Ilia's Corner

NPM debug and chalk packages compromised

If you deployed any Node app in the last 30 days you may be serving malware right now. Attackers slipped a back-door into chalk, debug, ansi-regex and 15 other ultra-popular packages that sit *everywhere* in your dependency tree. The poisoned versions look normal in CI logs but steal credentials and environment variables at runtime. One-liner to check your lockfile and the exact clean versions are inside. Patch today—your next npm install is too late. https://www.aikido.dev/blog/npm-debug-and-chalk-packages-compromised

aikido.dev · 3 min read

Immich – High-performance self-hosted photo and video management solution

Drop Google Photos without losing AI magic. Immich gives you face recognition, duplicate detection and mobile auto-backup on your own server (or a $5 VPS). All metadata stays yours, storage scales to whatever disk you attach, and the new CLI can bulk-import a 100 GB Lightroom catalog overnight. Great excuse to dust off that Raspberry Pi. https://github.com/immich-app/immich

hackernews · 2 min read

Experimenting with Local LLMs on macOS

Skip the API bills—run a 30 B-parameter model on your MacBook Air. This walkthrough shows how llama.cpp squeezes quantized models into 12 GB RAM so you can prototype RAG pipelines or private coding assistants with zero network latency. Includes ready-to-copy launch scripts and a one-line Swift shortcut that turns any selected text into a GPT-4-class rewrite while you stay completely offline. https://blog.6nok.org/experimenting-with-local-llms-on-macos/

hackernews · 4 min read

ASML becomes main shareholder in Mistral AI after €1.3 B deal

The only company that can build €150 M EUV lithography machines now owns 11 % of Europe’s flagship LLM house. Expect AI co-processors baked directly into the scanners that etch every 2 nm chip, slashing mask-set iteration from weeks to hours. For developers that means tighter feedback loops between silicon design and AI model training—your next hardware-aware quantization could ship inside the fab itself. https://www.techzine.eu/news/infrastructure/134395/asml-becomes-main-shareholder-in-mistral-ai-1-3b-euro-deal/

reddit · 3 min read

Poor man’s FlashAttention: Llama.cpp-gfx906 fork

Breathe new life into that forgotten AMD MI50 in the office closet. This fork grafts hand-tuned assembly kernels onto llama.cpp so a single 300 W GPU can serve 30 B-parameter models at 30 tokens/s—no cloud subscription, no Nvidia tax. Perfect for internal code-gen or private chatbots under a desk. https://github.com/iacopPBK/llama.cpp-gfx906

Local LLM inference on legacy AMD GPUs

C++ · 1.5 stars

ROCm 7 nightly for Ryzen AI laptops

Turn a Ryzen AI 395+ “Strix Halo” notebook into a portable ROCm workstation. One script Dockerizes Unsloth, bitsandbytes and llama-cpp so you can fine-tune 8 B models on the plane seat-back tray—no external GPU, no EC2 bill. https://github.com/shantur/strix-rocm-all

Mobile AMD AI dev stack

Shell · 0.3 stars

TableMind: AI that writes its own Python to reason over messy spreadsheets

Stop wrestling with pandas. TableMind ingests any CSV or SQL dump, writes exploratory code, runs it, debugs tracebacks and returns a clean notebook with charts and insights. Benchmarks show it beats GPT-4 + Code-Interpreter on real-world business questions by 18 %. https://arxiv.org/abs/2509.06278

Tool-augmented LLM agents · CUHK & Tencent · 5 min read

REMI: Long-term causal memory for lifestyle recommendation agents

REMI doesn’t just remember your step count—it builds a private causal graph linking late-night commits to next-day bug rates and coffee intake to pull-request merge time. Early demo shows 2× improvement over rule-based health bots for shift-work engineers. https://arxiv.org/abs/2509.06269

Personalized AI systems · HKUST & Huawei · 4 min read