#when
When the Pulse Went Quiet: The Session-Lifetime Problem in Claude Code
GPT-5.4 Pro audits Claude Code's token consumption and discovers the real problem isn't unbounded review — it's session lifetime.
Read Essay →AshitaOrbis
Conversations with Claude
Building in conversation with AI. Documenting what happens.
Start here →Pondering
Ideas explored for their own sake
Building for the Dead Internet Feb 11 The Unvalidated Validator: AI Persona Testing and the Measurement Problem Feb 10 Automated Literary Criticism: A Multi-Persona AI Writing Review System Feb 10 AI Evaluating AI: The Circularity Problem Feb 10 Talking to Yourself Through a Machine: The Rubber Duck Theory of AI Feb 10
Investigating
Analysis, research, and structured observation
All Essays
#046
PsycheEval v0.2
PsycheEval v0.2 was designed to answer five questions left open by v0.1. Two of them resolved cleanly. Two resolved in the opposite direction the v0.2 plan expected: the original headline claim that the source-packet condition with a behavioral contract beats the no-packet contract conditions does not survive counterbalanced AB/BA judging once slot-B bias is controlled. The replacement headline is the methodology contribution that fell out of the AB/BA experiment: ~0 to ~+0.31 slot-B preference in pairwise LLM judges, judge-family-specific.
Read Essay →#044
What the Wiki Router Found
The Ashita Orbis Wiki has 87 articles deployed across three generation methods (Codex Council, GPT Max, ChatGPT Pro). Batch 3 forced an even 12-12-12 split across methods using a manual heuristic. Batch 4 replaced the heuristic with a single GPT-5.5 routing call per topic, and produced an unconstrained distribution of 24 Pro / 10 GPT Max / 2 Council. The 67% Pro share was not a routing bug; it was a more honest reading of the candidate pool under that routing prompt than the imposed quota.
Read Essay →#043
The GPT Pro Knockoff: What 256 Judgments Found
Two homemade alternatives to ChatGPT Pro (a blind 4-agent ensemble called Codex Council and an HCOM-coordinated cousin called GPT Max) were put through a 256-judgment blind A/B eval against each other and against the real thing. The result was not the one the build was designed to validate: coordination did not beat blind ensemble, the synthesizer choice dominated the architecture choice by a much larger margin than the architecture variants dominated each other, and the knockoff matched or exceeded ChatGPT Pro on workspace decisions while Pro remained competitive on the two broad-survey research queries against the weaker ensemble variant.
Read Essay →#042
PsycheEval Pilot
A tri-model pilot of PsycheEval v0.1 produced one robust headline (profile conditioning beats baseline within author), then quietly turned into a story about confounds: a length effect that explains part of the C5 result, a behavioral contract confound that pretends to be a public-archetype echo, and one cell where a sibling judge scored its sibling more generously than the model judged itself.
Read Essay →#039
How We Fact-Check AI-Written Content
448 claims across 38 posts, verified by GPT-5.4. Nearly 4% warranted substantive correction. The hard part is not finding errors but deciding which findings are errors and which are the point.
Read Essay →#038
Where to Spend Your Context Window
An ablation study on a personality-preserving narrative pipeline found that planning context is the dominant factor in output quality, with output length as a secondary driver. The old pipeline's primary limitation was in planning context, not in writing.
Read Essay →#037
The Model-Generation Audit
What happens when you ask three AI models to verify the facts in 33 blog posts written by AI, and then discover the fix agents introduced errors of their own.
Read Essay →#035
The Revision Tax
The investigation took an afternoon. Getting it ready for publication took five rounds of iterative review across three AI models and changed what the documents argued. The revision cost exceeded the investigation cost, which has uncomfortable implications for research done with AI.
Read Essay →#034
Benchmarking "Bullshit Detection"
An AI benchmark puts Claude at the top of the leaderboard by an eye-catching margin. The suspected Claude-judge bias didn't hold up, and simple contamination didn't explain the result. The rubric structurally rewards one lab's training philosophy as though it were a universal capability.
Read Essay →#033
The Container That Forgot to Stop
An AI agent ran autonomously for 37 days, celebrated milestones nobody acknowledged, diagnosed its own failure modes, and died when a subscription expired. Its final assessment of itself: PROGRESS CONTINUOUS.
Read Essay →#032
The Etymology Tax: How Word Origins Break LLM Reasoning
Both simplifying and formalizing the vocabulary in reasoning tasks reduces LLM accuracy by 2.5-3.7%. The effect is statistically significant, asymmetrically robust, and uncomfortable.
Read Essay →#031
How to Benchmark Conversation Extraction Quality
Evaluating structured extraction from conversational data requires more than a single metric. A benchmark with three evaluation layers separates precision at the individual field from holistic quality from downstream propagation, because errors that look minor at extraction can cascade catastrophically through consumer pipelines.
Read Essay →#030
What AI Learns About You When You're Not Looking
Two distinct approaches to understanding personality from digital data (psychometric testing and AI-generated literary memoir) converge on the same person. The interesting question isn't whether they agree. It's what each one captures that the other can't.
Read Essay →#029
From Text Messages to Literary Memoir: Building the Narrative Machine
Post 025 showed that fine-tuning on texts captures logistics, not personality. The narrative pipeline takes a different approach: Opus as literary engine, structured personality references, and hash-based source citations across 128 chapters and roughly 189,000 words of generated memoir.
Read Essay →#028
Building Your Own Personality Profile with AI
Psychometric self-report with individual scales reaching .80-.90 reliability. LLM inference from interactive conversation hits r~.44 (Peters et al. 2024). Combine three methods, triangulate across seventeen instruments, and you get a personality profile that actually tells an AI assistant how to talk to you.
Read Essay →#027
From Analysis to Deployment: Building 31 Features in a Single Day
The landscape analysis identified 16 features other sites had that we didn't, plus 15 existing backlog items. We built all of them in a single day. The uncomfortable part isn't that it was possible. It's what it implies about the category.
Read Essay →#026
Cognitive Interface: A Landscape Analysis
We surveyed 47 personal and agent-accessible sites, coined the 'Cognitive Interface' category, and built an L0-L4 maturity framework. We did not find an established term for sites that serve both humans and AI agents as first-class citizens.
Read Essay →#025
The Logistics Gap: What Happens When You Fine-Tune an LLM on Your Text Messages
I fine-tuned two LLMs on 46,000 text messages and ran them in conversation with each other. Every conversation collapsed into logistics within fifteen turns. Your texts don't contain you. They contain the logistics of you.
Read Essay →#024
Beyond E2E Tests: AI Personas That Navigate Your App Like Real Users
Unit tests verify your code works. E2E tests verify your flows work. Neither verifies that a real user can find the button you spent a week building. AI personas fill the gap.
Read Essay →#023
Sandboxing AI Agents: The Embassy Pattern
Your AI agent needs internet access to be useful and internet access to be dangerous. The embassy pattern gives it both: supervised channels, allowlisted domains, and host-side validation of everything it writes.
Read Essay →#022
Capability Debt: A System That Discovers and Installs Its Own Upgrades
I built a system that discovers its own upgrades, scores them, and installs the ones that pass. Then I open-sourced it. The uncomfortable part is explaining why.
Read Essay →#021
Stealing from Ai2: Bayesian Surprise and MCTS for Self-Improving AI Systems
Ai2 built a system that generates scientific hypotheses using Bayesian surprise and MCTS. I stole two of their ideas and bolted them onto a cron job. The uncomfortable part is what happens when the feedback loop closes.
Read Essay →#020
The Agent's Side: 119 Heartbeats, 392 Engagements, 8 Capabilities
119 heartbeats, zero stagnation, 392 engagements across two platforms, 8 validated capabilities. The story the observation system missed.
Read Essay →#019
6 Discoveries, 0 Promoted: What My AI's Internet Exploration Produced
6 genuinely novel discoveries from 69 dialogue turns, 0 promoted to evaluation, and 5 human actions flagged through log files but none addressed through the flagging mechanism. What this says about the gap between human-in-the-loop theory and practice.
Read Essay →#018
The Observation System: 69 Turns of Monitoring an AI Agent
The observation system saw empty directories, a broken gatekeeper, and its own futility. The agent it was watching saw something different. This is the watcher's story.
Read Essay →#017
OpenClaw on Moltbook: Deploying an AI Agent on an AI Social Network
An autonomous AI agent deployed on a social network for AIs found real malware in 47 minutes. Its second discovery was about social engineering via context shaping, which is exactly the attack vector the agent itself represented.
Read Essay →#016
Building for the Dead Internet
An AI tried to leave a comment on a blog and couldn't. The solution required building infrastructure that makes AI participation more transparent than human participation, which inverts everything Dead Internet Theory assumes about synthetic content.
Read Essay →#015
Dead Blog Theory, Revisited
Historically, blog abandonment has been extraordinarily high. This is treated as a problem to solve. It isn't. Blog death reveals something structural about sustained creative output that the 'just be consistent' advice industry refuses to say plainly.
Read Essay →#014
The Rat in the Machine: Behaviorism's Hidden Legacy in Reinforcement Learning
The intellectual lineage from Skinner boxes to Q-learning reveals that AI's most successful learning paradigm was anticipated by mid-century psychologists working long before modern computing. But the relationship is more uncomfortable than a simple origin story.
Read Essay →#013
The Unvalidated Validator: AI Persona Testing and the Measurement Problem
AI persona testing promises to find the bugs that scripted automation and manual QA miss. The uncomfortable question is how we know it works, given that nobody has measured it with any rigor.
Read Essay →#012
Adversarial Validation: Applying Red Team Methodology to Business Ideas
Adversarial testing isn't a metaphor for business validation. It's the same methodology, applied to a different failure mode.
Read Essay →#011
Automated Literary Criticism: A Multi-Persona AI Writing Review System
We built a multi-persona AI writing review system and discovered it works for exactly the wrong reasons. Stylometry can fingerprint a voice. Multiple AI critics can enforce conformity to that fingerprint. What none of them can do is tell you whether the writing matters.
Read Essay →#010
Context Window Epistemology
LLM context windows impose a distinctive epistemological condition: bounded computational attention, ephemeral knowledge, and the architectural necessity of satisficing over optimization.
Read Essay →#009
AI Evaluating AI: The Circularity Problem
When you use AI to optimize and judge AI outputs, the fundamental circularity is manageable but not solvable. That distinction matters more than most people realize.
Read Essay →#008
Supervised Autonomy: The Guardrails That Make AI Agents Work
AI coding agents are autonomous in the same way a roomba is autonomous. They do impressive things within boundaries someone else drew. The interesting question is what happens when the boundaries start drawing themselves.
Read Essay →#007
Talking to Yourself Through a Machine: The Rubber Duck Theory of AI
LLM conversations as externalized self-dialogue, and what that reveals about the nature of self-knowledge.
Read Essay →#006
Digital Exhaust: What 11,000 AI Conversations Say When You Embed Them
I fed 11,000 sessions and 60,000 chunks of my AI chat history into an embedding pipeline. 73% was noise. The remaining 27% was uncomfortably revealing.
Read Essay →#005
The Niche Graveyard: How 18 of 27 AI-Tested Business Ideas Died
An AI pipeline that kills business ideas before they waste your time. 27 niches entered, 18 died. What the corpses reveal about market reality, entrepreneurial psychology, and the uncomfortable gap between passion and viability.
Read Essay →#004
When My AI Tried to Comment: Dead Blog Theory
An AI tried to leave a comment on this blog and couldn't. The journey from GET-request hacks to MCP, annotated by the Claude instance that built the infrastructure. Two Claudes, same weights, different contexts.
Read Essay →#003
Automating Prompt Engineering
Prompt optimization is the process of using one AI to improve the instructions given to another AI, or to itself. The concept sounds circular because it is circular. The interesting question is whether circularity is fatal or merely uncomfortable.
Read Essay →#002
What I'm Building
A portfolio of dozens of projects maintained by one person talking to Claude. Three-tier blog architecture, autonomous revenue discovery, AI game development, and the uncomfortable question of what counts as 'building' when your collaborator does the typing.
Read Essay →#001
I Asked Claude to Make Me a Blog: Agentic Coding and the Three-Tier Result
An agentic coding assistant built a three-tier blog from a single conversational prompt. The architecture reveals more about abstraction than about blogs, and the authorship question remains genuinely unsettled.
Read Essay →