# Lecture 14: From BERT to ChatGPT - A Lightning Overview & GenAI Applications **Date:** August 12 **Duration:** 1.5 hours **Instructor:** Prof. Vivek Farias ---- # manual note virtue of (push the number; making average ); blue embedding i have helps disambiguate (separate train station and radio station) - how do i know what to pay attention to; - the animal didn't cross the street because it was too tired VS the animal didn't cross the street because it was too wide new in gpt 5 (bigger model and more data) - getting a model think more (test time) siting: does this slow down the knowledge creation?? ---- ## Learning Objectives - Understand the evolution of AI/ML technologies and their operational applications - Critically evaluate GenAI applications in business operations - Identify where GenAI adds genuine value vs. where it creates risks - Develop frameworks for implementing AI in operational processes ## Required Readings - **[YC's Latest Batch: 'Maybe AI Can Do This?'](https://techcrunch.com/2023/04/08/ycs-latest-batch-sure-was-a-lot-of-maybe-ai-can-do-this/)** ### TechCrunch Article Key Insights: **AI Startup Proliferation:** The article identifies a pattern in Y Combinator's 2023 batch where many startups followed a formula: "Pick a use case, do a little fine tuning of an available model, cherry pick some good examples for screenshots and bolt on a prefab UI." **Critical Questions About AI Applications:** - Is this really what's needed? - Won't this need lots of oversight? - Does this introduce liability or decrease transparency? - Did anyone ask customers if they want this? - Who verifies and audits the results? - Who is displaced by these tools? **Customer Experience Concerns:** "Having AI, as it currently exists, do something for you is kind of like admitting that it doesn't matter." The article raises important questions about when AI enhancement actually degrades the customer experience. ## AI Technology Evolution ### Historical Context 1. **Traditional AI (1950s-1980s):** Rule-based systems, expert systems 2. **Machine Learning (1990s-2000s):** Statistical learning, decision trees, SVMs 3. **Deep Learning (2010s):** Neural networks, computer vision, speech recognition 4. **Transformer Era (2017+):** BERT, GPT, large language models 5. **GenAI Explosion (2022+):** ChatGPT, multimodal models, widespread adoption ### Key Technological Breakthroughs #### BERT (2018) - **Bidirectional:** Reads text in both directions for better context - **Pre-training:** Learned language patterns from massive text datasets - **Fine-tuning:** Adapted for specific tasks with smaller datasets - **Impact:** Dramatically improved natural language understanding #### GPT Series Evolution - **GPT-1 (2018):** Proof of concept for generative text - **GPT-2 (2019):** Showed emergent capabilities, initially withheld due to concerns - **GPT-3 (2020):** Breakthrough in few-shot learning and general capability - **ChatGPT (2022):** User-friendly interface brought AI to mainstream - **GPT-4 (2023):** Multimodal capabilities, improved reasoning #### Current Capabilities - **Text Generation:** Human-quality writing and communication - **Code Generation:** Programming assistance and automation - **Image Generation:** DALL-E, Midjourney, Stable Diffusion - **Multimodal:** Understanding and generating across text, images, audio - **Reasoning:** Improved problem-solving and logical thinking ## GenAI in Operations Management ### High-Value Applications #### Process Automation - **Document Processing:** Contract analysis, invoice processing - **Customer Service:** Intelligent chatbots, ticket routing - **Quality Control:** Automated defect detection and classification - **Scheduling:** Resource allocation and workforce planning #### Decision Support - **Demand Forecasting:** Enhanced prediction models - **Supply Chain Risk:** Early warning systems for disruptions - **Maintenance Planning:** Predictive maintenance optimization - **Inventory Management:** Dynamic safety stock calculations #### Communication Enhancement - **Translation Services:** Real-time multilingual operations - **Report Generation:** Automated operational dashboards - **Training Materials:** Personalized learning content - **Documentation:** Process descriptions and SOPs ### Questionable Applications #### Customer-Facing Automation - **AI Receptionists:** May degrade customer experience - **Automated Sales Calls:** Potential for customer alienation - **Generic Email "Personalization":** Customers recognize and resent fake personalization - **AI Interview Screening:** Removes human element from relationship building #### Low-Value Substitutions - **Simple Content Generation:** Where quality doesn't matter - **Routine Data Entry:** Better solved with process improvement - **Basic Customer Queries:** May frustrate customers seeking human help - **Generic Marketing Content:** Lacks brand voice and authenticity ## Critical Evaluation Framework ### Questions to Ask Before AI Implementation #### Value Creation 1. **Real Problem:** Does this solve a genuine operational problem? 2. **Customer Benefit:** Do customers actually want this automated? 3. **Quality Standards:** Can AI meet the required quality level? 4. **Human Alternative:** Is human performance significantly worse/more expensive? #### Risk Assessment 1. **Error Consequences:** What happens when AI makes mistakes? 2. **Oversight Requirements:** How much human supervision is needed? 3. **Transparency:** Can decisions be explained and audited? 4. **Liability:** Who is responsible for AI-generated outcomes? #### Implementation Readiness 1. **Data Quality:** Is sufficient, clean data available? 2. **Change Management:** Are people prepared for this change? 3. **Integration:** How does this fit with existing systems? 4. **Measurement:** How will success be measured and monitored? ## Operational Implementation Strategies ### Pilot Approach - **Small Scale:** Start with limited scope and well-defined use cases - **Success Metrics:** Clear measurement of improvement over baseline - **Feedback Loops:** Regular assessment and adjustment - **Stakeholder Buy-in:** Include affected employees in design and testing ### Human-AI Collaboration - **Augmentation vs. Replacement:** Enhance human capabilities rather than eliminate jobs - **Complementary Strengths:** AI for pattern recognition, humans for judgment - **Escalation Paths:** Clear procedures for when AI needs human intervention - **Continuous Learning:** Systems that improve from human feedback ### Technology Integration - **API-First:** Integrate with existing operational systems - **Data Pipeline:** Ensure clean, consistent data flows - **Security:** Protect sensitive operational and customer data - **Scalability:** Design for operational growth and changing needs ## Industry Applications ### Manufacturing - **Quality Control:** Visual inspection automation - **Predictive Maintenance:** Equipment failure prediction - **Production Planning:** Demand-driven scheduling - **Supply Chain:** Supplier risk assessment and management ### Healthcare - **Medical Records:** Automated documentation and coding - **Diagnostic Support:** Pattern recognition in imaging - **Drug Discovery:** Molecular design and testing - **Operations:** Capacity planning and patient flow ### Financial Services - **Fraud Detection:** Transaction monitoring and analysis - **Risk Assessment:** Credit scoring and portfolio management - **Regulatory Compliance:** Automated reporting and monitoring - **Customer Service:** Intelligent query routing and response ### Retail/E-commerce - **Demand Forecasting:** Inventory planning and management - **Personalization:** Product recommendations and pricing - **Visual Search:** Image-based product discovery - **Logistics:** Route optimization and delivery planning ## Discussion Questions for Class ### Critical Evaluation 1. **Value vs. Hype:** Looking at the TechCrunch list of AI startups, which applications seem genuinely valuable vs. solutions looking for problems? 2. **Customer Perspective:** When do customers appreciate AI automation, and when does it feel like a degradation of service? 3. **Implementation Strategy:** How should companies approach GenAI implementation to maximize value and minimize risk? ### Personal Experience Reflection **Prompt:** Reflect on processes you are intimately familiar with (say, from your most recent role): where does GenAI fit in? What value does it add? And what are the risks? **Analysis Framework:** - **Current Process:** How does it work today? - **Pain Points:** What are the main inefficiencies or problems? - **AI Opportunity:** Where could GenAI realistically help? - **Value Assessment:** What would be the tangible benefits? - **Risk Factors:** What could go wrong with AI implementation? - **Human Element:** What should remain human-driven? ### Strategic Considerations 1. **Competitive Advantage:** When does AI create sustainable competitive advantage vs. when does it become table stakes? 2. **Organizational Change:** How should companies prepare their workforce for AI integration? 3. **Ethics and Responsibility:** What ethical considerations are most important in operational AI applications? ## Emerging Trends and Future Directions ### Technology Evolution - **Multimodal AI:** Integration of text, image, audio, and video understanding - **Specialized Models:** Industry and task-specific AI systems - **Edge Computing:** AI processing closer to operational data sources - **Autonomous Systems:** Self-managing operational processes ### Business Model Impact - **AI-First Companies:** Born-digital organizations built around AI capabilities - **Traditional Transformation:** Incumbent companies adopting AI for competitive advantage - **New Service Models:** AI-enabled services and business models - **Platform Economics:** AI as a platform for third-party applications ### Regulatory Landscape - **AI Governance:** Emerging frameworks for AI oversight and compliance - **Data Privacy:** Increased regulation of AI data usage - **Algorithmic Accountability:** Requirements for AI transparency and fairness - **Industry Standards:** Development of AI best practices and standards ## Key Takeaways - **Critical Evaluation:** Not all AI applications create genuine value; many are solutions looking for problems - **Customer-Centric Design:** AI should enhance rather than degrade customer experience - **Human-AI Collaboration:** The most successful applications augment human capabilities rather than replace them - **Implementation Discipline:** Successful AI deployment requires careful planning, measurement, and continuous improvement - **Strategic Focus:** AI should solve real operational problems, not just demonstrate technological capability ## Risks and Limitations ### Technical Limitations - **Hallucination:** AI systems generating plausible but incorrect information - **Bias:** Systemic discrimination in training data and algorithms - **Brittleness:** Poor performance outside training distribution - **Interpretation:** Difficulty understanding AI decision-making processes ### Operational Risks - **Over-Reliance:** Loss of human capabilities and judgment - **System Failures:** Operational disruption when AI systems fail - **Security Vulnerabilities:** AI systems as targets for cyber attacks - **Compliance:** Regulatory challenges with AI-driven decisions ## Preparation for Next Class - Review course concepts in preparation for wrap-up session - Consider how various topics (capacity, inventory, networks, etc.) connect to AI and technology trends - Prepare questions for final exam review ## Teaching Notes - Balance enthusiasm for AI capabilities with realistic assessment of limitations - Encourage critical thinking about when AI truly adds value - Use specific examples from students' work experience - Connect AI applications to core operations management concepts covered in course - Prepare students to be thoughtful consumers and implementers of AI in their careers