Section 2 Quiz

--- shuffle_answers: true shuffle_questions: false --- ## DeepSeek V3 uses a Mixture-of-Experts (MoE) architecture with 671 billion total parameters but activates only 37 billion per request. What is the primary advantage of this approach? > Hint: Think about the relationship between total model size and per-request computational cost. - [ ] It makes the model smaller and easier to download > The total model size is still very large. MoE doesn't reduce the download size -- it reduces the computation per request. - [x] It provides the capability of a very large model while keeping per-request compute costs manageable > Correct! MoE architecture routes each token to specialized expert sub-networks, so you get the breadth of a 671B model but only pay the compute cost of a 37B model per request. This is why DeepSeek can offer frontier-level performance at dramatically lower cost. - [ ] It allows the model to run on consumer smartphones > Even 37B active parameters is too large for smartphone deployment. SLMs (1B-3B) are used for on-device AI. - [ ] It eliminates the need for GPU hardware entirely > MoE still requires GPU hardware for inference, just less of it per request than a dense model of equivalent capability. ## What is the key distinction between a foundation model and a fine-tuned model? > Hint: Consider how each type is created and what it's optimized for. - [ ] Foundation models are always open-source while fine-tuned models are proprietary > Licensing doesn't define the category. Both foundation and fine-tuned models can be open-source or proprietary. - [ ] Fine-tuned models always have more parameters than foundation models > Fine-tuned models are derived from foundation models and typically have the same parameter count. - [x] Foundation models are pre-trained on broad, diverse data for general tasks, while fine-tuned models adapt a foundation model to specific domains using targeted datasets > Correct! Foundation models like GPT-4o and Llama 3.3 are trained on massive general datasets. Fine-tuned models like Med-PaLM 2 or Qwen 2.5-Coder take a foundation model and further train it on domain-specific data for better performance on specialized tasks. - [ ] Fine-tuned models are built entirely from scratch without using any pre-trained weights > This describes custom-built models, not fine-tuned models. Fine-tuning explicitly leverages pre-trained foundation model weights. ## A company is evaluating whether to use a closed-source model (like Claude Opus 4) or an open-source model (like Llama 3.3 or DeepSeek R1). Which factor has changed most significantly in 2025/2026 compared to earlier years? > Hint: Think about what made open-source models less attractive in the past and what has changed. - [ ] Closed-source models have become much cheaper > While prices have decreased overall, this isn't the most significant change in the open vs. closed dynamic. - [x] Open-source models have closed the capability gap, competing with closed-source models on many benchmarks > Correct! The most significant shift is that models like DeepSeek R1 and Llama 3.3 now compete with proprietary models on many tasks. Organizations can choose based on data privacy, cost, and control rather than capability alone. - [ ] Open-source models have become closed-source > The trend is actually the opposite -- more models are being released with open weights. - [ ] Closed-source models now allow full weight inspection > Closed-source providers have not opened their model weights for inspection. ## Reasoning models (like o3, o4-mini, and DeepSeek R1) differ from traditional LLMs primarily because they: > Hint: Consider the fundamental difference in HOW these models generate their responses. - [ ] Have been trained on more data than traditional models > Training data volume doesn't define the reasoning model category. - [ ] Can process images and audio in addition to text > This describes multimodal models, not reasoning models specifically. - [x] Employ an internal deliberation process, breaking problems into steps and verifying their own work before presenting an answer > Correct! Reasoning models use "test-time compute" -- they spend more time thinking about complex problems, decomposing them into steps, evaluating approaches, and sometimes backtracking when they detect errors. This makes them dramatically better at math, coding, and multi-step analysis. - [ ] Respond faster than traditional models because of optimized architectures > Actually the opposite -- reasoning models typically take longer to respond because their deliberation process consumes additional tokens and time. ## A startup is building an offline translation app for humanitarian workers in remote areas with no internet connectivity. Given the providers and model types covered in this section, which approach should they choose? > Hint: Consider the deployment constraints (offline, on mobile devices) and match them to the model categories and providers discussed. - [ ] Use OpenAI's GPT-4o via API for its strong multilingual capabilities > GPT-4o has excellent multilingual performance, but it requires an internet connection to access via API. The scenario explicitly requires offline functionality in remote areas with no connectivity, making any cloud API approach unsuitable regardless of the model's language capabilities. - [ ] Deploy DeepSeek V3 (671B parameters) on a local server the team carries > DeepSeek V3 requires a cluster of GPUs to run, making it impractical for a portable field deployment. At 671B total parameters (37B active via MoE), it needs enterprise-grade infrastructure far beyond what humanitarian teams can transport to remote locations. - [x] Deploy a Small Language Model like Llama 3.2 (1B-3B) on the workers' mobile devices > Correct! This scenario requires offline functionality on portable devices -- exactly the use case Small Language Models are designed for. Llama 3.2 (1B-3B) can run on smartphones without internet connectivity, Meta's permissive license allows commercial use, and the 2025/2026 SLM landscape shows these models can match much larger models on focused tasks like translation. The trade-off in capability compared to frontier models is acceptable given the deployment constraints. - [ ] Fine-tune Claude Opus 4 on humanitarian terminology and deploy via Anthropic's API > While fine-tuning for the domain is a sound strategy, Claude Opus 4 is a closed-source model accessed only via API. It cannot be deployed offline or on mobile devices. The scenario's connectivity constraints make any API-dependent solution unworkable, regardless of how well the model is customized. ## A security team needs to evaluate an AI model before deploying it in a sensitive government application. Why would they favor an open-weight model over a closed-source API? > Hint: Think about what you can do with a model whose weights you can inspect versus one accessed only through an API. - [ ] Open-weight models are always more accurate than closed-source models > Accuracy varies by model and task -- this isn't a universal advantage of open-weight models. - [ ] Open-weight models don't require any computational resources > Open-weight models still require significant compute for inference, especially large ones. - [x] Open-weight models allow security teams to inspect model weights, audit behavior, and maintain full control over data processing within their own infrastructure > Correct! With open weights, security teams can examine the model itself, run it in air-gapped environments, and ensure sensitive data never leaves their infrastructure. This level of inspection and control is impossible with closed-source API-only access. - [ ] Open-weight models have built-in security features that closed models lack > Open-weight models may actually have fewer safety guardrails than carefully controlled closed-source models, as their weights can be modified post-release. ## Which statement about the current AI provider landscape is most accurate? > Hint: Consider the overall competitive dynamic in 2025/2026. - [ ] OpenAI dominates the market with no serious competition > The market is now highly competitive with multiple strong players. - [ ] Only US-based companies produce competitive AI models > This is factually incorrect -- DeepSeek (China), Qwen/Alibaba (China), and Mistral (France) are all competitive. - [x] The landscape is intensely competitive with significant contributions from companies across the US, China, and Europe, including both commercial and open-source offerings > Correct! The 2025/2026 AI ecosystem features strong competition from OpenAI, Anthropic, Google, Meta (US), DeepSeek, Alibaba/Qwen (China), and Mistral (France), with the open-source ecosystem blurring the line between commercial and community-driven development. - [ ] Open-source models have completely replaced commercial offerings > While open-source has narrowed the gap, commercial models still lead on some complex tasks and offer managed infrastructure.