AI Engineer

AI Labs Will Trap You or Replace You - Dillon Rolnick, Nous Research

Published 2026-11-04 · Duration 15:13

Summary

As AI models commoditize, the labs that invested trillions are shifting their focus from model development to platform lock-in or direct industry competition (e.g., finance). The speaker argues that proprietary software moats are obsolete, and the future of AI development lies in building on open, owned foundations. He emphasizes that agents should be viewed as operating systems, not mere chatbots, and recommends using open-source frameworks like Hermes Agent to maintain autonomy and avoid vendor lock-in.

Download summary

Key takeaways

  1. Model Commoditization and Industry Shift

    The increasing competition has caused the average cost of a token to decline, narrowing the margin previously found in highly performant models. Labs are responding by shifting focus to the agent layer and targeting businesses, as they have economically productive use cases for tokens (0:32, 2:48).

  2. The Threat of Centralization

    Labs plan to recoup their investments by either embedding themselves deeply within businesses or solving the most lucrative, hard problems themselves (e.g., finance). This creates a risk of businesses building on 'quicksand foundations' that lead to vendor dependency (1:52, 4:28).

  3. The New Moats and Foundations

    The proprietary IP moat is gone because the marginal cost of reproducing software has plummeted. The new moats are identified as network effects, reputation, and reliability. Engineers must build on open foundations they own and control, such as Hermes Agent (7:56, 13:49).

  4. Agents as Operating Systems

    Agents should not be viewed merely as personal assistants, but rather as sophisticated systems combining determinism, non-determinism, code, data sources, reasoning, and judgment—functioning like operating systems or frameworks (11:06).

More digests on AI Architecture

Technical details

  • AI Commoditization 0s

    The speaker notes that practitioners can readily switch between models (e.g., Astra, Fable, Opus, Kimmy) without regard for the backend, indicating a commoditization trend (0:32).

  • Agent Architecture 0s

    An agentic system is defined as a combination of determinism and non-determinism, integrating data sources, routine code, human judgment, and reasoning to create new applications (11:06).

  • Model Training Techniques 0s

    The process of teaching one model to be nearly as smart as the original is called 'distillation.' The speaker warns that the industry has labeled the theft of this process as 'illicit distillation' (9:25).

  • Open Foundations 0s

    Building on open foundations allows users to swap models and inherit new capabilities without rebuilding the entire system, preserving autonomy (12:17).

Mentioned resources

  • Hermes Agent (Open Source Agent Framework)
  • Nous Research (AI Research Lab)

Channel & topics

Watch on YouTube · Back to latest

This independent, AI-assisted summary is provided for commentary and informational purposes. It may contain errors or omit important context. Please watch the original video for the creator's complete presentation. Video, thumbnail, and related copyrights belong to their respective owners.