@Blum_OG: "everyone uses AI. almost nobody understands how it works." that gap is real - and it's the whole point here's what the…

X AI KOLs Timeline News
ai-education tokens attention-mechanism llm hallucination rag rlhf

Summary

An explanatory tweet thread breaking down how AI works, covering tokens, attention, parameters, context windows, hallucination, RAG, and RLHF to help users become sharper users of AI.

"everyone uses AI. almost nobody understands how it works." that gap is real - and it's the whole point here's what the article actually explains: > tokens are reusable building blocks - 1 token ≈ 0.75 words > attention lets every word look at every other word at once > this single idea unlocked modern AI > GPT-4 has ~1.8 trillion parameters - all from layered neurons with adjustable connections > context windows range from 4,000 tokens (early GPT) to 1,000,000 (Gemini 1.5 Pro) > the "lost in the middle" problem: models focus on the start and end - not the middle > hallucination happens because LLMs predict tokens, not truth > RAG fixes this - look it up first, then answer > RLHF is what makes models feel helpful, not just fluent most people think prompting is a trick it's not - it's just clear communication with a pattern-matching system once you see that, the whole thing clicks understanding the mechanics doesn't make you an engineer it makes you a sharper user
Original Article
View Cached Full Text

Cached at: 05/24/26, 12:17 AM

“everyone uses AI. almost nobody understands how it works.”

that gap is real - and it’s the whole point

here’s what the article actually explains:

tokens are reusable building blocks - 1 token ≈ 0.75 words attention lets every word look at every other word at once this single idea unlocked modern AI GPT-4 has ~1.8 trillion parameters - all from layered neurons with adjustable connections context windows range from 4,000 tokens (early GPT) to 1,000,000 (Gemini 1.5 Pro) the “lost in the middle” problem: models focus on the start and end - not the middle hallucination happens because LLMs predict tokens, not truth RAG fixes this - look it up first, then answer RLHF is what makes models feel helpful, not just fluent

most people think prompting is a trick

it’s not - it’s just clear communication with a pattern-matching system

once you see that, the whole thing clicks

understanding the mechanics doesn’t make you an engineer

it makes you a sharper user

Similar Articles