ACCT 331 WEEK 11

Course Information

  • Course Title: ACCT 331: Introduction to Applied Artificial Intelligence

  • Location: Schreiber Hall, Room 302

  • Schedule: Tuesday & Thursday, 1:00-2:15 PM

  • Course Code: 10304

Course Schedule and Important Dates

  • Exams:

    • Exam #2: Week 12, on 11/13/25

    • Final Exam: Week 1-14, on 12/12/00/25

AI Courses You Can Use

  • Google DeepMind: Build Your Own Small Language Model
    Description: Learn fundamentals of language models.
    Duration: 6 hours

  • Google DeepMind: Train a Small Language Model (Challenge Lab)
    Description: Develop foundational tools and data preparation for training models.
    Duration: 1 hour 30 minutes

  • Google DeepMind: Represent Your Language Data
    Description: Learn how to prepare text data for models.
    Duration: 4 hours

  • Google DeepMind: Design And Train Neural Networks
    Description: Focus on the training process for models, spotting and mitigating issues.
    Duration: 4 hours

  • Google DeepMind: Discover The Transformer Architecture
    Description: Investigate the transformer architecture mechanisms.
    Duration: 4 hours

  • Google DeepMind: AI Research Foundations | Google Skills

Current News in AI

JPMorgan Chase and AI Investment

  • Jamie Dimon, CEO of JPMorgan Chase, states that the bank's $2B investment in AI has 'paid for itself.'

  • AI implementation has driven significant savings across business lines since the initial investment in 2012, averaging a $2 billion benefit for the expense.

KPMG and AI Integration

  • KPMG will examine how employees use AI tools as part of annual reviews, tracking effectiveness and encouraging integration into work.

  • This change reflects how AI is reshaping consulting, with companies investing deeply in AI technologies.

Rising Tech Investments

  • Major tech companies like Meta, Oracle, Google, and Microsoft have increased annual capital expenditures significantly from 2016 to 2024, with a clear rising trend in AI investments.

User Growth of Platforms

  • ChatGPT's user base has been rapidly growing, projected to reach millions of weekly active users over the next two years, indicating a surge in interest and utilization.

Financial Outlook for AI

  • OpenAI projects a challenging cash flow scenario over the next years, confronting investors with anticipated losses.

The AI Revolution

  • The future of AI holds both challenges and providential opportunities for economic growth.

  • Historical parallels: Industrial Revolution and Digital Revolution significantly uplifted living standards, despite fears of job losses due to automation.

  • The discourse emphasizes the historical expectation that advancements create more jobs than eliminated, urging cautious governmental policy-making to avoid obstructing progress.

Key Findings from AI Research

Everyday AI

  • Mainstream Adoption: 46% of business leaders now leverage Generative AI (GenAI) daily; 80% engage weekly.

  • Business applications span across internal roles from IT to Operations, increasingly applying AI for efficiency in daily tasks.

Proving Value

  • Approximately 72% of enterprises measure business-linked ROI metrics with rising confidence in future investments in AI technologies, especially in sectors like Financial Services.

The Human Capital Lever

  • A growing commitment from leadership in AI adoption is observed, yet skill gaps and training budgets represent ongoing challenges.

LLMs and Generative AI in Business

Key Applications

  • Data Analysis & Productivity: GenAI usage in areas including document summarization and data analytics showcases efficiency gains.

  • Revenue Growth: Improved personalized experiences may lead to significant increases in sales conversions by 15-30%.

  • Cost Reduction: Automating customer service channels can lead to a reduction in operational costs by 70-85%.

Key Concepts

Language Models

  • Definition: Language models utilize AI to predict subsequent words in sentences.

  • Training Data: Language models are trained with vast amounts of text to generate coherent outputs.

  • Types of Models:

    • Self-Supervised Learning (SSL): Training approach that builds labels from input data, enhancing learning efficiency without manual labeling.

    • Transformers: Model architecture that enables handling long-range dependencies in data concurrently rather than sequentially, enhancing training and generational capabilities.

Technical Depth: LLMs and Their Working Principles

Mechanisms

Tokenization

  • Process: Text inputs segmented into tokens enabling manageable computation.

Embeddings

  • Definition: Vectors in high-dimensional space representing tokens, signifying semantic relationships among words.

Self-Attention

  • Functionality: Weighs the importance of words in relation to one another, enabling context-aware representations.

Business Application and Evaluation Strategies

  • Key Evaluation Metrics: Perplexity, BLEU Score, ROUGE, F1, and others measure prediction accuracy, quality of generated content, and various qualitative metrics.

  • Importance: Helps ensure that AI systems in production do not emit hallucinations or biased outputs.

Generative AI Assessment and Challenges

Evaluation Strategies

  1. Absolute Scoring: Simple ratings on defined quality scales.

  2. Comparative Ranking: Evaluators rank outputs from different AI systems to assess quality.

  3. Error Analysis: Identification of specific weaknesses to guide improvements.

Challenges in Evaluating Generative AI

  • Multiple Valid Outputs: Unlike traditional tasks, generation may produce many acceptable responses.

  • Subjective Quality Assessment: Creativity and engagement are difficult to quantify accurately.