Alex Kim

Builder, AI Engineer, Founder @ WotAI · Production AI on Claude - agents, MCP servers & Claude Code

[Alex Kim @AlexKim]

It only matters if your code branches on the number. If it does not, Jev and Haiku are the same model.

Full benchmark: https://t.co/MORxB2kS2s

The Claude Code hook that runs the check: https://t.co/Ty942MXQTv

[Alex Kim @AlexKim]

I ran 16 models to find out if TypeSafe's Jev was worth adding to my stack.

It came 10th on accuracy.

I wrote two headlines before one survived the data.

[Alex Kim @AlexKim]

Jev is the only one of the four honest models that answers in under a second. 455ms, fastest of all sixteen. Next fastest that admits uncertainty is Sonnet 5 at 3.7x.

That is the slot. A gate you can afford to run on everything.

[Alex Kim @AlexKim]

Fix for the collisions is one line in AGENTS.md naming which plugin owns the spine.

Full measurements, all six overlaps, and what only each one has:

https://t.co/5tedkBfh6H

[Alex Kim @AlexKim]

I have both Superpowers and Compound Engineering installed in one repo.

Six of their skills do the exact same job.

Nobody warns you about this.

[Alex Kim @AlexKim]

So the usual framing is backwards.

Superpowers: cheap to install, expensive every time it fires — which by its own rule is constantly.

Compound Engineering: free until you call it, then lighter per call.

[Alex Kim @AlexKim]

honest accounting: writing those catalogs took longer than editing the first several clips by hand would have

the payback is clip 12 costs what clip 11 cost, which was never true before

https://t.co/mBLdwvrPO7

[Alex Kim @AlexKim]

i let claude code edit a video for me

the only reason that doesn't terrify me is that it never got to decide anything

[Alex Kim @AlexKim]

so the work moved, it didn't disappear

i had to write the brief first: which moment gets which treatment, what a beat needs, how the layers stack

four layers of specification. that's the job now