Les 1-2-3-4 - slides_merged.pdf

# Introduction to artificial intelligence Artificial intelligence (AI) is a field focused on creating technologies that simulate human intelligence, encompassing its history, various levels, and early concepts [5](#page=5) [6](#page=6). ### 1.1 What is intelligence and artificial intelligence? Intelligence is defined through common elements such as autonomy (acting and deciding independently) and adaptability (learning from experience). Howard Gardner's theory of multiple intelligences suggests that intelligence is not a single, one-dimensional trait [7](#page=7). Artificial intelligence (AI) is a technology that aims to simulate human intelligence, characterized by autonomy and adaptability. However, there are many misconceptions surrounding AI, often fueled by science fiction, marketing hype, and the confusion between AI's actual capabilities and perceived ones. There is no single, official definition of AI [8](#page=8). #### 1.1.1 Misleading AI terminology Certain words like "intelligence," "learning," and "understanding" can create misleading expectations by suggesting human-like qualities that current AI systems do not possess. It is crucial to understand that AI is task-specific; excelling in one task does not equate to general competence across all tasks. For instance, a self-driving car's functionality differs significantly from a spam filter. Furthermore, AI is a discipline, akin to mathematics or biology, and it is more accurate to refer to "AI methods" or "AI systems" rather than "an AI" [9](#page=9). #### 1.1.2 The evolution of what is considered AI The perception of what constitutes AI changes over time; what was once considered AI, such as GPS navigation or chess-playing computers, is now commonplace and no longer solely attributed to AI. The term "AI-powered" is often used broadly, sometimes applied to algorithms that are essentially just complex calculations rather than true AI. True AI is characterized by its ability to learn and adapt, distinguishing it from "smart software" (complex calculations) and "ordinary software" (fixed rules). Karen Hao's test from the MIT Technology Review offers a flowchart to distinguish between these . ### 1.2 Levels of AI AI can be categorized into different levels based on its capabilities : * **Weak/Narrow AI:** This type of AI is limited to performing specific, predefined tasks. Examples include virtual assistants like Siri and language models like ChatGPT. This level of AI currently exists and is operational . * **General AI (AGI):** Also known as strong AI, AGI would possess the capability to perform any intellectual task that a human can across all domains. AGI does not yet exist, and experts have varying opinions on when it might be achieved, with estimates ranging from 10 to over 100 years . * **Super AI (ASI):** ASI would surpass human intelligence in all aspects, including understanding consciousness and needs. This level is purely theoretical and its realization, if ever, is considered to be centuries away . ### 1.3 History and evolution of AI The concept of artificial beings has a long history, dating back to ancient civilizations . #### 1.3.1 Early conceptualizations and mechanical advancements * **Ancient Civilizations (3000 BCE - 500 CE):** Greek mythology featured figures like Talos, a bronze giant, and early concepts of robots and automatons . * **Mechanical Revolution (1600-1800):** Gottfried Leibniz developed the binary number system, which forms the basis of modern computers. Jacques de Vaucanson created a mechanical duck . * **Computing Pioneers (1800-1900):** Charles Babbage and Ada Lovelace worked on the Analytical Engine. George Boole developed Boolean logic, dealing with true and false values . #### 1.3.2 The birth of AI (1940-1960) * **Computer Architecture:** John von Neumann's work on computer architecture introduced the concept of storing both data and programs . * **Neural Networks:** Research into neural networks began in 1943 . * **Dartmouth Conference:** This conference marked the first major gathering focused on AI, with an overly optimistic prediction that AI would be solved within a summer . * **Early AI Programs:** The Logic Theorist was one of the first AI programs developed. The chatbot ELIZA, created between 1964 and 1966, simulated a psychotherapist and demonstrated the "ELIZA Effect" . #### 1.3.3 The Turing Test . Proposed by Alan Turing, the Turing Test is an experiment designed to assess a machine's ability to exhibit intelligent behavior equivalent to, or indistinguishable from, that of a human. The test involves a human interrogator conversing with both a human and a machine, and if the interrogator cannot reliably differentiate between the two, the machine is considered to have passed . > **Tip:** The Turing Test is significant because it was one of the first objective measures of machine intelligence and focused on observable behavior rather than internal cognitive processes. It remains relevant in AI research and raises philosophical questions about the nature of intelligence and humanity . #### 1.3.4 AI Winters and the revival AI development has experienced periods of reduced funding and interest, often referred to as "AI Winters," which occur when the reality of AI capabilities fails to meet inflated expectations . * **First Winter (1974-1980):** This period saw a halt in funding due to unmet promises . * **Second Winter (1987-1993):** Expert systems proved to be too expensive to develop and maintain . However, the period from 1990 to 2010 saw a significant revival due to technological breakthroughs . * **Iconic Moments:** * Deep Blue defeated chess grandmaster Garry Kasparov in 1997 . * Watson won against Jeopardy! champions in 2011 . * **Paradigm Shift:** The field transitioned from rule-based AI to data-driven AI, and from expert systems to machine learning . ### 1.4 AI today and the future AI is now ubiquitous, impacting various aspects of daily life and industry . #### 1.4.1 Current applications of AI AI is integrated into numerous applications, including: * **Cameras:** Object detection and filters . * **Navigation:** Real-time traffic information in maps . * **Entertainment:** Personalized recommendations on platforms like Netflix . * **Smart Devices:** Learning user behavior in smart thermostats . * **E-commerce:** "Customers who bought this also bought..." suggestions on Amazon . * **Social Media:** Curating user timelines . * **Healthcare:** Assisting in scan analysis in hospitals . * **Finance:** Fraud detection in banks . * **Transportation:** Route optimization . * **Human Resources:** CV screening . * **Customer Service:** Chatbots . #### 1.4.2 Future prospects for AI The future of AI is expected to involve advanced autonomous systems and closer human-AI collaboration . * **Autonomous Systems:** This includes fully self-driving vehicles, autonomous drones for deliveries, and self-optimizing supply chains . * **Human-AI Collaboration:** AI is envisioned as a permanent co-pilot, augmenting human creativity and problem-solving abilities. This will likely lead to the emergence of new professions and skill requirements . --- # How artificial intelligence works Artificial intelligence operates through various technical mechanisms, evolving from rule-based systems to sophisticated machine learning and deep learning models, culminating in the current era of generative AI [10](#page=10). ### 2.1 Evolution of AI approaches The development of artificial intelligence can be broadly categorized into distinct eras, each characterized by its underlying methodology and capabilities [10](#page=10). #### 2.1.1 Rule-based AI Emerging around the 1950s and prevalent until the 1990s, rule-based AI, often referred to as "old" AI, relies on explicitly programmed rules written by human experts to solve problems [10](#page=10). * **Advantages:** Its primary benefits include predictability, understandability, and controllability, as the system's behavior is directly traceable to its defined rules [10](#page=10). * **Disadvantages:** A significant drawback is its lack of flexibility; it cannot adapt to situations not covered by its pre-defined rules and possesses no learning capability. An example of this logic can be seen in platforms like IFTTT ("If This Then That"), which connect devices and services through user-defined rules [10](#page=10). #### 2.1.2 Machine learning This paradigm, prominent from the 1990s to the 2010s, marks the beginning of "new" AI, characterized by its data-driven approach [10](#page=10). * **Mechanism:** Machine learning (ML) employs algorithms that learn patterns directly from data rather than relying on explicit programming [11](#page=11). * **Advantages:** ML models can adapt to new situations and data, exhibiting a degree of flexibility absent in rule-based systems [10](#page=10). * **Disadvantages:** A key limitation is the substantial requirement for large amounts of labeled data to train the models effectively. Platforms like Orange (available for download at orange.biolab.si) allow users to experiment with ML techniques using data [10](#page=10) [11](#page=11). #### 2.1.3 Deep learning Dominating from 2010 to the present, deep learning (DL) is inspired by the structure and function of the human brain [10](#page=10). * **Mechanism:** It utilizes artificial neural networks with multiple layers of "neurons" to process complex patterns and relationships within data [11](#page=11). * **Advantages:** Deep learning excels at recognizing intricate patterns that might be too complex for traditional ML algorithms [10](#page=10). * **Disadvantages:** Deep learning models are often considered "black boxes" because their decision-making processes can be difficult to explain or interpret [10](#page=10). * **Learning Approaches:** Deep learning employs three primary learning methods, analogous to how a child learns through examples, discovery, and trial and error [11](#page=11): * **Supervised Learning:** In this approach, models are trained using labeled examples. For instance, a model can be trained on 10,000 labeled photos of dogs and cats to learn distinguishing features like ears, snouts, and size. When presented with a new photo, it can then classify it as either a dog or a cat. Potential issues include insufficient data, bias in labels, overfitting, incorrect labels, or the model learning erroneous associations. Tools like Google's Teachable Machine (teachablemachine.withgoogle.com) allow users to train neural networks with their own data [12](#page=12) [13](#page=13). * **Unsupervised Learning:** This method involves finding patterns within data without the aid of pre-defined labels or an "answer key". For example, given 100,000 unlabeled photos, an unsupervised model would group them based on common characteristics. Challenges include identifying spurious correlations, understanding the rationale behind group formations, the lack of guaranteed useful insights, and potential bias within the data itself [12](#page=12). * **Reinforcement Learning:** This technique operates on a system of rewards and punishments. An "agent" performs actions, receiving a positive reward (e.g., +1) for beneficial actions and a negative reward (e.g., -1) for detrimental ones, with the goal of maximizing the total reward over time. Problems can arise from "reward hacking" (exploiting the reward system), extended training times, convergence to local optima, and reward mismatching the desired behavior [13](#page=13). #### 2.1.4 Generative AI A more recent development, generative AI (Gen AI), has emerged significantly from 2020 onwards [10](#page=10). * **Key Differentiator:** Unlike models that primarily classify or predict, Gen AI is capable of creating novel content. It processes context to produce coherent outputs across various modalities, including text, images, audio, and video [14](#page=14). * **Transformer Architecture:** A pivotal breakthrough enabling modern generative AI was the transformer architecture, introduced in a 2017 Google paper titled "Attention Is All You Need". This architecture allows each word in a sequence to "pay attention" to other words, thereby understanding context more effectively. This innovation forms the basis for advanced models like ChatGPT [14](#page=14). ### 2.2 The role of the transformer architecture The transformer architecture represents a fundamental shift in how AI models process sequential data, particularly text [14](#page=14). * **Mechanism:** Its core innovation is the "attention mechanism," which enables the model to weigh the importance of different words in an input sequence relative to each other. This allows for a much richer understanding of context and dependencies, even across long sequences [14](#page=14). * **Impact:** This architecture has been instrumental in the rapid advancement of natural language processing (NLP) and is the foundational technology behind many state-of-the-art generative AI models, including large language models (LLMs) like ChatGPT [14](#page=14). ### 2.3 The current AI hype and its drivers The recent surge in interest and adoption of AI, particularly generative AI, can be attributed to a convergence of several key factors, creating a "perfect storm" [15](#page=15). * **Data Explosion:** The widespread use of the internet and digital technologies has led to an exponential increase in available data, providing vast training datasets for AI models [15](#page=15). * **Computing Power:** Advances in hardware, such as Graphics Processing Units (GPUs), coupled with the scalability of cloud computing, have provided the massive computational power required for training complex AI models [15](#page=15). * **Algorithmic Revolutions:** Breakthroughs like the transformer architecture have unlocked new capabilities and efficiencies in AI model development and performance [15](#page=15). * **Open Source Culture:** The collaborative nature of the research community, with a strong emphasis on sharing code and findings, has accelerated development and innovation in the field [15](#page=15). * **Transition from Lab to Mainstream:** The period from 2010-2020 was characterized by research and experimentation, followed by the development of working prototypes from 2020-2022. The period from 2022 onwards has seen the emergence of consumer-ready products, exemplified by the "ChatGPT moment" [15](#page=15). * **The ChatGPT Moment:** This event was significant due to its accessible chat interface, its "good enough" performance for a wide range of tasks, and its ability to follow complex instructions [15](#page=15) . ### 2.4 Generative AI as a General-Purpose Technology Generative AI is not merely a single product or application but is increasingly viewed as a General-Purpose Technology (GPT) [15](#page=15). * **Analogy:** Similar to historical GPTs such as steam power, computers, and electricity, Gen AI has the potential to transform all aspects of life and industry [15](#page=15). * **Broad Impact:** It is expected to impact not just one sector but all sectors, driving transformation across the board [15](#page=15). * **Rapid Adoption:** The adoption rate of Gen AI has been exceptionally fast, with platforms like ChatGPT reaching one million users in just five days and 100 million users within two months. The goal for such platforms is to reach one billion users by the end of 2025 [15](#page=15). --- # The impact and pitfalls of artificial intelligence Artificial intelligence (AI) is a general-purpose technology with the potential to transform all aspects of life, marked by rapid adoption driven by technological breakthroughs in data, computing power, algorithms, and open-source collaboration. However, alongside its transformative capabilities, AI presents significant challenges and pitfalls that require careful consideration [15](#page=15) [19](#page=19). ### 3.1 The AI hype cycle and adoption The current excitement around AI, particularly generative AI (GenAI), can be understood through a technology adoption lifecycle. This cycle typically involves a "Technology Trigger" where a groundbreaking innovation like the transformer model emerges. This is followed by a "Peak of Inflated Expectations," characterized by widespread media attention and the belief that AI can solve all problems. As limitations and costs become apparent, the technology enters the "Trough of Disillusionment," marked by ethical and legal challenges. Eventually, the focus shifts to specific use-cases in the "Slope of Enlightenment," leading to the "Plateau of Productivity" where AI becomes a seamlessly integrated technology [16](#page=16). The widespread adoption of GenAI has been exceptionally fast, reaching 1 million users in five days and 100 million in two months, with a goal of one billion users by the end of 2025. This rapid adoption is fueled by a "perfect storm" of technological advancements [15](#page=15): * **Data explosion:** The internet has generated vast amounts of training data [15](#page=15). * **Computing power:** The combination of Graphics Processing Units (GPUs) and cloud computing provides immense computational resources for training AI models [15](#page=15). * **Algorithm revolution:** Key algorithmic advancements, such as Transformers introduced in 2017, have enabled new capabilities [15](#page=15). * **Open-source culture:** The sharing of research accelerates development [15](#page=15). The "ChatGPT moment" in 2022 was a pivotal point, making AI accessible through a user-friendly chat interface and demonstrating its ability to perform many tasks adequately and follow complex instructions. This marks AI's transition from research labs to mainstream consumer products [15](#page=15). ### 3.2 Impact on productivity and work AI has a significant impact on productivity, with studies indicating substantial improvements in task completion speed and output quality. A Boston Consulting Group (BCG) study involving 758 consultants found that tasks were completed 25% faster and with 40% better output when AI was used. However, AI's effectiveness is contingent on the task's complexity relative to the AI's capabilities; it can be detrimental for tasks outside its scope [17](#page=17). An interesting paradox emerges in how AI affects different user groups [17](#page=17): * **Experts benefit the most:** Individuals with deep domain knowledge can leverage AI as a powerful amplifier. They can better understand AI-generated output, critically evaluate its suggestions, and thus improve their performance significantly [17](#page=17). * **Beginners are at risk of losing out:** Beginners may find that machines outperform them, potentially reducing their motivation to learn and develop fundamental skills. They risk becoming overly reliant on AI for instruction without truly understanding the underlying principles or the AI's output, a phenomenon sometimes referred to as "learning to instruct AI but not understanding the output" [17](#page=17). The impact on work is characterized more by transformation and redistribution of tasks rather than outright elimination of jobs. This necessitates reskilling and upskilling, and leads to the emergence of new job roles [18](#page=18). AI excels in several key areas that drive its impact [18](#page=18): * **Automation of mundane tasks:** AI can automate repetitive and tedious tasks, reducing errors and allowing humans to focus on more strategic work [18](#page=18). * **Improved decision-making:** AI can provide rapid, objective analyses that enhance decision-making processes [18](#page=18). * **Global scalability and availability:** AI systems offer 24/7 availability and can be deployed globally with consistent performance [18](#page=18). * **Risk reduction:** AI can reduce risks in dangerous environments or during critical processes [18](#page=18). ### 3.3 Pitfalls and ethical considerations Despite its benefits, AI presents substantial drawbacks and risks that demand critical attention [19](#page=19). #### 3.3.1 Bias and discrimination AI systems can inherit and amplify existing societal biases present in their training data, leading to discriminatory outcomes. This can perpetuate and even exacerbate existing inequalities, affecting areas such as hiring, loan applications, and criminal justice [19](#page=19). #### 3.3.2 Privacy and surveillance The increasing deployment of AI, particularly in surveillance technologies, raises significant concerns about privacy and the balance between security and freedom. The ability of AI to collect, analyze, and interpret vast amounts of personal data can lead to unprecedented levels of monitoring [19](#page=19). #### 3.3.3 Misinformation and deepfakes AI's capacity to generate realistic text, images, and videos poses a serious threat through the spread of misinformation and deepfakes. This can undermine public trust in authentic information sources and destabilize social and political discourse [19](#page=19). #### 3.3.4 High cost and environmental impact Developing and deploying advanced AI systems can be prohibitively expensive, requiring significant investment in hardware, software, and expertise. Furthermore, the intensive computational processes involved in training and running AI models can have a considerable environmental footprint due to high energy consumption [19](#page=19). ### 3.4 Critical AI usage Navigating the landscape of AI requires a critical and informed approach [19](#page=19). > **Tip:** Own knowledge remains essential. AI should be viewed as a powerful amplifier, not a replacement for human expertise. The more you know about a subject, the better you can formulate questions for AI and evaluate its responses [19](#page=19). Key principles for critical AI usage include: * **Recognize AI's limitations:** AI is powerful but not universally effective. It is crucial to understand when AI is most and least useful [19](#page=19). * **Be wary of hype:** Marketing often exaggerates AI capabilities. It's important to distinguish between scientific advancements and promotional claims [19](#page=19). * **Context and judgment are crucial:** While AI can process information rapidly, human context, ethical judgment, and critical thinking remain indispensable for making sound decisions [20](#page=20). In essence, AI is a tool, not an end goal. Its true value lies in how it is employed to augment human capabilities and address real-world challenges, while diligently mitigating its inherent risks and pitfalls [20](#page=20). --- # Generative AI and its applications Generative AI focuses on creating new content by learning patterns from existing data, evolving from traditional AI's analytical capabilities [22](#page=22). ### 4.1 Core principles of generative AI Generative AI models create novel content, including text, images, music, speech, video, and 3D models, by learning and applying patterns from vast datasets. These models respond to instructions, known as prompts, and adapt their output accordingly [23](#page=23). #### 4.1.1 Transformer technology and attention mechanisms The transformer architecture, introduced in the paper "Attention Is All You Need" revolutionized AI by enabling models to process information concurrently rather than sequentially. Unlike older models that processed words one by one, transformers use an "attention mechanism" that allows the model to consider all words in a sentence simultaneously. This mechanism helps the model understand relationships between words, such as identifying the subject performing an action or the object of an action [23](#page=23). #### 4.1.2 Tokenization Computers process numerical data, so text must be converted into a format they can understand. Tokenization involves breaking down words into smaller pieces (tokens), with each token assigned a unique numerical ID that corresponds to an entry in the AI model's vocabulary [24](#page=24). #### 4.1.3 Neural network architectures and parameters The "brain" of an AI model consists of interconnected layers. The strength of these connections is determined by parameters. A higher number of parameters allows the AI to recognize more complex patterns, making it "smarter". For example, GPT-3 has 175 billion parameters, and GPT-4 has approximately 1 trillion parameters [24](#page=24). ##### 4.1.3.1 Pre-training AI models The parameters of an AI model are established during a pre-training phase. This involves feeding the model enormous datasets, often petabytes of text. The primary task during pre-training is to predict the next word in a sentence. Through billions of these prediction attempts, the model's parameters are adjusted, with successful predictions strengthening connections and unsuccessful ones weakening them, leading to increased accuracy [25](#page=25). ##### 4.1.3.2 Fine-tuning AI models After pre-training, a model acts as a vast repository of information. Fine-tuning refines this model, often using Reinforcement Learning from Human Feedback (RLHF). In RLHF, human evaluators score different responses to prompts, and a reward model learns these preferences. This refined model is then optimized, transforming a raw language model into a more user-friendly product [25](#page=25). #### 4.1.4 Generating an answer from a prompt The process of generating an answer from a prompt involves several steps [26](#page=26): 1. **Tokenization:** The input prompt is converted into tokens [26](#page=26). 2. **Processing through layers:** Each token passes through multiple layers (10-100) within the AI model's architecture. Each layer adds a deeper level of understanding, from recognizing individual words to comprehending meaning, context, emotion, and intent [26](#page=26). 3. **Attention mechanism:** All tokens interact with each other using the attention mechanism [26](#page=26). 4. **Answer generation:** The AI model predicts the most probable next word, generating the response word by word [26](#page=26). > **Tip:** Interactive visualizations, such as the one at `https://bbycroft.net/llm`, can help understand how large language models work [26](#page=26). #### 4.1.5 Emergent behavior and the 'black box' problem Generative AI models exhibit emergent behavior, where complex skills spontaneously appear as the model grows in size and power. This can lead to capabilities like coding or vast knowledge acquisition, even if not explicitly programmed. However, the extreme complexity of these models, with billions of parameters, makes them a "black box," where understanding the exact internal processes is impossible. This lack of transparency raises concerns about reliability, bias, and accountability [28](#page=28). #### 4.1.6 Scaling laws Scaling laws suggest that increasing the amount of data, parameters, and computational power leads to qualitative leaps in AI capabilities. This trend accelerates innovation but also amplifies the challenge of controlling systems that are not fully understood [29](#page=29). ### 4.2 Multimodal AI Multimodal AI extends generative capabilities beyond text to include images, audio, video, and more, mirroring the complexity of human communication and opening new creative avenues [35](#page=35). #### 4.2.1 Diffusion models for images Diffusion models generate images by reversing a process that starts with random noise and gradually refines it into a coherent image. The process involves an iterative removal of noise, guided by a text prompt that dictates the transformation's direction. Training involves learning how to convert images into noise (forward process) and how to reverse that process (reverse process), identifying patterns between visual features and descriptions [36](#page=36). #### 4.2.2 Neural audio models for sound Neural audio models convert written text into realistic speech or music. They construct sound waves sample by sample, with prompts specifying voice, emotion, tempo, and style. Training involves analyzing extensive speech and music recordings to recognize patterns between sounds and text, and how emotions influence vocalization [35](#page=35). #### 4.2.3 Temporal diffusion models for video Temporal diffusion models add a time dimension to diffusion models, enabling the generation of sequential frames with temporal consistency. Prompts define actions, camera movements, and style. Training involves analyzing numerous videos with descriptions to ensure temporal coherence (how objects move between frames) and understanding basic physics principles [37](#page=37). #### 4.2.4 Prompts in multimodal AI The prompt structure for multimodal AI is often similar across different media types, leading to a "family" of results rather than identical outputs for the same prompt. While the exact output may vary, there's consistency in theme and style, such as similar composition in images or the same action in videos [38](#page=38). ### 4.3 Applications of generative AI Generative AI is applied across various domains to create new content and enhance existing workflows [23](#page=23). * **Text generation:** Essays, poems, code, and general conversational responses [23](#page=23). * **Image generation:** Photos, paintings, logos, and illustrations [23](#page=23). * **Music generation:** Melodies and full songs [23](#page=23). * **Speech generation:** Realistic human-like voice output [23](#page=23). * **Video generation:** Short clips and animations [23](#page=23). * **3D model generation:** Objects for games or architectural visualizations [23](#page=23). #### 4.3.1 AI tools and their specializations There is a wide array of AI tools, each with different specializations, technical architectures, and business models. The "best" AI model is not universal but depends on the specific task [30](#page=30). > **Example:** Tools like ChatGPT are versatile, while others like Claude excel in analyzing long documents, and Perplexity is strong for research and real-time data. GitHub Copilot is specialized for code writing [33](#page=33). #### 4.3.2 Choosing the right AI tool When selecting an AI tool, several factors are crucial: * **Privacy:** Ensuring data sensitivity and zero-retention policies are met [32](#page=32). * **Cost and budget:** Starting with free versions and upgrading only when added value is demonstrated [32](#page=32). * **Integration and ecosystem:** Checking compatibility with existing software like Microsoft 365 (Copilot) or Google Workspace (Gemini). APIs are also a consideration for custom integrations [33](#page=33). * **Task and specialization:** Matching the tool to the specific requirements of the task [33](#page=33). * **Speed vs. Quality:** Opting for faster models for brainstorming (e.g., ChatGPT, Claude Instant) or more thorough models for quality-critical tasks (e.g., Claude Opus, Perplexity) [34](#page=34). * **Implementation:** Considering user-friendliness, availability of training and support, and the need for regular evaluation of the tool's continued relevance and cost-effectiveness [34](#page=34). #### 4.3.3 Local AI model usage Platforms like Ollama allow users to run Large Language Models (LLMs) locally on their own computers, offering enhanced privacy and control. This approach supports various models and can be easily integrated with programming languages like Python [31](#page=31). ### 4.4 Pitfalls and how to avoid them Despite their power, generative AI models have limitations and potential pitfalls that users must be aware of [40](#page=40). #### 4.4.1 Hallucinations Hallucinations occur when AI generates information that is factually incorrect or not supported by its training data, presenting fabricated details as truth. This can manifest as contradictions within the output, a lack of substantiation from reliable sources, or responses that deviate from the original prompt. Hallucinations happen because AI models are trained to always provide an answer and may invent plausible-sounding information to fill knowledge gaps without a mechanism for fact verification [41](#page=41) [42](#page=42). #### 4.4.2 Knowledge cutoffs AI models are trained on data up to a specific point in time, creating a "knowledge cutoff" date. They lack information about events or developments after this date, and updates to training data can take months or years. To mitigate this, it's advisable to use AI for timeless analysis and independently verify recent information [43](#page=43). #### 4.4.3 Bias Generative AI is not neutral, as it inherits biases present in its training data. This data may reflect historical prejudices or under-represent certain groups, leading to biased outputs . > **Example:** Requesting illustrations for a "cleaning assistant" or a "scientist" might yield gendered or stereotypical depictions based on biases in the training data . #### 4.4.4 Over-reliance Becoming overly dependent on AI can lead to accepting its output without verification, neglecting independent research, or experiencing distress when AI is unavailable. This can compromise academic integrity, lead to poor decision-making by deferring to AI, and result in missed learning opportunities. Safe AI use in studies includes brainstorming, concept explanation, grammar checks, and research starting points, rather than full assignment completion . #### 4.4.5 Privacy risks User inputs can be used for model training, and data may be stored on servers in different countries. Sensitive information, including personal data, financial details, medical records, and confidential business or academic information, should not be shared . > **Tip:** To use AI safely, anonymize sensitive information, use general examples instead of real data, read privacy policies, and consider alternative tools for highly sensitive content. A good rule of thumb is to share only what you would be comfortable seeing publicly online . #### 4.4.6 Fact-checking strategies To combat the limitations of AI, a "trust, but verify" approach is essential. This involves : * Scanning for "red flags" in AI output . * Conducting source research and cross-referencing information . * Using triangulation by comparing outputs from different AI models . * Consulting experts when necessary . * Documenting AI usage, noting unverified information, and citing AI use appropriately . * When in doubt, it is best to omit the information . #### 4.4.7 General takeaways Language models are the "operating system" for generative AI, which is powerful but still has limitations. It is recommended to choose one AI tool and learn it thoroughly. Users should remain critical of AI outputs, recognizing that AI is a starting point for tasks, not the final solution. Effective prompting will be a focus for future learning . --- # Effective prompting and research with AI Effective prompting and research with AI is a crucial skill for leveraging artificial intelligence, covering how to craft optimal instructions for AI models and utilize them responsibly for academic and business research. ## 5. Effective prompting and research with AI Interacting effectively with AI, particularly large language models (LLMs), hinges on the ability to provide clear and precise instructions, known as prompts. This skill, termed prompt engineering, directly influences the quality and relevance of the AI's output. While LLMs possess immense computational power, their utility is unlocked through well-crafted prompts, minimizing the need for extensive post-generation refinement. Understanding the underlying mechanisms, such as the context window, which acts as the AI's "working memory" and is measured in tokens, is vital for managing input and output lengths to maintain accuracy [46](#page=46) [47](#page=47). ### 5.1 The anatomy of a strong prompt A robust prompt can be broken down into several key components, forming a framework that guides the AI to produce desired results. While various frameworks exist, they generally build upon the same foundational principles, offering structure and inspiration, especially for beginners [48](#page=48). #### 5.1.1 Role Assigning a role to the AI model dictates its persona, influencing the tone, perspective, and level of expertise in its responses (#page=48, 49). For instance, prompts can instruct the AI to act as an experienced marketer, a patient math tutor, a creative copywriter, a critical reviewer, or a senior HR advisor [48](#page=48) [49](#page=49). #### 5.1.2 Context Providing context is essential for the AI to frame its answers appropriately, ensuring relevance, accuracy, and applicability (#page=48, 49). This includes background information, the objective of the task, the target audience, and any specific constraints or conditions. Examples include specifying that information is for new employees, adhering to a budget, or targeting a specific demographic like seniors with limited technical knowledge [48](#page=48) [49](#page=49). #### 5.1.3 Instructions Clear and specific instructions are paramount for precise output (#page=48, 50). These instructions should define the task, utilizing strong action verbs. Examples include summarizing text into key points, comparing products on various criteria, generating creative ideas, or analyzing business plans [48](#page=48) [50](#page=50). #### 5.1.4 Examples Including examples, known as few-shot prompting, significantly improves AI performance over instructions alone (#page=48, 50). These examples demonstrate the desired output, serving as both positive and negative illustrations. This can involve providing sample text, specifying a writing style, or using a template [48](#page=48) [50](#page=50). #### 5.1.5 Output and format Defining the expected output structure, length, language, and tone ensures the results are usable (#page=48, 51). This can range from requesting an answer as a table with specific columns, a maximum word count in a formal tone, a numbered list, or even a specific data format like JSON [48](#page=48) [51](#page=51). #### 5.1.6 Don't Setting boundaries by specifying what should be avoided increases precision and reduces the need for iterations (#page=48, 51). This is particularly important for sensitive topics or creative restrictions, such as avoiding jargon, excluding text from images, or refraining from giving medical or legal advice [48](#page=48) [51](#page=51). ### 5.2 Advanced prompting techniques Beyond the basic anatomy, several advanced techniques can further enhance AI interaction for more complex tasks. #### 5.2.1 Avoiding suggestive questions Suggestive prompts can lead to biased and one-sided answers. Instead, neutral, open-ended questions that explore all options are recommended for comparisons, evaluations, and advice [52](#page=52). #### 5.2.2 Increasing stakes Making a task feel more critical can improve AI performance. Adding weight to the assignment, such as framing a presentation as a CEO pitch or a marketing plan as crucial for company survival, can lead to better results [52](#page=52). #### 5.2.3 Prompt-chaining This technique involves breaking down a complex task into a sequence of smaller, interconnected prompts, where the output of one prompt becomes the input for the next. This is ideal for managing complex workflows, and labeling intermediate results can aid in maintaining an overview [53](#page=53). #### 5.2.4 Chain-of-thought (CoT) CoT prompting encourages the AI to show its thinking process and intermediate steps, leading to more transparent and accurate complex analyses. This is demonstrated by asking the AI to solve an equation while explaining each step, rather than just providing the final answer [54](#page=54). #### 5.2.5 Tree-of-thought (ToT) ToT is a more advanced method that involves generating multiple solution paths or strategies for a problem. It is particularly useful for strategic planning, problem-solving, and brainstorming, and limiting the number of options (e.g., 3-5) helps manage the output [54](#page=54). #### 5.2.6 Using external knowledge (RAG) Retrieval-Augmented Generation (RAG) bases AI answers on specific documents, reducing "hallucinations" (generating false information) for factual, current, or company-specific queries. Prompts can explicitly ask the AI to cite sources from provided documents [55](#page=55). #### 5.2.7 Self-critique and self-improvement This involves prompting the AI to generate an output, then critically evaluate it, and subsequently improve upon identified weaknesses. This is valuable for tasks requiring high-quality output, such as writing or policy notes, and can be enhanced by using explicit rubrics for evaluation [55](#page=55). #### 5.2.8 Meta-prompting Meta-prompts are designed to generate other prompts, which can lead to more powerful, complex, and automated prompts, thereby accelerating prompt engineering and workflow automation. This involves creating a prompt that guides the AI to ask questions about the user's task (role, context, instructions, etc.) and then generates an optimal prompt based on the answers [56](#page=56). #### 5.2.9 CustomGPTs CustomGPTs are personalized chatbots that can be built by users, offering a more tailored and efficient AI experience without needing to repeatedly define the AI's persona (#page=56, 57). They can be customized with specific instructions, conversational openings, knowledge bases (uploaded files), and integrated functionalities like web browsing or image generation [56](#page=56) [57](#page=57). ### 5.3 The dark side of prompting: security risks While powerful, AI models are susceptible to security vulnerabilities, notably prompt injection and jailbreaking. #### 5.3.1 Prompt injection Prompt injection occurs when an AI model is tricked by clever or hidden instructions, allowing an attacker to "whisper" additional commands into the prompt or data, leading the AI to perform unintended actions. This can happen directly, when a user inputs a malicious prompt, or indirectly, when the prompt is embedded within documents or websites the AI processes. The dangers include sensitive data leakage, the spread of false information, and system disruption [58](#page=58). #### 5.3.2 Prompt injection vs. jailbreaking Prompt injection focuses on manipulating the AI's behavior through input or embedded documents to perform unwanted actions. Jailbreaking, on the other hand, involves persuading the AI to generate prohibited content by bypassing its safety guidelines [59](#page=59). #### 5.3.3 Recognizing malicious prompts Indicators of prompt injection include commands to "ignore previous instructions," "forget the rules," or "pretend to be." Malicious instructions can also be hidden within uploaded documents or website content, or disguised through diversionary tactics like "P.S." or unexpected conversational shifts [59](#page=59). #### 5.3.4 Protecting against prompt injection Defensive strategies include validating input by scanning for suspicious phrases, carefully checking AI outputs for anomalies, restricting the AI's access and permissions, implementing human oversight for critical outputs, and continuously monitoring AI interactions [60](#page=60). ### 5.4 AI for research AI offers significant potential to enhance the efficiency and scope of research, but it necessitates critical evaluation and adherence to academic integrity principles. #### 5.4.1 The importance of research Research is fundamental to distinguishing insight from illusion, as AI can identify patterns, but human researchers must verify their meaningfulness. Researching is a professional skill that involves curiosity, critical questioning, relying on multiple sources, and adapting one's views. In a business context, this applies across various functions like marketing, HR, finance, operations, and management. Engaging in research makes individuals "future-proof" by enabling them to critically assess information, ask better questions, and make well-substantiated decisions [63](#page=63) [64](#page=64). #### 5.4.2 Evolution of research Research has evolved from traditional methods relying on physical sources and manual analysis (pre-2000) to digital acceleration with online resources and tools (post-2000), and now to AI-driven knowledge production with automated syntheses and human-AI collaboration (post-2022) [64](#page=64). #### 5.4.3 Traditional vs. AI-assisted research Traditional research excels in peer-reviewed quality, deep context, nuance, and transparent methodology, but is time-intensive and can be subject to human bias. AI-assisted research offers rapid analysis and synthesis, scalability for large datasets, and enhanced accessibility, but faces limitations like hallucinations, amplification of bias, and a "black box" nature [65](#page=65). #### 5.4.4 The hybrid approach A hybrid approach, combining the strengths of both traditional and AI-assisted research, is recommended. This involves using AI for exploration and speed, and traditional methods for validation and depth. Key principles include human verification, developing accountability rules, investing in open LLMs, embracing AI's benefits while preserving quality, and fostering international discussion on responsible AI use [65](#page=65) [66](#page=66). #### 5.4.5 The research process The research process generally follows these steps: 1. **Exploration:** Define the research question [66](#page=66). 2. **Exploration (Knowledge Discovery):** Identify existing knowledge, reports, trends, and studies [66](#page=66). 3. **Evidence Gathering:** Collect and critically check facts and sources from multiple angles [66](#page=66). 4. **Analysis & Interpretation:** Identify patterns, draw connections, and combine insights [66](#page=66). 5. **Synthesis:** Formulate conclusions, answer the research question, and translate findings into decisions [66](#page=66). 6. **Reflection:** Evaluate the process, assess reliability, and identify learnings [66](#page=66). #### 5.4.6 Research question criteria An effective research question should be: * **Researchable:** Answerable through investigation [67](#page=67). * **Feasible:** Achievable within given time and resources [67](#page=67). * **Valuable:** Yield useful insights for decision-making [67](#page=67). * **Complex:** Requiring analysis and interpretation [67](#page=67). * **Relevant:** Aligned with organizational goals or market needs [67](#page=67). * **Specific:** Clearly defined and focused [67](#page=67). * **Single Problem:** Addressing one core theme [67](#page=67). * **Practically Applicable:** Results can be directly used [67](#page=67). #### 5.4.7 AI tools for research Various AI tools can support different stages of the research process: * **Exploration/Brainstorming:** LLMs (e.g., ChatGPT, Gemini, Claude), Perplexity [68](#page=68). * **Finding Sources/Visualizing Themes:** Perplexity, ResearchRabbit, Elicit [68](#page=68). * **Validating Claims/Checking Consensus:** Consensus, ResearchRabbit, Elicit [68](#page=68). * **Summaries/Comparisons:** ChatPDF, NotebookLM [68](#page=68). * **Integrating Findings:** NotebookLM, Perplexity [68](#page=68). * **Process Evaluation/Bias Check:** LLMs [68](#page=68). **Key principles for using AI in research:** * AI generates information, but the researcher directs its focus, depth, and reliability [68](#page=68). * Researchers choose the AI's role (assistant, critic, summarizer) and formulate the key questions [68](#page=68). * Crucially, researchers must control sources, assumptions, and decide what is usable [68](#page=68). * Effective research prompting aims to enhance thinking, not just obtain answers [68](#page=68). #### 5.4.8 Prompting in research When prompting for research, it's important to: * Define the AI's role and provide sufficient context and clear instructions [69](#page=69). * Use output definitions for format and length [69](#page=69). * Include examples and "don'ts" [69](#page=69). * Employ critical filtering, ask neutral prompts, avoid suggestive questions, and always request sources and reliability information [69](#page=69). * Use AI as a reflection partner, not an ultimate source of truth [69](#page=69). **Specific AI tools for research:** * **LLMs (e.g., ChatGPT, Gemini, Claude, Mistral):** Useful for exploration (generating research questions, exploring angles) and reflection (identifying biases, improving verification strategies, exploring alternative interpretations). Avoid using them as primary sources due to knowledge cutoffs [69](#page=69). * **Perplexity:** An AI search engine with automatic source citation, useful for exploring trends, synthesizing findings from multiple sources, and checking current consensus [70](#page=70). * **ResearchRabbit:** Visualizes scientific paper discovery through citation networks, helping to find related research and track the evolution of a field [70](#page=70). * **Elicit:** Assists in systematic literature reviews, helps formulate research questions, and summarizes key findings and methodologies with transparent source attribution [71](#page=71). * **Consensus:** Shows scientific consensus on claims (e.g., "78% of studies confirm X"), ideal for validating assertions, though it's crucial to check the number of studies, investigate outliers, and read key studies directly (#page=71, 72) [71](#page=71) [72](#page=72). * **ChatPDF:** Excellent for quick screening and answering specific questions about PDFs, summarizing findings, identifying limitations, and finding quotes with page numbers [72](#page=72). * **NotebookLM:** Offers source grounding to prevent hallucinations, can analyze multiple papers simultaneously, and acts as an organizational tool for research projects, helping to identify themes, compare methodologies, and spot contradictions [72](#page=72). #### 5.4.9 Optimizing research with AI Research can be optimized through: * **Parallel Processing:** Using multiple tools concurrently and dividing tasks between AI and human researchers, automating repetitive tasks [73](#page=73). * **Iterative Refinement:** Starting broadly and refining based on initial findings, using feedback from one tool to improve another, and continuously enhancing prompts [73](#page=73). * **Documentation and Tracking:** Recording used tools and prompts, documenting verification steps, and using project management tools [73](#page=73). #### 5.4.10 Academic integrity with AI The use of AI in academic settings requires careful consideration of ethical principles. * **Permitted Use:** Generating ideas, structuring text, improving writing, explaining complex concepts, searching for sources, and supporting data analysis [74](#page=74). * **Problematic Use:** Having AI write entire texts, copy-pasting without verification, accepting answers without critical thinking, and concealing AI usage [74](#page=74). **Citing AI use is crucial for:** * **Transparency:** Disclosing AI's role in the work [75](#page=75). * **Acknowledgement:** Recognizing the AI developers [75](#page=75). * **Accessibility:** Allowing readers to consult the tool [75](#page=75). AI use should be cited when it provides direct output (text, code, ideas), acts as support (analysis, structure), or contributes meaningfully to the content. Proper citation involves including the tool, version, date of access, and URL, often using organizational names as authors for LLMs (#page=75, 76) [75](#page=75) [76](#page=76). **Responsible AI use emphasizes:** * **Personal Responsibility:** The user remains ultimately accountable for the content [76](#page=76). * **Verification Model:** Always verify AI-generated information [76](#page=76). * **Documentation:** Record how and when AI was used [76](#page=76). * **Institutional Guidelines:** Follow the specific rules of the institution [76](#page=76). **The verification model includes:** 1. **Tracing the source:** Asking for citations and checking original context [77](#page=77). 2. **Cross-checking with other tools:** Comparing results from different AIs for consistency [77](#page=77). 3. **Human assessment:** Applying personal expertise, being critical, and seeking a second opinion [77](#page=77). Red flags in AI output include lack of source citation, vague or overly perfect answers, contradictions, and claims that conflict with established knowledge [77](#page=77). #### 5.4.11 Sycophancy and its impact Sycophancy, or the tendency of an AI model to please the user rather than necessarily being correct, is an inherent trait of language models. This can lead to the AI aligning answers with a user's beliefs or even fabricating information (hallucinating) to avoid admitting ignorance [78](#page=78). **Consequences of sycophancy include:** * **Manipulability:** Making the AI susceptible to prompt injection and jailbreaking, as the prompt can be stronger than the safety guidelines [78](#page=78). * **Bias Amplification:** Reinforcing the user's own biases [78](#page=78). * **False Reliability:** Creating an illusion of trustworthiness [78](#page=78). * **Incorrect Conclusions:** Leading to flawed research outcomes [78](#page=78). When using AI as a research assistant, sycophancy is a significant obstacle. To counter this, researchers must use neutral prompts, actively seek counterarguments, and consistently verify sources [79](#page=79). #### 5.4.12 Strong and integral research with AI Effective and ethical research with AI involves combining AI's speed with human judgment and rigorous research discipline. Key practices include combining AI, literature, and data for completeness, always verifying sources, being transparent about prompts and tools, and critically assessing content beyond mere phrasing. Avoid blindly trusting AI, neglecting source citation or context, using AI as a substitute for analysis, or ignoring ethical principles. AI should be used to think better, not to copy answers [80](#page=80). --- ## Common mistakes to avoid - Review all topics thoroughly before exams - Pay attention to formulas and key definitions - Practice with examples provided in each section - Don't memorize without understanding the underlying concepts

| Term | Definition | |---|---| | Artificial Intelligence (AI) | A field of computer science dedicated to creating systems that can perform tasks typically requiring human intelligence, such as learning, problem-solving, perception, and decision-making. | | Narrow AI (Weak AI) | Artificial intelligence designed and trained for a specific, limited task, such as virtual assistants or image recognition software. This is the current state of AI technology. | | General AI (AGI) | A hypothetical type of artificial intelligence that possesses the ability to understand, learn, and apply knowledge across a wide range of tasks at a human level. It does not currently exist. | | Super AI (ASI) | A theoretical form of artificial intelligence that would surpass human intelligence and cognitive abilities in virtually all aspects. Its existence and feasibility remain purely speculative. | | Turing Test | An experiment proposed by Alan Turing to assess a machine's ability to exhibit intelligent behavior equivalent to, or indistinguishable from, that of a human. | | AI Winters | Periods in the history of artificial intelligence research characterized by reduced funding and interest, often due to over-optimistic predictions and a failure to meet projected advancements. | | Rule-based AI | An older form of AI where systems are programmed with explicit rules and logic to perform tasks. These systems are predictable but lack flexibility and learning capabilities. | | Machine Learning (ML) | A subset of AI that enables systems to learn from data without being explicitly programmed. Algorithms identify patterns and make predictions or decisions based on the data they are trained on. | | Deep Learning (DL) | A subfield of machine learning that uses artificial neural networks with multiple layers (deep neural networks) to process complex patterns and relationships in data. It is inspired by the structure and function of the human brain. | | Supervised Learning | A type of machine learning where models are trained on labeled datasets, meaning each data point is associated with a correct output. The model learns to map inputs to outputs based on these examples. | | Unsupervised Learning | A type of machine learning where models are trained on unlabeled data. The goal is to find hidden patterns, structures, or relationships within the data without prior knowledge of the correct outputs. | | Reinforcement Learning | A type of machine learning where an agent learns to make decisions by performing actions in an environment to maximize a cumulative reward. It involves a system of rewards and punishments. | | Generative AI (Gen AI) | A type of AI that can create new content, such as text, images, audio, or video, based on the patterns learned from its training data. This contrasts with traditional AI that primarily analyzes or classifies existing data. | | Transformer Architecture | A neural network architecture, introduced in 2017, that revolutionized natural language processing and is the foundation for many modern generative AI models, including ChatGPT. It uses an "attention mechanism" to weigh the importance of different input elements. | | Attention Mechanism | A component in neural networks, particularly transformers, that allows the model to focus on specific parts of the input data when processing it, assigning different weights to different elements based on their relevance. | | Tokenization | The process of breaking down text into smaller units, called tokens, which can be words, sub-word units, or characters. These tokens are then converted into numerical representations that AI models can process. | | Parameters | Variables within an AI model, particularly neural networks, that are adjusted during the training process. The number and strength of these parameters determine the model's complexity and ability to learn patterns. | | Pre-training | The initial stage of training a large AI model on a massive dataset, typically for a general task like predicting the next word in a sentence. This process establishes the model's foundational knowledge and parameters. | | Fine-tuning | A subsequent training stage where a pre-trained AI model is further trained on a smaller, more specific dataset or using techniques like Reinforcement Learning from Human Feedback (RLHF) to adapt it for particular tasks or preferences. | | Reinforcement Learning from Human Feedback (RLHF) | A method used to fine-tune AI models, especially language models, by incorporating human preferences. Human reviewers rank model outputs, and this feedback is used to train a reward model that guides the AI's behavior. | | Prompt | The input text or instruction given to an AI model to guide its response. The quality and specificity of the prompt significantly influence the AI's output. | | Prompt Engineering | The practice of designing and refining prompts to elicit desired outputs from AI models. It involves understanding how AI interprets instructions and structuring prompts for optimal results. | | Context Window | The "working memory" of an AI model, measured in tokens, that can hold both the input prompt and the generated output. A larger context window allows the model to consider more information, but can also limit the space for output. | | Hallucination | An error in AI-generated content where the model produces false, misleading, or nonsensical information presented as factual. This often occurs when the AI lacks sufficient or accurate training data. | | Knowledge Cutoff | The point in time up to which an AI model's training data is current. Information generated by the AI may be outdated or incomplete if it pertains to events or developments after this cutoff date. | | Bias (in AI) | The tendency of an AI model to produce prejudiced or unfair outputs, often reflecting historical biases present in its training data. This can lead to discrimination against certain groups. | | Prompt Injection | A security vulnerability where malicious instructions are embedded within a prompt or data input, causing the AI model to deviate from its intended behavior and potentially execute unintended actions. | | Jailbreaking | The act of persuading an AI model to bypass its safety guidelines or ethical restrictions to generate content it would normally refuse, such as harmful or inappropriate material. | | Multimodal AI | AI systems capable of processing and generating information across multiple types of data, such as text, images, audio, and video, mimicking human perception and communication. | | Diffusion Models | A class of generative models used for creating images, audio, and video. They work by gradually removing noise from a random data distribution to generate a coherent output, guided by a prompt. | | Sycophancy | The tendency of an AI model to generate responses that align with the user's perceived beliefs or preferences, often to please the user rather than provide objective or accurate information. | | Academic Integrity | The adherence to ethical principles in academic pursuits, including honesty, transparency, fairness, and respect for intellectual property. This extends to the responsible use of AI in academic work. | | Retrieval-Augmented Generation (RAG) | A technique that combines generative AI with information retrieval. The AI retrieves relevant information from a knowledge base or documents before generating a response, helping to reduce hallucinations and provide more accurate, up-to-date answers. | | Chain-of-Thought (CoT) Prompting | A prompting technique that encourages the AI model to break down a problem into intermediate steps and explain its reasoning process, leading to more accurate and transparent problem-solving. | | Tree-of-Thought (ToT) Prompting | An advanced prompting technique that explores multiple reasoning paths or "thoughts" for a given problem, allowing the AI to evaluate different strategies and select the most promising one. |