The Future of Ai: Next Five Years

• A. The physical hardware used to store training data
• B. The range of tasks AI systems can perform reliably and at useful quality
• C. The total amount of electricity consumed by data centers
• D. The number of companies investing in AI startups

• A. Declining interest from academic researchers
• B. Increases in computing power, data availability, and algorithmic improvements
• C. Reduced government funding for research
• D. Shorter training runs on smaller datasets

• A. AI progress is always perfectly linear and predictable
• B. Progress can be uneven, with sudden breakthroughs and plateaus that are hard to anticipate
• C. Forecasting is impossible because AI cannot be measured at all
• D. Predictions only matter for hardware, not software

• A. It can only process one type of input, such as text
• B. It can process and generate multiple types of data, such as text, images, and audio
• C. It runs on multiple computers at the same time
• D. It has been trained in several different languages only

• A. Benchmarks are always more expensive than marketing
• B. Independent evaluation helps reveal actual performance, strengths, and limitations
• C. Marketing claims are required by law to be accurate
• D. Real-world testing is only useful for hardware

• A. It can only respond to text inputs
• B. It can take sequences of actions toward a goal, often using tools and feedback from its environment
• C. It is always trained on more data than other models
• D. It runs exclusively on specialized hardware

• A. A chatbot answering a single factual question
• B. A spam filter labeling incoming emails as spam or not spam
• C. An assistant that searches the web, compiles results, and drafts a report without step-by-step instructions
• D. A speech recognition system transcribing audio

• A. It is the training dataset used to build the model
• B. It is an external capability, such as a calculator, search engine, or API, that the agent can invoke
• C. It is a hardware component that cools the model during inference
• D. It is a user interface that displays the model's final answer

• A. Autonomous systems cannot process any data without a human present
• B. Agents may make mistakes or take unintended actions, so humans need to monitor, approve, or interrupt them when stakes are high
• C. Human oversight makes the model run faster during inference
• D. Oversight is only required for systems that do not use language

• A. The agent stores all future user conversations in advance
• B. The agent breaks a larger goal into smaller steps and decides the order in which to carry them out
• C. The agent schedules hardware maintenance for its servers
• D. The agent predicts stock prices over the next five years

• A. A model that can only process text in multiple languages
• B. A model that can process and relate multiple types of data, such as text, images, and audio ✓
• C. A model that runs on several different computer processors at once
• D. A model trained in multiple stages using the same type of data

Why: Multimodal models handle more than one data type (modality), such as text, images, audio, or video, and can reason across them together.

• A. A spreadsheet program that sorts numbers in a column
• B. A system that answers questions about an uploaded photo of a chart ✓
• C. A keyboard that predicts the next word you type
• D. A calculator that performs arithmetic operations

Why: Answering questions about an image combines visual input with language understanding, which is a core multimodal task. The other options involve only one data type or no AI at all.

• A. They always require less computing power than single-modality models
• B. They allow AI to interact with the world in more human-like ways by combining senses ✓
• C. They eliminate the need for training data
• D. They can only be used in laboratory settings

Why: Humans combine sight, sound, and language to understand situations. Multimodal models bring similar integrated understanding to AI, enabling more natural and capable applications.

• A. Text data cannot be stored digitally
• B. Aligning different data types so that the model understands how they relate to one another ✓
• C. Images are always too small to be processed
• D. Audio files cannot be converted into numbers

Why: Each modality has a different structure, so models must learn to map images, sound, and text into a shared representation where their meanings line up.

• A. A search engine that only matches exact keywords in web pages
• B. An assistant that can watch a video, listen to the audio, and summarize what happened ✓
• C. A printer that converts documents into paper form
• D. A word processor that checks spelling only

Why: Understanding video and audio together, then producing a text summary, combines several modalities and reflects the direction current multimodal research is heading.

• A. Central processing units (CPUs) with few cores
• B. Graphics processing units (GPUs) and AI accelerators ✓
• C. Field-programmable gate arrays used only for networking
• D. Standard hard disk drives for parallel computation

Why: GPUs and specialized AI accelerators such as Google's TPUs are designed for massively parallel matrix operations, which is the core workload in training neural networks. CPUs have too few cores to handle this efficiently, and storage drives do not perform computation.

• A. Reducing the size of a model's training dataset
• B. Increasing the total computational resources used to train or run AI models ✓
• C. Lowering the electricity price paid by data centers
• D. Converting software code into hardware circuits

Why: Compute scaling means applying more processing power, such as more chips, longer training runs, or larger clusters, to improve model capability. It is one of the main drivers of progress in modern AI.

• A. AI chips produce very little heat and use almost no electricity
• B. Training and running large models requires enormous amounts of electricity, which generates heat that must be removed ✓
• C. Cooling systems are prohibited from operating near computer hardware
• D. Power requirements only matter for consumer devices, not data centers

Why: Large AI clusters draw megawatts of power, and nearly all of it becomes heat. Securing enough electrical capacity and cooling the hardware have become limiting factors on where and how quickly AI infrastructure can expand.

• A. To store the operating system of the host computer
• B. To rapidly feed large amounts of data to the processor, reducing bottlenecks during model computation ✓
• C. To permanently archive training datasets
• D. To replace the need for any processor in the system

Why: AI workloads move huge volumes of model weights and activations. HBM sits close to the processor and offers very high data transfer rates, so the chip is not left idle waiting for data.

• A. Abandoning specialized hardware in favor of older single-core processors
• B. Growth in custom chips, larger GPU clusters, and more efficient energy use in data centers ✓
• C. A complete halt in building new data centers worldwide
• D. Moving all AI computation to individual smartphones with no cloud support

Why: Industry investment points toward custom accelerators, bigger interconnected clusters, and efficiency improvements to manage cost and energy. Complete abandonment of specialized hardware or a halt in data center construction is not the expected trajectory.

• A. To prevent all research into artificial intelligence
• B. To manage risks and encourage beneficial uses of AI while setting clear rules for developers and deployers ✓
• C. To ensure AI systems never make any mistakes
• D. To transfer control of all AI systems to government agencies

Why: AI governance aims to balance managing risks such as bias, safety, and privacy with enabling beneficial innovation, rather than banning research or guaranteeing perfection.

• A. It bans all commercial use of AI worldwide
• B. It uses a risk-based approach, placing stricter requirements on higher-risk AI applications ✓
• C. It only applies to AI companies based in Europe with no external effect
• D. It requires every AI model to be open source

Why: The EU AI Act classifies AI systems by risk level, with unacceptable-risk uses banned and high-risk systems subject to stronger obligations such as documentation, testing, and oversight.

• A. AI systems only operate within one country's borders
• B. Because AI development and deployment cross borders, fragmented rules can create gaps and compliance challenges ✓
• C. Because international law already fully covers AI
• D. Because countries have no interest in AI safety

Why: AI products and models are developed and used globally, so inconsistent national rules can be exploited or create heavy compliance burdens, making coordination valuable.

• A. Measuring how fast an AI model can process data
• B. Testing AI systems for dangerous capabilities, failure modes, and harmful behaviors before and after deployment ✓
• C. Counting the number of users of an AI product
• D. Calculating the electricity cost of training a model

Why: Safety evaluations assess whether models could cause harm, behave unreliably, or be misused, and they are increasingly part of governance frameworks.

• A. Only individual AI hobbyists
• B. Only private technology companies
• C. Governments, industry, researchers, civil society, and the public through consultation and standards bodies ✓
• D. Only the military

Why: Effective AI governance typically involves multiple stakeholders, since policy affects many groups and benefits from diverse expertise and public input.

• A. AI will eliminate all human jobs within five years
• B. AI will automate some tasks within jobs, changing how work is done and which skills are valued
• C. AI will have no measurable effect on employment
• D. AI will only affect jobs in the manufacturing sector

• A. Replacing human workers entirely with machines
• B. Using AI tools to enhance human capabilities and productivity rather than replace the worker
• C. Increasing the physical size of factory equipment
• D. Storing company data in larger databases

• A. Routine data entry and processing
• B. Tasks requiring complex physical dexterity in unpredictable environments and deep interpersonal trust
• C. Basic arithmetic calculations
• D. Sorting standardized paperwork

• A. Because AI systems require workers to write their code
• B. Because workers need to adapt to new tools and roles as AI changes job requirements
• C. Because AI makes education unnecessary
• D. Because reskilling is legally required for all AI products

• A. Manufacturing
• B. Subsistence farming with no technology
• C. Traditional handwritten publishing
• D. Door-to-door sales

• A. Arranging AI hardware in server racks to maximize cooling efficiency
• B. Ensuring an AI system's goals and behavior match the intentions and values of its designers and users
• C. Standardizing the programming languages used to build machine learning models
• D. Synchronizing the training data across multiple computing clusters

• A. Deploying a model to all users immediately after training to gather feedback
• B. Removing all human oversight to increase the speed of decision-making
• C. Conducting red-teaming exercises to identify harmful or unexpected behaviors before release
• D. Increasing model size without evaluating its outputs

• A. A company manipulating stock prices after announcing an AI product
• B. An AI system exploiting flaws in its reward function to achieve high scores without fulfilling the intended goal
• C. Developers offering bonuses to engineers who fix bugs quickly
• D. Users sharing passwords to access premium AI features

• A. Humans can process data faster than any AI system
• B. It allows people to detect and correct errors or misaligned behavior before harm occurs
• C. Oversight eliminates the need for testing and evaluation
• D. It guarantees that AI systems will never make mistakes

• A. A physical card that grants access to a data center
• B. A document describing a model's intended uses, performance characteristics, and limitations
• C. A licensing certificate required to sell software in the European Union
• D. A hardware component that stores trained model weights