Summary

Dan Shapiro traces the evolution of AI interfaces through four stages: completion, chat, agents, and 'claws' — autonomous agents that learn and act on your behalf without being asked. After an initial security disaster with OpenClaw and a failed attempt at running one in a locked-down sandbox, he found a middle path: giving his claw carefully scoped permissions with clear guardrails. The result was something qualitatively different from prior AI tools — an autonomous assistant that mines contacts, drafts emails, and does background research. He concludes by announcing plans to deploy claws across his company Glowforge, building what he calls a 'claw printer.'

Key Insight

The next frontier beyond AI agents is autonomous 'claws' that learn and act on your behalf without prompting — but making them useful requires carefully scoped permissions rather than full autonomy or full lockdown.

Spicy Quotes (click to share)

• 8

  Completion finishes your sentence. Chat discusses things with you. Agents work for you. Claws work without you.

• 5

  A claw is an agent that learns and acts on your behalf without waiting for you to ask.

• 5

  The more autonomy you give them, the more they get both helpful and dangerous. The reverse is also true.

• 8

  It was safe. It was also useless. I built an extra-slow ChatGPT with a Mac Mini for a toupee.

• 8

  I do not need new ways to send emails I regret. I have achieved sufficient coverage in that market.

• 5

  I've gone from building a claw, to building a claw printer. Everyone deserves their own, custom claw.

• 7

  I learned this the dumb way, which is my preferred curriculum.

Tone

witty, reflective, experiential