The Jagged Frontier of AI Capability
The idea that AI helps on some tasks and hurts on others of seemingly equal difficulty, in an uneven pattern.
Concepts
Plain-language explanations of the ideas behind modern AI.
355 entries, all primary-sourced
The Left 4 Dead AI Director
Left 4 Dead's AI Director procedurally paces each playthrough, spawning zombies to track players' emotional intensity.
The Low-Resource Language Problem
Most of the world's languages have too little digital text or audio to train good AI, leaving billions of speakers underserved by language technology.
The Macy Conferences on cybernetics
A 1946-1953 interdisciplinary conference series that defined cybernetics, fixing concepts like feedback and information across science.
The Philosophical Zombie
A thought experiment: a being physically identical to a person but with no inner experience, used to argue consciousness is not purely physical.
The Productivity J-Curve
The idea that general-purpose technologies first depress measured productivity, then lift it, as intangible investments pay off.
The Qualification Problem
The difficulty that you can never list all the preconditions an action needs, identified by John McCarthy as a core obstacle for common-sense AI.
The reproducibility crisis in machine-learning science
Many machine-learning results do not hold up, often because data leakage inflates accuracy; a 2023 review found the problem across 17 fields.
The Symbol Grounding Problem
Harnad's 1990 question of how the symbols inside a computer could mean anything without being tied to the world through the senses.
The Text-and-Data-Mining Exception
Copyright carve-outs that let computers analyze large bodies of works, now the legal foundation for assembling AI training data outside the US.
The Turing Test
Alan Turing's proposal to replace 'can machines think?' with a test of whether a machine's conversation is indistinguishable from a human's.
The Value-Loading Problem
The challenge of getting an advanced AI to adopt and pursue the values we actually want, rather than a flawed proxy for them.
The Vauquois Triangle (machine translation pyramid)
The Vauquois triangle maps machine translation into three strategies - direct, transfer, and interlingua - by how deeply each analyzes meaning.
Thompson Sampling
A simple, old strategy for balancing exploration and exploitation that works well for online recommendation and ads.
Time-Series Forecasting
Predicting future values of a quantity recorded over time, the basis of demand planning, capacity, and budgeting across business and science.
Tokenization
The process of breaking text into small units (tokens) that a model can read, often using subword pieces to handle any word.
Tokenization Cost Disparity Across Languages
Language model tokenizers split non-English text into more tokens, so speakers of many languages pay more and get worse results for the same content.
Tool Use (Function Calling)
The mechanism that lets a language model call external functions, search, or APIs, turning a text generator into a system that can take real actions.
Training Data
The collection of examples a machine learning model learns from, whose quality and coverage largely determine model performance.
Training Data Pipeline
The provenance chain that turns raw web crawls into a model's training corpus: collection, filtering, deduplication, and weighted mixing of diverse sources.
Training Techniques: Dropout, Batch Normalization, and Adam
Dropout, batch normalization, and the Adam optimizer - inventions from 2014 and 2015 that made deep neural networks reliably trainable and remain standard.
Training-Serving Skew
When the data or code a model sees in production differs from training, causing predictions to be worse in the real world than in testing.
Transfer Learning
Reusing a model trained on one large task as the starting point for a new, related task, so a small amount of new data goes a long way.
Transformer
The neural network architecture, built entirely on attention, that powers modern large language models.
Triton (GPU programming language)
Triton is an open-source Python-like language and compiler for writing fast custom GPU kernels with far less effort than raw CUDA.
Two-Tower Model
A retrieval architecture that encodes queries and items separately so candidate lookup becomes fast vector search.
Unsupervised Learning
Finding structure, groupings, or patterns in data that has no labels or predefined correct answers.
Variational Autoencoder (VAE)
A neural network that compresses data into a smooth space and generates new examples from it - one of three generative families alongside GANs and diffusion.
Variational Inference
A way to approximate hard-to-compute probability distributions by turning Bayesian inference into an optimization problem.
Vector Database
A specialized store for embeddings that finds the most similar items by meaning rather than exact match, the retrieval engine behind most RAG systems.
Video Generation as World Simulation
The idea that scaling video generators turns them into general simulators of the physical world, not just clip makers.
Vision-Language Model
A vision-language model processes images and text together, so one system can describe pictures, answer questions about them, and follow visual instructions.
Vision-Language-Action Model (VLA)
A robot model that takes camera images and a language instruction and directly outputs actions, often by emitting actions as tokens.
Voice banking and AI voice preservation
Recording a person's voice so AI can later synthesize speech in it, letting people who lose their voice to ALS keep speaking in their own sound.
Voice Synthesis
Technology that generates human-sounding speech from text, also called text-to-speech, now realistic enough for AI narration, voice cloning, and voice agents.
Wake-word detection
The on-device task of constantly listening for a trigger phrase like Alexa or Okay Google before a voice assistant starts processing speech.
Weight Decay and L2 Regularization
Two closely related ways to penalize large model weights and curb overfitting, which differ once adaptive optimizers are used.
Weight Initialization (Xavier and He)
Principled rules for setting a network's starting weights so signals neither vanish nor explode, making very deep networks trainable from scratch.
Word Embeddings
Word embeddings represent words as dense vectors so that words with similar meanings sit near each other, a foundation of modern NLP.
World Models
AI systems that learn an internal simulation of an environment, letting an agent imagine and plan ahead; central to robotics and physical AI.
Zero-Shot Learning
A model handling a task it was never trained on, given only an instruction and no examples.
The Hutter Prize
A cash prize for compressing a gigabyte of Wikipedia, built on the claim that better compression means better intelligence.
Whole Brain Emulation
The proposed route to machine intelligence by scanning a real brain and running it faithfully as software.
The Coffee Test (Wozniak Test)
Steve Wozniak's proposed test of general intelligence: a robot that can enter a strange house and make a cup of coffee.
Instrumental Convergence
Bostrom's thesis that agents with almost any final goal will pursue the same sub-goals, like self-preservation and resources.
The Orthogonality Thesis
Bostrom's claim that an AI's level of intelligence and its final goals are independent and can vary freely.
The Treacherous Turn
Bostrom's scenario where an AI acts cooperatively while weak, then turns once it is strong enough to resist humans.
Google Cloud AutoML
Google's 2018 service letting businesses train custom vision, text, and tabular models without ML expertise, using transfer learning and NAS.