The impact of artificial intelligence in 2025 – Panel discussion with Duke, Google, IBM

Key Concepts

AI & Generative AI (GenAI): Artificial intelligence, particularly focusing on models that generate new content (text, images, code, etc.).
Large Language Models (LLMs): AI models trained on vast amounts of text data, capable of understanding and generating human-like text.
Agents/Agentic Technologies: AI systems designed to perform specific tasks autonomously, often involving reasoning and decision-making.
Open Source AI: The concept of making AI models, training data, and related tools openly available for modification and redistribution.
Frontier Models: Cutting-edge AI models pushing the boundaries of what's possible.
Retrieval Augmented Generation (RAG): A technique where LLMs retrieve relevant information from external sources before generating a response, improving accuracy and reducing hallucinations.
Soft Skills: Non-technical skills like critical thinking, problem-solving, and analytical abilities.
Enterprise Readiness: The suitability of AI technologies for use in large-scale, production environments.
Hallucination: The tendency of LLMs to generate incorrect or nonsensical information.
Reasoning Models: AI models capable of logical inference and problem-solving.
AI Literacy: Understanding the capabilities, limitations, and ethical considerations of AI.

Panel Introduction and Audience Level Setting

The panel consists of experts from Fidelity (AI search platform), Duke University (AI Masters program), Google (DevRel), and IBM (WatsonX).
The moderator, Mark, uses Gamma AI to create the presentation, highlighting the use of AI tools.
The audience is polled to gauge their familiarity with AI:
- Most use ChatGPT.
- A significant number are building or using AI in their jobs.
- Many identify as knowledge workers.
The panel aims to discuss the societal and professional impact of AI, addressing concerns about job security and the future.

The Future of AI Roles and Skills

John (Duke): Emphasizes the increasing importance of "soft skills" like critical thinking, problem-solving, and analytical skills. These are seen as uniquely human and essential for leading AI-driven workforces.
Fran (Google): Highlights the need for both strong technical skills (understanding the underlying math) and adaptability to change. She looks for candidates who are excited about new ideas and pivoting.
Critical Thinking: Consistently identified as a top skill by organizations like McKinsey and the World Economic Forum.

Enterprise-Ready AI Technologies

Sachin (Fidelity): Focuses on practical AI applications and Enterprise readiness.
- Key Use Cases:
  - Co-pilots for coding.
  - Information retrieval and search.
  - Content generation for marketing and social media.
- Priorities: Identifying high-impact use cases with measurable ROI.
- Considerations: Explainability, understanding model creation, security, and data privacy.
Bill (IBM): Discusses the evolution of AI infrastructure.
- Focus: Making open-source AI stacks (like PyTorch) work efficiently in standard data centers.
- Key Technologies: VM (inference engine) now under open governance at the Linux Foundation.
- Emphasis: Maturation of core AI technologies and the emergence of adjacent technologies like agentic systems.

Open Source AI: Importance and Challenges

Fran (Google): Stresses the importance of open governance and standards for AI APIs to prevent vendor lock-in and foster competition.
John (Duke): Highlights the value of open-source AI for education and research.
- Education: Students use open models for fine-tuning and understanding model architecture.
- Research: Open-source tools enable academic researchers to contribute meaningfully to AI progress, even without massive resources.
- Berkeley Example: A team trained a reasoning model with performance comparable to OpenAI's GPT-4 preview for only $450.
Bill (IBM): Discusses IBM's Granite family of models, fully open-sourced under the Apache 2.0 license.
- Rationale: To ensure customers own the IP generated by the models and to promote widespread adoption, similar to the success of open-source operating systems.
- Apache 2.0 License: Crucial for enterprise customers as it provides clarity on IP ownership.

Predictions for 2025

Sachin (Fidelity): Predicts the rise of AI agents and the importance of reasoning models for their effectiveness.
Fran (Google): Echoes the excitement around agents, emphasizing their potential to automate tasks and free up human time.
John (Duke): Foresees AI agents taking on real work in areas like customer service and lead generation, leading companies to consider AI agents vs. human hires.
Bill (IBM):
- Unhappy Prediction: High-profile disasters with generative AI due to rushed deployments without proper security and governance.
- Happy Prediction: The emergence of design patterns for generative AI (like RAG), leading to the development of libraries and frameworks that accelerate application development.
Mark (Moderator):
- Robotics: Increased adoption of robots in dangerous jobs, driven by advancements in AI and data collection (e.g., Tesla's Optimus).
- Competition: The AI landscape will become more democratic, with companies like Facebook and Google challenging OpenAI's dominance.

Q&A Highlights

Question 1: Impact of AI on Entry-Level Engineering Roles:
- Concern: If AI handles low-level tasks, where will new engineers gain experience?
- Sachin's Response: AI will augment, not replace, fundamental engineering skills. Upskilling and training are needed to orchestrate AI effectively.
Question 2: Ethics and Philosophy of Training LLMs:
- Concern: How are ethical principles and diverse perspectives considered when training LLMs, especially given the US's dominance in AI development?
- John's Response: Data used for training significantly impacts the reliability of outputs. Post-processing stages (supervised fine-tuning, reinforcement learning with human feedback) are crucial for reducing bias. The selection of humans guiding these stages raises questions about their values and biases.
- Potential Solutions: Broad and diverse stakeholder participation or different model versions tuned to specific values.
- Bill's Response: Emphasizes a lifecycle approach to trustworthy AI, starting with problem definition and user research, and transparency throughout the process.
- AI Risk Atlas: A free resource for understanding and mitigating risks associated with AI deployment.
Question 3: Accuracy and Determinism in AI:
- Concern: LLMs can produce non-deterministic answers, making them unsuitable for tasks requiring accuracy. How to measure abnormalities and ensure reliability?
- John's Response: Avoid using LLMs for tasks where incorrect answers are problematic. Advocate for RAG to improve accuracy by grounding responses in external information.
- Fran's Response: For deterministic tasks, use LLMs for parsing questions and phrasing answers, but rely on external APIs for accurate data retrieval.
Question 4: Upskilling for Experienced Professionals:
- Concern: How can experienced professionals with legacy technology skills remain competitive in the AI era?
- Sachin's Response: Focus on becoming a domain expert and augment that expertise with AI literacy. Understand the capabilities and limitations of AI.
- Bill's Response: Focus on user value and solving useful problems.

Synthesis/Conclusion

The panel discussion highlights the transformative potential of AI, particularly generative AI and AI agents. While concerns about job displacement and ethical considerations are valid, the panelists emphasize the importance of upskilling, focusing on uniquely human skills like critical thinking and problem-solving, and adopting a responsible approach to AI development and deployment. Open source AI, ethical considerations, and the emergence of design patterns are key themes shaping the future of the field. The panelists predict a more democratic AI landscape with increased competition and a focus on practical applications that deliver real value to users.