5sigmas · Artificial intelligence, without the noise

See AI.
Understand how it actually works.

Each topic connects a visual explanation, an interactive mechanism, technical depth and primary sources. Build intuition first, manipulate the mechanism, inspect its limits and then go to the evidence.

01 SeeOrient yourself around one idea
02 ManipulateChange variables and inspect the mechanism
03 Go deeperTechnical chapter and primary sources
The standard behind the name

Why 5sigmas?

In particle physics, a result is conventionally treated as a discovery only when it reaches five-sigma significance: a threshold stringent enough that an extreme background fluctuation has a probability of roughly one in 3.5 million under the usual Gaussian interpretation.

5sigmas applies the same editorial instinct to artificial intelligence: separate durable knowledge from daily noise, demand evidence, and build judgment before accepting a narrative, tool or promise.

5σdiscovery threshold ≈ 1 in 3.5Mextreme-noise probability Signal > noiseeditorial principle
Normal distribution with the five-sigma threshold separated from expected noise
Do not publish because something is new. Publish when the explanation survives evidence, context and its own limitations.

Choose how to continue

Core concepts

Direct answers to the concepts behind modern AI systems.

View all concepts →
01 LLMs What large language models are, how they learn and how they generate tokens. Models 02 Transformers Attention, representations and the architecture behind modern language models. Architecture 03 Reasoning What reasoning means for an LLM and what changes at inference time. Inference 04 Model evaluation How to measure real capability without confusing a benchmark with the objective. Evaluation 05 AI agents Goals, tools, state and boundaries that turn model responses into actions. Systems 06 Prompt injection Why untrusted data can influence instructions and which controls reduce the risk. Security

Visual explanations

Open the complete path →
Native English visual about reasoning-model failure modes
Visual · Reasoning
Why a convincing reasoning trace can still fail
 Native English visual about latency and streaming in reasoning systems
Visual · Interaction
Latency is part of model quality once a human is waiting
 Native English visual about risks from additional inference-time reasoning
Visual · Reliability
More inference compute is a control knob, not a guarantee

Choose a path

View all series →
01 AI and Generative AI Foundations The essential distinctions between software, learning, generation and AGI. 4 chapters 02 Multimodality Text, image, audio and video inside the same generative system. 5 chapters 03 Reasoning Models Capability, failure, cost and latency when models spend compute before answering. 5 chapters 04 Data Centers in Space The physics and infrastructure behind orbital compute. 4 chapters

Engineering

Technical note

Proactive and reactive agents and tool calls

How to separate the conversational contract, internal execution and deferred completion without turning conversation history into the system database.

Open the note →
Conversational runtime
Internal state
Asynchronous execution