From Sloppy to Spooky

Two workflows to test Fable and Sol with

If you were looking for proof that the AI models are spooky now, you could rely on the opinion of the federal government. After all, they won’t even let you have access to Anthropic’s smartest model, and Fable took weeks of negotiations to be fully released. OpenAI’s latest whizpopper of a model, GPT-5.6 Sol, took 2 weeks of government review before being released to the general public. That probably means these models are pretty good!

However, as a red-blooded, Bass Pro Shops-loving American who would rather die than take advice from anyone but my wife, I’m telling you to ignore the government, ignore the AI influencers, and to especially ignore the marketing from companies like OpenAI and Anthropic. You gotta figure out if this is true for yourself.

What makes this generation of models so special is that all the software and capabilities surrounding them has gotten really good. They can now (semi)reliably operate a web browser. They can connect to almost every source of data in your life and then understand what it means. They can create spreadsheets, PowerPoint decks, or even a poetry robot. Using ChatGPT or Claude, you can build specialized skills and looping workflows that will automate a decent chunk of your busy work.

So today, I want you to prove to you that my last paragraph wasn’t hyperbolic drivel. I want to lay out 2 ways someone should actually test these out. Let’s get it.

Let’s Loop It Up

Here’s a way these new models have wowed me. For months now, I’ve had an AI agent scan whatever academic research was published in the last week, surface 40ish of the most relevant papers, and I’ll then flip through the abstracts to see if there is anything that is relevant to my interests (or to you, dear reader).

This has only sorta worked. On Opus 4.8 or GPT-5.5, the results were mediocre. The sweep ran fine, but the selection was the problem, and for a while I misdiagnosed it. I kept tinkering with the relevance criteria I’d written, and despite the tinkering, performance barely moved. The reason, I eventually realized, is that my edits could only ever encode my theory of what I like. “Prioritize novel causal claims about AI and labor markets” is me guessing at my own taste, and humans are famously bad at articulating their own taste. What the system actually needed was my revealed preferences: my in-the-moment reactions to specific papers, at volume.

So when Fable landed, I rebuilt it as a loop. Every Friday morning, the agent still sweeps the week’s new papers and hands me 30 to 40 abstracts. This time though, I use Wispr Flow to audibly narrate my thoughts as I read. “Too incremental.” “Poorly constructed dataset, no causal claim.” “Novel claim I enjoyed.” The narration isn’t a productivity flourish. It’s the data collection: dozens of unfiltered verdicts on specific papers, captured while my taste is actually operating instead of reconstructed from memory. Then, crucially, I ask Fable to adjust its own instructions based on what I said. The improvement genuinely surprised me. Given the latitude to maintain its own rules from that stream of verdicts, it have gotten meaningfully better at knowing what I consider interesting than some human editors I’ve worked with.

To be entirely fair, this idea of loops has been in the water for a while. What has changed with these new models is they crossed some invisible threshold of capability. Now, the results of these AI-report-to-human feedback loops actually hold up. It has gone from a cool demo to something that I can actually use and recommend to y’all.

Here is how to build yours in 20 minutes. I think your first loop should be a report, some recurring piece of analysis you already wish you had. An easy place to start is a weekly monitor of your market, your competitors, or new research in your field. (Yes, I’m aware I’m teaching you to build a micro competitor to my own newsletter.)

1. Open the ChatGPT desktop app or Claude. No connectors needed for this one; web search is enough.

2. Paste this charter, edited for your beat:

You run my weekly monitor loop. Create a file called

monitor-loop.md with 2 sections: RULES and LEARNED.

RULES (I’ll edit these; you won’t):

- Once a week, scan for [new research in my field / news

  on these competitors / developments in my market].

- Deliver 5-10 items. Each: 1-line summary, why it

  matters to me, link to the primary source.

- Number every item. I’ll reply with verdicts: KEEP,

  KILL, or EXPAND, plus my reasons.

LEARNED (you edit this; I won’t):

- After each cycle, compare my verdicts to your picks.

  Write what you learned as short rules. Apply them

  next cycle.

Run your first cycle now.

3. I would then run the cycle 3 times before judging. Cycle 1 will be mediocre, maybe even bad. That’s ok! Don’t quit. The point is that you can braindump on why it is bad, so the agent can adjust. Do this until the result is good enough.

You can think about loops as being extraordinary about making volume of thought, but not quality of it. For my search, Fable and Sol frequently miss what an idea implies. It’ll overindex on flashy numbers or get lost down a rabbit hole. Sometimes I feel like an intellectual panhandler, sifting through the mud for gold flakes of knowledge. For the last few years this was more a time-losing hobby than actual workflow. But now, finally, the models are good enough to make this possible.

Where this becomes really magical though is when you give the models access to private data. That’s when they go from useful to spooky. Here’s how.