ET@Davos: Physics Over Predictions: Andrew McLaughlin On The Next Wave of AI

Key Concepts

• Quantitative AI (AQ): AI models built around physics equations, producing scientifically accurate and rigorous outputs, contrasting with the limitations of Large Language Models (LLMs).
• Large Language Models (LLMs): AI models trained on massive text datasets, capable of generating human-like text but lacking in mathematical, scientific, and numerical reasoning.
• Hallucination (in AI): The tendency of LLMs to generate factually incorrect or nonsensical information.
• Mediocrity (in AI output): The common issue of LLM-generated content being bland, unoriginal, and lacking in depth or insight.
• Numeracy Problem: The inability of LLMs to accurately perform mathematical calculations or understand numerical concepts.
• Transformer Architecture: The innovation enabling the training of large language models.
• Physics-Based AI/Math-Based AI: AI systems grounded in fundamental physical laws and mathematical principles, ensuring accuracy and reliability in quantitative domains.
• Generative AI: AI capable of creating new content, whether text, images, or solutions to complex problems.

The Limitations of Current AI & The Rise of Quantitative AI

The current wave of AI excitement is largely driven by Large Language Models (LLMs), but their limitations are becoming increasingly apparent. While LLMs excel at generating text, they struggle with tasks requiring mathematical reasoning, scientific understanding, or accurate problem-solving. As stated, “If you ask a language model to solve a math problem or to design a molecule or to answer a science question, you'll get garbage for the most part.” These models are prone to “hallucinations” (generating false information), “mediocrity” (producing bland and uninspired content), and a fundamental “numeracy problem” – an inability to work with numbers effectively. For example, deploying LLMs for customer service often results in “bad experiences” due to the complexity of human interaction and the models’ inability to accurately predict and respond appropriately.

A new approach, termed “Quantitative AI” (AQ), is emerging. AQ utilizes models built around physics equations, aiming for scientifically accurate and rigorous outputs. This isn’t a replacement for LLMs, but rather a complementary architecture suited for tasks where precision is paramount. “We believe very strongly that these kinds of models…are fit for many many use cases that language models can't solve.” AQ is envisioned to handle quantitative problems while LLMs can serve as a user-friendly interface (like chatbots).

AQ: How it Works & Potential Applications

AQ leverages the power of AI to rapidly and accurately solve complex quantitative problems. The process involves:

1. Defining the Problem: Identifying a scientific or mathematical challenge (e.g., designing a drug to bind to a protein).
2. Establishing Equations: Formulating the relevant physics equations to model the problem (e.g., protein shape, electron interactions).
3. Data Generation: Generating data from these equations.
4. Model Training: Training an AI model on this data.
5. Problem Solving: Using the trained model to leap to the solution, bypassing iterative trial-and-error processes.

This approach allows for tackling problems with numerous variables, such as predicting molecular interactions within the human body. Specific applications mentioned include:

• Drug Discovery: Precisely targeting drugs to kill cancer cells or block misfolded proteins causing Alzheimer’s.
• Materials Science: Designing new metal alloys and catalysts.
• Financial Risk Simulation: Modeling and predicting financial market behavior.
• Defense: Developing remote-controlled weapons and swarm technologies. “You all can see that remote controlled weapons are going to dominate the battlefield.”
• Geolocation: Utilizing magnetic fields for precise vehicle location using quantitative AI and sensitive sensors.

AI 2.0 or Incremental Progress? & The "Sexiness" Factor

The speaker frames AQ not necessarily as “AI 2.0,” but as an “additional wave of AI” alongside LLMs. While some aspects draw on traditional machine learning techniques, certain components represent genuine innovation. There’s a concern that current AI attention is overly focused on the “sexiness” of LLMs, potentially overshadowing deeper breakthroughs. “I don't want to overstate it, but I think that the um sexiness of language model type AI is a little bit inflated compared to the reality.” Many businesses experimenting with AI are finding limited real-world results, highlighting the gap between hype and practical application.

India’s Opportunity in the AQ Era

India possesses significant potential to become a global leader in Quantitative AI. The speaker identifies several key factors:

• Scientific Depth: India has an “unprecedented” and “unparalleled” depth of talent in math, physics, chemistry, and biology.
• Decreasing Cost of Innovation: The cost of compute and data center infrastructure is falling, making AI development more accessible.
• Interdisciplinary Training: A growing trend among younger Indians to combine expertise across disciplines (physics, computer science, mathematics).
• Government Support: Public-private partnerships and government investment in deep tech are crucial. The government can act as an early adopter of AQ technologies.

However, India needs to address certain challenges:

• Siloed Academic Disciplines: Breaking down barriers between academic departments to foster interdisciplinary collaboration.
• Patient Capital: Securing long-term investment for deep tech projects, which require significant upfront capital and have longer timelines for returns.
• Mindset Shift: Moving beyond outsourcing and focusing on original research, development, and intellectual property creation. “The real question is can India move quickly uh to take advantage of it and to provide that combination of infrastructure capital talent vision that's required to really build out those sectors.”

Strategic Implications & Governance of AI

AI is poised to become a strategic national asset, comparable to energy or defense. “Yeah, I mean it's an easy yes. [laughter] Well, it's not an easy guess. Well, I think it is I mean in the same sense as electricity and so forth, but so to me it's an easy yes.” This raises questions about governance and trust. The speaker advocates for international collaboration and transparency, emphasizing that countries should have the capacity to defend themselves and maintain national sovereignty. “I don't think that's a competitive thing. I'd like there to be quantitative AI companies in India uh that we work with and I like them to be companies we compete with as well.”

Playbook for India in the AI Game

The speaker’s key recommendations for India’s AI strategy are:

1. Foster Interdisciplinary Talent: Encourage collaboration between scientists, computer scientists, and mathematicians, creating “computational physicists, computational chemists, computational biologists.”
2. Focus on Quantitative AI: Prioritize the development of AI models grounded in physics and mathematics, rather than solely focusing on LLMs.
3. Leverage Quantum Physics & Math: Recognize the importance of the mathematical principles underlying quantum physics for solving complex problems in various fields.
4. Invest in Infrastructure & Capital: Provide the necessary resources (compute, data centers, funding) to support AI research and development.

Conclusion

The future of AI lies not solely in the ability to generate human-like text, but in the capacity to solve complex quantitative problems with scientific rigor. Quantitative AI represents a significant advancement, offering solutions in areas like drug discovery, materials science, and national defense. India, with its strong scientific talent pool, has a unique opportunity to become a global leader in this emerging field, but requires a strategic shift towards interdisciplinary collaboration, long-term investment, and a focus on foundational research. The key is to move beyond the “sexiness” of LLMs and embrace the power of physics-based AI to unlock real-world innovation.