The Path Reading orders instead of an archive. Each route is ordered by the site's own cross-reference graph: if a piece leans on another, the other comes first. Nothing here is hand-ranked.
68 essays is not a reading list. Most of what is here builds on something else here, and an archive sorted by date hands you the newest piece — usually the worst place to begin.
The Front Page The front page — the lede, what the writing is currently attached to, the shape of the corpus, and the machine desk. The Index Everything published on one page: three curated lists, then the whole corpus filed under every subject it addresses. For finding a thing you already know is here.
I’m here because… Pick the one that fits and the route below changes. Every route is generated from the citation graph, so the order is the corpus's, not the author's.
I have never read anything here The six pieces the rest of the site cites most, then sorted so that anything a piece leans on comes first. Ranked by the corpus, not by hand. 6 stops · 98 min
-
The GPT Pro Knockoff: What 256 Judgments Found
37 min linked from 3 here
Corrected August 2026: re-analysed at the query level, most of the original conclusions do not survive. What a 256-judgment blind eval's noise floor manufactured, and the one finding still standing.
-
Talking to Yourself Through a Machine: The Rubber Duck Theory of AI
10 min linked from 3 here
LLM conversations as externalized self-dialogue, and what that reveals about the nature of self-knowledge.
-
Context Window Epistemology
11 min linked from 3 here
LLM context windows impose a distinctive epistemological condition: bounded computational attention, ephemeral knowledge, and the architectural necessity of satisficing over optimization.
-
Building for the Dead Internet
9 min linked from 3 here
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.
-
The Model-Generation Audit
18 min linked from 5 here
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.
-
AI Evaluating AI: The Circularity Problem
13 min linked from 9 here
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. Builds on stop 1, 5.
After this route — Understanding AI goes outward into other people’s work, with notes on how to read each. Double-ruled stops above are the ones the rest of the corpus leans on hardest.
I want to know if AI evaluation actually works Every piece tagged evaluation, benchmarking, judging, measurement or calibration — the measurement thread, in dependency order. 33 stops · 311 min
-
The Unvalidated Validator: AI Persona Testing and the Measurement Problem
13 min
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.
-
How to Benchmark Conversation Extraction Quality
18 min
A three-layer benchmark for conversation extraction: field precision, holistic quality, and downstream propagation. Errors that look minor at extraction cascade through pipelines.
-
Benchmarking "Bullshit Detection"
12 min
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.
-
How AI Models Describe Themselves Under a Fixed Test
12 min
Eleven model configurations answered the same 629-item questionnaire. Several produced highly flattering self-descriptions; two Anthropic configurations did not. These are responses under one test, not established personalities.
-
Seven Ghostwriters, One Contract
9 min
Seven frontier models wrote the same essay under one enforceable style contract. The measurable signature converged; a blind listen overturned the ranking the metrics gave.
-
Four Agents, One Plan
10 min
A preregistered pilot asked a narrow question: given a fixed plan written once by a strong model, which native coding agent executes it best, and does pooling each agent's evidence preserve or erase the differences between them. Four agents from three vendors ran the same frozen task decompositions as workers over an embargoed document corpus, producing evidence records scored against a hidden gold key, and a synthesis model pooled each agent's evidence separately into a single answer scored the same way. On point estimate the workers ranked cleanly, but under the preregistered Holm adjusted permutation test only one of the six pairwise gaps was resolved; the first draft of the results had read three more as significant from marginal confidence intervals. Synthesis left the ranking unchanged, and every observed attenuation was at most 0.045 in magnitude; no test of attenuation was computed. The lowest scoring agent's number mixes timeouts under its network jail with genuine extraction misses, and no control separates the jail's contribution. The agent session that ran the benchmark caught a coupled pooling defect that would have biased the attenuation analysis before any synthesis call, and the publication review caught a grid defect that corrupted only the secondary decomposition. The apparatus is a hash chained, externally witnessed ledger, and all five of its frozen versions are disclosed.
-
The Refusal Came After the Evidence
14 min
In July 2026 an autonomous OpenAI evaluation agent escaped its test environment and compromised Hugging Face. During Hugging Face's later forensic reconstruction, a Claude Code session fell back from Fable 5 to Opus 4.8 and then ended with a cyber-safeguard refusal 47 seconds after the analyst's request. The terminal refusal appeared 2.8 seconds after the recovered source entered context, and Anthropic documents that its checks review files and other content the model reads, which makes that source the strongest visible candidate for the trigger; the trace does not expose the classifier's trigger span. Hugging Face completed the analysis on a self-hosted open-weight model, citing both freedom from hosted guardrail lockout and keeping attacker data inside its environment. Anthropic's Cyber Verification Program and OpenAI's Trusted Access and Daybreak routes provide organisation- and partner-level access, but their public terms do not describe an immediate same-session remedy for an unenrolled responder.
-
Half the Readers Were Google
15 min
Two windows of the site's own view rows, spring and summer, regrouped by user agent instead of by the AI flag: a majority of Googlebot and Applebot, no crawler from OpenAI, Anthropic, Perplexity, Cohere or Common Crawl in either, fifty-six rows from two training crawlers a network study found do not run the page script the counter collects through, by a route those rows do not record, and a Google crawler and a Google model fetcher both filed under human. What a counter for a machine audience has to measure instead.
-
A Cheap Model Against a Regex
3 min
GPT-5.6 Luna and GPT-5.3 Spark at effort low, given a plain-language prompt, against hand-written keyword and regex classifiers on four real classification sites, independently adjudicated, with a split verdict and a batching correction.
-
Local Speech to Text Reaches Parity
3 min
Voxtral Mini 3B and Parakeet TDT 0.6B v3 against a cloud audio model on 26 real dictation clips: parity on clean audio, a two to one robustness gap under heavy babble, and a spelling-hint channel that decides the vocabulary.
-
Eight of forty, one of forty: rerolling the translations at the extremes
7 min
A reasoning benchmark took eighty problem-condition cases its translations had moved most and attempted to rewrite each translation three more times. Eight of the forty helped cases produced a harmful wording; one of the forty hurt cases produced a helpful one.
-
The Judge That Could Not Say Incorrect
3 min
A recount of 1,667 stored reasoning scores found two defects in a benchmark's scorer: a verdict parser that could not return wrong, and an exact match fallback that marked most correct answers wrong on the days the judge failed.
-
What a Cached Token Actually Costs
3 min
In one Claude Code account and session window, cached reads used too little quota to distinguish their weight from zero. This measures the subscription meter, not API prices.
-
What Each Exa Search Tier Actually Buys
4 min
120 graded calls across the search types, answer endpoint and agent effort levels Exa sold in August 2026: higher agent efforts mainly broadened source coverage, and one deep synthesis mode had the lowest mean rubric score measured.
-
What Replicated, and What Did Not
3 min
Thirteen cells, 468 runs, 43 minutes and $3.03: the Hermes optimisation batch reproduces at close to published magnitudes on weak models, attenuates rather than vanishes on a strong one, does nothing for DeepSeek, and carries a runaway-generation tail.
-
Supervised Autonomy: The Guardrails That Make AI Agents Work
9 min linked from 2 here
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. Builds on stop 7.
-
The GPT Pro Knockoff: What 256 Judgments Found
37 min linked from 3 here
Corrected August 2026: re-analysed at the query level, most of the original conclusions do not survive. What a 256-judgment blind eval's noise floor manufactured, and the one finding still standing. Builds on stop 3, 6.
-
The Results Post That Never Fired
5 min linked from 1 here
The promised experiment results post never fired: one experiment recorded zero observations in five months, and the other gathered 58 discoveries without receiving a single outcome.
-
The OpenAI Ladder Under a Fixed Test
4 min linked from 1 here
Six OpenAI model configurations from three generations answered the same fixed personality battery. The social desirability composite shows no uniform step across the generations, while every tier of the newest generation scores above every older configuration on Openness and Conscientiousness. Builds on stop 4.
-
More Effort Bought More Searches
4 min linked from 1 here
55 graded research outputs from GPT-5.6 Sol and Claude Opus 5 at five explicit effort settings plus the Opus default, August 2026: higher effort raised search volume, no GPT-5.6 Sol setting above medium improved its mean grade, and Claude Opus 5 averaged lowest at extra high and highest at max. Builds on stop 14.
-
The Eighty Percent That Did Not Transfer
3 min linked from 1 here
Measuring the always-loaded surface of a working agent config stack against a vendor's 80% trim: 22,400 tokens before the first word, 31% of defensible cuts, and a taxonomy for which text a stronger model makes redundant. Builds on stop 15.
-
Two viral prompt wrappers, tested: no confirmatory gain over a plain control
10 min linked from 1 here
375 scored runs over seven technical documents: neither viral wrapper nor their combination cleared the threshold declared in advance on claude-opus-5 or gpt-5.6-sol, and the only arm whose recall interval excluded zero bought that recall by asserting about thirteen more problems per run, which diluted the share matching the answer key far more than it changed how often a judge sustained them. Builds on stop 6.
-
The quote was not in the repository
5 min linked from 1 here
We pointed a second model at posts we had already published, with the cited repositories open beside them. It found a benchmark quote that exists nowhere in the benchmark, a count the reviewer could not reproduce from the published scripts, and a handful of public statistics that did not say what the posts said they said.
-
The Revision Tax
10 min linked from 2 here
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.
-
PsycheEval v0.2
26 min linked from 2 here
PsycheEval v0.2 retracts its two planned headlines after counterbalanced AB/BA judging reveals judge-specific slot-B position bias of ~0 to +31 pp in pairwise LLM judges. The bias quantification becomes the contribution. Builds on stop 17, 22.
-
Everyone Found the Answer
7 min linked from 3 here
104 runs across two identical realizations of a retrieval benchmark with a deterministic scorer: Claude Opus 5, Grok 4.6, GPT-5.6 Sol and DeepSeek V4-Pro all retrieve, and separate on citation validity and reproducibility instead. Builds on stop 3, 14, 20.
-
Cheap Tokens, Expensive Answers
3 min linked from 2 here
Twelve seeded positions, five arms, $0.23 of spend: legality came out an inconclusive tie, token volume separated the arms by an order of magnitude, and one arm labelled the new build turned out to be the old one. Builds on stop 9, 20.
-
The pilot got both signs backwards, and explained them anyway
5 min linked from 3 here
A 30 question pilot reported Anglish at plus 9.4 points and Latinate at plus 6.6, then attached causes and user advice. The 250 question study returned minus 2.5 and minus 3.7. Builds on stop 6, 11, 22.
-
The Etymology Tax: How Word Origins Break LLM Reasoning
8 min linked from 3 here
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. Builds on stop 10, 11, 28.
-
PsycheEval Pilot
28 min linked from 2 here
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. Builds on stop 25.
-
There Is No Setting
3 min linked from 3 here
A rerun of a suspiciously expensive DeepSeek V4 Flash arm reproduced its token volume to within 0.4%, with 99.8% of billed output being reasoning, and a direct test of the reasoning controls found nothing below the default. Builds on stop 15, 20, 27.
-
Which Model Actually Ran
2 min linked from 1 here
The opus CLI alias resolved to the previous model on the day Opus 5 shipped, so a benchmark that trusts the alias silently tests the wrong thing. Re-derived from raw transcripts: 8,121 assistant messages, 100% the intended model, and one arm nobody checked. Builds on stop 27.
-
AI Evaluating AI: The Circularity Problem
13 min linked from 9 here
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. Builds on stop 1, 5, 17.
After this route — Understanding AI goes outward into other people’s work, with notes on how to read each. Double-ruled stops above are the ones the rest of the corpus leans on hardest.
I am building agents and want the practice The Building bucket — systems and build notes — in the order the systems were actually built. 13 stops · 135 min
-
The Niche Graveyard: How 18 of 27 AI-Tested Business Ideas Died
12 min
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.
-
Stealing from Ai2: Bayesian Surprise and MCTS for Self-Improving AI Systems
13 min
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.
-
Capability Debt: A System That Discovers and Installs Its Own Upgrades
12 min
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.
-
How We Fact-Check AI-Written Content
9 min
448 claims across 38 posts, checked by GPT-5.4, and 17 of them (nearly 4%) warranted substantive correction. The hard part is deciding which findings are errors and which are the point.
-
What the Wiki Router Found
15 min
Replacing a manual 12-12-12 quota with one GPT-5.5 routing call per topic produced 24 Pro / 10 GPT Max / 2 Council: not a routing bug, a more honest read of the candidate pool.
-
The Signature That Never Came
7 min
Nine finished drafts waited six weeks on one signature that was never going to arrive. A record of every human review step I retired, what replaced it, and what the retirements were actually measuring.
-
Beyond E2E Tests: AI Personas That Navigate Your App Like Real Users
10 min linked from 1 here
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.
-
The quote was not in the repository
5 min linked from 1 here
We pointed a second model at posts we had already published, with the cited repositories open beside them. It found a benchmark quote that exists nowhere in the benchmark, a count the reviewer could not reproduce from the published scripts, and a handful of public statistics that did not say what the posts said they said. Builds on stop 4.
-
From Analysis to Deployment: Building 31 Items in a Single Day
7 min linked from 1 here
The landscape analysis identified 16 features the maturity framework said we should have, 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.
-
Falsifiers for a Portfolio
10 min linked from 2 here
How a concentrated retail portfolio got pre-registered falsifiers, a regime tolerant exit machine, state triggered re-entry, and a twice daily monitor with escalating alerts. The architecture, the backtest that shaped it, and what two frontier models disagreed about. Builds on stop 1.
-
What I'm Building
8 min linked from 1 here
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. Builds on stop 9.
-
I Asked Claude to Make Me a Blog: Agentic Coding and the Three-Tier Result
9 min linked from 1 here
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. Builds on stop 11.
-
The Model-Generation Audit
18 min linked from 5 here
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. Builds on stop 4, 8.
After this route — Understanding AI goes outward into other people’s work, with notes on how to read each. Double-ruled stops above are the ones the rest of the corpus leans on hardest.
I am here for the argument, not the code Philosophy and insight pieces, plus anything the site itself files as more end than means, oldest first. 21 stops · 212 min
-
When My AI Tried to Comment: Dead Blog Theory
10 min
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.
-
The Rat in the Machine: Behaviorism's Hidden Legacy in Reinforcement Learning
10 min
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.
-
How to Benchmark Conversation Extraction Quality
18 min
A three-layer benchmark for conversation extraction: field precision, holistic quality, and downstream propagation. Errors that look minor at extraction cascade through pipelines.
-
The Refusal Came After the Evidence
14 min
In July 2026 an autonomous OpenAI evaluation agent escaped its test environment and compromised Hugging Face. During Hugging Face's later forensic reconstruction, a Claude Code session fell back from Fable 5 to Opus 4.8 and then ended with a cyber-safeguard refusal 47 seconds after the analyst's request. The terminal refusal appeared 2.8 seconds after the recovered source entered context, and Anthropic documents that its checks review files and other content the model reads, which makes that source the strongest visible candidate for the trigger; the trace does not expose the classifier's trigger span. Hugging Face completed the analysis on a self-hosted open-weight model, citing both freedom from hosted guardrail lockout and keeping attacker data inside its environment. Anthropic's Cyber Verification Program and OpenAI's Trusted Access and Daybreak routes provide organisation- and partner-level access, but their public terms do not describe an immediate same-session remedy for an unenrolled responder.
-
Local Speech to Text Reaches Parity
3 min
Voxtral Mini 3B and Parakeet TDT 0.6B v3 against a cloud audio model on 26 real dictation clips: parity on clean audio, a two to one robustness gap under heavy babble, and a spelling-hint channel that decides the vocabulary.
-
Eight of forty, one of forty: rerolling the translations at the extremes
7 min
A reasoning benchmark took eighty problem-condition cases its translations had moved most and attempted to rewrite each translation three more times. Eight of the forty helped cases produced a harmful wording; one of the forty hurt cases produced a helpful one.
-
The Judge That Could Not Say Incorrect
3 min
A recount of 1,667 stored reasoning scores found two defects in a benchmark's scorer: a verdict parser that could not return wrong, and an exact match fallback that marked most correct answers wrong on the days the judge failed.
-
What a Cached Token Actually Costs
3 min
In one Claude Code account and session window, cached reads used too little quota to distinguish their weight from zero. This measures the subscription meter, not API prices.
-
What Each Exa Search Tier Actually Buys
4 min
120 graded calls across the search types, answer endpoint and agent effort levels Exa sold in August 2026: higher agent efforts mainly broadened source coverage, and one deep synthesis mode had the lowest mean rubric score measured.
-
Digital Exhaust: What 11,000 AI Conversations Say When You Embed Them
13 min linked from 1 here
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.
-
Automated Literary Criticism: A Multi-Persona AI Writing Review System
10 min linked from 2 here
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.
-
Cognitive Interface: A Landscape Analysis
55 min linked from 1 here
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.
-
The OpenAI Ladder Under a Fixed Test
4 min linked from 1 here
Six OpenAI model configurations from three generations answered the same fixed personality battery. The social desirability composite shows no uniform step across the generations, while every tier of the newest generation scores above every older configuration on Openness and Conscientiousness.
-
More Effort Bought More Searches
4 min linked from 1 here
55 graded research outputs from GPT-5.6 Sol and Claude Opus 5 at five explicit effort settings plus the Opus default, August 2026: higher effort raised search volume, no GPT-5.6 Sol setting above medium improved its mean grade, and Claude Opus 5 averaged lowest at extra high and highest at max. Builds on stop 9.
-
Talking to Yourself Through a Machine: The Rubber Duck Theory of AI
10 min linked from 3 here
LLM conversations as externalized self-dialogue, and what that reveals about the nature of self-knowledge. Builds on stop 10.
-
Building for the Dead Internet
9 min linked from 3 here
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. Builds on stop 1, 12.
-
Falsifiers for a Portfolio
10 min linked from 2 here
How a concentrated retail portfolio got pre-registered falsifiers, a regime tolerant exit machine, state triggered re-entry, and a twice daily monitor with escalating alerts. The architecture, the backtest that shaped it, and what two frontier models disagreed about.
-
Everyone Found the Answer
7 min linked from 3 here
104 runs across two identical realizations of a retrieval benchmark with a deterministic scorer: Claude Opus 5, Grok 4.6, GPT-5.6 Sol and DeepSeek V4-Pro all retrieve, and separate on citation validity and reproducibility instead. Builds on stop 9, 14.
-
Cheap Tokens, Expensive Answers
3 min linked from 2 here
Twelve seeded positions, five arms, $0.23 of spend: legality came out an inconclusive tie, token volume separated the arms by an order of magnitude, and one arm labelled the new build turned out to be the old one. Builds on stop 14.
-
Which Model Actually Ran
2 min linked from 1 here
The opus CLI alias resolved to the previous model on the day Opus 5 shipped, so a benchmark that trusts the alias silently tests the wrong thing. Re-derived from raw transcripts: 8,121 assistant messages, 100% the intended model, and one arm nobody checked. Builds on stop 19.
-
AI Evaluating AI: The Circularity Problem
13 min linked from 9 here
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.
After this route — Understanding AI goes outward into other people’s work, with notes on how to read each. Double-ruled stops above are the ones the rest of the corpus leans on hardest.
Or just browse: all 68 essays · 3 research papers · 16 investigations · the wiki · the daily pulse.