frontpage.
newsnewestaskshowjobs

Made with ♥ by @iamnishanth

Open Source @Github

fp.

Open in hackernews

Show HN: Kore – Stack based language where compiler is the reward function

https://github.com/konf-dev/kore
1•processorx•1h ago
I built a stack-based programming language in Rust with Cranelift JIT, SPIR-V GPU backend.

Kore has ~140 tokens and 142 opcodes. Programs are sequences of stack operations — push values, apply functions, compose by concatenation. There's a proof checker that statically verifies stack effects and a capability system that gates side effects.

How I got here: I wanted to explore whether a compiler could serve as the reward function for training an LLM to write code. Most code-gen benchmarks use test suites or human eval. But if the compiler can verify correctness before execution — accept or reject in microseconds — you don't need a test harness at all.

That meant I needed two things: a language small enough for an LLM to learn the full vocabulary, and a security model that makes generated programs safe by construction.

The security model (P4: capabilities attenuate): Every side effect in Kore is gated by a capability — io for print/rand, fs for file access, exec for shell commands. The compiler auto-detects what a program needs and tags the binary. At runtime, you grant permissions explicitly (--allow io) or it won't run. When you spawn a sandboxed child, it gets parent_caps & requested_caps — it can never escalate. Pure programs (no capabilities) are deterministic and run on any backend.

This matters because an AI agent writing Kore tools gets compile-time proof that the tool can't touch the filesystem or exec shell commands. Not sandboxing — the proof checker rejects it statically.

I fine-tuned DeepSeek-R1-14B (QLoRA, Unsloth) on 478 curriculum tasks, it worked pretty well. Will probably try a MCTS+RL experiment to see if it can learn to write programs to solve games and optimization problems. The ultimate goal would be having it write its own compiler and self-host.

Flood Fill vs. The Magic Circle

https://www.robinsloan.com/winter-garden/magic-circle/
1•gyomu•4m ago•0 comments

Show HN: A CLI tool to automate Git workflows using AI agents

https://github.com/leochiu-a/git-pr-ai
1•leochiu-a•8m ago•0 comments

Use AI to find movies and TV shows on your streaming services

https://pickalready.com
1•hudgeon•9m ago•0 comments

Spec driven development doesn't work if you're too confused to write the spec

https://publish.obsidian.md/deontologician/Posts/Spec-driven+development+doesn%27t+work+if+you%27...
1•habitue•11m ago•0 comments

GenAI Go SDK for AI

https://50984e11.maruel-ca.pages.dev/post/genai-v0.1.0/
1•cpeterso•12m ago•0 comments

Show HN: I built an AI-powered late-night call-in radio show from my RV

https://lukeattheroost.com
1•lukemacneil•15m ago•0 comments

HeartMuLa: Open-source music foundation model achieving commercial-grade quality

https://heart-mula.com
1•hanruezz•15m ago•1 comments

An emotional app to figure out your next step

https://www.heyecho.app/
1•samxkoh•15m ago•0 comments

Show HN: I built a macOS tool for network engineers – it's called NetViews

https://www.bedpage.com/
2•n1sni•17m ago•0 comments

We chose a pipeline over speech-to-speech for evaluative voice AI

https://productfit.substack.com/p/why-speech-to-speech-apis-fail-when
1•niraj_kothawade•17m ago•0 comments

Show HN: BlazeMQ – 52KB Kafka-compatible broker in C++20, zero dependencies

https://github.com/awneesh123/Blaze-mq
1•awneeshtiwari•17m ago•0 comments

LLMs Refuse High-Cost Attacks but Stay Vulnerable to Cheap, Real-World Harm

https://expectedharm.github.io/
2•blackcat201•18m ago•0 comments

AI Doesn't Reduce Work–It Intensifies It

https://simonwillison.net/2026/Feb/9/ai-intensifies-work/
1•walterbell•18m ago•0 comments

Show HN: Decision Guardian – Surface past architectural decisions on GitHub PRs

https://decision-guardian.decispher.com/
1•iamalizaidi•22m ago•0 comments

Show HN: Open-source civic toolkit – 48 policies, 12 interactive tools, forkable

1•david_mchale•29m ago•0 comments

Can You Fly That Thing?

https://tomtunguz.com/can-you-fly-that-thing/
1•walterbell•30m ago•0 comments

Making a working intercom system from some old phone [video]

https://www.youtube.com/watch?v=mkJmT9kiu30
2•Refreeze5224•30m ago•0 comments

Trump says new US-Canada bridge won't open without Canadian concessions

https://www.freep.com/story/news/politics/2026/02/09/trump-says-gordie-howe-bridge-wont-open-with...
5•SilverElfin•31m ago•1 comments

Show HN: Multi-attribute decision frameworks for tech purchases

1•boundedreason•36m ago•0 comments

KiraStudio 1.0.0 – a lightweight, cross-platform music studio

https://kirastudio.org
4•ksymph•43m ago•0 comments

Can my SPARC server host a website?

https://rup12.net/posts/can-my-sparc-server-host-my-website/
4•pabs3•43m ago•0 comments

Report on the subject of Manufacturers (1791) [pdf]

https://constitution.org/2-Authors/ah/rpt_manufactures.pdf
4•pilingual•45m ago•0 comments

Explaining the PeV neutrino fluxes with quasiextremal primordial black holes

https://journals.aps.org/prl/accepted/10.1103/r793-p7ct
1•dataflow•45m ago•0 comments

Internet Background Noise

https://en.wikipedia.org/wiki/Internet_background_noise
2•tripdout•45m ago•0 comments

Show HN: Secure managed hosting for OpenClaw (free and BYOK)

https://openclaw-setup.me/
1•Gregoryy•48m ago•0 comments

Ask HN: How do you interpret P99 latency without being misled?

1•danelrfoster•51m ago•0 comments

Trapped Between Pitch, Disclaimer, and Confession

https://gilpignol.substack.com/p/trapped-between-pitch-disclaimer
1•light_triad•51m ago•0 comments

The Vocabulary Priming Confound in LLM Evaluation [pdf]

https://github.com/Palmerschallon/Dharma_Code/blob/main/paper/vocab_priming_confound.pdf
1•palmerschallon•54m ago•0 comments

The committee problem: why B2B demos die after the form

https://blog.skipup.ai/buying-committee-demo-scheduling-problem/
1•bushido•56m ago•0 comments

Show HN: Seedance2.live – One place to try many AI image and video models

https://seedance2.live
1•yuni_aigc•1h ago•1 comments