Hacker News: 44za12

New comment by 44za12 in "Qwen3.5 Fine-Tuning Guide"

44za12 — Wed, 04 Mar 2026 16:47:24 +0000

Specialised models easily beat SOTA, case in point: https://nehmeailabs.com/flashcheck

New comment by 44za12 in "Number Research Inc"

44za12 — Wed, 04 Mar 2026 05:14:34 +0000

All of us use the same keyboards more or less, maybe us randomly typing a large number is not as random as we would like to think. Just like how “asdf”, “xcyb” are common strings because these keys are together, there has to be some pattern here as well.

In Defense of Boring Technology

44za12 — Tue, 17 Feb 2026 06:43:27 +0000

Article URL: https://aazar.me/posts/in-defense-of-boring-technology

Comments URL: https://news.ycombinator.com/item?id=47044412

Points: 1

# Comments: 0

Show HN: RightSize CLI, Find the cheapest LLM that works for your prompt

44za12 — Wed, 28 Jan 2026 10:50:41 +0000

RightSize runs your prompt against candidate models (Kimi, GLM, Qwen, Gemma etc.) in parallel via OpenRouter. Then it uses a stronger model as a Judge to score accuracy against a baseline model.

Happy to answer questions.

Comments URL: https://news.ycombinator.com/item?id=46793694

Points: 3

# Comments: 0

New comment by 44za12 in "Tell HN: I cut Claude API costs from $70/month to pennies"

44za12 — Mon, 26 Jan 2026 13:24:03 +0000

Yes, I included a 'Model Selection Cheat Sheet' in the README (scroll down a bit).

I map them by task type:

Tiny (<3B): Gemma 3 1B (could try 4B as well), Phi-4-mini (Good for classification). Small (8B-17B): Qwen 3 8B, Llama 4 Scout (Good for RAG/Extraction). Frontier: GPT-5, Llama 4 Maverick, GLM, Kimi

Is that what you meant?

Show HN: LLM Sanity Checks – A practical guide to not over-engineering AI

44za12 — Mon, 26 Jan 2026 12:49:12 +0000

I keep seeing teams use frontier models for tasks a regex or a 4B model could do cheaper and faster.

This repo is a collection of opinionated patterns and heuristics to help you rethink the architecture of your AI workflows and ease the decision-making process while ensuring maximum efficiency.

It covers:

- A decision tree for architectural sanity checks. - Tradeoffs between JSON and delimiter-separated output. - Patterns for cascading models (verifying small models before calling big ones).

Open to feedback on other anti-patterns you've seen in production.

Comments URL: https://news.ycombinator.com/item?id=46765011

Points: 1

# Comments: 0

New comment by 44za12 in "Tell HN: I cut Claude API costs from $70/month to pennies"

44za12 — Mon, 26 Jan 2026 07:03:53 +0000

This is the way. I actually mapped out the decision tree for this exact process and more here:

https://github.com/NehmeAILabs/llm-sanity-checks

New comment by 44za12 in "Stop using JSON for LLM structured output"

44za12 — Sat, 24 Jan 2026 07:46:53 +0000

For simple extraction tasks, a delimiter-separated string uses 11 tokens vs 35 for JSON. Output tokens are the latency bottleneck.

Stop using JSON for LLM structured output

44za12 — Sat, 24 Jan 2026 07:46:52 +0000

Article URL: https://nehmeailabs.com/post/structured-output-overhead

Comments URL: https://news.ycombinator.com/item?id=46741819

Points: 2

# Comments: 1

FlashCheck-270M: Open weights for fact verification (Apache 2.0, WASM Demo)

44za12 — Mon, 15 Dec 2025 19:02:08 +0000

Article URL: https://huggingface.co/nehmeailabs-org/nehme-flashcheck-270m

Comments URL: https://news.ycombinator.com/item?id=46278797

Points: 2

# Comments: 0

New comment by 44za12 in "De-Googling TOTP Authenticator Codes"

44za12 — Mon, 01 Sep 2025 12:35:45 +0000

Love the minimalism.

New comment by 44za12 in "De-Googling TOTP Authenticator Codes"

44za12 — Mon, 01 Sep 2025 10:21:24 +0000

Shameless plug.

I’ve been using a cli tool i had created for over 2 years now, it just works. I had more ideas but never got to incorporate those.

https://github.com/44za12/horcrux

New comment by 44za12 in "[dead]"

44za12 — Sat, 30 Aug 2025 07:16:41 +0000

Have been using remove.bg for this for years now.

New comment by 44za12 in "Show HN: Rust macro utility for batching expensive async operations"

44za12 — Sun, 17 Aug 2025 15:03:17 +0000

Like a sempahore?

New comment by 44za12 in "Gemma 3 270M: Compact model for hyper-efficient AI"

44za12 — Thu, 14 Aug 2025 16:55:38 +0000

I’ve had great luck with all gemma 3 variants, on certain tasks it the 27B quantized version has worked as well as 2.5 flash. Can’t wait to get my hands dirty with this one.

New comment by 44za12 in "Benchmarking GPT-5 on 400 real-world code reviews"

44za12 — Fri, 08 Aug 2025 06:20:24 +0000

Can you benchmark Kimi K2 and GLM 4.5 as well? Would be interesting to see where they land.

New comment by 44za12 in "Vibechart"

44za12 — Thu, 07 Aug 2025 22:07:50 +0000

That was quick, vibe coded, I presume?

New comment by 44za12 in "C10kday"

44za12 — Tue, 05 Aug 2025 12:54:34 +0000

The ability to submit a story using a curl would have been fun.

New comment by 44za12 in "My bytecode optimizer beats Copilot by 2x"

44za12 — Sun, 03 Aug 2025 20:45:24 +0000

Tried that it’s taking exactly as much time as my program.

New comment by 44za12 in "My bytecode optimizer beats Copilot by 2x"

44za12 — Sun, 03 Aug 2025 17:26:45 +0000

I have breach parser that i had written to parse through over 3 billion rows of compressed data (by parsing i simply mean searching for a particular substring), I’ve tried multiple LLMs to make it faster (currently it does so in <45 seconds on an M3 pro mac) none have been able to do that yet.

https://github.com/44za12/breach-parse-rs

Feel free to drop ideas if any.