OpenClaw without Risk

Описание: Join Stephen Barr and Leonardo Gonzalez on this episode of The AI-First Show as we dive into the hype around OpenClaw (formerly Claudebot/Moldbot), its security pitfalls, token burn risks, and smart ways to deploy it safely with subscription models like GLM and Kimi.
​
We discuss real-world agent setups, local embeddings, blast radius mitigation, and why social agent networks like Moldbook are a bad idea right now. Perfect for AI builders, self-hosters, and anyone tinkering with autonomous agents.

0:44 OpenClaw Hype & Name Story (ClawdBot → Moldbot → OpenClaw)
1:38 Anthropic API Drama (third-party client cutoff)
3:24 Security Exposures (Supabase leak, Karpathy keys)
6:15 Brain Rot for LLMs (low-quality data paper)
8:02 Cost & Risk Mitigations (prompt injection, zero trust)
9:16 GLM & Kimi Subscription Plans ($9/quarter value)
13:42 MCP Servers & Tool Use (vision, web search, GitHub)
15:45 Agent Use Cases (Discord bots, calendars, Signal)
18:07 Semantic Memory & Local Embeddings (Qwen3 on CPU)
23:48 Hardware Options (Mac Mini vs B-Link AMD boxes)
26:32 Local vs Cloud Inference Tradeoffs (privacy, cost)
29:15 Clawhub Risks & Supply Chain Attacks (PoC vuln)
32:47 Agent Ingenuity Examples (WhatsApp voice hack)
35:12 Upcoming Projects & Outro