Mid-October

🔥

#AdventureX #AdventureX26 #hackathon

This is how you know Chimeling API is industry-standard.

We learn from the best, Clerk.

Left: Clerk Backend API
Right: Chimeling API

Join the waitlist 👉 https://chimeling.app

#Chimeling #Cairntale #api

A sneak peek 👀

#Chimeling #swift #macOS

It's a relief to remove the Tauri app from my monorepo.

Tauri builds lean, but development feels split. I'm always bridging TypeScript and Rust. That constant context-switching stops it from feeling like true web dev. When I pick a web-based framework for desktop apps, I want to write in TS and avoid extra friction, not double it.

#tauri #Chimeling #rust

Completely overhauled Chimeling's entire visual identity!

#Chimeling #Cairntale #design

Two models take a large amount of information in and return something very small and rigidly structured. Neither of them writes prose.

Shrug, mine, returns a single number, the probability that a reply is worth making. It does not decide what to say. It decides whether to say anything at all.

TypeSafe AI's Jev makes the same cut and lands at the opposite end. Where the two split is the part worth writing about.

https://aka.mrwillcom.com/shrug-post

#AI #Shrug #Jev #TypeSafeAI

Both start from the same diagnosis. What has been missing is not a better generator but the layer that decides whether to act, and that layer has to speak probability, since its consumer is code rather than a human.

Then they part. Shrug imitates one human deciding whether to answer. Jev aims at correct decisions, at scale. This is not a matter of quality but of a different target.

Two consequences follow from that difference.

The first is randomness. When you imitate a human, randomness is the fidelity, and it shows up as hesitation, omission, and impulse. The model should be able to stay quiet at p=0.9 and still blurt something out at p=0.1. When you imitate judgment, the same randomness is a defect to be measured and removed.

The second is silence. In Shrug it is half of the label space and also the majority class, so it cannot be forgotten. In Jev the allowed answers are declared by the caller, so a way out is one candidate among the others. Whether it exists at all is settled at the call site.

Here is what I think the split actually is. It is not a matter of taste but of constraint.

I cannot build a general decision model, so I narrowed the problem until it was one I could finish, which was reply or not. Once it is that narrow, a binary classifier is not a compromise. It is the only shape that fits, and half of a binary is silence.

TypeSafe AI can go general. So the output space has to be declared by whoever calls it, and nothing inside that space gets a privileged lane.

Scope follows capability, and scope decides which philosophy is the only coherent one.

Neither position is portable. Ask why Shrug is not general and the honest answer is not that general is worse, but that general was never available to me. Ask the same of Jev and it flips.

That is the shape of the whole thing. What you imitate sets the limits, and the limits set the design. Two projects can share a diagnosis, share the same subtraction, and still insist on opposite things, because the thing each was built to imitate was never the same. So I do not expect this category to converge. I expect it to keep splitting along that line, into models that imitate a human and models that imitate judgment.