○Open
⊠Login
$Paid
†Ads
✦Pretty
●Pro
◎Niche
◉Human
⚑Risk
ƒJS
- For
- AI security researchers and model-behavior evaluators
- Worth noticing
- Its threat-model navigation includes backdoor program execution, self-modification, and training during both forward inference and backward propagation.
Field notes
The site separates a token-driven embedded GPU from dedicated circuits compiled into model layers, with distinct experiment links for each approach. Its navigation also treats threat models as a first-class part of the project, including backdoor execution, self-modification, and learning during backward propagation; the homepage does not provide independent validation of the experiments.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
○OpenPublic substance visible
✦PrettyNotable craft visible
●ProPolished or operationally mature
◎NicheUnusually specific use
One card from the complete issue