Our group, the Cognitive Anteater Robotics Laboratory (CARL), creates brain-inspired neural networks to control robots. We do this for two reasons: 1) By embodying the artificial brain in a robot, we can study the robot’s behavior similar to a psychologist examining a person’s or animal’s behavior. Unlike psychologists and neuroscientists, we can record from every aspect of this artificial brain during its lifetime. 2) Despite the amazing progress in AI recently, artificial intelligence is no match for biological intelligence. Taking inspiration from biology may lead to better applications in many areas (healthcare, disaster relief, exploration, climate change, just to name a few). That being said, our field requires us to use the latest AI architectures and methods.

The goal in our research group is to understand how language works: how it is that humans can hear or read language and, in real time, compute the meaning underlying that language; and how humans can start with a meaning they want to express and quickly compute the right words in the right order to express that meaning. From the evolution of the human species until about 2020, there was only one device in the universe that could understand and produce language at the level that humans do: the human brain itself. Now we have another one: large language models. So now a good research question is: is there some commonality between how the human brain processes language and how systems like ChatGPT work? It turns out there is and there isn’t. The internal structure of systems like ChatGPT isn’t much like the human brain. But at a higher level, there is a connection: for both the human brain and ChatGPT, a fundamental task is prediction of what is coming next. Your brain is always trying to anticipate what is going to happen next in any circumstance, and while you’re reading or listening to language is no exception: you are constantly guessing at the upcoming words, and this process of prediction turns out to be essential to how we understand language. ChatGPT is the same: the basic system is trained to do nothing but predict the next word in a text, like a gigantic auto-complete. So AI for us is an inspiration for understanding how language works for humans.

For the last few years, I’ve been working on a book - Intellivision: How a Videogame System Battled Atari and Almost Bankrupted Barbie® (MIT Press), which comes out in the fall. In the book, my co-author and I examine the impact of home computing on the rise of videogames, and how video games sought to make TV intelligent within the huge limits of computation. In a sense, this is an alternative history of AI. Now, I’m switching my focus to the metaverse and climate change. In June, it will be my 20th anniversary of starting research in the virtual world Second Life, and I’ve done research in several other virtual worlds as well. So it’s going to be fun for me to return fully to virtual worlds for this new project. I’m really interested in how we might make virtual worlds more climate friendly and sustainable by reducing their energy and resource use, and simultaneously how we might use them to reduce climate change and carbon impact in the physical world. I’m also interested in using VR (“virtual reality,” which I’d prefer to call “sensory immersion”) and AI to explore how climate refugees are using online spaces to create and build community as they lose their physical spaces to rising sea levels, as is the case with Tuvalu. So a few of the ways that I’m seeing AI show up in these spaces include interactions with non-player characters (NPCs) that are becoming more interactive and can remember and respond better when engaged; in-game AI that can help users build within the virtual world so that these spaces are more accessible to non-programmers; and apps and games that allow users build their own worlds.

There are both positives and negatives to how recent advancements in AI technology are reshaping the way we live and work. On the positive side, AI can identify patterns in highly complex, dynamic processes. For example, AI-driven climate models can be more accurate in their weather predictions than physics-based models; AI-driven drug discovery can be extremely impactful and even creative; and large language models have massively increased the speed with which we can do new scientific analyses by writing helpful code, for example. But on the negative side, these models take massive amounts of energy to train and run—and that has a serious climate impact. They also become another crutch for us to turn to when we don’t want to think for ourselves: it’s very tempting to just type a question into the chatGPT prompt and get an authoritative-sounding, clear answer. Why do any thinking for yourself, or learn anything at all, if the answers are always at your fingertips? The danger here is that we will begin to forget how to be creative, if we don’t ever have to produce anything new for ourselves. And finally, while these models are useful, many of them are locked away behind paywalls, which creates an equity issue. If you can afford to have access to these tools, you increase your competitiveness for high paying jobs, educational opportunities, and other drivers of upward socioeconomic mobility—a self-perpetuating cycle.

AI has already transformed various sectors, including transportation, healthcare, and software development, and its influence is expected to grow. Innovations such as autonomous vehicles are revolutionizing transportation, while in healthcare, AI enhances diagnostic accuracy. In the area of software engineering, tools like GitHub Copilot and ChatGPT are changing the development process. Research indicates that these AI coding assistants can reduce coding time by up to 50%, making software development more efficient. From my own experience, leveraging these AI tools has enabled me to undertake complex programming projects that were previously beyond my reach.

Yes and no. On the one hand, there has been rapid development, and for someone who has been involved with AI over several decades, it has been impressive to see how far AI has come. On the other hand, AI’s abilities have been overhyped. I roll my eyes every time I hear or read that AI is or is becoming “sentient”. True sentience or consciousness (whatever that is!) requires a living system. These are not living and do not have the same motivations as living organisms.

I study consciousness (in humans) for a living, so like Jeff I also roll my eyes when I hear that AI is just about to become sentient because it seems that way. I don’t think I agree that it requires a living system, but I do agree that there are a lot of limitations in current and near-future AI systems that make it unlikely to outpace us and take over the world, at least not like Skynet “did” in the Terminator franchise. One of the biggest limitations of our current AI is that it doesn’t know what it’s wrong. It can be easily tricked by so-called “adversarial attacks” — cases where you change just a few pixels in an image, and suddenly your fancy AI model decides that an image of a panda is definitely a gecko instead. Large language models show the same kind of problem: when they don’t “know” an answer, they will just make something up, and if you call them out on it, they often say, “Oh sorry, here’s another answer I made up” if they say something true and you accuse them of falsehood, they’ll even change their answer to accommodate you! Not all AI works like this of course, but these kinds of failures indicate that none of our current AI systems really “knows” anything. They’re maybe quite good at predicting what is going to happen next, but there is no “core knowledge base” that they “believe” like a human would, and no artificial system has any core motivations. A quote I like by Daniel Dennett (personal communication to a colleague, shared publicly with permission in a paper in 2021) sums this up nicely: “How do we go from doing things for reasons to having reasons for doing things?” Humans have reasons, but so far, AI systems do not. Maybe, given the rapid pace of advancement, this will change sooner rather than later, but I think that a shift in how these models work at their core will be necessary to make these kinds of changes happen.

Yes, AI is a very fast moving field. Admittedly, I was a little skeptical two years ago about some of the performance claims of large-language models such as ChatGPT but now I am quite impressed by the progress of ChatGPT on various performance benchmarks. As pointed out by MP, it is true that in some cases, generative AI models such as ChatGPT do not have a good understanding of their own knowledge limits. However, we were surprised in our own research to find that in some particular cases, ChatGPT actually does have some knowledge about what it knows and doesn’t know. For instance, analysis of ChatGPT's internal representations reveals a correlation between its confidence levels and the accuracy of its responses suggesting that there is some rudimentary knowledge about its own knowledge boundaries. However, a significant challenge arises when ChatGPT articulates its confidence when providing explanations and rationale for its answers. For reasons we do not yet fully understand, ChatGPT explains itself in an overly confident manner that is not aligned with its own internal beliefs. This misalignment highlights a broader issue within AI development—achieving general metacognitive abilities and a theory of mind. These aspects involve understanding the bounds of one's knowledge and distinguishing between one's beliefs, knowledge, and goals versus those of others. Addressing these limitations is crucial for enhancing AI's reliability and trustworthiness, and I am optimistic that we will see significant advancements in these areas in the coming years.

One issue is that AI is often used in ways that its developers might not have anticipated. Deepfake technology has a lot of useful applications, but then it gets used primarily to create deepfake pornography. I think a good solution is to have regulatory agencies - similar to the EPA - that are able to respond flexibly with new policy to deal with emerging AI-related issues.
 

If generative AI ends up playing a large role in everyone’s day-to-day lives, then the nature of the AI models is hugely important from an ethical perspective. For example, in a world where most text is produced by generative AI or with the aid of generative AI, if that generative AI only has one worldview, then it would become difficult to write anything that goes against that worldview. In a world where most media is generated by AI, if the AI model is biased toward some worldview, then that worldview would become “locked in” in almost all media in a deep way. You may already have encountered this kind of thing dealing with ChatGPT, where you ask it a question and it refuses to answer, or when it refuses to generate certain images according to its makers’ idea of what is harmful. From this perspective, an important question is who trains the AI, because the AI will end up reflecting their values. I think we should be extremely skeptical of people who claim they want to control AI in the name of “safety”: granting any group power over the contents of generative AI would be granting that group an enormous amount of power. I think the solution lies in the inherently democratic nature of statistical AI systems, which become more powerful as computation becomes cheaper. Currently only large corporations with access to large computing clusters can train a system like ChatGPT, but as computation becomes cheaper, it will be possible for more people to create models which reflect their own values. If we regulate AI prematurely, or set limits on who may develop these systems, we run the risk of creating a scenario where only large corporate entities can create powerful AI models, and I believe this is profoundly dangerous.

When the internet first got going, it was mostly a tool for universities, the government and military. There weren’t Amazons or Googles or much commerce of any type in the early days. A huge difference with the metaverse, virtual worlds and AI is that they are forming now, in the mid-2020s, and 99% are owned by big powerful corporations. It’s a very different world when for-profits are overwhelmingly running the tools that we’re all now relying on. Regulation is going to be a very important ethical consideration, but also too are the creative possibilities and constraints. There are a lot of cool things AI can do that might not be profitable, and so a corporation with a narrow vision on dollars may not explore some of these uses that could have great benefits for society. This is why I’m really interested in one of the smaller virtual worlds out there called Wolf Grid, which is a rare non-corporate owned world where some really interesting builds and developments are happening. Of course, another ethical consideration we need to always keep in mind and address is the way bias can creep into AI and algorithms. A lot of great researchers are already tracking how these forms of bias get encoded and how they can be challenged.

Absolutely. As I mentioned in other answers, “AI” and machine learning in general can discover patterns in very complex, dynamically behaving systems that humans either may never discover, or that it would take us a very long time—and lots of expensive experiments—to work out. We already see successes in medicine, climate model predictions, and so on. If we can build AI models that we can trust, and we can work together with them as partners in scientific and technological discovery, I am extremely excited to think what we will discover together. This is one place where alignment may also play a role. Alignment refers to working to ensure that an AI system makes decisions using similar internal representations and decision policies as a human would. In other words, we want to make sure the AI makes decisions how we do, not just the same decisions that we make. Often, alignment would be pursued in cognitive reasoning, classification, or similar kinds of domains where we’re trying to build AI that can help solve problems. This could help make AI a useful partner in science or technological development, because it would “think” more like a human partner. But we could also use alignment approaches to ensure that the value an AI system places on e.g. equitable treatment or access to resources is as high as (or higher than) the value assigned by humans. Establishing guidelines for AI development that targets ethical or moral alignment with values shared across cultures, and those which preserve the habitability of our planet, may help steer us in a positive direction.

I think biologically inspired AI. It is already happening. Current AI is reaching a plateau and people are looking to biology to make their systems learn faster, learn more, and adapt to change.

I think big breakthroughs in AI are going to come through big breakthroughs in hardware: especially chips that can run the computations needed for AI in a way that is much more energy-efficient than the existing hardware. We know that it’s possible: the human brain uses much, much less energy than any graphics processing unit.

There’s an emergent area exploring how AI is actually used. A lot of research is on what it does, but we also need to track how it’s actually used. When we look at the history of technology, we see over and over again that a technology designed for X purpose gets used for that purpose, but then gets used for Y and Z as well—new uses never imagined at the outset. After all, the internet wasn’t originally intended for Netflix, and yet, here we are! So it will be interesting to see the unexpected ways AI will be used.

The pace of AI development is truly, gobsmackingly astonishing. With jaw-dropping announcements coming from all angles at rapid pace, it’s beginning to look like the dawn of some of our favorite science fiction narratives: powerful systems that can help solve problems humans could never hope to understand, or which can facilitate discoveries that can make all our lives better. But it’s scary, too. I’m not scared that AI is going to “go bad” and turn us all into batteries like in The Matrix or anything, but I am worried that its continued advance will both exacerbate the societal and environmental problems we have now while also slowly chipping away at our creativity, critical thinking skills, and independence. It’s already bad enough for us all when telecommunications systems, transportation, or other utility infrastructure goes down for a few hours. Should we become even more completely dependent on AI for everything—education, healthcare, communications, etc.—if it fails to behave as we expect, we might not know how to do much for ourselves anymore.

Sure. We have collaborative research with Liz Chrastil in UCI’s Biological Sciences and Doug Nitz in UCSD’s Cognitive Science on how people, animals and robots find their way around. You may see some of our robots roaming around Aldrich Park. We are working in the area of neuro-inspired computation. Just recently our team was granted a Beall Applied Innovation Proof-Of-Product award and invited to participate in the National Science Foundation’s I-Corps program to investigate the commercial viability of our software framework. Lastly, I have been working with an enthusiastic interdisciplinary group of undergraduates who are creating a really cool intelligent robot. We are in stealth mode currently, but expect to hear more from them this Spring!

In a new project with the Honda Research Institute, I am broadening the scope of my research on human-AI collaboration to include aspects of well-being as well as efficiency and accuracy. Autonomous agents will play a large role in mobility in future societies through autonomous vehicles and robots, and it will become increasingly important to cooperatively adapt the behavior of these autonomous agents to promote prosocial behavior that can affect overall wellbeing. The project's goal is to investigate the tradeoffs between prosocial behavior and autonomous agent efficiency. Our long-term goal is to understand how prosocial behavior of autonomous agents can improve prosocial behavior among humans and create better societies.