1. Get a LessWrong post
2. Configure the two-pass ollama pipeline
Which models, and why two passes
Utah teapot named the tension herself: “for better writing styles you need older models but for higher reasoning abilities you need newer models.” She is right, and that is exactly why MoreRight is a pipeline, not a prompt. Inverting an argument is a reasoning task — you have to find the load-bearing claims, negate them without strawmanning the structure, and re-read the evidence against the original conclusion. Rendering the result is a voice task — fluent, period LessWrong prose. One model is asked to do neither job; each pass gets the model class that is good at its half.
Pass 1 — Invert (newer reasoning models)
- Default:
qwen3:14b— strong reasoning, ~9 GB at ollama's default 4-bit quantization (sizes approximate). Fits fully on the RTX 4080 laptop's 12 GB VRAM, and runs on a 780M iGPU ranch node with 32 GB shared RAM.ollama pull qwen3:14b - Fallback / explicit-reasoning option:
deepseek-r1:14b— the R1-Distill-Qwen-14B; its visible chain-of-thought makes inversion maps easy to audit.ollama pull deepseek-r1:14b - Small-node option:
qwen3:8bordeepseek-r1:8b(~5 GB) — the pick for 16 GB 780M nodes.ollama pull qwen3:8b - Bigger-laptop stretch:
qwen3:32b— better maps on long, dense posts; will spill past 12 GB VRAM into system RAM on the 4080 laptop, so expect it slower, not impossible.
Pass 2 — Render (older, fluent, period-voice models)
- Default:
llama3.1:8b— fluent, steerable, ~5 GB; the safe house style.ollama pull llama3.1:8b - Period-voice alternative:
mistral:7b— Mistral 7B (v0.3), an older-generation model with a drier, more blog-era cadence.ollama pull mistral - Deep-cut experiment:
llama2:13b— Llama 2 era prose; ollama's library notes 13B models want ~16 GB RAM. The 7B (llama2:7b, ~8 GB) is the small-node version. Funniest output, weakest instruction-following — which is part of the experiment.ollama pull llama2:13b
The process
- Fetch. Parse the post ID from the LessWrong URL and pull the post from LessWrong's GraphQL endpoint. On any failure, paste the text instead.
- Invert (Pass 1). The reasoning model returns an inversion map only: flipped title, flipped thesis, each load-bearing claim negated, the original's own evidence re-read against its conclusion, and the original structure preserved.
- Render (Pass 2). The style model rewrites the post from the map alone — same structure, opposite argument, fluent LessWrong-like voice.
- Label. The output is shown beside the original, in inverted colors, under the parody banner. The banner is part of the output, not decoration: it travels with the copy button.
Honest limits
- Long posts need chunking by section; this POC sends up to ~14,000 characters and tells you when it has truncated. A production version inverts section by section and stitches.
- Inversion quality is bounded by Pass 1. A weak map makes a confident, fluent, wrong-shaped mirror — Pass 2 cannot fix a bad map, it can only make it prettier.
- Satire of an argument is not a refutation of it. MoreRight output is a joke with a method, and the parody labelling is non-negotiable.
- Fetching is one post at a time, at human pace, from the public GraphQL endpoint.