概念

CW Context WindowThe model's working memory — from 2K tokens to millions in five years. DM Diffusion ModelsGeneration by denoising — the engine of images and video. MO Mixture of Experts (MoE)Huge capacity, small bills: only a few experts fire per token. MU MultimodalityOne model, many senses: text, image, audio, video in a single network. NN Neural Networks & Deep LearningThe substrate: layered functions tuned by gradient descent. OV Open vs Closed WeightsThe industry's defining strategic split. PA ParametersThe learned numbers — and why 'bigger' stopped being the whole story. SS State-Space Models (Mamba)The main architectural challenger: linear-time sequence processing. T& Tokens & TokenizationThe unit AI is bought, sold and measured in. TR TransformerThe 2017 architecture everything on this site exists to run. WM World ModelsLearning how reality behaves — the claimed next frontier.

AA AI AgentsModels that act: plan, call tools, iterate — the 2025 product wave. DI DistillationTeaching a small model to imitate a big one. FT Fine-tuning & LoRAAdapting a trained model cheaply — the small-compute on-ramp. HA HallucinationConfident fabrication — the failure mode that shapes deployment. IN InferenceRunning the model — now the majority of AI compute and the focus of the buildout. PT Post-training (RLHF, RLVR and after)Where a raw predictor becomes an assistant — and, lately, a reasoner. QU QuantizationFewer bits per weight: how models shrink to fit real hardware. RL RLHFReinforcement learning from human feedback — the technique behind ChatGPT's manners. RA Retrieval-Augmented Generation (RAG)Look it up, then answer — how enterprises ground models in their data. SL Scaling LawsThe empirical curves that justify the capex on this site. SD Synthetic Data & the Data WallWhen the internet runs out, models train on model output. TT Test-time Compute & ReasoningThe 2024 discovery: models get smarter by thinking longer. T& Training & Pre-trainingMonths of thousands of GPUs predicting the next token, trillions of times.

CH ChipletsMany small dies, one big chip — yield economics beat monoliths. EC Export Controls & the Chip WarCompute as statecraft: the US–China technology conflict. F& FLOPS & Precision FormatsHow compute is counted — and why the same chip has five different numbers. ML Moore's Law & Its AfterlifeTransistor scaling slowed; the industry routed around it. P& PUE & Data-Center EfficiencyPower Usage Effectiveness: how much electricity feeds chips vs overhead. SA Sovereign AINations buying their own stack: compute as critical infrastructure.

AS AI SafetyThe umbrella: from jailbreaks and misuse to catastrophic risk. AL AlignmentMaking systems that pursue what we actually intend. ER Existential Risk (x-risk)The argued worst case — and the industry's strangest consensus statement. IN InterpretabilityOpening the black box: reading a model's internal features.

AR AI RegulationThe EU legislates, the US pivots, everyone watches compute. CG Compute GovernanceRegulating AI by regulating FLOPs, chips and data centers.

AE AI's Energy DemandFrom 1.5% of world electricity toward — how much? The load-growth question. JP Jevons ParadoxWhy efficiency can increase total consumption — the argument of 2025. J& Jobs & AutomationWhat the models do to work — early evidence, loud predictions. TB The Bubble QuestionRailroad boom, dot-com rerun, or the real thing? The arguments, attributed. TE The Economics of AITrillion-dollar capex against fast-falling token prices.

MC Machine ConsciousnessCould there be something it is like to be a model? Serious people now ask. MP Moravec's ParadoxHard things are easy, easy things are hard — still true, still shaping robotics. TC The Chinese RoomSearle's 1980 challenge: does symbol manipulation ever become understanding? TT The Turing Test1950's benchmark, quietly passed — and quietly retired.

A( AGI (Artificial General Intelligence)The destination everyone names and no one defines the same way. AT AGI Timelines & PredictionsWho predicts what, by when — on the record. S( Superintelligence (ASI)Beyond human-level — the explicit goal several labs now name. TS The Singularity & Recursive Self-ImprovementThe feedback-loop hypothesis: AI improving AI, faster each cycle.