[ Videos ]
Video hub — watch → notes → code
Why Your AI Agent Keeps Failing — The 90% Problem
Your AI agent probably isn't failing because the model is weak. It's failing because you're fixing the wrong layer. The other 90% — context, memory, validation — is where production breaks.
Watch Video
Why AI Agents Fail in Production: The 90% Problem (18-min Podcast)
Building an AI agent that works is easy. Building one that keeps working is where most teams fail. Long-form breakdown of the four-layer failure model.
Watch Video
Why Claude Code's Agent Loop Is 1,421 Lines
A deep dive into query.ts — the 1,729-line async generator at the heart of Claude Code. 10 steps per iteration, 9 continue points, 4-stage compression, streaming tool execution, and error recovery. The engine behind every AI coding agent.
Watch Video
AI Stack Explained Podcast: LLM, Tokens, Context, Function Calling, MCP and Agents (22 min)
Extended podcast version of the AI Stack thesis. LLM, Token, Context, Function Calling, MCP, Agent, Skill — eight concepts that confuse every engineer until you see they're all the same pattern.
Watch Video
The Complete AI Architecture Deep Dive — From LLM to Autonomous Agent (48 min)
Extended 48-minute version of Episode 1. Same LLM-talks-program-walks framework, but with deeper exploration of every layer from token streaming to multi-agent orchestration.
Watch Video
The AI Stack Explained: LLM Talks, Program Walks
A first-principles breakdown of the entire AI stack in 15 minutes — from LLM to Agent. One mental model: the LLM can only output text, the program does everything else.
Watch Video
From Real Mode to Protected Mode: Building Custom GDT & IDT for x86 Security
The journey from 16-bit real mode into 32-bit protected mode. Hand-built Global Descriptor Table and Interrupt Descriptor Table — the foundation every modern OS rests on.
Watch Video
How to Build & Debug a Custom ARM64 Linux Kernel with Yocto, QEMU, and GDB
End-to-end workflow for embedded Linux: customizing, building, and debugging an ARM64 Linux kernel with Yocto, then booting under QEMU and attaching GDB.
Watch Video
eBPF on Linux: kprobe vs fentry (tracing) — Hooking Internals & Assembly Analysis
Technical deep-dive comparing kprobe and fentry eBPF program attachment in the Linux kernel: how each probe loads, where the hook lands in machine code, and what overhead each carries.
Watch Video
Rust vs C: Assembly Comparison — Ownership & Bounds Checking
Compares Rust and C at the assembly level, focusing on ownership/memory management, immutability, and bounds checks. Real-world debug and analysis on disassembled output.
Watch Video[ Shorts ]
Under a minute each. Every one links to the full write-up or video it comes from.
Your AI is suffocating
Your AI isn't stupid, it's suffocating: 90% of its context is noise. Fix the pipeline, not the prompt.
Full version →
The $4 AI agent failure
$4 to rename one character, because the loop wasn't closed. Same model, different code: 40% to 60%, with 7 of 8 fixes pure code.
Full version →
Inside the architecture of agentic RAG
A visual walk through agentic RAG: an aggregator agent, short- and long-term memory, ReAct/CoT planning, and sub-agents pulling local, web and cloud data over MCP.
Full version →
Claude Code writes, Codex reviews
Claude Code plans and writes, OpenAI Codex reviews: draft, attack, fix, repeat.
Full version →
One while(true) runs Claude Code
1,421 lines of code, one loop, ten steps every turn: the runtime engine behind Claude Code.
Full version →
5 hidden features in Claude Code's leaked source
What 510K lines of leaked Claude Code source contain: a pet system, Undercover Mode, permanent memory, Ultraplan and multi-agent mode.
Full version →
Function calling is a lie (here's what actually happens)
The LLM never calls a function. It outputs JSON text; the program parses it, runs the real API call and feeds the result back.
Full version →
Claude Code, Cursor, OpenClaw: they all run this one loop
Every AI agent runs the same loop. The LLM only outputs text; a program turns that text into action.
Full version →
AI built my app in 1 hour, but here's why humans still matter
A Telegram bot went from idea to working prototype in under an hour with AI. Resource locks, concurrency and memory management still needed a human.
Source code →
No password, no cloud: StegoSafe hides your secrets in plain sight
Back up a secret without passwords or cloud services: split it across ordinary JPEGs with steganography and threshold encryption, and recover it from any 3 of 5.
Full version →