Summary
Nick Saraev walks through everything that happens between typing a prompt into ChatGPT and receiving a reply, aimed at viewers who use large language models daily but have never seen what happens behind the scenes.
The journey begins with the app itself. ChatGPT is a consumer application wrapped around a model, and it bundles your message with a system prompt, custom instructions, saved memories, attached files, and the entire conversation so far. Because models have no built-in memory, this growing bundle helps explain “context rot,” the drop in reliability as inputs get longer. Next, smaller safety classifier models screen the exchange. Saraev cites Anthropic’s classifier reducing successful jailbreaks from nearly 90% to about 4 to 5% at under 1% of the compute cost, and notes that OpenAI may route some requests to different models.
The video then explains tokenization, where text is split into pieces averaging roughly 70% of a word. OpenAI’s vocabulary runs to about 200,000 tokens, and examples such as “strawberry” and “hello world” in English versus Hindi show why models stumble on math and why English is cheaper to process. From there, tokens travel to Nvidia GPU data centers, where requests are batched with other users’, and pass through layers built on the attention mechanism from Google’s 2017 paper “Attention Is All You Need.”
📺 Source: Nick Saraev · Published October 19, 2026
🏷️ Format: Deep Dive
