Chapter 40: Quiz

Test your understanding of the future directions in AI engineering.

Question 1. Which of the following best describes the concept of "test-time compute scaling"?

(a) Training the model for more epochs to improve accuracy (b) Using larger batch sizes during inference to improve throughput (c) Allocating additional computation at inference to improve per-query accuracy (d) Reducing model size to speed up inference

Question 2. In the inference scaling law $\text{Accuracy}(c) \approx A - B \cdot c^{-\alpha}$, what does $c$ represent?

(a) The number of model parameters (b) The amount of training data (c) The inference compute (FLOPs or tokens) (d) The learning rate

Question 3. A world model approximates which function?

(a) The reward function $r(s, a)$ (b) The policy $\pi(a|s)$ (c) The transition dynamics $s_{t+1} = f(s_t, a_t)$ (d) The value function $V(s)$

Question 4. What is the primary challenge with multi-step world model predictions?

(a) The model becomes too large to run (b) Prediction errors compound over time (c) Quantum decoherence (d) The training data becomes stale

Question 5. In neurosymbolic AI, what advantage does a symbolic component provide over a purely neural system?

(a) Faster training on large datasets (b) Better image recognition accuracy (c) Formal guarantees and compositional generalization (d) Lower memory requirements

Question 6. Elastic Weight Consolidation (EWC) prevents catastrophic forgetting by:

(a) Replaying stored examples from previous tasks (b) Adding separate parameters for each new task (c) Penalizing changes to parameters important for previous tasks, using the Fisher information matrix (d) Freezing all model weights after the first task

Question 7. Which of the following is NOT a family of continual learning methods?

(a) Regularization-based methods (b) Replay-based methods (c) Architecture-based methods (d) Augmentation-based methods

Question 8. AlphaFold2 is a landmark AI application in which scientific domain?

(a) Weather forecasting (b) Drug synthesis (c) Protein structure prediction (d) Theorem proving

Question 9. In the context of AI agents, what is a "multi-agent system"?

(a) A single model with multiple output heads (b) Multiple specialized agents that coordinate to solve complex tasks (c) A model that runs on multiple GPUs simultaneously (d) An ensemble of identical models

Question 10. Which safety concern is MOST specific to autonomous AI agents (as opposed to static models)?

(a) Training data bias (b) Overfitting to the test set (c) Taking irreversible actions in the real world (d) High memory usage

Question 11. In the graduated AGI framework, a "Level 0" system is best described as:

(a) A system with no AI capabilities (b) A system that is superhuman at one specific task (c) A system that matches average human performance across all tasks (d) A system that exceeds expert human performance across all tasks

Question 12. What is a "variational quantum circuit" (parameterized quantum circuit)?

(a) A quantum circuit with fixed gates optimized during manufacturing (b) A quantum circuit with learnable parameters adjusted by a classical optimizer (c) A classical neural network simulating quantum behavior (d) A random quantum circuit used for benchmarking

Question 13. The "barren plateau" problem in quantum machine learning refers to:

(a) Quantum computers overheating during computation (b) Exponentially vanishing gradients in random parameterized quantum circuits (c) Insufficient training data for quantum models (d) The inability to measure quantum states without destroying them

Question 14. A "T-shaped" AI engineer profile means:

(a) Deep expertise in every area of AI (b) Broad knowledge across the field combined with deep expertise in one or two areas (c) Expertise only in the theoretical foundations (d) Knowledge only in practical deployment

Question 15. Which of the following is the BEST strategy for navigating AI hype cycles?

(a) Adopt every new technology immediately to stay ahead (b) Ignore all new developments and focus on proven methods (c) Read original papers, check evaluations, reproduce results, and wait for independent replication (d) Follow social media influencers for curated information

Answer Key

  1. (c) Test-time compute scaling allocates more computation during inference, not training, to improve accuracy on individual queries.

  2. (c) In the inference scaling law, $c$ represents inference compute, measured in FLOPs or tokens allocated to a single query.

  3. (c) A world model approximates the transition dynamics, predicting the next state given the current state and action.

  4. (b) Compounding error is the primary challenge: small per-step errors accumulate over multiple prediction steps, causing trajectory divergence.

  5. (c) Symbolic systems provide formal guarantees (verifiable correctness) and compositional generalization (systematic recombination of learned components).

  6. (c) EWC adds a regularization term weighted by the Fisher information matrix, which penalizes changes to parameters that were important for previous tasks.

  7. (d) The three main families are regularization-based, replay-based, and architecture-based. "Augmentation-based" is not a standard category of continual learning methods.

  8. (c) AlphaFold2 solved the protein structure prediction problem, predicting 3D protein structures from amino acid sequences.

  9. (b) A multi-agent system consists of multiple specialized agents (e.g., planner, researcher, coder, critic) that coordinate to solve complex tasks.

  10. (c) Autonomous agents with tool access can take irreversible real-world actions (deleting files, sending emails, making purchases), which is a unique safety concern not present in static models.

  11. (b) Level 0 (Narrow AI) describes systems that are superhuman at one specific task but cannot generalize to other tasks.

  12. (b) A variational/parameterized quantum circuit has tunable gate parameters that are optimized by a classical optimizer, analogous to trainable weights in a neural network.

  13. (b) Barren plateaus occur when the gradients of the cost function vanish exponentially with the number of qubits, making optimization intractable.

  14. (b) T-shaped means broad knowledge across the full AI landscape (horizontal bar) with deep expertise in one or two specializations (vertical stroke).

  15. (c) The most reliable strategy is to engage with primary sources, verify claims through evaluation and reproduction, and wait for independent confirmation.