Section 2 Quiz

--- shuffle_answers: true shuffle_questions: false --- ## A Mixture-of-Experts (MoE) model has hundreds of billions of total parameters but activates only a small fraction of them 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 on disk 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 buys the capability of a very large model at a manageable per-request cost > Correct! MoE routes each token to a handful of specialized expert sub-networks instead of running the whole network, so a model holding hundreds of billions of total parameters may activate only a small fraction of them per request. You get the breadth of the full model at a fraction of the per-request compute -- which is how providers with smaller budgets have been able to compete at the frontier. - [ ] It allows the model to run on a consumer smartphone > Even the activated fraction of a frontier MoE model is far too large for a phone. On-device AI uses Small Language Models, at the bottom of the size spectrum. - [ ] It removes the need for GPU hardware altogether > 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 and fine-tuned models proprietary > Licensing doesn't define the category. Both foundation and fine-tuned models can be open-source or proprietary. - [ ] Fine-tuned models invariably carry more parameters than foundation ones > Fine-tuned models are derived from foundation models and typically have the same parameter count. - [x] Foundation models are pre-trained broadly; fine-tuned models adapt one to a domain > Correct! Foundation models are trained on massive general datasets. Fine-tuned models like Med-PaLM 2 or Qwen-Coder take a foundation model and further train it on domain-specific data for better performance on specialized tasks. - [ ] Fine-tuned models are built from scratch, using no 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 proprietary API-only model or a self-hosted open-weight model. Which factor has changed most significantly in recent years? > Hint: Think about what made open-source models less attractive in the past and what has changed. - [ ] Closed-source models have become dramatically cheaper > While prices have decreased overall, this isn't the most significant change in the open vs. closed dynamic. - [x] Open-weight models have closed the capability gap on many benchmarks > Correct! The most significant shift is that leading open-weight models now compete with proprietary ones on many tasks. Organizations can choose based on data privacy, cost, and control rather than capability alone. - [ ] Open-source models have themselves become closed-source > Not as an ecosystem trend -- more capable open-weight models are available now than ever. Individual vendors do reverse course, as Meta did when it moved its frontier work to a proprietary successor, which is exactly why the section frames open weights as a bet on a vendor's strategy. But that is a vendor-level risk, not the direction of the field. - [ ] Closed-source models now permit full weight inspection > Closed-source providers have not opened their model weights for inspection. ## Running a request with extended reasoning enabled differs from a standard request primarily because the model: > Hint: Consider the fundamental difference in HOW the response gets generated. - [ ] Has been trained on more data than it uses for standard requests > Extended reasoning is a setting applied at request time. The same trained model handles both; nothing about its training changes. - [ ] Can process images and audio as well as plain text > That describes multimodality, a separate capability. A model can be multimodal with reasoning turned off, or text-only with it turned on. - [x] Deliberates internally, working in steps and checking itself before answering > Correct! This is "test-time compute" -- the model spends extra tokens on private working-out, decomposing the problem, evaluating approaches, and sometimes backtracking when it detects an error, before writing the visible answer. It buys markedly better results on math, coding, and multi-step analysis. - [ ] Responds faster, because reasoning uses an optimized architecture > The opposite. Those intermediate reasoning tokens take time to generate and are billed like any other, so extended reasoning is both slower and more expensive than a standard response. ## 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 a frontier model through a cloud API, for its multilingual strength > Frontier models have excellent multilingual performance, but reaching one requires an internet connection. 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 a frontier-scale open-weight model on a server the team carries > A frontier-scale model needs a cluster of GPUs, making it impractical for portable field deployment. Even with MoE reducing the active parameter count, it requires enterprise-grade infrastructure far beyond what humanitarian teams can transport to remote locations. - [x] Deploy a Small Language Model, at the low end of the range, on the workers' phones > Correct! This scenario requires offline functionality on portable devices -- exactly the use case Small Language Models are designed for. A model at the small end of the SLM tier runs on a smartphone with no connectivity at all, several are available under genuinely permissive licenses that allow commercial use, and small models hold up well on focused tasks like translation. The capability trade-off against a frontier model is acceptable given the deployment constraints. - [ ] Fine-tune a frontier model on humanitarian terminology, served by the vendor > While fine-tuning for the domain is a sound strategy, a proprietary model is 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 consistently more accurate than closed ones > Accuracy varies by model and task -- this isn't a universal advantage of open-weight models. - [ ] Open-weight models require no computational resources to run > Open-weight models still require significant compute for inference, especially large ones. - [x] Security teams can inspect the weights, audit behaviour and keep data in-house > 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 ship with 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 today. - [ ] One vendor dominates the market, with no serious competition > The market is now highly competitive with multiple strong players. - [ ] Only US-based companies produce genuinely competitive models > This is factually incorrect -- DeepSeek (China), Qwen/Alibaba (China), and Mistral (France) are all competitive. - [x] Intensely competitive, with US, Chinese and European labs all contributing > Correct! The AI ecosystem features strong competition from OpenAI, Anthropic, Google and Meta (US), DeepSeek and Alibaba/Qwen (China), and Mistral (France), with the open-weight ecosystem blurring the line between commercial and community-driven development. - [ ] Open-source models have entirely replaced commercial offerings > While open-source has narrowed the gap, commercial models still lead on some complex tasks and offer managed infrastructure.