Generate 30 questions (10 easy, 10 medium, 10 complex) along with answers from this PDF
Here are 30 questions (10 Easy, 10 Medium, 10 Complex) based on the provided PDF content, along with their answers and page references:
Easy Questions
What term describes AI models like GPT-4 and Gemini?
Answer: Foundation Models.
Page: 9
What enables language models to scale into LLMs?
Answer: Self-supervision.
Page: 6
What are the two main types of language models?
Answer: Masked language models and autoregressive language models.
Page: 4
What is the basic unit of a language model?
Answer: Token.
Page: 3
Name one common AI engineering technique for adapting models.
Answer: Prompt engineering, RAG, or finetuning.
Page: 11
What is the most popular AI use case according to surveys?
Answer: Coding.
Page: 20
What does "human-in-the-loop" mean?
Answer: Involving humans in AI decision-making processes.
Page: 31
What metric measures the time to generate the first token?
Answer: TTFT (Time to First Token).
Page: 33
Which company launched the code-completion tool GitHub Copilot?
Answer: GitHub (owned by Microsoft).
Page: 20
What does LMM stand for?
Answer: Large Multimodal Model.
Page: 9
Medium Questions
Why do language models use tokens instead of words or characters?
Answer: Tokens reduce vocabulary size, handle unknown words, and capture meaningful components (e.g., "cook" + "ing").
Page: 4
How does self-supervision overcome data labeling bottlenecks?
Answer: It infers labels from input data (e.g., predicting next tokens in text), eliminating manual labeling costs.
Page: 6–7
What distinguishes foundation models from traditional task-specific models?
Answer: Foundation models are general-purpose, multimodal, and adaptable to diverse tasks.
Page: 10
What are the three factors enabling AI engineering's growth?
Answer: General-purpose AI capabilities, increased AI investments, and low entry barriers.
Page: 12–14
How did the MIT study (2023) show ChatGPT impacted writing tasks?
Answer: Reduced time by 40%, increased output quality by 18%, and narrowed skill gaps between workers.
Page: 23
What is the "Crawl-Walk-Run" framework for AI automation?
Answer:
Crawl: Human involvement mandatory.
Walk: AI interacts with internal employees.
Run: AI interacts directly with external users.
Page: 31
Why are internal-facing AI applications (e.g., knowledge management) deployed faster than external-facing ones?
Answer: Lower risks (data privacy, compliance, failures) while building expertise.
Page: 19
What challenge does AI's open-ended output pose for evaluation?
Answer: Lack of predefined ground truths makes measuring correctness difficult (e.g., for chatbots).
Page: 44
How did prompt engineering affect Gemini's MMLU benchmark performance?
Answer: Using CoT@32 (32 examples) instead of 5-shot boosted Gemini Ultra from 83.7% to 90.04%.
Page: 45
What are the three competitive advantages in AI startups?
Answer: Technology, data, and distribution.
Page: 32
Complex Questions
Why do larger models require more training data?
Answer: Larger models have higher capacity to learn; more data maximizes performance (not efficiency).
Page: 8
Explain how AI engineering workflows differ from traditional ML engineering.
Answer:
ML Engineering: Data → Model → Product.
AI Engineering: Product → Data → Model (due to pre-trained models enabling rapid iteration).
Page: 47 (Figure 1-16)
What ethical concern arises from AI-generated SEO content farms?
Answer: Proliferation of low-quality, automated content risks degrading trust in online information.
Page: 24
How did Goldman Sachs Research quantify AI investment growth by 2025?
Answer: $100B in the US and $200B globally.
Page: 13
What inference optimization challenges exist for autoregressive models?
Answer: Sequential token generation causes high latency (e.g., 100 tokens take ~1 second at 10ms/token).
Page: 43
Why might GPU vendor restrictions pose a "fatal" risk for AI products?
Answer: Bans on GPU sales (e.g., due to regulations) can cripple compute-dependent applications overnight.
Page: 35
How does the "data flywheel" create a competitive moat for AI startups?
Answer: Early market entry gathers usage data → insights improve products → attracts more users → reinforces data advantage.
Page: 32 (Footnote 21)
Contrast pre-training, finetuning, and post-training.
Answer:
Pre-training: Training from scratch (random weights).
Finetuning: Adapting a pre-trained model (application-level).
Post-training: Further training by model developers (e.g., instruction-tuning).
Page: 41–42
What are the three layers of the AI engineering stack?
Answer:
Application development (prompts, interfaces).
Model development (training, datasets, optimization).
Infrastructure (serving, compute, monitoring).
Page: 37
How did the 2023 Eloundou et al. study measure "AI exposure" in occupations?
Answer: % of tasks where AI reduces completion time by ≥50%. Fully exposed jobs included mathematicians, tax preparers, and web designers.
Page: 17
These questions and answers comprehensively cover key concepts, use cases, and technical nuances from Chapter 1 of the provided text.
No comments:
Post a Comment