Blinded By Science Episodes ▶ CIUT 89.5 FM

AI, in plain English

Background for the show: what the words mean, and who actually supplies whom in the AI industry. Written for listeners, not engineers.

Explore

AI ecosystem relationship matrix

Who supplies data, models, compute, investment and distribution to whom, across thirteen of the companies that shape AI. Filter by relationship type, swap the axes, select any cell to read what the relationship is.

Open the matrix →

Glossary of AI terms

43 terms that come up on the show — from tokens and training to hallucination, RAG and scaling laws — each explained in a sentence.

Jump to the glossary →

Glossary — 43 terms

The words that come up on the show, each in a sentence.

AGI (Artificial General Intelligence)
AI that can do any human-level task (not yet real).
AI (Artificial Intelligence)
Computers acting smart like humans (learning, chatting, deciding).
AI Agent
AI that acts on its own (e.g., books flights for you).
Algorithm
Step-by-step recipe for solving problems (AI's "instructions").
ANN (Artificial Neural Network)
Web of digital "neurons" spotting patterns in data.
Autonomous Agent
AI that runs solo without human help.
Bias
AI unfairness from bad training data (e.g., stereotypes).
Chatbot
AI for conversations (like ChatGPT).
Chips / GPUs (e.g., H100)
Super chips powering AI math (1 million = global AI muscle).
Cognitive Computing
AI mimicking human thinking (reasoning, memory).
Compute / Compute Capacity
Total chip "brainpower" for AI (grows 3.3x yearly since 2022).
Context Window
How much info AI "remembers" at once (e.g., 1M tokens = a book).
Custom GPT
Tweaked ChatGPT for specific jobs (e.g., recipe generator).
Data Augmentation
Faking extra data to train AI better.
Dataset
Big pile of data for AI training.
Deep Learning (DL)
Advanced ML with layered "neural nets."
Embeddings
Numbers representing word meanings for AI search.
Fine-Tuning
Tweaking a trained AI for niche tasks (less compute than full training).
FLOP (Floating Point Operations)
Math calculations AI does (GPT-4 ~2×10²⁵).
Generative AI
AI creating new stuff (text, images, music).
GPT (Generative Pre-trained Transformer)
OpenAI's chat AI family (GPT-4 > Llama compute).
GPU
Graphics chip repurposed for AI (e.g., H100 clusters).
Hallucination
AI confidently making up facts.
Inference
Using trained AI for real work (less compute than training).
LLM (Large Language Model)
Huge text AI like GPT, Llama, Claude.
Machine Learning (ML)
AI learning from data, no coding rules.
MCP (Model Context Protocol)
Linking AI to tools/data (e.g., in Claude).
Natural Language Processing (NLP)
AI understanding human language.
Overfitting
AI memorizing data too well, bad at new stuff.
Post-Training
Refining AI after basics (key for GPT-5 efficiency).
Pre-Training
Initial massive data feed (huge compute for GPT).
Prompt
Instructions you give AI (e.g., "Explain simply").
Prompt Engineering
Crafting perfect AI instructions.
RAG (Retrieval-Augmented Generation)
AI pulling real facts to avoid hallucinations.
Reinforcement Learning (RL)
AI learning via rewards/punishments (RLHF tunes chatbots).
RLHF (Reinforcement Learning from Human Feedback)
Humans rating AI outputs to improve it.
Scaling Laws
More compute/data = smarter AI (drives 3.3x growth).
Supervised Learning
Training AI with labeled examples (input + right answer).
Synthetic Data
Fake data made by AI to train more AI.
Tokens
Text chunks AI processes (~4 letters each).
Training
Teaching AI with data (GPT-4 ~25k GPUs for months).
Transformer
AI architecture behind LLMs (changed everything).
Unsupervised Learning
AI finding patterns without labels.