Tiny Brains, Big Gains: No-GIL Python, $300k OSS, and Europe’s Chip Power Play

October 14, 2025

What we covered this week: check out YouTube, listen wherever you get your podcast or read below 👇

Less is More: Recursive Reasoning with Tiny Networks | https://alexiajm.github.io/2025/09/29/tiny_recursive_models.html 

A tiny 7M‑parameter Tiny Recursion Model (TRM) iteratively updates a latent state and its own answer, showing 45% on ARC‑AGI‑1 and 8% on ARC‑AGI‑2. It challenges “bigger is better” by using recursion to extract more reasoning from small nets. If cheap, capable small models pan out, startups can build clever agents without big‑tech budgets.

Python 3.14 will change the way you parallelise code | https://valatka.dev/2025/10/11/on-python-3-14-parallelization.html 

Python 3.14’s free‑threading (no GIL) lets plain threads run CPU workloads in parallel, not just I/O. Expect wins for model inference, data work, and latency‑sensitive services; throughput isn’t everything. The release also adds parallel subinterpreters; together they could shift Python’s architecture/cost math more than any autoscaling tweak.

The dev behind lightGallery reports $300K+ via dual licensing: GPL/AGPL for open use, paid commercial for closed‑source. Strategy: contributor agreements, a clean major‑version switch for relicensing, and a few exclusive features to push upgrades. For businesses, the copyleft trade‑off makes buying easier; for maintainers, it’s a pragmatic path to sustainability.

In rare move, Dutch government takes control of China-owned chipmaker Nexperia | https://www.reuters.com/world/china/dutch-government-intervenes-chinese-owned-computer-chip-firm-nexperia-2025-10-12/ 

The Netherlands intervenes at Nexperia, citing governance failings and economic‑security risks. Using the rarely invoked Availability of Goods Act, ministers can suspend directors and reverse decisions deemed harmful. Wingtech’s Shanghai‑listed shares fell ~10%. A signal of Europe’s tougher tech‑sovereignty stance over critical chip assets.

Simon Willison coaxed Claude’s Code Interpreter to zip its public skills folder, then open‑sourced the prompts and scripts it revealed. The package includes Word/PowerPoint/Excel/PDF helpers and a pypdf form‑filler CLI—handy for power users, but it also prompts questions about transparency and opsec around hidden agent capabilities.

LSEG and Microsoft will let Copilot Studio agents (inside Microsoft 365) access licensed LSEG data via an LSEG‑managed MCP server using the Model Context Protocol. Promise: fewer brittle glue scripts, more governed connectivity to Workspace and Financial Analytics, with a phased rollout and a deep catalog of datasets.

Inside The Controversy Surrounding Strava’s New Run-Coaching App—And What We Thought After Trying It | https://www.womenshealthmag.com/fitness/a68152711/runna-app-controversy-review-i-tried-it/ 

After six weeks with Strava’s Runna, a tester liked tailored paces, reminders, and the clean UI. Amid TikTok critiques of aggressive mileage, Runna says experts design plans while AI monitors progress and suggests adjustments. Verdict: solid for experienced runners, but not a full replacement for a human coach—and beginners should dial back.

PostHog outlines pitfalls when applying AI to big, messy codebases: don’t treat legacy monoliths like toy projects; provide context and guardrails (.cursor/rules, specs, subagents); avoid tasks AI handles poorly. Expect more reviews and tests, and set operational guardrails—or you’ll ship more code and more bugs, faster.