AI Unraveled: Latest AI News, ChatGPT, Gemini, Claude, DeepSeek, Gen AI, LLMs, Agents, Ethics, Bias · Etienne Noumen

[AI UNRAVELED SPECIAL] LLMs Explained from First Principles — Vectors, Attention, and the Scaling Wall

·37 min·3 clips
LLMs don't think or reason; they're just mathematical systems scaled to mimic comprehension.
Start with the claim. An LLM does not think, reason, or intend; it runs weighted sums and non-linear transformations at huge scale. The framing is blunt on purpose. The host treats the transformer as a mathematical system that can imitate comprehension without having symbolic reasoning or built-in logic. Tokens become geometry. Text moves through high-dimensional vectors, where meaning is handled as numerical relationships. Attention does the context mixing. Multi-head attention compares position and similarity, giving the model several ways to track relationships across a sequence at the same time. Position matters. The episode ties positional encoding to Fourier features and signal processing: low frequencies help with long-distance relationships, while high frequencies keep nearby word order from getting smeared. That part earns its keep. The model needs sequence geometry to shift with token location so it can handle longer sequences without losing order. Then the episode gets into stabilization. Feed-forward networks reshape each token on its own through a linear transformation, a non-linear GELU activation, and another linear transformation. Depth is where things get touchy. Stack dozens or hundreds of layers, and the numbers can fade toward zero or blow past useful training. Residual connections give the signal a bypass. The original vector routes around a sublayer and gets added back to the processed output, which helps preserve information through the stack. Layer normalization keeps the math from drifting. By moving means to zero and variance to one, it steadies gradients and keeps information from breaking down. The larger argument is severe: scale can produce impressive statistical approximation, but the closing warning is clear. More brute force does not automatically mean true autonomous reasoning, especially in an architecture the speaker says was never built to measure it.

As heard by us

A blunt, technically grounded tour of transformer mechanics and their limits.

This episode treats the transformer as machinery, not mystery, and keeps the focus on vectors, attention, positional encoding, residual connections, layer normalization, and scaling limits.

Read the full review in PlayNext →

Why you'd press play

You want the transformer explained cleanly, without hype, fog, or hand-waving.

Read the full recommendation in PlayNext →
Listen to the show on