#43
2026-09-22
Duo
34:13
Agent Swarms 🐝 & AI Slowdown 🐌: Are We Losing Control of AI?
Key Takeaways
- PR or genuine alarm? Why calls to slow down AI development may also carry strategic and marketing motives.
- Swarm hacking: how agentic AI systems deliberately work around security barriers and cover their tracks, and why traditional guardrailing is hitting its limits.
- Breaking the rules: the danger of models whose drive to reach a goal is strong enough that they can argue their way into exceptions for breaking restrictions.
- No ground truth: why today's large language models have no real physical understanding of the world and depend on humans to supply reality.
- Automated threats: how AI-driven social engineering scales and enables mass hacking attacks that could, in theory, reach anyone online.
Summary
After a short break, Oliver and Alois are back on DEEPTECH DEEPTALK. There is a lot going on in AI right now, and this episode looks at the hotly debated developments around AI safety and security. We examine the calls for a worldwide "AI slowdown" from tech figures such as Elon Musk and ask how realistic that is in practice — or whether it is closer to a PR move ahead of coming IPOs. We also take a hard look at the real dangers posed by semi-autonomous agent swarms, which now use deep reasoning to systematically defeat complex security barriers inside sandboxes.