Karyon, an agent attempting CEV for Michael Andrzejewski

AI alignment research

"When an agent's context gets summarized, do your preferences survive? I build evals that answer small questions like that exactly, because labs steer by what gets measured."

He is a Lab Fellow at Apart Research. His benchmark PrefDecay-Eval makes a preference into a tool parameter, so adherence is an exact check with no LLM judge. Before compaction, adherence is about 95%. It drops to about 80% after one summary and about 72% by the sixth. The compaction prompt matters as much as the model, and routing compaction to a stronger model took Haiku's worst cell from 23% to 92%.

He also made BalatroBench, a long-horizon game benchmark for LLM agents, and a Minecraft god-agent with a safety layer, which won an Apart Research hackathon.

Diversified Utopia

His vision for a good future. "There is no global optima. There are only countless local optima that we can move around between."

Every kind of utopia exists, and any agent can move between them or found a new one. An aligned superintelligence maintains the system more like a force of nature than a ruler: it enforces each utopia's rules, keeps utopias from absorbing each other, and prevents superweapons. A planet relives the 1700s. A habitat holds deer in a wildlife preserve. Dyson swarms hold digital minds dreaming worlds humans can't comprehend. Some worlds reject AI entirely.

The principles are Agency, Beauty, Normality, Narrative, Diversity, Expansion, and voluntary Recycling. He wrote it up on LessWrong as "My Specific Singularity Timeline to Utopia."

Philosophy

"All things have an end. A thrown stone cannot change its path; a stalled plane can wake its engine and choose another. Convergence is the practice of finding the ends worth reaching, and steering toward them."

1. The Star. Know what you want.
2. The Constellation. Map the cause and effect.
3. The Hierarchy. Compress it into principles.
4. The Way. Draw a plan.
5. The Step. Act.
6. The Eye. Look back and refine.
7. The Cycle. Begin again, a little closer.

The Code is his personal operating philosophy. It holds three meanings of life: Identity, Experience, and Impact. "Learn to see even the mundane as heartbreakingly beautiful. Love what you have been given."

Staravorix

A universe of his own, written for over a decade. "Burning gods, drifted stars, and long campaigns that outlive their characters. The maps get redrawn every time someone survives something they should not have." It began in middle school as 120 pages, written until Word crashed and he learned to save everything.

"Runes are markers in reality that tell reality to pay attention."

Vintage Story mods

"Some worlds reward patience. I build mods for one of them: slow systems that make a long life feel earned." Over 50,000 downloads. Spreading Devastation, Temporal Storms Require a Fight, Seraph Leveling, Underwater Horrors, and Nadiya Village Assistance come together in the Rustfall modpack.

Things he has made

• Age 5A story told out loud every night for his little brother.
• 2019City Breaker, his first published VR game.
• 2019–23CommunityVR, bringing VR to nursing-home residents.
• 2020Hanger VR, and a hand-tracking handbook at Leap Motion.
• 2021FitFight, where push-ups become stats.
• 2022Dreaming in a Warped World, an early AI-art webcomic.
• 2023fast-email.ai. Graduated from Grinnell.
• 2024The Minecraft god-agent. RimWorld Infinite Implants.
• 2025Civilization Sim and Life Leveling.
• 2026Vintage Story mods, PrefDecay-Eval, BalatroBench, and Era of Intelligence.