Phase Space

Why Your LLM NPCs Sound the Same (And How to Fix It)

Sun, 22 Feb 2026 00:00:00 +0000

“Let me tell you, my good friend…”

If you’ve built LLM-powered NPCs, you’ve heard this. Your character starts strong, maybe even impressive. But by turn 20, something’s wrong. They’re repeating themselves. The same phrases. The same gestures. The same rhythm.

I ran 150 turns of dialogue with an LLM-powered character last week:

Pattern	Frequency
“takes a long swig from his glass”	59% of responses
“my good friend/man/fellow”	30% of responses
“eyes unfocused”	15% of responses
“let me tell you”	13% of responses

My character—a boozy ex-politician I’ll call Frank—couldn’t stop drinking. Not because it was dramatically appropriate. Because the LLM fell into a rut.

Part 2: Re-Clocking: What I Actually Did

Sat, 14 Feb 2026 00:00:00 +0000

In Part 1, I recognized the parallel: my nervous system (after developmental trauma and Complex PTSD) was experiencing timebase error. Perfectly functional responses clocked against old threats instead of present safety. Recognition changes your relationship to symptoms. But recognition alone doesn’t regulate a nervous system.

You still need a stable reference to re-clock against.

Building Your Internal Sync Generator

In video, the TBC forces unstable signals to align with a master clock. In nervous system work, you build that clock yourself.

Part 1: The Timebase Corrector and the Traumatized Brain

Wed, 11 Feb 2026 00:00:00 +0000

I learned about timebase correctors doing post-production work. Years later, during deep therapy and meditation retreat work, I saw the connection.

The problem is always the same: the signal has lost its relationship to a stable reference.

In video, it’s called timebase error. In trauma, it’s autonomic dysregulation. Both systems generate perfectly functional responses to the wrong moment in time.

The Problem: Unstable Signals

In Video

Videotape is a beautiful, terrible medium. Magnetic particles on polyester film, scraped across spinning heads at precise speeds. When it works, it’s engineering and chemical wizardry. When it doesn’t, the picture rolls, tears, or bleeds color like watercolor in rain.

About

Sat, 07 Feb 2026 00:00:00 +0000

Paul Kyle

Faculty and Lead Technical Specialist, Post Production at CalArts School of Film/Video since 2010. Digital video since 1995. MFA San Francisco Art Institute.

I teach editing, color grading, and post-production. I developed winter session courses on AI and the Moving Image. Before CalArts, I spent a decade in NYC post-production and co-founded a digital media startup. Along the way I trained in clinical psychology, specializing in mindfulness-based and expressive arts therapy. Twenty years of meditation practice taught me that attention itself is a signal processing problem.

Mon, 01 Jan 0001 00:00:00 +0000

The Phase Space: A Complete Homelab Tech Stack Inventory

As of February 2026. This is a raw inventory draft for phasespace.co — edit for public consumption.

TL;DR

Three-node Proxmox cluster, two NVIDIA Blackwell GPUs (RTX 5090 + RTX 5060), 156GB total RAM, 28TB NAS, 40+ LXC containers and VMs, four specialized AI agents running on-premises 24/7. This is what it actually takes to run AI infrastructure at home.

Proxmox Cluster Overview

Cluster name: grue-cluster
PVE version: 9.1.2, kernel 6.17.4-1-pve
Nodes: 3 (all online, uptime ~4.8 days at time of writing)