Two stories set today's tone: a 17GB open-weights file doing near-frontier work on a laptop, and Stripe reportedly paying over $7B for OpenRouter, the metering layer between everyone else and those models. Around them, the plumbing of AI access is argued over in public — watermarks, published system prompts, a gray market in resold tokens, and the power and cooling bills underneath.
In this episode:
- Qwen 3.8 27B is excellent, and overthinks wildly by default
- Stripe reportedly buying OpenRouter for more than $7B
- Anthropic's published Claude system prompts hit the HN front page
- "A perversion of writing": Claude's text watermark takes fire
- A Trinidad embedded engineer answers "RISC-V: they should have known better"
- "LLMs can't jump": DeepMind on the missing abductive leap
- Who are the token brokers? Inside the AI credit resale economy
- Cloudflare injects its analytics beacon when you switch nameservers
- Ling 3.0 support merged into llama.cpp
- What it takes to run Qwen 3.8 27B at home
- Goldman pitches the AI productivity trade, away from infrastructure
- The 'AI stock god' had 56% of the fund in two names before it blew up
- Grimes County meets Terafab: the world's largest building as a chip plant
- Dario Amodei: AI's image problem is a crisis of trust
- Reticulum, the cryptography-first mesh stack, back on the front page
- Buf's Protobuf language server
- 800VDC pushes AI power design from the grid to the gate
- Coolant quality becomes a reliability signal for AI hardware
- Goldman and Morgan Stanley both call the Fed on hold
- Workday downgraded on valuation as a Silver Lake buyout looms
