AI Q&A: Anna Jaffe

As the CEO of Mobi, Anna Jaffe’s technical expertise as an MIT graduate and industry leader has shaped her vision to use technology to solve large-scale, intractable problems. Her academic research focused on biomimetic design and using natural systems to help people solve the most pressing issues of our time. She is particularly interested in the intersection between technology and humanity—and how technology can increase our empathy, sensitivity, and decision-making capabilities, especially in high-risk situations. Anna’s passion for exploring how AI can shape the future for the better is defining how Mobi impacts the world.

1. What do you think is one of the biggest misconceptions about AI?

I think one of the biggest misconceptions is that people think AI is artificial human intelligence, but machines and people have really different types of intelligence. Making machines behave like people shouldn’t be the goal of AI. You don’t want to encode the biases we have. There are things we understand and feel deeply, but there are other things that we struggle to make sense of. So much of the world is changing all the time. It is impossible to know what is happening to every person alive in Tokyo right this second—that is not a human-level problem. Humans can’t even understand why people act the way they do in small groups, much less at a city-scale. It's a real problem to only imagine and shape AI in our own image.

There is a quote by the physicist Richard Feynman: “To every man is given the key to the gates of heaven. The same key opens the gates of hell. And so it is with science.” AI can absolutely be a key to heaven or hell. For me, AI is a tool to understand causality, and therefore to make more impactful and informed decisions. We as humans usually consider two or three options, but the advantage of AI is that it allows us to consider a billion scenarios. The magic here is that even if there’s only one good outcome out of 25 billion outcomes, it can find that one good outcome, give you information about it, and leave the human free to make a decision. The risk is that if you ask it to find a path to a devastating future, it can find a path to that too. So far, we’ve tried to shape AI in our own image, rather than asking it to collaborate with us to understand the world in ways we otherwise can’t. This needs to change.

2. What are some problems that AI can solve that would be impossible to solve otherwise?

Understanding causality. AI can help us understand causality in a way that’s impossible for humans otherwise because AI can try all of the possible scenarios and then see which one is the best fit for what you want. It can imagine (by reasoning through) all of the options and their results—and “think” not just one step forward but tens, if not hundreds, of steps forward into the future. This is the real power of this technology: showing the future impact of people’s actions. We started Mobi in part because, historically, the best large-scale planning and optimization technology was 2-5% better than an expert operator. With good Human-Collaborative AI you can accomplish 20-30% more: this is tremendously valuable for anyone who operates mission-critical infrastructure (water, power, telecom, first responders), particularly during a crisis.

Another advantage of AI is that it can acknowledge bias and the differences between people. Every person has their own rationale for why they do things. As rational or irrational as they are, they can be modeled and reasoned against. Think of imagining future scenarios: if you model everyone as the same, then all your scenarios will be wrong. Not all people are the same. It’s hard to know a human in real time, but it actually is something you can model in real time—AI can do what’s called contextual identity. We can begin modeling the world as it actually is. For most companies, CRM (Customer Relationship Management) systems are the primary tool to hold and manage customer insight, but human identity and preferences are much more fluid than our CRM models represent (see Chapter 4 of Banerjee and Duflo's Good Economics for Hard Times). Imagine in economics if you didn’t need an equation to limit our understanding of people, but could instead represent people as they really behave—imagine what you could forecast!

3. How would you define Collaborative AI?

There are two dimensions to defining Collaborative AI:

First, technically, Collaborative AI is about using the right tools and techniques for the right problem. You have clear collaboration across technological solutions and between models. When we work with AI, we’re trying to understand things as they are, so we use a combination of search, math, language and vision models, expert systems and rules-based systems, constraint programming, etc. Collaborative AI can help us understand, for a particular moment in time, the right way to get to the most honest, accurate, and human-relevant answer. When you’re trying to understand something about the world or supply chain or whatever it is you’re trying to understand, Collaborative AI means that you’re picking the right collaborative set of tools to do that.

The other half of the answer is the collaboration between the technology and the person. We think on different temporal and spatial scales. There are a bunch of good examples of this in David Mindell’s book Our Robots, Ourselves, for example when Armstrong landed on the moon using both the heads-up display and his own training and perception of the lunar surface. It’s this notion of understanding how you can set people up to spend more time in the physical world, talking to each other and doing the things people do to feel alive, and how we can let technology disappear into the background, while giving us a deeper awareness of what is true and helping us make the decisions we need to make, either for ourselves, our work, or for our communities.

There are things that technology is good at and there are jobs that it can do well. And then there are things that people excel at. It’s not just that humans and technology work together, but that together they can solve problems they couldn’t solve separately, and usually that has to do with decision-making so that you can take the right path confidently and with purpose.

4. How do you work with AI at Mobi?

I used to be entirely responsible for product, but today our company has an amazing product team and engineering team. It isn’t just Peng (Peng Yu, Mobi CTO), Yinxiang (Yingxiang Yang, Mobi Chief Data Scientist), and I, the way it used to be. The organization is bigger now, so day-to-day my job is less about defining exactly what we’re doing and more focused on why we’re building what we’re building, and with whom, and what problems they’re solving. We don’t want to use AI just to sell stuff. We’re really interested in what we’re using it for and what it’s doing.

There’s a good Edwin Land quote that says “Don't undertake a project unless it is manifestly important and nearly impossible.” The right thing to do isn’t always the easiest thing to do. Some of the most beautiful things we’ve made as people weren’t the easiest or most obvious things to make. I was in a conversation with Amar Bose before he died, and he said “you should do the things where you’re in a race with no one,” where you are so sure that you can solve this better than anyone on the planet that you’re completely committed to it because you deeply believe that it deserves to exist. That’s my primary job at Mobi: to make sure that we deeply believe that the things we’re building deserve to exist, and then that we understand what the path is to bring them to life, while maintaining their integrity and impact.

5. What’s unique about how Mobi is working with AI?

The first thing that’s unique is that we’re focused on problems dealing with space and time. For large language models, this is a big weakness. They’re exceptional at creativity for image or text generation and interpretation, but they’re not good at the context of a living, breathing, moving world because they’re focused on the position of words in a sentence or paragraph. We’re focused on things that move and change and where the future state is unpredictable—but that can be influenced with intelligent decision-making.

The second thing is that we’ve built a really comprehensive toolbox of techniques. We have deep expertise in language models, computer vision, constraint programming, linear optimization, etc. We’re a deep, deep tech company and still small enough where we don’t have individual groups focused on particular techniques. We’ve built a complete set of tools and, even within an individual engineer or researcher on the team, they have these different techniques integrated into their set of skills. This is what sets us up to solve an entire person’s job or team’s job or company’s job with unimaginable impact. We are unique in that we bring a suite of tools and technology that can sit on top of a large company’s base platform and complete the job to be done. We can build products and solutions with our partners and say with a high degree of confidence that we’ll succeed together, and we’ll have the tools that we need to get there—because we do.

6. What are the biggest challenges of working with AI right now?

One of the bigger challenges is tied to the misconception around AI. Almost everyone I talk to who isn’t an AI researcher says that currently AI, and by AI they mean large language models, feel magical. For any problems they can’t solve today, it’s unclear whether that problem will be solved tomorrow by AI (LLMs). So they can’t make decisions about what tech to use because everyone around them is saying that this tool tomorrow will solve all of your problems even if it doesn’t work today.

AI as a field is as diverse as, let’s say, a bookstore. Saying that all AI is the same as language models is like saying all books are cookbooks. But in a bookstore there are also fantasy books which help you imagine a world you’ve never been in, history books that tell you about the past, etc. Similarly, all of the different types of AI can solve different problems. There’s so much noise and hype around generative AI and LLMs that it can be hard working with partners and getting them to make good decisions: often they just want to use the most recent exciting thing, because they don’t have enough depth of expertise in this subject.

As a result, we try to do “white box AI” instead of “black box AI,” where we really explain all of our tools, including their strengths and limitations. That way we can confidently explain each tool and why we’re picking it, what it will do for the partner, and how it will work, so that they can make good decisions.

7. What is the most exciting future change that could come from AI?

How it will help us make human, sensitive, intelligent, informed decisions during a fast-moving crisis.

To me personally, I want all the volunteers who show up in the face of natural disasters to get to contribute all of their energies to make the biggest difference. The best of humanity shows up in those moments, so if we could actually take advantage of that then we could make a huge difference. With Human Collaborative AI you can have knowledge about everything in that moment that is needed and changing: someone just got saved by someone in a boat, so now that boat needs to go somewhere else, for example. Everything is constantly changing in the first two to three days after a disaster: the people who are available, the available resources, who needs to be rescued, where people are sent, etc. That is a whole year’s worth of activities happening over the course of three days, so if you could position everyone where they are most needed then just imagine what could be accomplished. For example, if a doctor is moving sandbags, and a construction worker is driving the water distribution truck—well ideally they should be at the places where they can do their work the best. Ideally the construction worker is building something, the doctor is in a hospital, and the volunteering grandparent can drive a vehicle, let's say. It’s finding the right person for the right task at the right time.