Video

Talks and technical walkthroughs. Playback stays on this page.

7 published

  1. Muse, Grok, and the Agent Harness

    Separates the model, the harness, and the permissions that turn a request into real work.

  2. Why Your Local LLM Slows Down as the Conversation Grows

    Separates model weights, the KV cache, prefill, prefix reuse, and token generation.

  3. Why Adding More Agents Can Make a System Worse

    Handoffs, queues, duplicated work, and shared-state conflicts can slow a multi-agent system.

  4. Prompt Injection: When a Document Becomes an Instruction

    Follows an untrusted instruction from a document toward a tool call, and where permissions can stop it.

  5. LLM Inference Explained: Where Every Millisecond Goes

    Maps prefill, decode, batching, queues, KV cache, memory bandwidth, and GPU execution.

  6. MCP vs APIs vs CLI: Which Interface Should AI Agents Use?

    Traces one GitHub issue workflow through MCP, a direct API, and the command line.

  7. How AI Agents Call Tools: Model, Runtime, and Service

    The model proposes arguments. The runtime validates and routes. A service performs the action.