samchon

Joined 1 year ago

Karma
3
Submissions
45
Member since
Mar 5
https://github.com/samchon
View on Hacker News

Recent submissions

Comment · 7 months ago
I was personally invited by the Qwen team to speak at Qwen Meetup Korea, and got to present locally here in Korea yesterday — pretty honored to have been reached out to directly.

The talk was about how I got function calling to work reliably on deeply recursive union types — the stuff the industry generally says doesn't work. With `qwen3-coder-next`, first-try success rate was 6.75%. And the entire Qwen 3.5 model family was hitting 0% on union types due to a consistent double-stringify bug. Both ended up at 100%.

Slides (PPT) are also available in the link — speaker notes are written inside as slide notes if you'd like the full narrative behind each slide.

## TL;DR

1. *AutoBe* — AI backend auto-generation agent. Not text code, but AST data via function calling. 4 AST types + 4-tier compiler validation + self-healing loops. 2. *Typia* — The infrastructure that turns 0% into 100%. A single type automates schema, parser, validator, and feedback generator. Lenient JSON parsing + type coercion + precise validation feedback. 3. *In Praise of Function Calling* — Types eliminate ambiguity. Schemas constrain through absence, not prohibition. Model-neutral, mechanically verifiable, deterministically convergent. Applicable to all engineering domains with validators. 4. *Qwen* — Small models are the best QA engineers. They expose system vulnerabilities large models silently paper over. 5. *6.75% is not failure — it's the first input to the loop.* If you can verify, you converge.

Comment · 7 months ago
Next subject: function calling for domains where the final answer is unverifiable.

Legal rulings, business decisions: you can't verify the conclusion. But you can verify the process. Model every mandatory reasoning step as typed schema properties, put the verdict last.

CoT becomes 100% structurally guaranteed. Streaming + property order validation catches the AI skipping steps in real time, mid-generation. DeepPartial<T> fills corrections progressively as output streams in.

If you can verify the process, you can harness any domain.

Comment · 7 months ago
Just allowing Yolo, and sometimes do rolling back
Comment · 11 months ago
Amazing

Live from the official Hacker News API

A modern reader for Hacker News · SvelteKit + Skeleton