Ilia's Corner

China’s 17 MW floating turbine just rewrote offshore wind economics

Think offshore wind stops where the continental shelf drops? China’s first domestically-built 17 MW floater is already powering 40,000 homes from waters deeper than 50 m—no foundations, no land footprint, just a 260-m rotor spinning above the abyss. For anyone building energy-hungry data centers, edge grids, or island micro-clusters, this is a plug-and-play blueprint for gigawatt-scale power delivered far beyond cable reach. If your roadmap includes hydrogen plants, AI training farms, or desalination, bookmark the specs: the same floater arrays can be moored, plugged, and cash-flowing in 18 months instead of the decade a traditional wind farm needs.

Sustainability Times · 4 min read

Cognitive Load: the invisible bug in every architecture decision

Every `if` statement, every micro-service, and every cleverly-named variable quietly taxes your team’s collective RAM. This GitHub repo translates cognitive-load theory into a checklist you can run on pull requests—spot the modules that will burn out juniors before version 2.0. If your sprint velocity mysteriously flat-lines while story points balloon, run the audit here before you hire another dev.

hackernews · 6 min read

Andrew Ng: stop optimizing code, start optimizing decisions

The new constraint in AI startups isn’t how fast you ship features—it’s how fast you decide which features matter. Ng’s takeaway: LLMs slash coding time, so the bottleneck moves upstream to product management. Translation: if your team is still debating ticket priority in Jira while competitors iterate daily with GPT-4, you’ve already lost. Re-org for rapid hypothesis testing or become the next cautionary case study.

hackernews · 3 min read

Why your AI model needs its own virtual machine

Running LLMs inside generic Linux containers is like running Photoshop on DOS. This Sigplan post argues for an AI-native VM layer that exposes tensor registers, memory-tile locality, and sparse-attention kernels directly to the model—cutting inference latency by up to 60 %. If you’re shipping on-prem GPU clusters or edge TPUs, this is the missing HAL between your model and bare metal.

hackernews · 7 min read

The $3.1 trillion ETF pile now permanently props up NVIDIA, Microsoft, Broadcom

Index funds have quietly parked $3.1 trillion in the top ten US stocks, turning mega-caps into yield-generating utilities. For developers, that means steady R&D budgets at the AI chip kings and long-term job security for anyone who can speak CUDA, ROCm, or Triton. If you’re negotiating stock compensation, remember: your RSUs now ride on ETF inflows as much as product revenue.

hackernews · 2 min read

Red: ship cross-platform GUI apps from a 1 MB script

Forget Electron bloat and cloud build pipelines. Write once in Red, get native Windows, macOS, Android, and iOS binaries from a single 1 MB file—no runtime, no royalties, no kidding.

cross-platform desktop & mobile apps

Red · 36 stars

Plebbit: P2P Reddit that even China can’t block

A Reddit clone that runs over IPFS/pubsub with no servers, no mods, and no takedowns. Spin up your own “subplebbit” in five minutes; users fetch posts from whoever is closest, making censorship impossible and latency minimal.

decentralized social media

JavaScript · 18 stars

Spectrum: catch Clojure spec violations before production

Turn clojure.spec annotations into compile-time guarantees. Spectrum piggybacks on the compiler to surface nil-pointer bugs and malformed data before your REPL even restarts.

static analysis

Clojure · 11 stars

China’s 1 MV “electricity highway” moves gigawatts 1,500 km with 75 % less loss

Ultra-high-voltage lines at 1,000 kV AC / 800 kV DC are no longer experimental—they’re a working grid moving hydro power from Tibet to Shanghai in real time. If you run compute in regions with stranded renewables, start lobbying for a UHV intertie; it’s cheaper than batteries.

energy transmission · Our China Story · 5 min read

Intermittent fasting linked to 135 % spike in fatal heart events

A 19,000-person study found that compressing daily eating into an 8-hour window correlates with dramatically higher cardiovascular risk. The takeaway for tech teams: skipping breakfast to ship code faster may literally kill you.

health · BBC · 3 min read