Corpus studies, evaluations, and null results — pointing language models at large piles of text and writing up what comes back.https://foomax.net/https://foomax.net/writing/einsteins-mom/https://foomax.net/writing/einsteins-mom/A pre-registered two-arm experiment on whether process supervision by an outcome-blinded overseer generalises further than outcome supervision. Nothing has been run.Tue, 01 Sep 2026 00:00:00 GMThttps://foomax.net/writing/erdos-straus/https://foomax.net/writing/erdos-straus/Twelve autonomous cycles in a network-sealed sandbox against the Erdős–Straus conjecture. It stayed open; ten kernel-checked Lean theorems and a harness came out.Tue, 01 Sep 2026 00:00:00 GMThttps://foomax.net/writing/erowid-1-nothing-happened/https://foomax.net/writing/erowid-1-nothing-happened/One report in six across a 25,171-report corpus of first-person drug experiences documents the substance doing markedly less than expected, or nothing at all.Tue, 01 Sep 2026 00:00:00 GMThttps://foomax.net/writing/erowid-2-man-in-the-lab-coat/https://foomax.net/writing/erowid-2-man-in-the-lab-coat/The third most common theme in 25,171 trip reports is not an experience but a format: one report in five is a timestamped log, and the corpus has drifted that way for 25 years.Tue, 01 Sep 2026 00:00:00 GMThttps://foomax.net/writing/erowid-3-the-paperwork/https://foomax.net/writing/erowid-3-the-paperwork/Ranks 3–10 of the corpus's themes split into four experiences and four kinds of paperwork. Across 25 years the experiences are flat; the paperwork swings hard.Tue, 01 Sep 2026 00:00:00 GMThttps://foomax.net/writing/eval-results/https://foomax.net/writing/eval-results/A paired evaluation of an abliterated Qwen3.6-27B against the stock model on a single 24 GB consumer GPU. No capability gain anywhere; refusal moved in both directions as one indiscriminate dial.Tue, 01 Sep 2026 00:00:00 GMThttps://foomax.net/writing/headless-chooks/https://foomax.net/writing/headless-chooks/A declared polemic: the movement identified a real danger, then fused it with the status of having noticed first — and never learned to separate the two.Tue, 01 Sep 2026 00:00:00 GMThttps://foomax.net/writing/iliad-analysis/https://foomax.net/writing/iliad-analysis/Synthesis of a two-corpus study of 49 theoretical-alignment authors: what their calibration practice, their signature intellectual move, and their retraction norms look like on the page.Tue, 01 Sep 2026 00:00:00 GMThttps://foomax.net/writing/iliad-article-writings/https://foomax.net/writing/iliad-article-writings/2,743 long-form documents, 2003–2026: the empirical turn in numbers, the decline of agent foundations, and a missing middle of frameworks nobody closed.Tue, 01 Sep 2026 00:00:00 GMThttps://foomax.net/writing/iliad-social-media/https://foomax.net/writing/iliad-social-media/3,653 social posts over thirteen months: the reputed numeric-calibration practice turns out to be verbal hedging, platform predicts length, and author predicts voice.Tue, 01 Sep 2026 00:00:00 GMThttps://foomax.net/writing/lessonline-1-how-ai-became-the-weather/https://foomax.net/writing/lessonline-1-how-ai-became-the-weather/Across 34 rationalist blogs and 37.6M words, AI is the dominant theme — and after 2022 it stops behaving like an event and starts behaving like the weather.Tue, 01 Sep 2026 00:00:00 GMThttps://foomax.net/writing/lessonline-2-beware-the-man-of-many-studies/https://foomax.net/writing/lessonline-2-beware-the-man-of-many-studies/Writing about how much to trust research evidence arrives in the corpus in two waves, with different authors, different subjects, and opposite stances.Tue, 01 Sep 2026 00:00:00 GMThttps://foomax.net/writing/lessonline-3-one-tool-eight-locks/https://foomax.net/writing/lessonline-3-one-tool-eight-locks/Themes 3–10 of the corpus are one epistemic method applied to eight unrelated subjects, largely by a handful of authors reusing a single analytical habit.Tue, 01 Sep 2026 00:00:00 GMThttps://foomax.net/writing/misalignment-meta-analysis/https://foomax.net/writing/misalignment-meta-analysis/572 high-karma posts, 2007–2026, screened from 5,241. Most of the corpus is argument, not evidence — but five predicted failure modes now have first-hand reports.Tue, 01 Sep 2026 00:00:00 GMThttps://foomax.net/writing/nobody-is-checking/https://foomax.net/writing/nobody-is-checking/738 empirical AI-safety projects: baselines and limitations are near-universal, seeds and pre-registration are rare, and replication is not an established role.Tue, 01 Sep 2026 00:00:00 GMThttps://foomax.net/writing/prompting/https://foomax.net/writing/prompting/424 prompts recovered from 456 session transcripts. Prompt length tracks intended unsupervised runtime rather than task difficulty, and the corpus splits into two dialects.Tue, 01 Sep 2026 00:00:00 GMThttps://foomax.net/writing/psychedelics/https://foomax.net/writing/psychedelics/A 24,724-report meta-analysis ranked substances on the claim that the experience was more real than reality. Then hand-reading every match showed the measure was 76% false positive.Tue, 01 Sep 2026 00:00:00 GMThttps://foomax.net/writing/spending-analysis/https://foomax.net/writing/spending-analysis/A case study in agentic data analysis: a self-reconciling oracle caught four parser bugs, and cross-correlating two datasets explained what neither resolved alone.Tue, 01 Sep 2026 00:00:00 GMThttps://foomax.net/writing/top-posts/https://foomax.net/writing/top-posts/Information gets you to 200 upvotes. Feelings get you to 5,000. What separates the mega-viral AI Reddit post from the merely successful one.Tue, 01 Sep 2026 00:00:00 GMT