Welcome to Signal AI. This is your weekly signal in the noise.

The Big Story: AI Goes Local

Llama 3.1 70B on a single RTX 3090 — Researchers demonstrated running a 70-billion parameter model on consumer hardware, bypassing the CPU entirely via NVMe-to-GPU direct memory access. This isn’t a prototype; it’s a glimpse at a future where AI inference lives on your desktop, not in the cloud.

Why it matters: Every dollar spent on API calls is a vote for cloud dependency. As local inference improves, we’ll see a bifurcation — heavy lifting in the cloud, everyday inference on-device.

Karpathy’s “Claws”: The New Agent Pattern

Andrej Karpathy flagged “Claws” as a new architectural layer for LLM agents — giving models persistent UI control beyond chat. Think: an AI that can click, scroll, and act rather than just respond.

The Hacker News discussion (600+ comments) shows the tension: some see it as the future of computing, others as a security nightmare waiting to happen.

Signal take: We’re heading toward AI that lives in your OS, not just your browser. Buckle up.

Europe’s Edge: AI on a Microcontroller

zclaw — a project running a personal AI assistant on an ESP32 microcontroller in under 888 KB. That’s smaller than most podcast cover images.

The developer hints at “something ridiculous coming soon.”

Quick Hits

• Claude Opus 4.6 dropped with hybrid reasoning (instant or extended thinking) — now competitive with o1 on coding benchmarks
• Anthropic raised $30B at $380B valuation — 10x annual revenue growth signals massive runway for model development
• Microsoft added MIDI 2.0 to Windows 11 — niche, but shows OS-level AI integration accelerating

Coming Wednesday

Education deep-dive: What is inference, actually? The difference between training and inference, and why it matters for your wallet.

Signal AI — European-focused AI intelligence. One coffee/month. Invite a friend, get 2 months free.

Reply to tell me what I missed.