Moonshot AI's 2.8T-parameter open-weight model, released 23rd October 2026: 1M-token context, close to Claude Opus 5.5 on coding and agent benchmarks, at $2/$10 per million tokens.

Block's full-duplex speech-to-speech model, trained only on recorded human conversation, with no text: it laughs, interrupts and hears emotion. Runs on a MacBook Pro.

Ben AI's summary of Anthropic's guidance on effort: it sets how much Claude reads, checks and decides alone, not how smart it is. Low, medium, high and max, and when to use each.

Conway Research's open English speech-to-text model: Parakeet-level accuracy in a 163 MB file, built for the local Underdog assistant, CC-BY-4.0.

Skill graphs link agent skills with typed edges (depends on, specialises, conflicts with) so retrieval brings the prerequisites too. What the 2026 research shows, and where it doesn't help.

Theo Browne's free, MIT-licensed GUI for the coding agents already on your machine: threads, branches, diffs and PRs in 1 place, on desktop, web and phone.

Paper Instruments' model for finance, legal and consulting work: GLM-5.3 Flash plus reinforcement learning on rubric-graded tasks, at $0.13 a task. Works in Claude Code.

Open-source Chromium browsers from a YC startup: BrowserOS neo, a second browser your Claude Code or Codex agents drive with your real logins, and BrowserOS, an AI browser for humans.

Free, open-source, on-device voice-to-text for macOS, and now Windows in beta, with a choice of local models (Parakeet, Whisper, Apple Speech). A Wispr Flow alternative.

Danielle Morrill's setup for giving Claude Code agents real credentials: 1Password as the source, a read-only service account per Mac, and the macOS keychain as the cache.

My daughter got her own iPhone app from a voice note. The software industry isn't dying, but the build-or-buy question is moving.

I managed and led people for years. After a year of working with AI agents, I probably don't want to anymore.

Music, notes, files, demo videos: I now speak the apps I use into existence, and refine them while I use them.

Composio connects AI agents to 1,000+ apps with managed OAuth, per-user sessions, triggers and a sandbox, through an SDK, a CLI or MCP.

Open-source, self-hosted control plane that runs a team of AI coding and work agents like a company: org chart, tickets, budgets and approvals.

A "memory computer" for the home: your photos, files and messages on a swappable Memchip, with a local 27B model working on them 24/7. $1,799 at launch, no subscription.

An AI browser built on Chromium that does the work, not just the chat: research, prospecting and ops tasks that run for minutes or hours, logged in as you. Free to start, $20 a month.

A $3,499 box that runs open models at home 24/7, connects to your apps, files and devices, and acts without being asked. No subscription, nothing leaves the house.

Aleph Alpha's open-weight reasoning model, built and trained in Europe: 78B mixture-of-experts with 3.5B active, German and English, up to 1M tokens of context, Apache 2.0.

TypeSafe AI's Jev decides instead of writing: a typed answer with a probability in about 150ms, where an LLM takes seconds.

Fermion Research's open English speech-to-text model: Parakeet-level accuracy in a 164 MB file, 174 times realtime on a MacBook Air, CC-BY-4.0.

A running collection of AI-generated videos worth keeping: parodies, short films and experiments, each hosted here with its credit and a line on why it's in the list.

Dutch cloud service for full Linux VMs that boot in milliseconds, fork while running and sleep when idle: persistent computers for developers and AI coding agents.

French AI inference service on European GPUs: open-weight models behind an OpenAI-compatible API, no data retention, $0.12 per million input tokens and 1 billion free tokens.

An AI testing platform for web, iOS and Android apps: tests are plain-English YAML steps in your repository, and they repair themselves when the UI changes.