Ilia's Corner

Ship an entire AI workflow in one sentence—no code, no $200/mo middleware

AutoAgent (https://github.com/HKUDS/AutoAgent) turns plain English into production-grade LLM pipelines. Need a multi-step agent that scrapes GitHub, summarizes issues, and emails the team every morning? One prompt, one file, zero glue code. You keep the keys to your own API accounts and run it anywhere Python runs—laptop, CI, serverless. The framework auto-generates retries, caching, and guardrails, so you look like a 10× engineer even if you just learned what LangChain is.

github_trending · 3 min read

Turn a $499 robo-mower into a centimeter-accurate lawn Roomba

OpenMower (https://github.com/ClemensElflein/OpenMower) rips out the toy-grade GPS and firmware from off-the-shelf mowers and bolts on RTK-GPS, ROS, and open-source autonomy. The result: military-grade row-by-row mowing on a weekend budget. You get the same 2 cm accuracy the pros charge $5 k for, plus a community hacking app that keeps your data (and your lawn) local. Perfect excuse to finally play with PID controllers and SLAM without buying a full-size tractor.

reddit · 4 min read

Copy-paste 200 battle-tested Vercel recipes and skip the yak shaving

Vercel’s own examples repo (https://github.com/vercel/examples) is the fastest way to bootstrap anything that needs auth, payments, AI, or edge functions. Each folder is a one-click deploy with real env-vars, TypeScript, and README that tells you exactly which $99/month SaaS you can now delete. Great cheat sheet for your next side-project or for impressing the boss with a working demo by lunch.

hackernews · 2 min read

Claude now ships actual files, not walls of text

Anthropic just upgraded Claude (https://www.anthropic.com/news/create-files) to output ready-to-use Excel, PowerPoint, PDF, and Google Docs. Ask for a quarterly KPI dashboard and you get the .xlsx—formulas, charts, and corporate blue included. Early users report cutting Friday report time from 3 h to 15 min. If you hate building the same slide deck every week, this is your new favorite intern.

hackernews · 3 min read

Go JSON v2: finally, no more silent data loss

The experimental encoding/json/v2 package (https://go.dev/blog/jsonv2-exp) fixes the 15-year-old landmines that swallowed your int64s and unknown fields. Explicit errors, struct tags that actually work, and streaming API that doesn’t allocate the GC to death. Opt-in today with Go 1.25 dev branch; your future self will thank you when the prod outage pager stays quiet.

backend

go · 115 stars

From-bash-to-go, part II: turn shell hacks into portable CLIs

This repo (https://github.com/go-monk/from-bash-to-go-part-ii) walks you through rewriting common Bash scripts (find, grep, curl pipelines) into single static binaries that run on Alpine, Windows, and your CTO’s Mac without coreutils drama. Perfect onboarding ramp for shell scripters who keep hearing “Go is faster” but don’t want to read the entire spec.

CLI

go · 44 stars

R-Zero: make any open LLM reason better—zero data needed

R-Zero (https://arxiv.org/abs/2508.05004) spawns two copies of the same 4–14 B model: a Challenger that writes problems and a Judge that scores answers. Repeat overnight and wake up to a model that beats GPT-4 on math benchmarks—no human labels, no prompt engineering religion. Code drops next week; prime your GPU box now.

AI self-training · HKUST & CUHK · 10 min read

mmBERT: one encoder, 1,800 languages, ModernBERT speed

mmBERT (https://huggingface.co/blog/mmbert) scales the lightning-fast ModernBERT architecture to low-resource tongues using only open corpora. If you build global search, support, or moderation tools, you can now swap the bloated multilingual BERT for a 131k-context model that actually fits on a T4. Benchmarks show +7 % F1 on Swahili and +11 % on Tagalog with zero extra data.

multilingual NLP · Hugging Face & Mila · 8 min read