On July 17th, 2025, the USC Viterbi Information Sciences Institute (ISI) and the Stanford Institute for Human-Centered AI (HAI) came together to spotlight the 2025 AI Index Report. Held at ISI’s Marina Del Rey headquarters, the event drew a packed crowd of tech enthusiasts filling every seat in the house. 

The AI Index, published annually by Stanford HAI, is one of the most comprehensive, data-driven views of artificial intelligence, recognized as a trusted resource by global media, governments, and leading companies. Its mission is to provide unbiased, rigorously vetted, broadly sourced data in order to help people develop a more thorough and nuanced understanding of the complex field of AI. Read the 2025 AI Index Report here. 

The event’s panel brought together leading voices from across the AI ecosystem, including Ray Perrault from SRI International, who serves as Co-Chair of the Stanford AI Index; Yolanda Gil from ISI, who also serves as Co-Chair of the AI Index; Karl Jacob from LoanSnap and ISI; and Karina Montilla Edmonds from SAP. The discussion was expertly moderated by Vanessa Parli from Stanford HAI.

Together, they unpacked the report’s key findings and broader implication, discussing the state  of AI models, diverging global approaches to AI policy, and what the future might hold for increasingly autonomous systems. Read highlights from the panel below.

Parli: The AI Index talks about foundation models driving breakthroughs. How do you see this field evolving? 

Yolanda Gil:

Foundation models are typically large language models that learn sequences of words to predict words, generate text, and correctly summarize documents. The same techniques are being used in science to build foundation models to understand scientific data. We can use these techniques to look at time series data; for example, temperatures in different locations around the world, different variables that are evolving over time. If we can look at large amounts of time series data, we can build AI models that know so much about how the world works that they can make better predictions about what may happen.

Beyond time series, foundation models are being developed in many areas, including medicine, biology, and geosciences; for example, making predictions about proteins. In climate science, I’ve worked with hydrologists who would tell me, I need data for 30 years in this location before I can make predictions about flooding. Now, they’re looking at foundation models that have so much knowledge about how water works, how climate works, that they don’t need 30 years of data anymore. These models are changing the game and how science is being conceived. I think that in the next few years, we’ll see more scientific discoveries everywhere, guided by these foundation models that require much less additional data.

Parli: AI has had many exciting advancements and systems are exceedingly doing better against benchmarks. Given this success, what are some of the remaining challenges that AI systems face?

Ray Perrault: To me, the main challenge in deploying AI is ensuring its reliability matches the expectations of the user. If you ask ChatGPT to write a poem, there are few constraints on the quality of what you get back—except that it should be amusing. But if you ask it to solve a complex mathematical problem or retrieve data from a database, the criteria are much stricter. These are problems where “approximate” isn’t good enough.

LLMs are systems that predict the next word in a sequence; they weren’t designed to do math. So it’s kind of amazing that they can do math problems as well as they do. But the question is: why is that the case, and is it good enough?

It’s increasingly clear that these systems fail on reasoning problems—that is, those that require applying rules systematically, especially on problems of arbitrary size. My classic example is multiplying two 20-digit numbers. Any elementary school kid learns how to do this. All you have to do is give them a pencil, and they’ll probably get the answer right once trained.

In contrast, LLMs can multiply small numbers fine, but as the numbers grow, they start increasingly failing. This is just one of many examples of reasoning problems where you can find the boundary of where the systems fail. So-called reasoning LLMs perform slightly better, but they still break in the same way—they just break a little further along.

So, what can we do about this? There are different views. One approach is to let LLMs call symbolic systems to solve reasoning problems that are embedded in natural language, compute an answer, and give it back. For example, if you want to do a database query, the LLM will create an SQL query, hand it off to an SQL system, and the answer will come back. Then comes the challenge of reintegrating that answer into the overall task. There’s a lot of ongoing work connecting language models with symbolic reasoning systems in a composable way.

Some argue that’s still not enough—we need a unified architecture that handles everything in one system. I agree. But people much smarter than I have worked on this for years, and we still haven’t solved it. Until then, forget about AGI. If you can’t reason, you’re not AGI—you’re not even close.

Parli: For workers, there is a growing anxiety about losing your job to AI. How do you respond to the fear that AI will replace us?

Karina Montilla Edmonds: I always say it’s not AI that’s going to take your job—it’s someone who knows how to work with AI. At SAP, we recently surveyed over 4,000 managers and employees, and the study showed that those colleagues who were more literate in AI showed a higher optimism for using AI. So, somewhat counterintuitively, the anxiety over AI in the workplace is often driven by a lack of understanding.

This presents a great area for collaboration between academia and industry. Both are focused on upskilling and reskilling people to use AI effectively and responsibly. There are also major opportunities for co-creation, collaboration and innovation, especially around the development of AI fairness, safety, and ethics.

Another opportunity I see is creating new ways to teach AI. The rapid pace of AI development creates a real challenge for education. In academia, launching a new course can take time—too much time. If you design a static AI course, it could be outdated before the first lecture. We’ve found that more modular, flexible training—delivered in smaller, dynamic pieces—is more effective and engaging.

Here, gamification could play a role; and for AI nothing beats experiential learning. We must constantly apply the skills we learn. People love learning through interactive formats, and that’s something we can lean into as we think about the future of AI education. There’s so much opportunity—if we move fast and work together.

Parli: The AI Index tracks the international AI landscape. Can you provide some perspective on different AI strategies around the world? 

Yolanda Gil:

The international perspective that the AI Index provides is very important—it helps us see how AI can emerge in different ways across regions. One significant trend I’ve noticed is the number of countries developing their own AI models, particularly language models. While you can translate across languages from any model, having a model that is native to a specific language offers unique advantages.

For example, Korean researchers have created an AI model that isn’t just about Korean language—it’s grounded in Korean customs, society, and culture. The model interacts and behaves very differently, and it may have a different performance from state of the art models in other languages, but it’s very much customized for Korean needs and Korean society. I find that really fascinating. They have developed their own benchmarks on how polite the model is and how much it understands Korean society. I think that’s very important, and we’ll see more of these kinds of specialized models moving forward.

Another important point the AI Index highlights is the ongoing challenge of AI policy. There’s a lot of debate around how to regulate AI and what principles should guide responsible AI development. Sometimes different governments propose different regulations. In the U.S., for example, we’ve seen various legislative proposals in the House and Senate, though few actually pass. Still, these efforts are reflective of the desire to regulate this technology.

The AI Index highlights a promising global trend: groups of countries coming together to define shared values and aspirations for AI, in Asia, Africa, Latin America, and so on. They are agreeing to what responsible AI should be, how AI should behave, and how AI can be more respectful of people and societal norms.

These kinds of international collaborations are significant. It’s beginning to reflect a universal desire to establish boundaries for how AI systems are developed and used—boundaries that align with societal values, and those differ regionally. AI regulations may still take time, and we have a lot to learn about the broader AI landscape. But I see it as a positive sign that governments are beginning to think more deeply and collectively about the nature of AI in our society and the policies that will come.

Closing on a high note 

Following questions from the audience, the AI Index panel event wrapped up with a rooftop networking reception overlooking Marina del Rey, where attendees continued the conversation over refreshments and sunset views. It was a perfect close to an afternoon focused on the most critical trends in AI.

Published on August 11th, 2025

Last updated on August 13th, 2025