The field notes
The Daily AI Drop - Morning Edition, October 20, 2026
Alex and Jordan cover four overnight drops: Mercury 3 diffusion LLM testing, Microsoft on why AI agents cost so much, a free local AI agent for your smart TV, and reinforcement learning explained through self-driving cars.
Four videos dropped overnight, and they accidentally tell one big story: the infrastructure of AI is changing under our feet — the model architecture, the economics, and where the thing actually runs.
- Diffusion at Scale: Testing Mercury 3 by Inception (Fahd Mirza) — a hands-on test of the most capable diffusion-based LLM, which refines whole answers at once instead of predicting token by token. If diffusion models crack long-form reasoning, inference economics could shift entirely.
- Why AI Agents Cost So Much (Microsoft Developer / Azure Essentials Show) — Takuto Higuchi talks with Sindhura Bandhakavi (Director of Engineering, Microsoft CoreAI) about "tokenomics": the emerging discipline of treating agent cost as an architecture design constraint, not a finance afterthought.
- Control Any Smart TV with Free Local AI Agent (Hermes + Ollama) (Fahd Mirza) — a hands-on tutorial for driving any LG, Samsung, or Android TV with an AI agent that runs entirely locally and free. A genuinely fun first Ollama project with a visible payoff.
- Reinforcement Learning Simplified: How AI Learns to Drive (Priyanka Vergadia) — the core concepts of RL (exploration vs. exploitation, reward signals, policy learning) explained in plain language through a self-driving car example — including the post-training step behind modern chatbots.
Diffusion at Scale: Testing Mercury 3 by Inception — Fahd Mirza
Mercury 3 from Inception is billed as the most capable diffusion-based large language model, and Fahd puts it through its paces. Diffusion models generate text by refining a whole answer from noise at once — like an image generator sharpening a blurry picture — rather than predicting one token at a time. That matters because parallelism could mean dramatically faster inference and holistic answer editing. Watch this if you want an early read on the architecture bet that might reshape what "fast" means for LLMs.
Why AI Agents Cost So Much — Microsoft Developer (Azure Essentials Show)
AI agents deliver business value fast, but as usage scales, organizations are discovering that intelligent experiences come with real operational costs. Host Takuto Higuchi sits down with Sindhura Bandhakavi, Director of Engineering at Microsoft CoreAI, to explore AI token economics — "tokenomics." The key insight: model routing, reasoning-step budgets, and context caching become architecture decisions. If you're building or buying agent workflows, this is your cue to start measuring cost per task now.
Control Any Smart TV with Free Local AI Agent (Hermes + Ollama) — Fahd Mirza
A hands-on tutorial for controlling any LG, Samsung, or Android TV with an AI agent that runs entirely locally and for free — Hermes as the agent brain, Ollama running the models on your own machine. No cloud subscription, no voice data leaving your house. Beyond the fun factor, it makes the local-agent pattern concrete: once a capable agent runs free on your hardware, every device becomes a candidate for an AI interface.
Reinforcement Learning Simplified: How AI Learns to Drive — Priyanka Vergadia
How does AI learn to drive, make decisions, and improve from its mistakes? Priyanka uses a self-driving car as the running example — from taking random actions and crashing to learning the best strategy through rewards and experience. She breaks down the core RL concepts in simple terms, and along the way explains the post-training technique that made modern chatbots actually useful. A great door into the machinery between a base model and the assistant you talk to.