Ilia's Corner

OpenSSH just shipped post-quantum crypto by default: rotate your keys this weekend

Every fresh SSH session you open today already negotiates a hybrid ML-KEM + X25519 key agreement. That means the packets leaving your laptop right now are immune to the quantum computer that doesn’t exist yet but could show up in 5–10 years. No config flags, no recompile—just upgrade to the new OpenSSH and run ssh-keygen -o -a 100 to re-encrypt your private keys with the new algorithms. Your future self (and your compliance team) will thank you.

hacker_news · 3 min read

GitHub folds into Microsoft CoreAI: what it means for your repos

The era of “independent GitHub” is officially over. Thomas Dohmke’s resignation triggers a full absorption into Microsoft’s CoreAI division. Expect tighter Copilot integration, faster Azure DevOps merges, and—if history repeats—new rate limits once the AI compute bills land. If you rely on Actions minutes or free private repos, start mirroring to a second forge (GitLab, Codeberg, Bitbucket) while the migration dust settles.

hackernews · 2 min read

Ford’s $30k electric pickup is a blueprint for cheaper EVs everywhere

Ford teased a mid-size truck due 2027 that drops 20 % of parts and 15 % of assembly time by using a modular skateboard platform. Translation: the same cost-slashing techniques—shared battery trays, swappable motor mounts, firmware-defined drivetrains—will trickle down to suppliers and open-source EV kits. If you’re hacking on CAN-bus controllers or building battery-management firmware, download Ford’s open specs when they land; they’ll become the new reference design.

reddit · 4 min read

Claude Code proves you don’t need another Electron monster

A single 8 MB binary drops Claude into your terminal exactly where vim, tmux, and grep already live. The demo shows scaffolding a Go microservice in 47 seconds without ever leaving the shell. If your team is tired of 500 MB “AI IDEs,” curl the binary, alias it to c, and regain half your RAM overnight.

hackernews · 3 min read

Chrome DevTools now flags CSS baseline support

Hover any property in DevTools and you’ll instantly see if text-wrap, color-mix, or @layer is Baseline (cross-browser ready), Limited, or Experimental. Ship with confidence or add the right fallback without tab-hunting on caniuse. Update Chrome and open DevTools → Elements → Styles to test drive it.

Frontend QA

DevTools · 2 stars

Ollama + GGUF: why your 20-billion-param model suddenly crashes

A new tensor type introduced by Mistral’s MoE layers hit Ollama’s parser and seg-faults on load. The fix—updating GGUF schema and re-quantizing—just landed on main. If you self-host LLMs, pin to the latest commit or re-pull your model files to avoid silent corruption.

Local AI

Go · 74 stars

AI compute jumped 15× in one year—here are the numbers

OpenAI burned 200,000 GPUs in 60 days to train GPT-5, and total compute is up fifteen-fold since GPT-4. If you’re budgeting cloud spend, plan for $0.05-$0.08 per 1k tokens on frontier models through 2026 unless you roll your own quantized stack.

Infrastructure · OpenAI · 5 min read

ThinkTuning teaches LLMs to self-reflect without distillation

Forget expensive chain-of-thought fine-tunes—this paper shows you can train reflection directly into smaller models using reinforcement learning on synthetic tasks. Drop the open-source code into your own RL pipeline to get 4× fewer hallucinations on math benchmarks without extra GPU memory.

Model Training · MIT, Google DeepMind · 8 min read