News for hackers. Stuff that matters.

gokhalh

Re:Transform your fund into something you can talk to on Thursday July 23, 2026 @01:19PM (#49025086)
This is sooo cool! So excited to see this!

Show HN: Humanize – Best static AI text humanizer, based on 50 research papers

https://github.com/harshaneel/humanize

I thought humanizing AI writing was straightforward. Just delete the em dashes, adjust the tone, and remove a few obvious words.However, as I "delved" into research on AI versus human writing, I discovered there was way more to it! Factors like sentence rhythm, repetition, hedge words, paragraph structure, punctuation habits, synonym cycling, and the "helpful assistant"...

( Read More... | no comments | 2 points | on HN )
Thanks, that's exactly the gap I was trying to fill.

On function-calling determinism: you're right that small-model JSON adherence drifts even at temp 0. localaik does not fully solve this. Two paths that help:

1. Constrained decoding. llama.cpp supports GBNF grammars for shape-strict generation. Passing a grammar via the upstream `grammar` field gives you JSON that parses and matches the schema by construction. It does not guarantee field VALUES (the model can still pick "wrong" entity names), but it eliminates the parsing-flake class of failures. Adding first-class support for this in the translation layer is on the roadmap.

2. Assertion strategy. Test the SHAPE of the tool call (was the right tool invoked? did it receive an arg named X?), not exact arg values. This is good practice for any LLM CI regardless of localaik. The "rerun on flake" anti-pattern is what I'm trying to avoid by being deterministic in the proxy layer, not the model layer.

GHA service container: glad it's useful. The 5-min cold-start budget (--health-retries 30) is the one tuning knob most folks miss. Steal away, that's why it's MIT.