Physical AI Arrives; Python Gets Lazy Imports; International Criminal Court Backs Open Source

Hi all, welcome to our week 46 overview 👋

Below what we discussed during the pod plus some other interesting links. Thanks for following and see y’all next week!

XPENG Shares Achievements in Physical AI Emergence: Unveils XPENG VLA 2.0, Robotaxi, Next-Gen IRON, and Flying Car | https://www.xpeng.com/news/019a56f54fe99a2a0a8d8a0282e402b7 

XPENG used its 2025 AI Day to pivot to “Physical AI,” unveiling VLA 2.0 (vision→action model), a Robotaxi program, next‑gen IRON humanoid, and a flying car. VLA 2.0 targets end‑to‑end deployment in vehicles/robots, trained on ~100M unannotated clips, backed by a 30k‑card cloud and a 72B base model; an on‑vehicle Ultra model runs on XPENG’s 2250‑TOPS Turing chip. Volkswagen is the launch customer; pioneer users by end‑Dec 2025, broader rollout from Q1 2026.

GitHub - topoteretes/cognee: Memory for AI Agents in 6 lines of code | https://github.com/topoteretes/cognee 

Cognee is an open‑source memory layer for AI agents, blending vector search with a graph DB to keep facts searchable and linked. It proposes modular ECL (Extract‑Cognify‑Load) pipelines as a RAG alternative, with quickstart (pip install cognee), a minimal async flow (add → cognify → memify → search), CLI helpers, and Apache‑2.0 licensing—aiming to give agents durable, self‑hosted memory in “6 lines of code.”

PEP 810 – Explicit lazy imports | https://peps.python.org/pep-0810/ 

Python will add explicit lazy imports, letting modules load on first use (lazy import … / lazy from … import …). The opt‑in feature targets Python 3.15, promises faster startup for CLIs (often 50–70% speedups) and lower memory in long‑running apps, while keeping standard imports unchanged. Migration aids (e.g., lazy_modules) and incremental adoption are supported for backward compatibility.

ICC to replace Microsoft Office with European open-source platform | https://dig.watch/updates/icc-to-replace-microsoft-office-with-european-open-source-platform 

The International Criminal Court will migrate from Microsoft Office to Open Desk, a German‑built, EU‑backed open‑source suite tied to Europe’s digital‑sovereignty push. Reporting cites Handelsblatt and EU coverage; the move reflects public‑sector interest in reducing dependence on U.S. cloud/office ecosystems and investing in homegrown platforms like Zendis’s Open Desk.

moonshotai/Kimi-K2-Thinking · Hugging Face | https://huggingface.co/moonshotai/Kimi-K2-Thinking 

Moonshot AI’s Kimi‑K2‑Thinking is an open “thinking” MoE model built for long‑horizon agentic workflows and tool use. The card cites 1T total parameters (~32B active), 256k‑token context, native INT4 support, and stable tool orchestration across 200–300 consecutive calls—positioning it for complex reasoning that interleaves planning, code, and tool chains.

Archive.today: FBI Demands Data from Provider Tucows | https://www.heise.de/en/news/Archive-today-FBI-Demands-Data-from-Provider-Tucows-11066346.html 

Heise reports a purported U.S. court order directing Canadian registrar Tucows to disclose customer/account/connection/payment data tied to Archive.today (archive.is / archive.ph). The outlet notes the PDF’s authenticity isn’t independently verified, raising questions about jurisdiction, the operator’s identity, and potential chilling effects on web archiving.

The Smol Training Playbook: The Secrets to Building World-Class LLMs | https://huggingface.co/spaces/HuggingFaceTB/smol-training-playbook 

Hugging Face’s Smol team released a practitioner‑focused, 200+ page playbook on training compact LLMs, distilling lessons from building SmolLM3. It covers when to train vs. adapt, data curation, loss/monitoring, infra and debugging, post‑training and reliability—rich with checklists, pitfalls, and trade‑offs to help teams ship small, strong models.

Former Meta employees launch Stream, a smart ring that takes voice notes and controls music | https://techcrunch.com/2025/11/05/former-meta-employees-launch-stream-a-smart-ring-that-takes-voice-notes-and-controls-music/ 

Sandbar, founded by ex‑Meta/CTRL‑Labs designers, unveiled Stream: a gesture‑activated smart ring for discreet voice notes and media control with an AI companion app. Preorders are $249–$299, shipping planned for Summer 2026; a Pro subscription follows a 3‑month trial ($10/mo). The startup raised $13M from True Ventures, Upfront, and Betaworks.

Wispr Flow | https://wisprflow.ai 

Wispr Flow pitches “effortless voice dictation”—claiming up to 4× faster than typing—plus AI commands, auto‑edits, and cross‑app control. It offers native apps for Mac and Windows and an iPhone app, with simple onboarding to try or download from the site.

Sam Altman / X post (status 1986514377470845007) | https://x.com/sama/status/1986514377470845007 

Sam Altman stated that OpenAI does not have or want government guarantees for its datacenters, arguing governments shouldn’t pick winners or bail out private firms. He suggested governments could build/own public compute whose upside accrues to the state, and clarified that prior loan‑guarantee discussions concerned semiconductor fabs, not OpenAI datacenters.

PhD training needs a reboot in an AI world | https://www.nature.com/articles/d41586-025-03572-w 

Nature essay argues PhD training must reboot for an AI‑first era: teach students to critique and collaborate with AI outputs, rethink assessment and supervision, and prepare for AI‑augmented research practice. It frames AI as reshaping core doctoral skills from literature review and analysis to writing and reproducibility.