Section 1 Quiz

--- shuffle_answers: true shuffle_questions: false --- ## What was the key breakthrough introduced by the Transformer architecture in 2017? > Hint: Think about how earlier models processed text compared to Transformers. - [ ] It dramatically reduced the volume of training data required > Not quite. Transformers actually require vast amounts of training data. Their breakthrough was in how they process that data. - [x] It let models process a whole sequence at once, using attention for context > Correct! Transformers introduced attention mechanisms that allow the model to analyze relationships between all words in a sequence at once, rather than processing text sequentially like RNNs and LSTMs. - [ ] It removed the need for GPU hardware during model training > Incorrect. Transformers are computationally intensive and rely heavily on GPU acceleration. - [ ] It allowed models to think and reason the way humans do > Not accurate. While Transformers produce impressive results, they predict patterns rather than truly "thinking." Reasoning capabilities came later with specialized reasoning models. ## Which of the following best describes the distinction between multimodal capability and reasoning capability? > Hint: They describe two different things about a model. Consider whether a single model could have both. - [ ] Reasoning models are invariably larger than multimodal ones > Size doesn't define either capability. A small model can be given a high reasoning effort setting, and a very large model may not be multimodal at all. - [ ] Multimodal means images only; reasoning means text only > This is an oversimplification. Multimodal means handling several input and output types together, and reasoning refers to deliberating before answering -- neither restricts the model to a single data type. - [x] Multimodality is about input and output *types*; reasoning is about *deliberation* > Correct! These are two independent axes, not competing categories. Multimodality describes what a model can take in and produce -- text, images, audio, video. Reasoning describes how much intermediate "thinking" it generates before its visible answer, which is increasingly an adjustable effort setting rather than a separate product. - [ ] Reasoning is the next generation, and replaces multimodal models > Incorrect. They aren't sequential versions of one another. Today's frontier models are typically multimodal *and* support adjustable reasoning effort simultaneously. ## A company's AI chatbot confidently cites "Dr. Sarah Chen's 2024 study on neural scaling laws in Nature" -- but no such paper exists. What type of error is this? > Hint: Consider whether this error stems from training data or from the model's generation process. - [ ] A bias error inherited from the training data > Not quite. Bias errors reflect systematic patterns in training data, not fabricated citations. - [ ] An error caused by outdated training data > Incorrect. The issue isn't that the information is outdated -- the citation was never real. - [x] A hallucination -- a convincing but nonexistent source, fabricated > Correct! This is a classic hallucination: the model generated a specific, plausible-sounding citation that has no basis in reality. Hallucinations are particularly dangerous because they can be very detailed and convincing. - [ ] A prompt injection attack against the model > No. Prompt injection involves malicious user input manipulating the model's behavior, not the model spontaneously fabricating information. ## A healthcare company is deploying a customer-facing AI chatbot to answer patient questions about symptoms and medications. Which AI limitation should concern them MOST when prioritizing their risk mitigation strategy? > Hint: Consider which limitation could cause the most severe real-world harm in this specific scenario, not just which limitation is most common. - [ ] Bias in the training data, leading to unfair outputs across demographics > Bias is a real concern for healthcare AI, and biased medical advice could disproportionately affect certain populations. However, in a customer-facing symptom and medication chatbot, the more immediate and severe risk is the model confidently providing fabricated medical information that patients act on -- a single hallucinated drug interaction could cause direct physical harm. Bias mitigation is important but is a systemic issue addressed through training data curation, not the most urgent deployment risk. - [x] Hallucinations -- convincing but false medical claims that a patient may act on > Correct! Hallucinations pose the greatest risk in this scenario because fabricated medical information (e.g., invented drug interactions, nonexistent dosage recommendations, or false symptom-condition associations) could directly cause patient harm. The evaluation criteria: (1) consequence severity -- acting on false medical advice can cause physical injury or death; (2) difficulty of detection -- hallucinated medical facts often sound authoritative and plausible to non-experts; (3) user trust -- patients tend to trust chatbot responses as vetted medical information; (4) frequency -- LLMs regularly hallucinate specific details like dosages and interactions. While bias and prompt injection are real concerns, hallucination risk demands the highest priority because the harm is direct, immediate, and potentially irreversible. - [ ] Prompt injection attacks that manipulate the chatbot's medical responses > Prompt injection is a genuine security concern, and a manipulated medical chatbot could provide dangerous advice. However, prompt injection requires a malicious actor to actively exploit the system, whereas hallucinations occur spontaneously during normal operation -- every patient interaction carries hallucination risk without any attacker involvement. For a customer-facing medical chatbot, the baseline risk of hallucinations affecting all users outweighs the targeted risk of prompt injection attacks. - [ ] Adversarial inputs that cause the model to behave unpredictably > Adversarial inputs are a valid concern but represent a targeted attack vector requiring sophisticated expertise. In a patient-facing chatbot, the far more pressing risk is hallucinations that occur naturally during routine use. Every conversation carries the risk of fabricated medical information, while adversarial attacks require deliberate, skilled intervention and affect fewer interactions. ## Small Language Models (SLMs) are significant because they: > Hint: Think about what problem SLMs solve that large models cannot. - [ ] Are consistently more accurate than large models > Incorrect. SLMs trade some capability for efficiency -- they excel at focused tasks but may not match frontier models on the most complex problems. - [ ] Require no training data at all in order to function > All language models require training data. SLMs are trained on carefully curated datasets. - [x] Bring capable AI to devices with no connectivity, at a fraction of the cost > Correct! SLMs typically run in the 1B-15B parameter range, small enough for phones, laptops, and edge devices, and on narrow well-defined tasks a good small model often holds its own against a much larger one. Note that none of this makes them safer: safety training is generally weaker, compression can degrade it further, and an attacker holding the device holds the weights. - [ ] Have completely replaced large models for every use case > SLMs complement rather than replace large models. They excel at specific tasks, while frontier models handle the most complex analysis. ## Which of the following is NOT a vulnerability or limitation of current LLMs? > Hint: Consider which item doesn't represent a genuine risk or weakness. - [ ] Biases inherited from training data, leading to unfair outputs > This is a real limitation -- LLMs can reflect and amplify biases present in their training data. - [ ] Susceptibility to prompt injection, where malicious input steers output > This is a real vulnerability -- prompt injection is a well-documented attack vector against LLMs. - [x] Autonomously updating their own weights mid-conversation to improve > Correct! This is NOT a real capability or limitation. LLMs do not update their weights during inference (conversation). Weights are fixed after training. This is a common misconception. What appears as "learning" in conversation is actually context management within the conversation window. - [ ] Generating hallucinations -- convincing but false information > This is a real limitation -- hallucinations remain one of the most significant challenges with LLMs. ## Your team finds that an internal coding assistant recommends the same non-existent Python package on every run. Why does that repeatability make it a SECURITY problem rather than only a quality one? > Hint: Ask what an attacker can do with a wrong answer they can predict in advance. - [ ] Repeated wrong answers indicate the model is overfitting, so its weights need retraining > Repeatability here is a property of how the model samples likely tokens, not evidence of overfitting -- and retraining is not the lever. The security consequence does not depend on why the name recurs. - [x] An attacker can register the hallucinated name in advance and wait to be recommended > Correct! Consistent fabrication is predictable fabrication. Registering the package a model reliably invents turns a correctness defect into a supply-chain compromise -- slopsquatting. Because different models converge on the same invented names, switching model is not a defence. - [ ] Repeated output means the assistant has cached a poisoned response and is replaying it > Nothing is being replayed from a cache. The model regenerates the same name because that name is the likely continuation, which is why the behaviour survives a cache flush and a fresh session. - [ ] A recurring wrong answer will be caught in review faster than a varying one > This gets the direction backwards, and it is the intuitive trap. A name that appears every time reads as corroboration rather than as error, so repetition lowers scrutiny instead of raising it. ## The evolution of AI can be summarized as progressing through distinct stages. What is the correct ordering? > Hint: Think about how each stage built on the previous one's capabilities. - [ ] Deep learning, rule-based systems, machine learning, then Transformers > The order is wrong. Rule-based systems came first, predating machine learning. - [ ] Machine learning, Transformers, deep learning, then rule-based systems > Incorrect. This reverses the actual progression. - [x] Rule-based systems, machine learning, deep learning, then Transformers > Correct! AI evolved from rigid rule-based systems to data-driven machine learning, then to deep learning with neural networks, and finally to the Transformer architecture that enabled the generative AI revolution we see today. - [ ] Transformers, deep learning, rule-based systems, then machine learning > This reverses the actual historical progression entirely.