Q&A with Dr. Amaury Lendasse
|
A Kummer College newsletter addressing current events of economic, technological, geopolitical, risk and regulatory affairs of goods and services.
|
|
|
A milestone for artificial intelligence (AI) will be reached this year as there have been announcements that the world’s first gigawatt data centers will attain two gigawatts levels. Missouri S&T is leading practical implementation of AI in education and research through the MinerAI programs and initiatives to define AI competency across campus.
In this Q&A session, Kummer College Dean Jim Sterling chats with the Engineering Management and Systems Engineering Chair and Professor Amaury Lendasse, an expert in machine learning and AI.
|
|
|
|
Why should AI research begin by defining the persona of the AI to be prompted?
|
|
|
|
Defining a persona is key because it narrows the model’s response space. Specifying a role, like “Senior Quantum Physicist” or “Distinguished Mathematician," sets expectations for tone, depth, vocabulary, and assumptions. It keeps the model within the right cognitive frame, reducing generic or misaligned answers. A persona should go beyond a job title, reflecting the discipline, reasoning style and boundaries of expertise.
|
|
|
|
Can you explain why claims are being made that "math" is being solved by AI? What benchmark shows this?
|
|
|
|
Claims that AI has “solved math” refer to strong performance on competition-style tasks, not research-level problems. Models now do well on GSM8K and the American Invitational Mathematics Examination, or AIME, especially with tools like Python. But on FrontierMath, which tests new, unpublished research-level problems, state-of-the-art systems solve fewer than two percent, showing math is far from solved.
|
|
|
|
What are other criteria than benchmarking that show the progress in training of AI?
|
|
|
|
Benchmarks are useful, but progress is also tracked through scaling, efficiency and generalization. Scaling laws show how performance improves predictably as compute, data and model capacity increase. Efficiency measures performance per unit of training or inference compute, including latency, throughput, memory, and energy. Inference-time computation can improve correctness on hard tasks. Generalization, especially zero-shot performance, shows whether a model can handle tasks it wasn’t explicitly trained on. Together, these criteria measure not just score, but how reliably, efficiently and broadly a model’s capability applies.
|
|
|
|
What is explainable AI and what recent progress can be mentioned that shows progress is rapid and substantial?
|
|
|
|
Explainable artificial intelligence, or XAI, makes AI systems auditable: you can trace what they did, why, what signals drove decisions, and whether they stayed within policy. Explainability is shifting from “nice-to-have” to operational accountability, especially for agentic systems that take actions. Progress is rapid, moving from post-hoc explanations to continuous monitoring, logging and policy enforcement, driven by governance expectations such as the National Institute of Standards and Technology, or NIST, and the European Union’s AI Act. Real advances now provide end-to-end traceability, enabling organizations to defend decisions, detect misbehavior and ensure compliance.
|
|
|
Missouri University of Science and Technology
|
228 Engineering Management 600 W. 14th St. | Rolla, MO 65409 US
1-800-522-0938
|
| |
|
Manage your preferences | Opt Out using TrueRemove™
Got this as a forward? Sign up to receive our future emails.
View this email online.
This email was sent to .
To continue receiving our emails, add us to your address book.
|
|
|
|
