Essential AI Knowledge
What this domain actually tests
The Essential AI Knowledge domain checks whether you can reason about how AI workloads consume compute, memory, and data - not whether you can derive backpropagation by hand. You should be able to distinguish training from inference, explain a forward and backward pass, and describe why the backward pass costs roughly twice the forward pass.
Core concepts to own
Know the learning paradigms (supervised, unsupervised/self-supervised, reinforcement) and the major architectures: CNNs for spatial data, RNNs for sequences, and transformers with self-attention for language and multimodal models. Understand the training loop vocabulary: epoch, batch size, step, learning rate, loss, overfitting, regularization, and validation split. Be fluent in transfer learning and parameter-efficient fine-tuning (LoRA adapters), plus retrieval-augmented generation (RAG) as a way to add knowledge without retraining. Distinguish parameters from tokens - parameters describe model size; tokens describe workload volume and billing.
Precision formats matter. FP32 is the legacy default, TF32 accelerates FP32 matmuls on Ampere and later, BF16/FP16 are standard for mixed-precision training, and FP8/INT8 drive inference throughput. Lower precision cuts memory and raises throughput but can cost accuracy.
Traps
Questions mix up latency (time per request) with throughput (requests or tokens per second), assume training and inference have identical memory needs, or imply a bigger batch always improves latency.
Worked example
Estimate memory: weights bytes = parameters × bytes per element. A 7B model in FP16 is 7e9 × 2 B ≈ 14 GB. Mixed-precision Adam training needs roughly 16–20 bytes per parameter (weights, gradients, two optimizer states, activations), so ~112–140 GB - clearly more than one 80 GB GPU. Inference needs ~2 bytes per parameter plus KV cache.