Three Reasons AI Pessimism Misses the Mark

Introduction

Steve Denning's recent Forbes article, ["3 Reasons Why The Promised Gains From AI Are Still Faraway,"](https://www.forbes.com/sites/stevedenning/2024/07/25/3-reasons-why-the-promised-gains-from-ai-are-still-faraway/) presents a pessimistic view of AI's capabilities and potential, relying on outdated information and misunderstandings. While Denning raises valid concerns, it is crucial to address his points with the latest advancements and a more nuanced perspective on AI's progress.

1. AI's Math Prowess: A Quantum Leap Forward

Denning's claim that AI struggles with simple math needs to be updated. While it's true that language models like OpenAI's GPT have historically had issues with basic arithmetic due to their probabilistic nature, recent developments paint a different picture. OpenAI's GPT-4 achieved an impressive 84.3% accuracy on the MATH dataset using code-based self-verification ([source](https://arxiv.org/abs/2308.07921)).

Moreover, as highlighted in a recent NY Times article, "AlphaProof, a New A.I. from Google DeepMind, Scores Big at the International Math Olympiad," a new breakthrough from DeepMind competed in the Math Olympiad and solved four out of six problems, achieving a performance on par with a silver medalist. This isn't just simple mathematics, but some of the most complex mathematical puzzles devised.

In the very article Denning referenced (https://www.nytimes.com/2024/07/23/technology/ai-chatbots-chatgpt-math.html), it shares an example of how AI can solve math problems: Khan Academy's AI-powered tutor, Khanmigo, effectively leverages a calculator app for numerical problems, ensuring high accuracy while maintaining the benefits of AI-driven tutoring. This hybrid approach demonstrates a practical and scalable solution to AI's arithmetic limitations.

Dismissing AI's potential based on math limitations is shortsighted. Humans, on average, are not exceptional at math; a random sample would likely score far lower than 84% on the MATH dataset. Yet, we've developed sophisticated mathematical techniques and capabilities. How? By building tools—calculators, computers, machine learning, deep learning, and now LLM models—that leverage all previous innovations. For instance, ChatGPT solved a complex number theory puzzle about prime factorization and calculating divisors by writing code to determine how many numbers from 1 to 100 have an odd number of divisors. Is that cheating? Perhaps, but the LLM then explained the thinking behind the answer, going far beyond a simple calculator. This approach mirrors our human journey with mathematics, combining tool use with conceptual understanding to tackle increasingly complex problems.

2. Taming the Data Beast: AI's Role in Data Management

While messy and inconsistent data in large firms is a long-standing issue, it's not insurmountable. The concept of "technical debt" is real, but AI and machine learning techniques are increasingly adept at managing and even leveraging imperfect data. Recent advancements in data preprocessing, cleansing, and integration tools are making it easier for organizations to prepare their data for AI applications.

Andrew Ng's research and frameworks emphasize the importance of iterative data improvement and the deployment of AI in practical, incremental steps. By adopting these strategies, organizations can progressively enhance their data quality and AI readiness, mitigating the risks Denning outlines. Additionally, AI-driven tools for data integration and anomaly detection are continuously evolving, providing firms with the means to address data issues more effectively than ever before ([source](https://www.datacamp.com/blog/what-is-gpt-4o)).

3. Breaking Silos: AI as a Catalyst for Collaboration

New organizational frameworks and collaborative tools are actively addressing the challenge of siloed expertise in organizations. Novartis's successful case, where data scientists were paired with business staff to integrate AI into operations, illustrates how cross-functional teams can drive significant AI gains.

Modern AI models are designed to be more user-friendly and accessible, allowing non-technical stakeholders to interact with and benefit from AI systems. Tools that facilitate better communication and collaboration across departments are becoming integral to AI deployment strategies. Companies that embrace these changes and foster a culture of continuous learning and collaboration are already reaping the benefits of AI ([source](https://openai.com/index/improving-mathematical-reasoning-with-process-supervision/)).

GenAI's Breakthrough: From Arithmetic to Advanced Mathematics

Recent advancements in generative AI challenge the notion that AI is limited to narrow applications. As reported in a recent NY Times piece, ["AlphaProof, a New A.I. from Google DeepMind, Scores Big at the International Math Olympiad,"](https://www.nytimes.com/2024/07/25/science/ai-math-alphaproof-deepmind.html) Google DeepMind's AlphaProof has achieved a silver medal performance at the International Mathematical Olympiad, solving complex mathematical problems at a high level of proficiency. This milestone demonstrates AI's growing capability to handle tasks that require both creative and logical reasoning.

Conclusion

While critical evaluation of AI's capabilities is essential, it's equally crucial to acknowledge the significant progress being made. AI is not a panacea, but its potential far exceeds Denning's pessimistic assessment. By focusing on the latest advancements, practical applications, and collaborative frameworks, we can better appreciate AI's current and future potential. The AI landscape is rapidly evolving, and those who recognize and adapt to these changes will be well-positioned to harness its transformative power.