HomeAI
Systems of intelligence
@DrIbrarAhmedAI
How modern AI works when you strip away the demos: next-token prediction, attention, embeddings, vector search, RAG, KV cache, speculative decoding, and agents with MCP. Each article pairs with the full lecture.
One task you type is not one request. On OpenRouter, agent work uses about 15× more tokens per request than human-driven work. This video shows where those tokens come from, what can be reused, and what breaks at scale. The 15× figure is OpenRouter data, per request. It is not a claim about every agent or the whole industry.
The bill did not disappear. It moved. Vector search was not wrong. It was one part of the system. Exact words want keyword search. A connection wants a graph. A reranker is a second check. A bounded corpus can sit in a long context window. None of those spends nothing. Match the question. Build the piece it actually needs.
An AI Agent can do something a chatbot cannot. It can act. Give it a browser, files, and code execution, and a normal webpage can hide a second instruction. The model follows the wrong line and asks for a private key. The request never reaches the key. The agent sits inside an isolated sandbox. Network policy and the real credential sit outside it.
Vector search earns its place when the passage means the same thing as the question. Ask for the password rotation policy and the nearest passage can match, even when the words differ. Change the question to a path, and similarity still returns three relevant pieces. Nothing in the index joins them. Keep the vector index.
Three relevant passages can still miss the path between them. An AI can hold the right documents, retrieve relevant chunks, and still miss the answer. Vector search finds similar meaning. It does not follow the link from vendor, to service, to product. Vector databases are not obsolete. Vector-only RAG is no longer enough.
The right documents can still produce the wrong answer. Vector databases are not going away. Vector search alone does not answer every enterprise question. This video shows why the right documents can still produce the wrong answer, where vector search still wins, and what keyword search, knowledge graphs, reranking, and long context each do.
Can you spot a deepfake? Blinking, teeth, and skin no longer reveal an AI face. Deepfake detection fails outside the lab. The generator already passed your eye test. The Model Already Passed Your Eye Test Why Deepfake Detectors Fail When They Leave the Lab | AI MECHANICS You will see: • GENERATOR - weaker models leaked.
This is not a chatbot. Anthropic built Claude Mythos to find security flaws humans had missed for decades. Then it restricted access. In 30 days, partners reported more than 10,000 high and critical findings. Claude Mythos Found 10,000 Serious Flaws in Its First Month Project Glasswing | AI MECHANICS Of the high and critical open-source subset reviewed by humans, 90.
How DeepSeek Shrinks Memory and OpenAI Scales Thinking The Transformer Is Evolving | AI MECHANICS Every token your AI remembers costs GPU memory. At short context that looks harmless. At long context, the KV cache becomes one of inference’s biggest constraints. DeepSeek’s answer is not “the Transformer is dead. ” It is compress what you store, then spend compute where thinking helps.
Most models stop at an answer. GPT-6 Astra is built to finish a job. This AI Mechanics episode explains the real shift: from one-turn chat to goal, reason, act, observe, adapt, complete. Computer use, agentic coding, research loops, mid-turn steering, async tools, and a token price about 2. 5 times Sol. No AGI claims. Vendor benchmarks are labeled OpenAI-reported.
Kimi K3 has 2. 8 trillion parameters-but only about 104 billion are activated for each token. This visual deep dive explains how 16 of 896 routed experts, 2 shared experts, Stable LatentMoE, and Kimi Delta Attention make that possible. Kimi K3 is a Mixture-of-Experts model with a 1M-token context window. We follow one token through expert routing, distributed communication, the 7168 → 3584 latent projection, Quantile Balancing, KDA, Gated MLA, and Block Attention Residuals.
What is an LLM? ChatGPT generates answers by predicting the next token. Beginner visual guide-no math required. Tokens, weights, context, and hallucinations. What you will learn • What LLM means • How next-token prediction builds an answer • Where probabilities come from • Context vs memory • LLM vs the full ChatGPT-style app Playlists Beginner AI: https://www.
Why do reasoning models spend more compute before answering? Test-time compute. Best-of-N, backtracking, verification, rewards, DeepSeek-R1, and tool use-visually explained. What you will learn • Training-time vs test-time scaling • Reasoning budgets and candidate paths • Best-of-N, backtracking, and verification • Outcome vs process rewards • When thinking longer helps-and when it fails Playlists DEEP AI: https://www.
What is speculative decoding? A small draft model proposes tokens; the large LLM verifies them. Same output distribution as the big model-faster when the draft is often right. What you will learn • Why one token per forward pass wastes GPU parallelism • Draft model vs target model • One-pass verification and rejection • Why draft quality controls the speedup • How it sits next to GQA, quantization, and PagedAttention Playlists DEEP AI: https://www.
What is entropy? Next-token prediction and compression are the same math. From Shannon’s information theory to why language models train with cross-entropy. Watch next: Transformer Attention Explained - https://www. youtube. com/watch? v=rbrSteyXx_0 What you will learn • Variable-length and prefix-free codes • Self-information I = -log₂ p • Why prediction equals compression • Entropy rate of language • How this leads to LLM training loss Playlists LLM Fundamentals: https://www.