Glossary
Short, plain-language definitions of the AI/ML terms I keep bumping into - explained the way I wish someone had explained them to me. One tight paragraph each, links to related terms, and never any math you have to do. Use the filter to jump straight to one.
No terms match that.
Core concepts
- AblationRemoving a model component on purpose to measure how much it mattered - the standard “ablation study.”
- AGIHypothetical AI that matches humans across essentially all intellectual tasks - a contested, moving goalpost.
- Context windowHow much text a model can hold in view at once, counted in tokens - input and output share the budget, its cost grows with the square of its size, and ‘in the window’ doesn’t guarantee ‘actually attended to.’
- DimensionsIndependent axes of variation; ML vectors live in hundreds or thousands of them.
- EmbeddingsDense vectors where distance & direction encode meaning - the backbone of semantic search and RAG.
- Epsilon gateA convergence threshold (eps) used as a hard pass/fail gate - and the silent failure it causes when the same code runs on different hardware.
- Gradient descentThe downhill-walking algorithm that trains models - nudge every parameter to shrink the error, then repeat.
- HallucinationWhen a model states something false as fluently and confidently as something true, because its job is to predict a plausible next token, not to say only true things - and it has no built-in signal for ‘I don’t know.’
- Latent spaceThe hidden vector space where geometry encodes meaning.
- LoRALow-rank adapters; parameter-efficient fine-tuning.
- Machine learningSystems that learn patterns from data instead of hand-written rules.
- Open weightsA model whose trained parameters are published for download, so you can run it offline and fine-tune it - narrower than ‘open source,’ which would also include the training data and recipe.
- ParametersLearned weights; the count is model size & memory cost.
- TemperatureThe sampling knob that sets how random a model’s word choices are - low is safe and repeatable, high is loose and surprising.
- TensorN-dimensional numeric array; the core ML data structure.
- TokenThe subword chunk a model actually reads and writes; text is split into these before the model sees it, and context limits and pricing are counted in them.
- Validation lossHeld-out validation error; the overfitting tripwire.
Local inference & formats
- Floating-point precisionHow many bits each number uses - fp32 vs fp16 vs bf16. Halving precision halves the footprint, and is not the same thing as quantizing.
- GGMLC/C++ tensor library powering llama.cpp; runs GGUF.
- GGUFllama.cpp’s single-file format for quantized local LLMs.
- llama.cpp vs OllamaThe local-inference engine vs the wrapper built on it.
- MLXApple-silicon ML framework; the Mac answer to GGUF.
- QuantizationStoring each weight as a low-bit integer index into a handful of buckets, with a scale per block to map it back. The 4x-8x shrink that puts a big model on a small machine, and what it quietly costs you.
- RSS samplerA lightweight monitor that polls a process’s Resident Set Size on a timer to chart its real-RAM footprint over time - so you see the memory cliff coming instead of OOM-ing into it.
GPU compute & backends
Architectures
- AttentionHow tokens weigh each other (query · key · value) - the heart of the transformer.
- GANGenerator-vs-discriminator generative architecture.
- GPTGenerative pre-trained (decoder-only) transformer LLM.
- Mixture of expertsA transformer that holds many ’expert’ sub-networks but routes each token to only a few, so a model with hundreds of billions of parameters runs at the cost of a much smaller one - cheap in compute, expensive in memory.
- QwenAlibaba’s family of open-weight LLMs - decoder-only transformers you can download, run offline, and fine-tune, from phone-sized up to huge.
- TransformerThe self-attention architecture behind modern LLMs.
Building blocks
- GELUThe smooth activation gate used inside GPT- and BERT-style transformers.
- Norm placementWhere LayerNorm sits in a transformer block - pre, post, or sandwiched - and why it decides whether a deep model trains at all.
- ReLUThe simplest activation gate: zero out negatives, pass positives through.
- Residual connectionsSkip-ahead wiring that lets networks go very deep without the signal falling apart.
Safety & alignment
- AlignmentGetting a model to do what you meant, not necessarily what you literally said - the gap that widens as the model gets more capable.
- Model welfareWhether an AI model can have experiences that matter morally - and what a lab owes it if the answer might be yes. Not a claim that models are conscious; a refusal to assume they aren’t.