--- # Learning objectives - Desribe the concept of agents - Use LLMs effectively by selecting appropriate prompting techniques - Discuss interaction modalities In the exam, you will be provided with a scenario and asked to write a prompt, identify the technique or interaction modality, and explain why this choice can be expected to be effective (ideally with reference to research studies). Students are provided with a conceptual overview of the prompting techniques and interaction modalities (teacher-centered instruction). Selected research papers are used to understand under which conditions they are effective (self-studying in small groups with short summaries). Examples are used to practice the application of prompts and assess their effectiveness (hands-on practice with short cases). Bild des Flickenteppich (wir wissen wenig, oft vermeintlich Wiederspruch)

TODO : start with Gartner and select genAI/agentic? - Generative AI and agentic AI applications are increasingly adopted in the industry (Gartner). https://www.gartner.com/en/articles/hype-cycle-for-artificial-intelligence

**TODO: SHORTEN/select**

BairdMaruping2021 (delegation) Note: bots vs. AI (technology is often hard to distinguish by users, may be a combination of rule-base, ML, reinforcement learning, LLM) - start with algorithmic management (role of (anthropomorphous) bots? / uncanny valley) - different forms of bots and capabilities (chatbot, ... )?

- Management by the IT department

tbd: handbook/obsidian/git Thought: LLM as a powerful capability, but its effectiveness depends on engineering the right environment https://www.gartner.com/en/articles/hype-cycle-for-artificial-intelligence?utm_source=chatgpt.com https://www.ibm.com/think/topics/shadow-ai Highlight LLMs as a key capability (connecting to the next slide) https://zapier.com/mcp Civilization advances by extending the number of important operations which we can perform without thinking about them. – Alfred North Whitehead, 1911 https://www.agentic.ai/

(provide summary afterwards?) - selected papers + summaries (whiteboard) Critically evaluate the strengths and weaknesses of interaction modalities (should). Refer to research studies to justify prompt designs (can). • Overriding AI (Fuegener) • Difference in experience (DellAqua) • Ghost-writing vs. sounding boards (ChenChan)

• Group work: analysis of selected research papers • Short presentations for the other groups • Summary slide Key references: - DellAqua - Fuegener - ChenChan

- Non-determinism (temperature parameter) ## 🧩 Foundation Models - Transformer-based neural nets pretrained on massive unlabeled text/data to predict next tokens - GPT‑3 debuted May 2020; GPT‑4 released March 14 2023; ChatGPT launched November 30 2022 - GPT‑4o (multimodal) released May 2024; GPT‑4.5 “Orion” launched Feb 27 2025 ## 🎓 Training & Tuning - **Pretraining**: Learn language patterns via self-supervised token prediction - **Fine-tuning**: Supervised tuning + RLHF (Reinforcement Learning from Human Feedback) on GPT‑3.5+ to align outputs ## 🌫️ Diffusion vs. Autoregressive Models - Standard LLMs are autoregressive (predict next token sequentially). - **Diffusion language models** offer parallel generation & editability; emerging work aims to match AR performance ## 📊 Benchmarking (Hugging Face Overview) - Key tasks: MMLU, MT‑Bench, Chatbot Arena; Hugging Face leaderboard aggregates publicly reported scores - Metrics include accuracy, latency, inference cost—use multiple benchmarks to avoid overfitting. This represents a significant departure from MLMs and has substantial consequences for the application of CLMs. In MLMs, since subsequent tokens are not random and are influenced by the target token, the use of tokens that follow the target token can lead to target variable leakage during training. ## 🎯 Summary | Component | Role | |---------------------|----------------------------------| | Transformer | Foundation of LLM architecture | | Self-supervised pretraining | Builds general language understanding | | Fine-tuning + RLHF | Aligns model with human values | | Benchmarking tools | Compare model strengths and trade-offs | TBD: Shortcomings (statistical parrot, reasoning, ...) Adatped version: Blended: watch youtube LLM-video in preparation and summarize at the beginning (Ergebnisse sichern)

Distinguish prompting techniques and interaction modalities (must) • Clarity, context and role assignment, output constraints • One-shot/few-shot, retrieval-augmented generation, reasoning chains

--- ## Exercise 3: Interpreting and Applying Handbook Practices **Scenario:** You’ve read part of your organization’s handbook about documentation practices, but you’re unsure how to apply them to your current project. You want to use an LLM to explain and contextualize these rules. **Task:** Design a prompt that gives the LLM the handbook section and your project description, asking for a contextual interpretation. **Draft Prompt Solution:** ```text I will provide a paragraph from our organization's internal handbook on documentation practices, followed by a brief description of my current project. Help me interpret the rules and give me specific suggestions for how to apply them in my context. ``` --- ## Exercise 4: Preparing a Slide Deck from a Meeting Transcript **Scenario:** You conducted a 1-hour strategy meeting and have a transcript or rough notes. You need to prepare a 5-slide summary presentation. **Task:** Design a prompt that guides the LLM to generate a slide outline with suggested titles and bullet points. **Draft Prompt Solution:** ```text Here is the transcript from our 1-hour strategy meeting. Create a 5-slide presentation summarizing the key points. Each slide should have a title and 3–5 concise bullet points. Emphasize decisions, next steps, and strategic themes. ``` --- ## Exercise 5: Research Task Delegation — Exploring a New Tool **Scenario:** You heard about a new productivity tool (e.g., Obsidian, Notion AI) and want a concise, opinionated overview to decide whether to invest time in exploring it. **Task:** Design a prompt that instructs the LLM to give you a short, actionable overview focused on features, trade-offs, and organizational use cases. **Draft Prompt Solution:** ```text Give me an opinionated overview of [Tool Name] for knowledge workers in organizations. Summarize its main features, strengths, limitations, and the kinds of workflows it supports. Focus on whether it’s worth adopting for managing tasks, notes, or team knowledge. ``` *Tip: When evaluating your prompts, consider if they provide the LLM with enough context, specify the desired format, and include role or tone instructions.* TODO : share solutions via rustpad --- Example: writing a literature review thesis (many different types, including ....., you decided to write a realist review) Reflection: data privacy, ...? Question: how would you use ChatGPT/LLMs? When would you not use it?