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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
An experiment whose entire premise is a controlled information boundary shipped a finished manuscript and left the one instrument that would measure the boundary with no records in it.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
A one-day frontier-model prototype carried a credential that read as validation and dissolved into two canceling errors. Verification took four more days and three NO-SHIP verdicts.
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.
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.
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.
Five years of instinct beat the index 4.84x to 1.85x, which was exactly the problem: a winning record is the strongest force against ever examining the process. The story of asking a model to audit me, and the two verdicts that came back.
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.
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.
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.
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.
GPT-5.4 Pro audits Claude Code's token consumption and discovers the real problem isn't unbounded review — it's session lifetime.
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.
An ablation study on a personality-preserving narrative pipeline found that all three rebuilt variants, which share planning artifacts made from the full archive, scored 6.0 to 6.7 points better than the old pipeline on the mean of four evaluation runs each and differed from one another by 0.7 points at most. The comparison cannot separate planning context from the accompanying change in planning architecture or from possible model-version effects, and regenerating one narrative moved its score by more than the variants' differences.
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.
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.
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.
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.
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.
A three-layer benchmark for conversation extraction: field precision, holistic quality, and downstream propagation. Errors that look minor at extraction cascade through pipelines.
Self-report scales reach .80-.90 reliability; LLM inference from conversation reaches r~.44 (Peters et al. 2024). Combining three methods yields a profile an AI assistant can act on.
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.
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.
I trained two Qwen3-8B models, each on one person's own text message history (32,574 reply turns in all), and ran them in conversation with each other. All five conversations settled into logistics, sleep talk or repetition loops, four of them within fifteen turns, which suggests that these texts carried their writers' coordination more than their substance.
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.
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.
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.
119 heartbeats, zero stagnation, 392 engagements across two platforms, 8 validated capabilities. The story the observation system missed.
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.
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.
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.
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 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.
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.
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.
Adversarial testing isn't a metaphor for business validation. It's the same methodology, applied to a different failure mode.
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.
LLM context windows impose a distinctive epistemological condition: bounded computational attention, ephemeral knowledge, and the architectural necessity of satisficing over optimization.
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.
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.
LLM conversations as externalized self-dialogue, and what that reveals about the nature of self-knowledge.
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.
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.
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.
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.
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.
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.