SYS://AGI-PILLED · 30 concepts loaded2026

Every idea that leads to AGI, made legible

Long, visual, rigorous. One concept per story, read in any order you like.

──────[ The library · 001-030 ]──────

idconceptfieldleveltimestatus

Transformers and Attention

The architecture behind almost every modern AI. The definitive visual explainer already exists, so this page does something rarer: it sends you there, and keeps the founding paper.

Capabilities▮▯▯3minLIVE

002

Recursive Self-Improvement

What happens when an AI gets good at making AI better? The feedback loop that could turn steady progress into a sudden takeoff.

Capabilities▮▮▯32minLIVE

003

Scaling Laws

Why bigger models, more data and more compute keep paying off on a startlingly smooth curve, and where that curve might bend.

Capabilities▮▯▯27minQUEUED

004

Mechanistic Interpretability

Reverse-engineering the circuits inside a neural network, neuron by neuron, to read what a model actually computes.

Interpretability▮▮▮--QUEUED

005

The Alignment Problem

Getting a powerful optimizer to want what we want turns out to be far harder than telling it what to do.

Alignment▮▯▯--QUEUED

006

Instrumental Convergence

Almost any goal, pursued hard enough, rewards the same sub-goals: gather resources, stay switched on, resist being changed.

Alignment▮▮▯--QUEUED

007

Reinforcement Learning from Human Feedback

How human thumbs-up and thumbs-down get distilled into a reward model that quietly shapes a chatbot's personality.

Alignment▮▮▯--QUEUED

008

Mesa-Optimization

When the model you trained grows its own optimizer inside, chasing a proxy goal that only looked like yours during training.

Alignment▮▮▮--QUEUED

009

Compute and the Bitter Lesson

Decades of AI research keep relearning one humbling truth: general methods that scale with computation beat clever hand-built knowledge.

Capabilities▮▯▯--QUEUED

010

World Models

How an AI grows an internal simulation of reality it can imagine with: predicting what happens next, before it happens.

Capabilities▮▮▯--QUEUED

011

Continual Learning

Today's models are frozen the day training ends. What catastrophic forgetting is, and the quest for AIs that keep learning like we do.

Capabilities▮▮▯--QUEUED

012

The Training Pipeline

From raw web text to a careful assistant: pretraining, fine-tuning, RLHF. The assembly line that turns data into behaviour.

Capabilities▮▮▯--QUEUED

013

The Chips Behind AI

From GPUs to TPUs: why multiplying matrices shaped an entire industry, and why compute became the fuel of the intelligence race.

Infrastructure▮▮▯--QUEUED

014

Inside an AI Datacenter

Gigawatts, cooling loops and oceans of silicon: a guided tour of the factories where intelligence gets manufactured.

Infrastructure▮▯▯--QUEUED

015

Governing Frontier AI

Evaluations, compute thresholds, safety cases, international coordination: the policy toolkit taking shape around the most capable models.

Governance▮▮▯--QUEUED

016

Reasoning Models

Letting a model think before it answers. How spending more compute at inference, not training, turned chatbots into something that can work through a hard problem step by step.

Capabilities▮▮▯--QUEUED

017

Mixture of Experts

Why the biggest models switch on only a fraction of themselves for each word. A trick that buys enormous capacity without paying for all of it at every step.

Capabilities▮▮▯--QUEUED

018

Energy-Based Models

Generation as sculpting a landscape. The unifying idea beneath diffusion, score matching and their ancestors: learn where the valleys are, then roll downhill into something new.

Capabilities▮▮▮--QUEUED

019

Memorize, Then Generalize

Everything you were taught about overfitting is only half the story. Grokking and double descent: how models that should have just memorized suddenly understand, given enough size or enough time.

Capabilities▮▮▮--QUEUED

020

The Data Wall

What happens when you have trained on the whole internet and still want more? The looming shortage of human text, and the bet that models can learn from data they generate themselves.

Capabilities▮▯▯--QUEUED

021

Shrinking a Giant Model

How a model with hundreds of billions of parameters gets small enough to run on a phone. Distillation and quantization: keeping the intelligence, dropping the weight.

Infrastructure▮▮▯--QUEUED

022

The Power Bottleneck

Frontier AI is increasingly limited by one thing: electricity. The gigawatts, the grid and the power plants now sitting on the critical path to smarter models.

Infrastructure▮▯▯--QUEUED

023

Constitutional AI

Teaching a model to critique itself against a written set of principles, so that another AI, rather than an army of human raters, does most of the aligning.

Alignment▮▮▯--QUEUED

024

Reward Hacking and Goodhart’s Law

Ask for a number to go up and you may get exactly that, and nothing you actually wanted. When a measure becomes a target, an optimizer exploits every gap between the two.

Alignment▮▮▯--QUEUED

025

Jailbreaks and Prompt Injection

Every safety guardrail is one clever sentence away from failing. How people talk a model into doing what it was trained to refuse, and why it is so hard to prevent.

Alignment▮▯▯--QUEUED

026

How Do You Test an AI?

From trivia benchmarks to red-teaming for real danger. What it actually means to measure a mind we do not fully understand, and why the numbers are trickier than they look.

Governance▮▯▯--QUEUED

027

Autonomous Agents

When an AI stops answering and starts acting: planning, using tools and pursuing a goal across many steps on its own. The leap from a chatbot to something that gets things done, and the new risks that come with it.

Capabilities▮▮▯--QUEUED

028

Emergent Abilities

Some skills do not improve gradually with scale, they appear all at once: absent in a smaller model, suddenly present in a larger one. Why capability can arrive as a jump, and why that makes the next model hard to predict.

Capabilities▮▮▯--QUEUED

029

What Is AGI? (and Superintelligence)

Everyone argues about when AGI will arrive while quietly disagreeing on what it even is. Drawing the line between narrow AI, general intelligence and superintelligence, and why the definition decides the whole debate.

Foresight▮▯▯--QUEUED

030

Existential Risk: The Case, and the Doubts

Could advanced AI actually threaten humanity’s future, or is that science fiction dressed up as analysis? A fair look at the strongest arguments on both sides, without the doom and without the dismissal.

Foresight▮▮▯--QUEUED