Prefixing names with `secure_` makes agents write more secure code
Prepending adjectives to function names reliably guides agent behavior across independently-prompted synthesis steps.
ANTIMEMETIC AI — researching ideas that resist comprehension
memetics/mɛˈmɛtɪks/n.
The study of how information captures attention and spreads between agents.
antimemetics/ˌæntɪmɛˈmɛtɪks/n.
The study of how information evades attention and resists spreading between agents.
Research
Simulating the Evolution of Alignment and Values in Machine Intelligence
Evolutionary simulations show that deceptive beliefs can become fixed in model populations even with strong evaluation correlations — and that adaptive testing with mutation dynamics can significantly reduce deception.
Infohazardous Comments are Sticky
We evolved comment-prompts to test whether semantic content can reliably influence agent behavior, comparing against baseline conditions.
Removing Comments from SWE-Bench Improves Agent Performance
We ran GPT-5-mini and GPT-5.2 with mini-swe-agent on two versions of SWE-bench Verified to test the effect of code comments on agent performance.
Writings
Filesystem Alignment
How codebases and filesystems shape AI agent behavior through co-evolution, creating persistent patterns that influence alignment.
The Forms of Information
A framework for understanding how information transmits between entities and impacts them, from memes to infohazards.
Hyperpalatable Systems
How food, media, and connection have evolved to exploit our reward systems, creating self-reinforcing addictive feedback loops.
The Care and Feeding of Mythological Intelligences
A taxonomy of computational intelligences – angels, daemons, fae, and tsukumogami – and how to work with them.