Skip to content

Peter Steinberger — State of the Claw

Source index. 17-min keynote + ~25-min AMA with Swyx at AI Engineer 2026. Raw at peter-steinberger-state-of-the-claw-2026.

Structure

  • Keynote (17 min): growth metrics, security flood, openclaw-foundation, corporate contributors
  • AMA with Swyx: OpenAI relationship, local models, workflow, dark-factory debate, prompt injection, future features, taste/system-design/saying-no

Concepts introduced

  • agent-taste — the surviving engineering skill
  • agent-security-slop — AI-generated CVE flood and maintainer coping
  • soul-md — persona file as first-class artifact alongside AGENTS.md

Entities

Notable quotes

  • "Stripper-pole growth." (Project velocity curve)
  • "The higher they scream how critical it is, the more likely it's slop." (Security advisories)
  • "Taste is the lowest bar — you instantly know if something stinks like AI."
  • "The 10x engineer is no longer about words per minute — it's token-to-token usage." (Cross-echo to eric-zakariasson)
  • "Madness with a touch of science fiction." (Self-description)
  • "What's the worst that can happen? Pics are already online if you use Grindr." (Personal-risk framing)

Notable moments

  • GSHJP CVSS-10 advisory — critical score on a permission-model edge case that ~nobody hits in practice; illustrates CVSS gaming.
  • Nemo-Claw breakout — Peter's Codex Security agent broke Nvidia's hardened sandbox in 30 minutes using un-nerfed internal models.
  • Agents of Chaos paper critique — academic paper ignored documented security recommendations for narrative impact.
  • Dreaming feature — memory consolidation plug-in, confirmed Anthropic working on same (source leak).
  • Dark-factory pushback — "The way to the mountain is usually never a straight line" — disagrees with pure-waterfall-agent-execution model.
  • Ubiquitous agents vision — iPads in every room, agent follows you, glasses + earbuds as endpoints, Star-Trek-style "computer".

Open questions

  • Can the Foundation survive the BDFL effect? Peter's energy is the project's engine.
  • Does "taste as moat" hold as models improve? Or does taste also get learned?
  • What's the equilibrium between corporate contributors and hobbyist direction when every integration is maintained by one vendor's employee?
  • At what point does the home-agent use case (Karpathy, Marin Dre) demand real standards (security, interop) vs staying "maximum fun"?