Matt.W | AI

thats insane.. Stanford just dropped a paper about loop engineering and self-improving agents. act -> reward your own rollouts -> train on them -> repeat let an agent train on its own reasoning with no hard check, and the failures compound: it gets more confident and less correct at the same time. that's the difference between a loop that compounds work and one that compounds error. read the paper first, then the article below.

how a model's /effort level works effort decides how long the model thinks before it answers, from low up to max. think of it like asking a person a question, low effort is the quick answer off the top of their head, high effort is them sitting down and working it out step by step before they reply the 4 effort levels: > low: fast and cheap, obvious stuff, sorting a lead or fixing a typo > medium: the balanced middle, a post from a brief, a standard email, summarizing a call > high: careful and thorough, a competitor teardown or planning a campaign > xhigh/max: the heavy one, slow and pricey, the gnarly problem high could not crack, untangling a messy funnel or a full launch plan more effort stops paying for itself past a point. on stet's public benchmark of 26 gpt-5.5 coding tasks, high roughly tripled the review quality of low, but going up to xhigh cost over 2x more for a gain that rarely earns it back. Anthropic says the same in their own effort docs, the sweet spot is usually high, as max can tip into overthinking. so default to high, or whatever your tool's baseline is the model you pick matters just as much, the same docs show Fable 5 on lower effort often beats older models even at xhigh this is effort routing, the same skill as choosing the right model for a job, one level down the setting is buried and named differently in every tool, high by default on Claude, medium on Codex, max off by default in a lot of apps. so you guess, or leave it on one level for everything you set it with a /effort command in claude code and the codex TUI, a flag like codex -e high, a line in your config, or the effort parameter on the API

An asian guy just found a way to use Fable 5 forever, even though Anthropic is going to make it access-limited after 12th July. He used Opus 4.8 to host the operating manual of fable 5 with just one single prompt. Access to the best model is never guaranteed. It disappeared once already this year. But you can use it forever. Here's how: 1. Extract the operating manual of Fable 5 2. Create an .md file and save it on your device 3. Create a new project "Fable 5 Brain" in Claude 4. Paste the instructions/ .md file straight into it 5. Run a pressure test to verify Save and bookmark this no matter what Full extraction prompt and practical guide are in the article below: ↓

I gave Fable one /goal prompt, and came back to a business. And it wasn't as expensive as you think because of the way I had Fable orchestrating a bunch of cheaper workers. Give this a 4 min read.