offbyone42

I feel like this misses how LLMs work.

Yes, you’re adding this layer of verification, but LLMs don’t think in ASTs or use formal logic.

They are statistical predictors, just predicting what the next token will be.

There is a reason they perform best with TS/PY and not Haskell. The difference in size of the code corpus for each language.

The premise behind this seems to ignore all of that.

I just wish product listings were clear and actually followed the specs.

I got extremely lucky. I bought a Kindle recently and decided to get one by BestBuy, just trying my luck to see if they had old stock. The firmware was 5.18.0.2 and I was able to jailbreak it.

If anyone is looking for a Kindle, try non-Amazon stores. Might have older stock with older and jailbreak-able firmware.

I'm working on a part 2 with a bit more detail for that skill! It just has to go through a couple more drafts.

I don't use AI for these. My friends edit/give feedback, so the cycle is slow.

I have a `SessionEnd` hook that summarizes the conversation with a bunch of metadata, but it doesn't always fire 'cause of my habits.

So, the `SessionStart` hook was meant to catch any conversation where the `SessionEnd` hook wasn't fired.

Edit: I do realize a batch job is better. That's what I do now.

How did you not get caught??!?

Meanwhile Claude Code starts with Bun and 400MB min instead of...I don't know, C? C++? Rust? Go? Hell, even Python?

I guess the findings of this paper make sense, but Claude has gotten into a habit of telling me to sleep when I'm pissed cause its being stupid.