G

GAN (Generative Adversarial Network)

• Category: Models & inference
• What people say: Two neural networks competing during training.
• What it actually means: A generator network tries to create realistic data while a discriminator network tries to tell real from fake. They train together: the generator gets better at fooling the discriminator, and the discriminator gets better at detecting fakes.
• Related terms: Loss Function, Latent Space, Diffusion Model

Goodput

• Category: Infrastructure & serving
• What it actually means: The rate of completed requests that satisfy defined service constraints, such as both time-to-first-token and per-token latency objectives, under a stated workload.
• Why it matters: Raw throughput can rise while users experience more slow requests. Goodput counts only work that meets the service contract.
• In practice: Declare the request distribution and latency thresholds, count only compliant completions, report percentiles beside the aggregate rate, and avoid comparing systems under different objectives.
• Common confusion: Goodput is not all completed throughput and is not a universal property of a model. It depends on workload and success thresholds.
• Learn it: Inference Metrics and Goodput
• Related terms: Service Level Objective (SLO), Time to First Token (TTFT), Time per Output Token (TPOT), Cost per Successful Task
• Sources: DistServe

GPT

• Category: Models & inference
• What people say: A generic name for any chatbot.
• What it actually means: Generative Pre-trained Transformer, a family label for generative transformer models pretrained on sequence-prediction objectives and adapted for downstream use. Product names and model architectures should not be treated as interchangeable.
• Why it's called that: Generative describes output production, pre-trained describes the initial broad training stage, and transformer identifies the architecture family.
• Related terms: Transformer, Autoregressive, LLM (Large Language Model)

Graceful Degradation

• Category: Reliability & operations
• What it actually means: Preserving a bounded core service when capacity or dependencies are impaired by reducing optional quality, features, freshness, or workload instead of failing every request.
• Why it matters: AI systems often depend on several slow or fallible components, so an explicit reduced mode can protect essential user outcomes during partial failure.
• In practice: Predefine which capabilities may be disabled, keep the reduced mode visible to operators, protect safety checks, test the fallback under dependency failure, and restore full service deliberately. Tell users when correctness, safety, freshness, or a promised contract materially changes.
• Common confusion: Graceful degradation is not silently returning a worse answer as if nothing happened. Operators always need visibility; users need disclosure when the reduced mode materially changes the result or service contract.
• Learn it: Production LLM Application
• Related terms: Circuit Breaker, Load Shedding, Model Router, Availability
• Sources: Google SRE: Addressing Cascading Failures

Gradient

• Category: Math & training
• What people say: The slope of the loss.
• What it actually means: A vector of partial derivatives pointing in the direction of steepest increase. In ML, you go opposite to the gradient (gradient descent) to minimize the loss.
• Common confusion: Optimizers can transform, average, clip, or adapt gradients instead of taking a plain negative-gradient step.
• Related terms: Backpropagation, Gradient Descent, Optimizer

Gradient Accumulation

• Category: Math & training
• What it actually means: Summing or averaging gradients from several microbatches before performing one optimizer update.
• Why it matters: It lets you approximate a larger effective batch when one device cannot hold all examples and activations at once.
• In practice: Scale the loss consistently, call the optimizer only after the chosen number of microbatches, and measure whether normalization or distributed synchronization changes behavior.
• Common confusion: Gradient accumulation reduces per-step activation memory, but it does not reproduce every property of processing the full batch simultaneously.
• Related terms: Batch Size, Mixed Precision, Optimizer, Backpropagation
• Sources: PyTorch AMP examples: Gradient accumulation

Gradient Clipping

• Category: Math & training
• What it actually means: Limiting gradient values or their combined norm before an optimizer update when they exceed a chosen threshold.
• Why it matters: It can prevent an unusually large gradient from destabilizing a training step and producing non-finite values.
• In practice: Log unclipped norms, clip after unscaling mixed-precision gradients, and investigate repeated clipping instead of treating it as a substitute for diagnosing instability.
• Common confusion: Clipping controls update magnitude; it does not repair invalid data, a broken loss, or a consistently unsuitable learning rate.
• Related terms: Gradient, NaN (Not a Number), Mixed Precision, Learning Rate
• Sources: On the difficulty of training recurrent neural networks

Gradient Descent

• Category: Math & training
• What people say: Walking downhill on the loss surface.
• What it actually means: A family of optimization updates that move parameters using the negative gradient of an objective, usually estimated from batches rather than the entire dataset.
• Related terms: Gradient, Learning Rate, Optimizer

Grounding

• Category: Retrieval & generation
• What it actually means: Connecting a generated answer or action to evidence, state, or observations that the system can identify and check.
• Why it matters: Grounding gives the system a basis beyond unconstrained generation and makes unsupported claims easier to detect.
• In practice: Retrieve a policy section, require the answer to cite it, and reject claims that the cited passage does not support.
• Common confusion: Adding documents to a prompt creates an opportunity for grounding. It does not guarantee the model will use them correctly.
• Learn it: Retrieval-Augmented Generation
• Related terms: RAG (Retrieval-Augmented Generation), Hallucination, Verification Gate, Reranker

Guardrails

• Category: Evaluation & safety
• What people say: Safety filters around a model.
• What it actually means: System controls that constrain inputs, tool use, outputs, permissions, and escalation. They can include schemas, policy checks, classifiers, allowlists, sandboxing, approvals, and post-action verification.
• Why it matters: No single filter covers all failure modes, so controls should be layered according to risk.
• Common confusion: Guardrails reduce risk; they do not prove that an AI system is safe.
• Learn it: Guardrails
• Related terms: Least Privilege, Approval Gate, Sandbox, Evaluation (Eval)