Penn's Kording Explains How LLMs Work: From Next-Token Prediction to Agents

On August 26, Penn professor Konrad Kording posted a long X thread explaining how LLMs work for informed lay readers, simultaneously shared by Yoav Goldberg. Both noted the piece was co-written with Tim, inspired by a CIFAR LMB discussion, improved by Brett Mensh, and will become part of an upcoming paper.

Confirmed

Core mechanism: the neural network generates answers word by word, with each word's probability conditioned on preceding words (m8). The core task is turning prompts into answers via two objectives—language imitation and human quality ratings—implemented by inference (answer-producing) and training (answer-improving) components (m9).
2022-era LLMs, though impressive, were essentially "stochastic parrots"; the field has since moved fast beyond that stage (m6).
Three ways to improve answers: "inner speech" as hidden scratchpad (m7); tool use such as databases, web search and calculators (m4); and agents that coordinate LLM components, memory, databases and tools (m5).
Training improvements: push toward ground truth in domains like math and coding; use LLM self-evaluation in open-ended domains (m3).
Overall conclusion: today's interactive "thinking" machines are the old recipe—next-token-prediction training—extended with inner speech, tool use, agent coordination and better evaluation (m2, m10).

Why it matters

Most LLM explainers target either total novices or pure specialists; this one offers technically literate non-AI readers a unified component-based framework covering inner speech, tool use and agents, explaining both current strengths and remaining weaknesses. It will be incorporated into a formal paper, giving it citable value.

2026-08-26 ~ 2026-08-26 · 11 related posts

Primary sources

Neuroscientist Kording shares an explainer of how LLMs work for an informed general audience — KordingLab · 2026-08-26 00:12 (UTC+08:00)
Yoav Goldberg publishes LLM explainer for an informed general audience — yoavgo · 2026-08-26 16:10 (UTC+08:00)
Today's LLM Strengths and Weaknesses Trace Back to Component Ideas — KordingLab · 2026-08-26 00:13 (UTC+08:00)

[source] Neuroscientist Kording shares an explainer of how LLMs work for an informed general audience — KordingLab · 2026-08-26
LLMs explained: Neural networks that convert prompts to answers via training and inference — KordingLab · 2026-08-26
LLM Core Mechanics: Probability-Based Word-by-Word Generation — KordingLab · 2026-08-26
From 'Stochastic Parrots' to Rapidly Evolving Modern LLMs — KordingLab · 2026-08-26
Improving LLM Answers: The First Idea, Inner Speech — KordingLab · 2026-08-26
Enhancing LLM Answers: The Second Idea, Tool Use — KordingLab · 2026-08-26
Generating Answers: The Third Idea, Agents — KordingLab · 2026-08-26
Improving LLMs: Steering Towards Truth and Self-Evaluation — KordingLab · 2026-08-26
The Evolution of Modern Interactively 'Thinking' Machines — KordingLab · 2026-08-26
[source] Today's LLM Strengths and Weaknesses Trace Back to Component Ideas — KordingLab · 2026-08-26
[source] Yoav Goldberg publishes LLM explainer for an informed general audience — yoavgo · 2026-08-26