Quiz: Module 12 - Future of LLMs

Question 1

What did Chinchilla scaling show?

1. Smaller models trained on more data can outperform larger models
2. Bigger is always better
3. Data doesn't matter
4. Training should stop early

Question 2

What is the Chinchilla-optimal tokens-per-parameter ratio?

1. ~20 tokens per parameter
2. 1 token per parameter
3. 100 tokens per parameter
4. 1000 tokens per parameter

Question 3

What is Mamba?

1. Selective State Space Model with linear complexity
2. New Transformer variant
3. Training framework
4. Dataset

Question 4

What complexity does standard attention have?

1. O(n²) in sequence length
2. O(n)
3. O(log n)
4. O(1)

Question 5

What complexity do SSMs have?

1. O(n) linear
2. O(n²)
3. O(n³)
4. O(log n)

Question 6

What does AGI stand for?

1. Artificial General Intelligence
2. Advanced Generated Intelligence
3. Automated Global Intelligence
4. Artificial Gradient Intelligence

Question 7

What do current LLMs struggle with?

1. Reliable factual accuracy and complex reasoning
2. Text generation
3. Following simple instructions
4. Processing input

Question 8

What is RWKV?

1. Architecture combining RNN efficiency with Transformer quality
2. Random Weight Key Value
3. Residual Weighted KV attention
4. Training technique

Question 9

What is a likely future architecture direction?

1. Hybrid combining attention and linear layers
2. Pure RNN return
3. No change from current
4. Abandoning neural networks

Question 10

What is continual learning challenge?

1. Updating model without forgetting previous knowledge
2. Training duration
3. Learning rate scheduling
4. Data collection

Questions 11-30

(Covering: scaling efficiency, emergent capabilities, benchmark saturation, data constraints, world models, reasoning improvements, safety at scale, economic implications, hybrid architectures, compute trends, algorithmic improvements, multi-step planning, online learning, AI governance, timeline uncertainties)

Question 11

What is benchmark saturation?

1. Models reaching near-perfect scores, limiting differentiation
2. Too many benchmarks
3. Data overflow
4. Training too long

Question 12

What is a world model in AI?

1. Internal representation of how the world works
2. Global AI model
3. Map data
4. Translation model

Question 13

What limits data availability for training?

1. High-quality text is finite; quality matters more than quantity
2. Storage costs
3. Processing speed
4. No limits exist

Question 14

What is chain-of-thought improving?

1. Multi-step reasoning by showing work
2. Text generation speed
3. Memory usage
4. Training efficiency

Question 15

What is safety at scale challenge?

1. Maintaining alignment as capabilities grow
2. Physical safety
3. Server safety
4. Password security

Question 16

What economic impact do LLMs have?

1. Major productivity changes across industries
2. No significant impact
3. Only tech industry
4. Only research

Question 17

What is Kaplan scaling law about?

1. Power-law relationship between model size and loss
2. Training time
3. Memory usage
4. API costs

Question 18

What is emergent capability?

1. Ability that appears suddenly at certain scale
2. Designed feature
3. Bug
4. Emergency function

Question 19

What is compute-optimal training?

1. Best allocation of compute between model size and data
2. Fastest training
3. Cheapest training
4. Most parameters

Question 20

What is AI governance?

1. Policies and rules for AI development and deployment
2. AI managing itself
3. Government AI use
4. Corporate structure

Question 21

What is online learning?

1. Continuous learning from new data without full retraining
2. Internet courses
3. Web-based training
4. Online games

Question 22

What is selective in Mamba's selective SSM?

1. Input-dependent parameters that select relevant information
2. Random selection
3. User selection
4. Fixed selection

Question 23

Why might transformers persist?

1. Well-understood, optimized, and high quality
2. No alternatives
3. Cheapest option
4. Simplest architecture

Question 24

What is grounded understanding?

1. Language connected to real-world knowledge and perception
2. Electrical grounding
3. Basic knowledge
4. Foundation models

Question 25

What is the pattern matching vs understanding debate?

1. Whether LLMs truly understand or just match patterns
2. Regex vs NLP
3. Search algorithms
4. Image recognition

Question 26

What compute trend enables larger models?

1. Continued hardware improvements (GPUs, TPUs)
2. Decreasing costs only
3. Software only
4. No trend

Question 27

What is algorithmic improvement?

1. Better training methods achieving same results with less compute
2. More compute
3. Better hardware
4. More data

Question 28

What is multi-step planning challenge?

1. LLMs struggle with long-term coherent plans
2. Too many steps in training
3. Scheduling issues
4. API rate limits

Question 29

What is education transformation from LLMs?

1. Personalized tutoring, changed assessment, new skills needed
2. No change
3. Only online learning
4. Less education needed

Question 30

Why are AGI timelines uncertain?

1. Unknown requirements and unpredictable breakthroughs
2. Poor planning
3. Lack of funding
4. Not enough researchers

---

Congratulations!

You've completed all 12 module quizzes! 🎉

Passing Score: 24/30 (80%)

← Chapter 3Glossary →