Claude Code ports C&C, Mechanical Turk fades, and Baidu's OCR scales up

Claude Code ported a 2003 RTS to iOS in hours, Amazon winds down Mechanical Turk, and Baidu's Unlimited OCR tops benchmarks with flat memory use.

The big picture

A mixed batch — one genuinely impressive agentic coding demo, one slow-motion sunset for a platform that trained half the early ML datasets, and a useful OCR research result buried under a pile of slower news.

Claude Code keeps racking up legitimate demos

A Google DeepMind developer used Anthropic’s Claude Code alongside the Fable 5 model to port Command & Conquer: Generals Zero Hour (the 2003 real-time strategy game) to native iOS. The first working build took 40 minutes, the full port took a few hours, and the source code is already on GitHub. The Decoder

This is the kind of demo that actually moves the needle. Porting legacy C++ to a mobile platform involves platform APIs, build systems, and a dozen small fires — the kind of tedious multi-step work that previous coding assistants fumbled badly. A 40-minute first build is legitimately fast for a human expert, let alone an AI-assisted session. The fact that the developer published the source means you can dig into what Claude got right and wrong, which is more valuable than any benchmark.

Separately, Simon Willison dropped sqlite-utils 4.0rc3 this weekend, noting he worked through a backlog of issues and PRs with a combination of Claude Fable 5 and GPT-5.5. The changelog ballooned enough that stable 4.0 got delayed — biggest additions are compound foreign key support and case-insensitive column matching throughout. Simon Willison’s Weblog

Two data points in one weekend of a developer shipping real code with these tools, not synthetic benchmarks. The compound foreign key support in sqlite-utils is a legitimately useful SQLite feature that’s been missing for years. Worth upgrading when stable lands.

Amazon quietly closes the Mechanical Turk on-ramp

Amazon is stopping new customer signups for Mechanical Turk, its two-decade-old platform for paying humans to do small tasks that computers handle poorly. Existing users aren’t immediately cut off, but the door is closing. TechCrunch

This one deserves a moment. Mechanical Turk was the unglamorous backbone of early AI development — image labeling, sentiment annotation, training data curation. The humans on that platform did the work that made a huge chunk of modern ML possible, often for fractions of a cent per task. The irony that AI is apparently what’s killing it off is hard to miss. If you’re still using MTurk in any data pipeline, now is a good time to evaluate alternatives like Scale AI, Labelbox, or even synthetic data generation.

Baidu’s OCR research is worth a look

Baidu published research on what they’re calling “Unlimited OCR” — a document reading system that can process dozens of pages in a single pass. Previous approaches capped out around ten pages. The key technical move is a modified attention mechanism (attention controls which parts of the input a model focuses on) that keeps memory use flat regardless of page count. It currently holds the top position on a major OCR benchmark. The Decoder

Flat memory scaling on document processing is a real engineering problem, not a press release problem. If you’re building any pipeline that ingests dense PDFs — contracts, research papers, financial filings — watch for this to show up in production APIs. The “treating memory like human forgetting” framing in the title is marketing, but the underlying benchmark result is concrete enough to take seriously.

Quick hits

Hugging Face dropped major updates to its Kernels library (GPU kernel management for ML inference), though the announcement post shipped with no actual content — check the Hugging Face blog directly once it populates.
Google ran a Workspace ad imagining the founding fathers using Gemini to draft the Declaration of Independence; Ben Franklin apparently used suggestion mode, which is the least historically plausible thing about it. The Verge
Wealthy families are paying companies like Alpha School and Forge Prep tens of thousands of dollars per year to use AI tutors instead of traditional schooling, with Silicon Valley VCs among early adopters — a niche trend, but an early signal for how AI edtech will be stress-tested before it hits public schools. The Verge