👋 Hello hello,
❗Before diving in, this is for the business owners and entrepreneurs: we're building AI automation workflow guides: step-by-step, click-by-click, tool-by-tool so the repetitive work runs itself. Tell us what you'd automate.
Fable's been out a few weeks and it's already splitting the people we follow into two camps. One camp is still prompting it like it's last year's model, spelling out every step like they're talking to a nervous intern. The other figured out something different, and their demos look like they're cheating. This week two of them wrote up exactly how they do it, and their playbooks barely overlap. Also, a Stanford professor at MIT explained why human eyes still beat any AI model at reading a graph.
Let's get into it.
🔥🔥🔥 Three Brain Food Items
The trend right now is prompting Fable with a big, underspecified goal instead of a step-by-step script, the way you'd hand a project to someone you trust rather than someone you're micromanaging. Fence it in with a few house rules it can never break, then let it figure out the rest.
The part worth stealing is the grading trick. Whatever built something should never be the thing that judges whether it's done. Spin up a fresh instance with no memory of the build to check the output against a hard bar, then loop the original one against that bar until it actually clears it.
The counter-argument making the rounds is that Fable is good enough now that the real limit is how many "unknowns" you're carrying into a prompt without realizing it. It breaks down into four buckets: what you know you want, what you know you haven't figured out, what's so obvious you'd never think to write it down, and what you haven't even considered.
The fix is to make Fable surface those before you write the real prompt. Ask it to interview you, brainstorm a few wild prototypes with you, or run a "blind spot pass" on a part of the problem you don't know well.
A Stanford professor's MIT talk on human vision is going around, and it reframes something we take for granted. Humans have been sketching the invisible into something visible for tens of thousands of years, from cave walls to Feynman diagrams, constantly deciding what detail to include based on what needs communicating, without ever being taught the theory.
AI still can't fully replicate that instinct. Leading multimodal models can match human accuracy on reading a chart and still get there through completely different mistakes, which suggests they're not "reading" it the way we are.
Help shape Practically AI
Tell us what's working. Your feedback shapes what lands in your inbox, and it takes less than 2 minutes.
🔥🔥 Two Tools Worth Knowing
1. 💰 MDOTM's Sphere
MDOTM is an AI platform built for asset and wealth managers that automates the parts still stuck in spreadsheets, like rebalancing portfolios and drafting client commentary, while keeping a human sign-off on every call. It already runs over $100 billion in assets for 60-plus firms including Morgan Stanley, and just raised $27M to keep scaling. Best for investment teams, not solo traders.
2. 🏗️ Higharc
A homebuilding AI that turns a flat 2D floor plan into a full 3D model, complete with construction documents and real-time cost estimates, no CAD work required. It just partnered with US LBM, the largest private building materials distributor in the US, to bring the same tech to suppliers, backed by a fresh $95M raise. Best for homebuilders and dealers still relying on manual takeoffs.
🔥 Things You Didn't Know You Could Do With AI
This one is going to make all the Harry Potter fans happy. A creator on X open-sourced a build that turns a reMarkable Paper Pro into an working version of the enchanted Tom Riddle diary from Harry Potter.
Put a reMarkable Paper Pro into developer mode (the project's companion tool, remagic, does this and installs the launcher in one command).
Download the prebuilt release bundle and copy it onto the tablet's AppLoad folder over SSH. Here’s how it works:
❝pen (raw evdev, full 4096-level pressure, hardware event rate)
│ strokes
▼
riddle ── idle 2.8s → commit page → PNG ──► oracle (resident LLM process,
│ streams reply sentence-by-sentence)
▼ strokes (Dancing Script → skeletonized to single-pixel pen paths)
display backend
├── qtfb — windowed, inside xochitl (AppLoad app)
└── quill — full takeover: xochitl stopped, vendor e-ink engine
driven directly for instant ink (lowest latency there is)Drop in an API key for any OpenAI-compatible backend (OpenAI, OpenRouter, Groq, or a local server) or skip that step to run it through a different backend instead.
Open AppLoad on the tablet, reload, and launch "The Diary."
Write on the page with the stylus. After about three seconds of no movement, it commits the page as an image, sends it to the model, and the reply streams back stroke by stroke in a flowing hand before fading away.
Help shape Practically AI
Tell us what's working. Your feedback shapes what lands in your inbox, and it takes less than 2 minutes.
Don't forget to rate today's post
Until next time,
Team @PracticalyAI
