Subtitle: It’s a feature of the architecture Summary: Hallucination in LLMs is not a data quality problem. It is not a training problem. It is not a problem you can solve with more [RLHF](https://en.wikipedia.org/wiki/Reinforcement_learning_from_human_feedbac), better filtering, or a larger context window. **It is a structural property of what these systems are optimized to do.** I have held this position for months, and the reaction is predictable: researchers working on retrieval augmentation, fine-tuning pipelines, and alignment techniques would prefer a more optimistic framing. I understand why. What has been missing from this argument is geometry. Intuition about objectives and architecture is necessary but not sufficient. We need to open the model and look at what is actually happening inside when a system produces a confident wrong answer. Not at the logits. Not at the attention patterns. At the internal trajectory of the representation itself, layer by layer, from input to output. That is what the work I am presenting here did.

Anurag Beniwal (Member of Technical Staff at ElevenLabs) breaks down the real-world challenges of building voice agents—from latency, transcription accuracy, and turn-taking to the tradeoffs between cascaded systems and end-to-end speech models. The conversation explores why production systems rely on “constellations” of models, how to design for non-technical users (especially in customer support), and why voice unlocks richer context—but introduces far more complexity than chat. Ultimately, it’s a deep dive into making voice AI practical, reliable, and usable at scale.

A deep dive into the practical limitations of agent protocols like MCP and A2A for low-level tasks, and why the "Linux philosophy" of using a raw command-line interface provides a more lightweight, composable alternative for local development, paving the way for an Agent OS.