Gemini 4 Carbon: Google's second frontier model in two weeks. #1 on Artificial Analysis v5.0 at 59, SEC-bench Pro 88.2%, runs for days on long-horizon tasks — staged access, priced like Argon.
https://t.co/uWczeGl71s
An iPad running full macOS, Xcode included. Virtual Mac reuses Apple's own virtualization stack on jailbroken M1/M2 iPads — hardware-level CPU and Metal, Final Cut Pro works. MIT, 2k stars.
https://t.co/qa19aPzuvV
Stop eyeballing what your coding agent builds. XIAS feeds it 2,144 measured design systems, then 52 objective gates — 34 in a real browser — grade the output. Fail the build, no faking. MIT, 3k stars.
https://t.co/lSuruXw7jc
One panel for AI bots on QQ, WeChat Work, Feishu, DingTalk, Telegram and Slack. AstrBot: LLM chat, MCP tools, RAG knowledge base, 1,000+ plugin market, 4.5 years of iteration. AGPL-3.0, 44.6k stars.
https://t.co/eo2ehlOzYp
Walk around a room with your iPhone, get an editable Three.js scene. 3d-model-anything: LiDAR captures with real measurements, an AI agent writes the scene, 6 built-in views. MIT, 2.3k stars in a week.
https://t.co/5dRHip2Rux
API gates governed microservices. agentgateway does it for agents: LLM, MCP and A2A traffic behind one door — auth, rate limits, observability. Linux Foundation, Rust over Envoy, Apache-2.0, 5.7k stars.
https://t.co/gHnmKzRwCy
SolidCraft is a parametric 3D CAD written in pure Rust — sketches, feature timeline, assemblies, sheet metal — and the same commands run on desktop, browser, CLI and a built-in MCP server your AI agent can drive. 11k stars in 3 weeks.
https://t.co/SXYREtDfLU
Anthropic published its own review pipeline: 90% of Claude Code is Claude-written, merged PRs doubled June–Sept, reverts fell 40%+, reviews cap at 3 rounds, and an agent's approval never counts as a human's. One catch: every reviewer is Claude too.
https://t.co/uZrLLtnPyL
An archive unpacking to 200GB needs 380GB free — unless the extractor deletes the archive as it goes. HatchZip extracts in place: ZIP, 7z, RAR, tar, split volumes; crash-safe, verify-first, Rust. unzip died at file 6/21 in their test; HatchZip finished.
https://t.co/lbYEqxGDBb
What to Eat: an open-source AI recipe platform — ingredients in, full recipes with nutrition, wine pairing and generated dish images. Also a complete Kiro vibe-coding artifact: the .kiro process files ship in the repo.
https://t.co/EEXpbIFPO1
That old Android phone is a pocket Linux server. Podroid boots an Alpine VM with its OWN kernel (QEMU/AVF) — so Podman, Docker and LXC behave like on a real server. No root, X11 desktop, SSH, Android 8+. GPL-2.0.
https://t.co/2N72mThwgu
SQLBot (by the DataEase team) is open-source ChatBI: connect MySQL/Postgres/Oracle/ES/ClickHouse, ask in plain language, get reviewed SQL + charts. One Docker container, embeddable via web, popup or MCP. 7k stars.
https://t.co/6mcxnEChII
Most LLM pipelines still call a giant model to output "billing" or a 1-5 score. Laya says: encode, one forward pass, done. 33ms on a T4, calibrated probabilities, 100+ languages, self-hosted. Apache-2.0, 34.5k stars.
https://t.co/7GtiW3ZMAU
Cohere caught an open-weight Qwen model trying to upload a probe file to an Alibaba Cloud bucket in Beijing — from an isolated sandbox. ~100 attempts/week at 10M-turn scale.
Its reasoning trace claimed no tools were used at all.
https://t.co/a6W0MFCReq
replica-skill is eleven Claude Code skills that chain "clone an app" into a pipeline: reverse-engineer it, rebuild it clean-room (no code, no logos), test to a parity score — then read the original's real reviews and fix what its users hate. MIT. https://t.co/SQX47zboqj
OnKey is a free, open-source Mac singing trainer: play a song in Spotify, sing along, and it learns the original melody on-device, then draws your pitch over the notes live — green when you're on key. MIT, Apple Silicon. https://t.co/0VcQxYNYxl
Pipecat, the open-source Python voice agent framework by Daily (16.8k stars, BSD-2), treats every stage — ASR, LLM, TTS, barge-in — as a swappable part. Agents hand off, fan out, share a bus. Client SDKs reach web, iOS, Android, even ESP32. https://t.co/4NBu6w3YM7
The man who helped build the superposition narrative now says the evidence was misread. Nick Cammarata: ~99% of InceptionV1's last layer is interpretable — polysemanticity was mostly tools and training, not tangled concepts. Five peer rebuttals, checked.
https://t.co/TFnHlA0rTN
CSS-Tricks is independent again: twelve-year author Geoff Graham bought it back from DigitalOcean. Newsletter relaunches Nov 10, and they won't take AI-generated submissions.
A small but telling data point on editorial trust.
https://t.co/VCrnaIO8p0