Ilia's Corner

OpenAI drops an open-weights gift basket: 20B & 120B MoE models you can actually ship

Remember when "open" AI meant reading a blog post and praying for an API key? That era just ended. OpenAI quietly shipped GPT-OSS—fully Apache-2.0 licensed, 20B and 120B-parameter MoE models that run on a single H100 or even a 16 GB gaming laptop (https://github.com/huggingface/transformers/releases/tag/v4.55.0). AMD already has step-by-step guides for Ryzen AI laptops and Radeon cards (https://www.amd.com/en/blogs/2025/how-to-run-openai-gpt-oss-20b-120b-models-on-amd-ryzen-ai-radeon.html), so you can prototype at the coffee shop and deploy at the edge without talking to a cloud vendor ever again. Translation: no usage quotas, no surprise bills, no rate limits when your demo hits Product Hunt.

github.com · 4 min read

KittenTTS squeezes studio-grade speech into 25 MB—run it on your grandma’s laptop

If you’ve ever priced TTS by the million characters, KittenTTS feels like daylight robbery in reverse. The model packs ElevenLabs-level quality into a 25 MB binary that loads in <50 ms on any x86 CPU (https://github.com/KittenML/KittenTTS). Embed it in an Electron app, ship it on a USB stick, or spin up 1 000 container instances without sweating egress fees. Perfect for kiosk apps, indie games, or that side project you swore would never need a cloud budget.

hackernews · 3 min read

Claude Opus 4.1: the first AI that actually refactors like a senior dev

Benchmarks are cute—shipping without regressions is better. Claude Opus 4.1 scores 74.5 % on SWE-bench Verified, but the real win is surgical multi-file rewrites that don’t explode your build. If your team spends Fridays untangling 3-year-old JavaScript, let Claude open the PR and keep humans for code review (https://www.anthropic.com/news/claude-opus-4-1).

hackernews · 3 min read

Reality check: AI is not making you a 10× engineer (yet)

Colton’s contrarian take (https://colton.dev/blog/curing-your-ai-10x-engineer-imposter-syndrome/) is the pep talk every dev needs. Generative tools still save the most time on boilerplate and test data—not architecture decisions—so budget your sprints accordingly. Stop feeling guilty about "only" shaving 15 % off story time; that compounds to a paid-off tech-debt lane by Q4.

hackernews · 4 min read

GitHub UI getting sluggish? It’s not just you

Yoyo-code’s deep dive (https://yoyo-code.com/why-is-github-ui-getting-so-much-slower/) shows how incremental React bloat turned everyday actions into 5-second stutters. If your team lives in PRs, consider filing internal issues or switching to the CLI for bulk operations until GitHub ships fixes.

hackernews · 3 min read

llama.cpp now runs GPT-OSS in 4-bit float

A new PR (https://github.com/ggml-org/llama.cpp/pull/15091) adds MXFP4 support, letting you slot the shiny new OpenAI weights into llama.cpp and hit 70 tok/s on an RTX 4090. Perfect for weekend experiments when you’d rather not rent an A100 in the cloud.

Local LLM inference

C++ · 38 stars

World’s fastest VIN decoder: 30 ms lookups from 21 MB

Cardog squeezed a 1.5 GB government database into 21 MB using perfect-hash tables and some clever bit-packing (https://cardog.app/blog/corgi-vin-decoder). If your product deals with vehicle history reports, you can now do offline VIN decoding faster than most REST calls.

Automotive APIs

Rust · 39 stars

Error Detection & Correction for Math in LLMs

EDCIM (https://arxiv.org/abs/2508.03500) is a lightweight framework that retrofits smaller models with symbolic rule checks, giving you GPT-4-level accuracy on math-heavy tasks without the GPT-4 bill. If your startup lives or dies on correct calculations, this is free guardrails.

LLM reliability · D. Singh et al. · 5 min read

Evo-MCTS designs 1000× faster gravitational-wave detectors

No PhD in astrophysics required: this method (https://arxiv.org/abs/2508.03661) uses LLM-guided search to auto-evolve detection pipelines that outperform hand-tuned baselines. The takeaway for devs: evolutionary + LLM hybrids can solve optimization problems that used to take domain experts months.

AI+science · H. Zhao et al. · 6 min read