Stratechery
10:01 PT
App Snore (This Week in Stratechery)
So what
App Store economics just reset in your favor, while Stripe–OpenRouter signals model access is becoming an aggregated, commoditized layer.
Apple settled with the EU and cut U.S. App Store fees, but Ben Thompson's read is that the App Store war he has covered for a decade now feels incidental beside the AI dealmaking reshaping the stack.
- Apple settles with the EU, lowers U.S. App Store fees, and adjusts ATT rules in Germany after years of regulatory pressure
- Stripe is reportedly acquiring OpenRouter, an implicit bet on a multi-model future and a shot at Aggregation
- Nvidia backs an OpenAI data center, another chip-vendor-to-frontier-lab deal; Anthropic's revenue growth "continues to amaze"
- Google buying Spirit Airlines data suggests data may finally be oil
- Sharp Tech asks whether DeepMind turnover is good news for Gemini; Sharp China covers Wang Yi in Seoul before Xi's September D.C. visit
Ben's Bites
07:08 PT
How I built this
So what
Your team's agent stack may be over-engineered: plain markdown files, git history and skills replace most of what memory vendors are selling.
Ben Tossell rebuilds his personal AI agent from scratch, showing it's just files, folders and instructions — and argues against elaborate auto-memory.
- Structure: AGENTS.md for instructions, plus code.md, todos.md and a memory.md pointer; CLAUDE.md holds only "read @AGENTS.md" since Claude ignores AGENTS.md
- Dropped the planned log.md after Codex noted git commit history already versions every change
- Rejected auto-saving memory: his old setup over-steered answers and shut down brainstorms, so he keeps small hand-edited files
- Past sessions are already searchable in dotfolders like ~/.claude/ and ~/.codex/; he floated an SQLite index of threads
- Named bots ("Emily" for email) are just chat sessions with instructions; repeatable jobs belong in skill.md skills
Superhuman
06:11 PT
Claude gets its own learning hub
So what
Dual Anthropic and OpenAI IPO clocks plus $3T in off-balance-sheet AI commitments reset how you price vendor risk and model lock-in.
Anthropic pulled 355 tutorials into a Claude Academy hub and is reportedly filing IPO paperwork this month, while OpenAI CFO Sarah Friar told staff its own listing waits until 2027.
- Claude Academy covers Claude, Cowork, Code, Tag and the API; Concise Mode ships for Claude Code
- OpenAI adds an Apple Messages plugin to ChatGPT Work and Codex, so users search and send iMessages in-app
- Reddit citations in ChatGPT fell 86% since mid-July, from ~4% to ~0.5% on Aug 14, per Promptwatch
- Alibaba Qwen topped 3B open-weight downloads in six months, ahead of Google and Meta
- Etched raised $700M at $21B, double its $10.3B July mark, after shipping its first rack to Jane Street
AI with Allie
07:15 PT
3 learnings you need from the AI Agent Workshop
So what
Framing AI as an operating system rather than a chatbot changes what non-engineer teams expect from your product's automation surface.
Allie K. Miller distills three takeaways from her AI agent workshop with Mark Cuban, arguing the people pulling ahead spent one afternoon building agent systems that now run without them.
- Treat AI as an operating system, not a chatbot: a chatbot waits for your next question, an OS runs your tools while you are elsewhere
- Work backwards from the problem and tell the AI your "why" — given the desired end state it can figure out the path
- If you can describe a task in plain English, you can have an agent doing it today, no technical background needed
- The bulk of the issue sells her $397 lifetime AI-First Academy: 52 chapters, six hours, alumni from Adobe, Cisco, PwC and Charles Schwab