Transcript

AAAI Presidential Panel Discussion: Perceptions vs. Reality

From AAAI · Episode page · Watch on YouTube

· 1 h 2 min · 11,531 words · 75 timestamped sections

This is the full, timestamped transcript of "AAAI Presidential Panel Discussion: Perceptions vs. Reality", an episode of AAAI published June 23, 2025 (1 h 2 min, about 11,531 words). It mentions 19 people, works and ideas, including Rodney Brooks, AGI, Gary Marcos, Peter Stone, TripleAI.

Who speaks in this episode?

What is discussed, and when

  1. – 10:00: Rodney Brooks, Gary Marcos, Peter Stone, Franchesca Rossi
  2. – 20:00: AI Winters
  3. – 30:00: Gary Marcos
  4. – 40:00: Chinese Room, Turing Test
  5. – 1:00:00: Neurosymbolic AI, AI Index, TripleAI, UT Austin
  6. – 1:01:59: Tom Dietrich

Notable quotes from "AAAI Presidential Panel Discussion: Perceptions vs. Reality"

  • Work at IBM research and I'm the past president of tripleAI and I would like to welcome you to the triple AI panel discussion on AI perception versus reality. before we get to…
    about Franchesca Rossi · at
  • Claims that are made about this technology. in fact in the community survey of the report of the triplei report we asked if the current perception of AI capabilities matches…
    about TripleAI · at
  • Thanks Tom. Gary Marcos, professor emeritus of psychology and neural science at NYU and co-founder of geometric intelligence robust AI. Thanks Gary for being here. And Peter Stone,.
    about Gary Marcos · at
  • Linear threshold neuron just like we have today hundreds of billions in the language models. his report had on the cover introducing the perceptron. He only had one of these…
    about Frank Rosenblatt · at
  • The biggest risk is the misallocation of resources. you know the this if you if you really think I mean the premise of open AI and some of these other companies is that they…
    about AGI · at
  • Know a mismatch with reality on the negative side where people and that has sort of historically people talk about this sort of AI winters in artificial intelligence where…
    about AI Winters · at
  • Robotics is also robotics subject to hype and if so what is the biggest topic of hype in robotics the biggest topic of hype in robotics right now is humanoid robots And I like…
    about Rodney Brooks · at
  • Do what's you know what everybody else is working on. and so you know and that's going to happen. I think we need to actively prop up the people who have a taste for trying different things. Yeah. I'm gonna have to jump off the call. So I.
    about Peter Stone · at

Who and what is mentioned in this episode?

People

  • Rodney Brooks — Person · Panelist (referenced expert; main speaker on robotics hype) · 2 mentions · first at
    Explains historical AI hype cycles (perceptron era) and argues current hype confuses performance with competence; also critiques humanoid-robot hype from a robotics practitioner perspective.
  • Gary Marcos — Person · Panelist (critic of hype; media incentives commentator) · 3 mentions · first at
    Argues media and money amplify AGI narratives; critiques unaccountable predictions and highlights distribution shift and robustness limits in LLMs; discusses demos vs production reality.
  • Peter Stone — Person · Panelist (AI researcher; education/AI literacy advocate) · 1 mention · first at
    Frames hype/AI-winter dynamics, argues for matching expectations to reality, and emphasizes AI literacy education and methodological diversity beyond LLMs.
  • Tom Dietrich — Person · Panelist (academic perspective on evaluation and adoption) · 2 mentions · first at
    Discusses how hype misallocates resources and distorts regulation; critiques evaluation paradigms like the Turing Test and highlights limits of statistical learning for causal understanding.
    Of Iroot, founder, chairman and CTO of Refing Robotics. Te thanks Rodney for joining us. Then we have Tom Dietrich, professor emeritus of Oregon State University and former president of TripleA.
  • Franchesca Rossi — Person · Host / Moderator · 1 mention · first at
    Introduces the AAAI report on the future of AI research and moderates the panel on AI perception vs reality, guiding discussion toward hype sources and mitigation.
  • Frank Rosenblatt — Person · Referenced historical researcher (perceptron hype example) · 1 mention · first at
    Serves as a historical anchor for recurring AI hype cycles; his perceptron was marketed with human-mind-like claims, paralleling modern LLM hype.
  • Kevin Roose — Person · Referenced journalist (media incentive example) · 1 mention · first at
    Cited as an example of media narratives that may favor AGI imminence framing, contributing to public hype and miscalibrated expectations.
    AGI and so he wants it to be true and everything that he writes is sort of in hopes that it's true. He's not really writing objective journalism. He's actually a columnist. but…

Companies & orgs

  • TripleAI — Company · Business focus · 2 mentions · first at
    TripleAI is the organization hosting the panel discussion and publishing reports on AI research.
  • UT Austin — Company · Affiliation · 2 mentions · first at
    Mentioned as the institution where one of the panelists is a professor.
    Gary for being here. And Peter Stone, professor of computer science and director of robotics at UT Austin, study panel chair of the first report from 100 years study of AI and…
  • AI Index — Company · Referenced expert · 2 mentions · first at
    Referenced as an initiative providing factual data on AI capabilities to counter hype.
    Yeah, there are also initiative from the AI community like the AI index for example that show the current capabilities no and limitations of AI.
  • IBM — Company · Affiliation · 1 mention · first at
    Mentioned as the employer of one of the panelists, providing context to their expertise.
    Hi everybody. I'm Franchesca Rossi. I work at IBM research and I'm the past president of tripleAI and I would like to welcome you to the triple AI panel discussion on AI perception versus.
  • MIT — Company · Affiliation · 1 mention · first at
    Mentioned as the institution where one of the panelists is a professor emeritus.
    What we're doing right now. So let me introduce the panelist for this discussion. we have Rodney Brooks, professor emeritus of robotics at MIT, co-founder and CTO of Robust AI…
  • Oregon State University — Company · Affiliation · 1 mention · first at
    Mentioned as the institution where one of the panelists is a professor emeritus.
    Refing Robotics. Te thanks Rodney for joining us. Then we have Tom Dietrich, professor emeritus of Oregon State University and former president of TripleA.
  • NYU — Company · Affiliation · 1 mention · first at
    Mentioned as the institution where one of the panelists is a professor emeritus.
    Thanks Tom. Gary Marcos, professor emeritus of psychology and neural science at NYU and co-founder of geometric intelligence robust AI. Thanks Gary for being here. And Peter Stone,.

Concepts & topics

  • AGI — Concept · Core topic · 4 mentions
    Artificial General Intelligence is a central concept discussed in the panel, focusing on its perception versus reality.
  • Neurosymbolic AI — Concept · Key argument · 2 mentions · first at
    Neurosymbolic AI is presented as a promising approach that combines neural networks with symbolic AI, contrasting with the hype around LLMs.
    Neurosymbolic AI. I think neurosymbolic AI is quietly ascendant right now. alpha fold is a neurosymbolic system, alpha proof, alpha geometry. I think we're going and the use…
  • AI Winters — Concept · Key argument · 3 mentions · first at
    Referenced as historical periods of reduced funding and interest in AI, serving as a cautionary tale about hype.
  • Turing Test — Concept · Key argument · 1 mention · first at
    The Turing Test is discussed as a benchmark for AI that has been both influential and misleading in assessing AI capabilities.
    And so we with in the case of the touring test systems were able to pass the touring test in quite strictly and yet were completely useless because the underlying system didn't actually have a very deep understanding.
  • Chinese Room — Concept · Key argument · 1 mention · first at
    Discussed in the context of understanding language and AI, contrasting with the capabilities of LLMs.
    Turing test. Okay. It passed the Turing test. That was a surprise. it became the Chinese room which was a common thinking tool about understanding language and it is the Chinese room and it.

Full transcript with timestamps

Hi everybody. Uh I'm Franchesca Rossi. I work at IBM research and I'm the past president of tripleAI and I would like to welcome welcome you to the triple AI panel discussion on AI perception versus reality. uh before we get to the actual discussion, let me say more generally that uh as you may be aware AI capabilities have evolved uh significantly and very rapidly in the last few years and also AI research as a consequent has evolved a lot. uh so to understand the evolution of AI research along many dimensions who is doing the research what topics is the research about how AI research is conducted and even the geopolitical scenarios triple AI in March of this year published a report on the future of AI research which covers 17 topics around the transformation in AI research and for each one of them uh goes into detail details very succinctly in in three or four pages about uh the open challenges the

history and the context of that topic. Uh the study was conducted by 25 prominent researchers and supported also by a community survey with about uh almost 500 respondents and it has been written in a very non-technical way intentionally to uh reach out to audiences that are not AI researchers including experts of other disciplines, policy makers, funding agency, media and also the general public. Uh as I said the report included 17 topics. One of the topics was the perception uh versus reality of AI. Uh and this is what we are going to discuss today. Indeed over the last 70 years of AI research uh AI has produced a lot of uh new and important technology but also many times uh these have generated some excessive hype. uh like other technology these hype trends have followed the general gardener hype cycle characterization.

Um and while as I said AI has been hyped also in the past, the current generative AI hype cycle is the first time that AI has been introduced to the majority of people of the world which of course are less familiar with the technology many of them and so sometimes do not have the tools to understand the validity of many claims that are made about this technology. Um in fact in the community survey of the report of the triplei report we asked if the current perception of AI capabilities matches the reality um and 79% of the responses disagreed or strong strongly disagreed.

Uh we also asked if the mismatch is hindering AI research and 90% of the responses agreed with this uh statement saying that the direction of AI research are sometimes driven by the hype that theoretical AI research is suffering as a result and less students are interested in academic research. Um overall most respondents felt that the current perception of AI capability is overblown and this is a real impact having a real impact in the field and we should find a way to educate people about the reality of AI. In fact 74% of the respondents also said that there should be communitydriven initiative to organize public debates on AI perception of a reality with video recordings to be made available to all. And so this is what we're doing right now. So let me introduce the panelist for this discussion. Uh we have Rodney Brooks, professor emeritus of robotics at MIT, co-founder and CTO of Robust AI, founder

of Iroot, founder, chairman and CTO of Refing Robotics. Te thanks Rodney for joining us. Then we have Tom Dietrich, professor emeritus of Oregon State University and former president of TripleA. Thanks Tom. Gary Marcos, professor emeritus of psychology and neural science at NYU and co-founder of geometric intelligence robust AI. Thanks Gary for being here. And Peter Stone, professor of computer science and director of robotics at UT Austin, study panel chair of the first report from 100 years study of AI and chief scientist of Sony AI. So let's start immediately with trying to um eliminate some you know give an idea of what is this uh perception of AI and possible hype around AI. So uh I'll start with Rodney.

So there is a wide agreement Rodney the recent advances of AI has been accompanied by a heavy dose of hype. Can you explain what the hype is about? Is it about AI capabilities or AI risks or both or something else? Yeah. Well, let me start by saying this is not new. Um, the current language models use neural models and neural models have been around for a long time and back in 1957 when Frank Rosenblat had his first linear threshold neuron just like we have today hundreds of billions in the language models. his report had on the cover introducing the perceptron. He only had one of these uh perceptrons. A machine which senses, recognize, remembers and responds like the human mind. Um so, and that was one neuron compared to the hundreds of billions we have today. So, it was hyped um and it's uh cycled. And I think there's two things going on with the current uh hype. One is that the confusion between

performance and competence. The performance of the language models is astonishing to all of us. I think uh that's a fair thing to say. But people then believe that there's a competence a generality behind that uh which is equivalent to a human being. And I think that's one of the problems. The other problem is RC. Clark said many years ago any sufficiently advanced technology is indistinguishable from magic. If you don't have a mental model of how it works, you don't know what the limitations are. So people go off in two directions. One direction is it's going to do everything and us humans will be then, you know, be able to write poetry.

Oh, except the LLMs are better than better at that than they are and and eat grapes. And the other direction is, oh, it's going to kill us. Um, because we have no understanding of what the limitations are at the moment. So both sides are pushing too far. Thanks, Rodney. Uh, Peter, Gary, or Thomas. Yes, Gary. I'll I'll add a little bit. Um, going back to the thing that Rod just mentioned in 1957, both the New Yorker and I believe the New York Times took it in Hookline and Sinker and they had huge headlines about that. In fact, Rod might have actually been reading from one of those headlines if I remember correctly.

Um, and so the media has been a big part of this since the beginning. The media is very uncritical and wants to tell this story. It's much more popular to tell a story about how AGI is imminent. So we have like Kevin Roose at the New York Times has a book coming out about AGI and so he wants it to be true and everything that he writes is sort of in hopes that it's true. He's not really writing objective journalism. He's actually a columnist. Um but people don't understand that distinction and you know he's got this vision. He wants to sell this vision. Thomas Freriedman the same way now. So there's there's a lot of incentive to say hey we're in this really important moment in history.

the media often repeats the you know CEOs of open AI etc without any question whatsoever what they're saying um or Jeff Hinton was repeated saying you know all the jobs are going to disappear which is a form of hype right and he I just read a piece about him in entrepreneur magazine and it didn't even mention that he predicted that all radiology jobs would disappear by 2021 there was no accountability to the um past history so the media has played a big role and I'll just wrap money has played a big role too. So people have investments at stake. They want to make it sound like these things are great so that they can get the next round of investment. So there's a lot of dynamic and I think the money and media dynamics are much stronger than they have been before. And so I think that reinforces everything or makes everything that Rod is describing even more potent and dangerous right now. Yeah. And you can

criticize the the media um for it. But then there's also people whose job it is to do this. science fiction writers and and movies about AI have been around for, you know, longer than than AI. And that's also contributes to the hype. People often have learned or, you know, think they've learned about this through science fiction. Um, where, you know, the job is to hype it up in some sense and that's not as as uh I don't think, you know, to be criticized. That's sort of what the the whole goal is of science fiction.

Tom, do you want to add anything? I guess I mean I I think academics uh also have motivation to hype things because that's how they they uh uh recruit students, raise money for their research and so on. So uh while we're talking many many orders of magnitude less than what Gary was mentioning, uh I have yet to see any academic who gets, you know, hundred million dollar or a billion dollar or $10 billion grants, but uh but there's some of the same dynamic at work even in the academic community. So, but Tom, why does it matter if AI gets hyped? Doesn't every new technology require strong selling to get people to invest in it and adopt it? So, what's the problem? Well, I think uh there it's a long list and maybe my the other panelists will help but uh the I guess the biggest risk is the misallocation of resources. uh you know the the this if if you if you really think I I mean the premise of open AI and some of these

other companies uh is that they are going to produce uh a single sort of all-encompassing technology that will revolutionize everything and uh they have thereby attracted billions tens of billions hundreds of billions of dollars in investment is that really the best use of our investment funds um there's also the question of what we should be regulating. A lot of the focus in regulation uh has been about preventing some sort of artificial super intelligence from taking over as opposed to maybe more concrete problems of privacy, security uh a and correctness, liability say in med medicine or law. So uh we've also seen uh companies particularly I think the leadership in companies really drinking this Kool-Aid and forcing all of their employees to use AI. Uh and that's going to uh because the technology is not very mature that's going to lead to a lot of people being very unhappy and will actually prevent the technology from

being used appropriately when when when it is more mature. I was I guess going to address this question in a moment, but I I'll jump in now because um I think this is worth emphasizing that all of us on the panel actually really like AI, right? We're kind of disgusted by the hype and we think it's having negative consequences on the field of research of AI. All of us were brought into the field because we're interested in the research questions. I think all of us care about money to, you know, greater lesser degrees, but we were really fundamentally driven by a set of research questions. And I think one of the biggest ways that this hype is perverting things is it's perverting the research landscape. Um, and it's also perverting the policy landscape as as Tom um was pointing out. So, in the research landscape, it makes it very hard to do something that isn't an LLM thing. If you're building a company and

it isn't an LLM thing, they say, "Why aren't you working on LLMs?" and you like try to patiently explain that many of them are not really you know technically um conversant enough to understand the limitations which goes back to one of Rod's earlier points um or if you're trying to get grants and it's not LLMs you know it it becomes harder and then on the policy side as Tom said a lot of it then is built around these fantasies so another fantasy I think Tom didn't mention explicitly is the whole race with China you know if it turns out that LLMs are actually the answer which I don't think any of us on the panel really think it is then that race is going to be a tie and So, we're putting all of this money into a race that cannot even be won because everybody just has the same recipes and so it just, you know, it's up to the size of the servers and everybody's competing on the side of the

server. There's not even going to be a victory. But all this policy is made around what in that case isn't a winnable war. It might be a winnable war if it's not LLMs, but that would require researching other things, but that doesn't fit in the narrative that, you know, a lot of people who are well um well endowed economically are trying to sell. And so it really perverts both research and the policy. So Peter um can you reflect a bit more about what happened in the past hypes cycles of yeah you know Rodney mentioned perceptron in the late 50s but also expert assistance in the 80s. What happened then? Was it a good or a bad thing for AI research AI companies and society at large? Yeah it's a great question. I I mean I think it's it's actually there is some good and uh you know it's and it's sort of understandable why people um hype uh you know there's also what what hype really is is a mismatch with reality but on the

positive side on the overblown side um there is an equal danger of sort of you know a mismatch with reality on the um on the negative side where people and that has sort of historically people talk about this sort of uh AI winters um in in artificial intelligence where everybody believes that that the time of artificial intelligence has come and gone and it's been proven to to fail and sort of you know in some sense that has been what has happened from the um you know the hype in in the past is that expectations get overblown and then inevitably they're not met and in the long run people say oh I guess you know it was a failure and then you get into periods where um there's sort of the opposite effect where there's not enough there's not as much funding as there should be or there's not enough as many companies investing in it as there should be and um and that's equally a problem. So you know when you look at it

from that side then there is the positive of then you need to build up people's expectations again what would you know would be healthy for the field for the companies the researchers for society if we could keep expectations matching um matching reality uh much better and then you wouldn't get these sort of wild um wild sort of swings and um one and you know just to to wrap up on this you know you also talk about things like you know perceptron but but also expert systems from the 1980s The perception is that that was a failure.

Um, and you know that there was an AI winter that became became of it. But actually in reality, expert systems are still there. They're still out there in products. There's, you know, there's the equivalent of expert systems in your dishwashers and microwaves that are basically if then rules of if it's a potato, this is how much you should, you know, heat it for and things like that. a lot of what has um you know a lot of AI technologies from the past uh have stopped being considered by the public to be AI. Um but they're still there.

They're still in you know and um it sort of leads to the the some people say that you know AI is trying to get computers to do the things they can't do yet. Once you they can do it then it's engineering. Um and you know this sort of plays into the hype as as you relates to the hype as well because um you know people forget the uh the past it's always the new thing um but promises get get over inflated and then you get the the sort of trough of of um of under uh negative hype I guess is what I would refer to it. But I think ultimately even though there is some short-term gain to uh you know getting more funding or getting the you know as a as a CEO getting the venture capitalist you know hooked you're making a promise that if you don't live up to it it's going to you know it's going to come back to bite e either you or at least the field in the long run. And so um in the long run it's it really is a harmful thing to to

overinflate these expectations. If I could just if I could just uh step in for a second. Um the earlier AI winters were basically about funding. So funding dried up after people were disappointed by the results. I think that there's a second risk here. So that might happen again. Um particularly if these very high valuations don't pan out. If open AI ends up looking like a wei work, then a lot of people run for the hills and then investment actually dries up for a while. You know, LLMs are not going to disappear, but the high valuations might. So one side of it is economic and financial. The other side is we're starting to see some public backlash and I think that the policies of the current administration are actually playing into that. So it's again a short-term long-term thing. The short term is well we won't have any regulation on our companies because we need to catch up with China or whatever. But they're

likely to be some fairly bad outcomes. And already the public is feeling like AI is underregulated. So now if we have some massive cyber attack or some, you know, a whole bunch of driverless cars go wrong in some horrible accident or whatever, the public is going to be really pissed. They're going to feel like we were promised that this was great and it's not and it's causing, you know, harm to society and the government's doing nothing about it. The companies told them not to do anything about it. This is a mess. I would not be surprised if the 2028 election is the first in which AI is really a big issue.

Hardly anybody talked about AI in 2024, but I think in 2028 because of labor um and possibly because of overpromising and so forth, AI might actually be a big political issue. Yeah. So, it's interesting that you say there are things here now about the current hype that are a bit different from the hype of the 80s or the 50s, which is first of all that is not just hype in the positive sense as Peter said, but also hype about the risks. No and second there is much more public engagement and use of AI you know so those two things make the scenario of the hype and the consequences I think a little bit different so that uh we cannot really apply or learn from what happened from the past but we have to understand what's the current scenario is but let me ask one last questions about hype and then we'll go into uh other topics uh um um about influencing the hype which is uh I'll ask Rodney first when we talk

about AI hype usually we think about hype around AI software like LLM's for example capabilities but what about robotics is also robotics subject to hype and if so what is the biggest topic of hype in robotics the biggest topic of hype in robotics right now is humanoid robots um And you know, I like to compare uh the the hype about LLMs is because it emulates our language, which is very important to us. The hype about humanoids, I think, is driven by having the human form. Um, and the valuations there are crazy. Uh, people are promising that the humanoid robots will be able to do everything next week. Uh, and there'll be quick deployment. I'm in the weird position. I've been building humanoid robots since 1992. Uh, and I have actually sold thousands of upper torso humanoid robots. Um, no one else has sold any in the current companies.

And that company went bankrupt, by the way. But, uh, I'm not bitter. Um, but I think it the promise of the human form is people see it and think it will be able to do everything. And I think there's two uh two u problems with current humanoids. One is the way they walk is not the way we walk. Um, and so, uh, whenever it's about a robot, a humanoid robot is about to fall, a lot of, uh, energy gets pumped into the system to try and catch its balance. And if it does fall, that kinetic energy has to go somewhere. So, the only humanoid robots you see close to people are actually little short ones right now.

Um, and people think, oh, we'll just be able to make them bigger. that does the cube square law. The amount of as as you change the the height, the um amount of energy goes up with with the cube of that. So, they're just going to get more dangerous. So, we have to have a totally different way of walking for them to be safe. And the second thing is um no one has really figured out how to make them dextrous. Um, and there's a a lot of money being spent on farms of people showing robots how to do stuff, but I think uh none of that has demonstrated that's an an effective strategy to this point.

I'll add a third issue, which is situational awareness, common sense reasoning, and so forth. So, it's all well and good to have a demo in some factory floor that's been carefully constructed and so forth. But if you really want to bring these things into a home, you know, everybody's home is different and you have animals running around and little children and skateboards on the floor and so forth. We we don't really have the software yet to get humanoid robots to to really understand that. Um Rod is is the author of the best seller or co-founder of the bestselling robot of all time, the Roomba, which is a domestic robot that actually works in a limited circumstance, but it doesn't promise to, you know, do everything in your household in the same way. And it looks like he wanted to say something else.

Over 50 million runes have been sold. Hundreds of millions of other knockoffs. Um, but they don't promise to do much and they're small and intrinsically safe because they don't have much kinetic energy in them. Um, so yeah, and I mean we we should be drawing the distinction though between, you know, there are robots and there you Rod, you talked about humanoid robots as the as what's being overhyped. I mean there's there's definitely an increase in um in potential I think for you know robotics in general. Um and you you have a new company that you're yeah you're working on it now. I think, you know, it's people do often ask about um, you know, whether there's an application that you could do with a humanoid robot that you can't do with a robot, you know, that that's somehow non-human in form. Like why two arms, why not three or why not one? Or why, you know, why two legs? Why not four or

why not wheels? and and um but I think you know a lot of it is this uh you know as Rod started up by saying that the the allure of being like us or to you know and it's you know anthropomorphification or anthropomorphizing robots makes then us then you know sort of project a lot of human capabilities onto them even if they don't have them they're promising more than they can deliver. Yeah. To be clear I I think it's a golden age for robotics in general. I just don't think humanoid forms are going to be the winner. we have so much more computation we can put on board. We can run so many neural models and and get some awareness around the robot that we couldn't possibly do before. So it it really is a golden age. Uh but it's it's going to be very uh directed at particular tasks is where it's going to play out. There's a parallel with the AGI kind of debate um where there there's sort of two

arguments people have. One is are LLMs useful and the answer is absolutely yes not for every purpose and they have their limitations and so forth and the other is are they the sort of allpurpose magic silver bullet that we call AGI artificial general intelligence and the answer for now is clearly no my personal view is that LMS will never be that and there's something similar in robotics like one question is are they going to be more useful and widespread in the next decade 100%. The other is are they going to be universal helpers that can do anything in your home or whatever?

And the answer is no. So there's a bit of a parallel there. Yeah. And just to respond briefly to one of the questions in the the from the the audience, I think you know it is it is useful to think of robots as um embodied artificial intelligence that that uh you know um autonomous robots you know chat GPT and large language models are disembodied. They don't interact with the world. Now we can put large language models onto robots and then make, you know, um embodied uh large language models that that can interact with the world. So um I I just saw that somebody was asking whether we should use the word robot or whether we should call them physically embodied AI. And I think that's that's a you know that is a um a useful way of thinking about it. It's it's an auton a good autonomous robot is an embodied artificial intelligence.

Yeah. Let let me let me just say then then people get the idea that okay because you could can put an LLM on a robot that's what everyone should do. I had this um you know I'm putting robots in warehouses. We have them in warehouses in multiple countries and and and and uh they're for fulfill fulfillment. And someone from a the largest seller of GPUs in the world said, "Why don't you put an LLM on it and then you can tell the robot to go to Isisle 3 and get a tube of toothpaste?" And they said, "No, every every few hours we get a wave, it's called a wave of 3,000 orders that have to be on a truck in two hours. I'm not going to talk to the robot about it. I'm going to use data processing to push out the instructions." So, you know, it's just not appropriate at scale. It's might be fine for a demo. Uh, and I'm not saying not not ever have LLMs on robots, but it's not the solution to every robotics

problem. Let me just throw in one thing. Um, from talking to Rod over the years and other people, but primarily from Rod, I learned very much the difference between a demo and a finished product. And I think a lot of what goes on is that someone like Elon Musk shows a demo or figure or something like that. Could be robotics. can also be AI and it's easy to make a demo. It's hard to make something production grade in the real world at the scale that you need it. And so partly the difference of demos is like just the instruction that you need, but part is like robustness doing it throughout the whole world in which you need to do it. And the naive audiences often just don't understand the difference. So like if you look at almost any robot demo and there's a bunch of objects, those objects will be far apart and the robot will pick them up and do something clever with some instructions. Now you do it in the real

world where the objects are closer together for example and the vision doesn't work as well, the manipulation doesn't work as well, maybe you have to pick up one object from under another and suddenly it doesn't work nearly as well. Um, if you just look at a bunch of demos like on YouTube of robots and you start to develop a sensitivity to like what they do in these sterile environments and what the real world would be like and there's often a huge gap. Okay. So now we understand what the hype is around AI around robotics and so on. But and we also uh we also understood and talked a little bit about what generates this hype but let's go a little bit deeper in that. So Tom uh the perception of AI and possibly also the hype is also influenced by experimental results of tests of AI behaviors. So what do you think has been the influence for example of the touring test or also of modern performance tests for AI

systems? Well, uh I think it really comes back to a bit of what we've been saying. The touring test was originally called the imitation game and it really rewarded people who could imitate human behavior include or computer programs that could imitate human behavior including human human flaws uh and typos and things like this. And uh and if we look at uh what statistical machine learning is, it is essentially a technology for training a system to imitate the its training data. And so we we with in the case of the touring test uh systems were able to pass the touring test in in quite strictly and yet were completely useless because the underlying system didn't actually have a very deep understanding of the world. uh uh and similarly we see with these large language models that uh that they can easily pass the touring test uh and and do all kinds of wonderful language behavior but uh the

the un in some sense they're missing the underlying causal structure of the world and I think part the fundamental problem is that they're built on statistical learning and let's make it clear my entire career has been devoted to studying statistical machine learning um and so It's uh it's a kind of bitter pill to realize that that it is also the source of a lot of the shortcomings. Yeah. A good example. So I think of statistical learning as essentially like a correlation seeking tool or I sometimes call it an information seeking missile. Um it looks for correlations between the inputs and the outputs in in the data that's being being used to train the system. And uh there was a very nice demo of where this goes wrong where a computer vision system was shown images of wolves and images of dogs and was asked to distinguish between them.

But the wolves were all photographed against a snowy background, the dogs against green grass or green forest. And the machine learning algorithm, of course, the simplest, easiest correlation to find was, oh, if there's snow, well, the computer doesn't know it's snow, but that white stuff in the background, wolf. Green stuff in the background, dog. Um, and so the problem is that that a correlation is not causation. And it's been very hard to discipline our machine learning algorithms to only learn causal relationships. In computer vision, we're often able to do this by uh detecting and cropping out the background so that it only gets to see the dog or the wolf.

But uh but but each new case uh requires more intervention like that and we don't have any technique that really scales. So I think the um we we've been playing statistical learning has been playing the invitation game at scale um and we still don't have ways of learning uh underlying causal invariance. Thank you. Anybody wants to chime in? Um 100% agree. But Gary, so um you know that Apple and you wrote about it of course recently published this paper by now kind of you know very visible but also very controversial called the illusion of reasoning showing that LLMs have issues reasoning about simple problems and these issues increase with the complexity of the of the problem instance. So is this kind of analysis useful uh to give a realistic expectation of the capabilities of current AI systems?

There's no way to understand the capabilities except to do empirical research. I don't think it's a perfect paper. Um and in a way I don't think it's a surprising paper because there are other papers um from RAW and and from my own work over the years and so forth that showed similar results. This one was much more popular and that's actually important. um it was much more popular because Apple published it. But I think the general conclusion of the paper is right that these systems are not really doing at least what I would call reasoning that they're not very robust. Maybe another way of putting it is they're not very robust at um executing algorithms. So you know one of the clearest examples in the paper was Tower of Hanoi which is a simple game with three pegs and a bunch of rings.

And if a human learns the algorithm they can do it with a problem of arbitrary size more or less. We could talk about memory and attention and so forth. Um, but these machines could do it for seven and then they failed at eight rings, which should not be that hard. And there was this paper called the illusion of the illusion of thinking that tried to make it sound like this was a limit on number of tokens, but it's not actually an output limitation on the system, right? The the problem uh with eight discs only requires 255 steps. They easily could have put the output there and they didn't. Um, and that shows you that you don't really know what they're doing. And it it's also an engineering issue, right? If you think that you've tested your system on a small version of the problem and it's going to work on the large version of the problem, if you have a genuine recursive algorithm, you can actually be confident you can do

verification and proof or whatever. With these systems, you really can't and you really don't know. The broadest limit, and I see broad wants to go, I'll just wrap it up. The broadest limit that's at the core of their paper is really what many of us now call distribution shift. It's been something that I've harped on throughout my career, which is these systems are very good when things are close enough to what they've been trained on, but they have they struggle when things get further away. And what counts as further away is more complicated with earlier models and in fact we can't even tell um from the outside. But all of the failures that we see or many of the failures that we see over and over again are really about generalizing beyond the distribution.

And it goes to Tom's point about statistical learning. There may be some form of statistical learning that we develop that has more causal understanding, but the way that we're doing now is really between very low-level features and that's never given us the robustness to generalize as much as we should. Yeah. So before I give the word to Rodney, let me correct what I said because I said the paper was titled illusion of reasoning while the title is the illusion of thinking of course. Uh, and so, uh, just want for the audience to be correct so that they can find it if they didn't see it already. Rodney, go ahead. Yeah. So, I I I I just want to talk about LLMs and people's perception of them a little.

You know, LLMs have have gotten rid of or or struck at the intellectual basis of a lot of stuff and it was surprising. Turing test. Okay. It passed the Turing test. That was a surprise. uh it became the Chinese room which was a a common thinking tool about understanding language and it it is the Chinese room and it just showed up uninvited. And the third one is um linguistics and um the the work on uh uh from uh Chsky and that there has to be this innate structure to learn language and the LLMs don't have that innate structure. So there's questions uh but at the same time I think people who understand and who work in AI don't think that LLMs have intent but somehow they want to find a way that they they are reasoning through language. Um so that's a that and I don't know whether the rest of the panel agrees. Do do people do do academics ever think that LLMs have intent? Uh I'm not so sure how they do. Well, I I I

mean I think if you give them a prompt, you can try to give them intent, right? Uh and and obviously this whole move toward agentic systems, which is more or less programs written using LMS as components. Uh the intent comes from the designer and the user, not from the LM itself. I mean, the LM is some is sort of a transducer between that and behavior, right? Yes, but there's some confusion about this in the AI safety literature where people are starting to write about how the LLMs are trying to deceive people and so forth. And I think a more correct way of interpreting that is they know from their training data lots of cases where humans have deceived one another, for example. And what they're doing is essentially they're pastiching texts together that they've seen before. Um, and we could have an argument about that, but I think as a rough uh approximation, they're past teaching things they've seen before and they've

seen people lying. They have seen science fiction stories about computers not shutting down and so forth. Um, so I don't think they're intending to lie to the user or to resist shutting down, but they will say stuff like that and it does pose a practical problem, right? Even if there isn't any intent, even if you take out the anthropomorphic understanding of them, which I think you should, I think people should not anthropomorphize them, you still have a practical problem that they don't do the things that we ask them to do. Um, my view is that LLMs will never be fully aligned with human intent and that that's a serious problem. One way that this has manifested is people now write these system prompts, which is a prompt that the user doesn't see because it's executed before every session or every query, whatever. And they have things like don't tell the user how to make biological weapons, but that they can

easily be jailbroken. They don't treat that as a logical constraint the way that a classic um constraint satisfaction system would work. Um it's just a bunch more text that sort of primes this text and not that text and whatever. Or they're say told don't use copyrighted materials or don't hallucinate, but they don't actually have the the deep conceptual understanding to follow those system prompts. And so there is a huge problem of alignment even if there isn't literal intent. that just the way the system works out doing its bestishing means that it doesn't always align with what it's been told right and and I mean this tendency of people of anthropomorphizing AI uh is very present you know especially now that AI can write for example like a human being and this is a reality is not you know uh then then depends why it writes like a human being you know What is behind that? Then we can discuss and

and we know that it's not reasoning similar to a human being. But the tendency of automorphizing really uh makes people think that it has also all the other characteristic of human being. So he it knows the truth but it tells you something else. So it's a lie but it's not a lie because LLM doesn't know the truth. Right. So, so how so Peter, how would you see this as big be playing a big role in this problem with AI hype and does it generate over trust in AI? Yeah, it's a it's a great way of thinking of it is as people we need to generalize about other people. When I meet a new person, it makes sense that I should get a small amount of data about that person and then generalize that because I know people in general, I'm going to be able to predict uh what they do. We leverage this. You know, we give 16-year-olds driver's licenses after a 10-minute road test. Um because, you know, we know that it's a human driving.

They have I mean, we can argue about whether it's they're safe drivers afterwards, but um but you know, we know it's a person. They have a brain. They perceive like us. we we ascribe um personhood to them and then generalize aggressively and we need to do that. Um and it's you know it goes back to what we were just talking about as well. If somebody says something to you that that indicates that they have a survival instinct, we assume that they do have a survival instinct because they're like a person. But um it's very natural for us then when something presents to us as humanlike for us to use this same kind of um uh reasoning or same kind of response um even though they it isn't human. So it would be a mistake for us to watch an autonomous car driving for 10 minutes uh safely and then give it a driver's license, right? and and you know the car companies thankfully are are you know Whimo did six months of

testing in on the streets of Austin before they put the first person inside. Um but it's it is a tendency uh for people when they see something that looks humanlike to um to anthropomorphize and then trust very very much that it will be and like I say I think it is a rational response for us when we are dealing with real people but when we apply it to non- people it can lead us um lead us astray very quickly and I think that is that that does play a big role here. Yeah. And I think that's partly because uh people uh do generalize very smoothly, right? So if you if you see that they're driving safely if for the first 15 minutes, it's a very safe to to extrapolate that to to other situations.

But uh but our software systems do not have that behavior. Uh and and and you know, it's been described as sort of jagged intelligence or jagged knowledge. Um and it's very easy to for to to end up in a situation where where the systems do the wrong thing. Uh I think this also relates to to how we evaluate our AI systems, right? When we evaluate humans, we have these intelligence tests where we measure a very small fraction of behaviors, often very artificial ones, and then extrapolate to say, well, these people will uh you know, be are are very intelligent and will do very well with more training and so on. Um but but I sort of uh make the analogy that you know you can estimate the strength of a person by measuring their biceps but measuring those biceps of a robot doesn't really tell you anything about its strength right and it's it's this it's the same failure that the the underlying causal link between the test

you're giving and the true task is is missing uh in these AI systems there are as a result I mean there are people who are I think you know saying that there should be um very strong uh I don't know either legislation or regulations or something that mean that makes it so that people can very visibly detect when something is um an autonomous agent be it uh you know a chatbot or an autonomous car or something like that that you you know that you can tell visibly that it's not a human behind the wheel but it's an autonomous car because you want to to bring different um predictions to how it's going to behave. Um and uh you know so so um and same thing yeah same thing with large language models if you think you're talking to a person um then uh you may resp and you're not you may respond inappropriately and so you know I think it is reasonable that people are saying you it should be very clear to

you when you're talking to a person and when you're talking to a large language model so you can predict appropriately um Thomas so there is also another article published recently called titled AI as a normal technology technology and this article claims that it will take many years for AI to be adopted in industry in industry successfully and in a useful way while many companies think the opposite that AI is ready to be used to be deployed uh in many internal processes and also solutions for their clients. So he do you think that this mismatch is generated by AI hype and if so why and by whom? Well, so yes, I think uh that hype is playing a role here. Um as I've as I said before, I've I've heard executives and companies uh talking about some problem and saying, well, when AI when AGI is achieved, then this problem will be solved and and they have magical faith that that AGI will somehow solve all kinds of problems. Um

but I think uh what the the sort of normal uh uh technology article uh argues is that like all previous technologies it takes 10 to 20 years between the uh sort of creation of the technology and how we and and are figuring out how to use it. You know uh Kerszswhile talked about the idea of a of a technological singularity meaning that that change happens so fast that we can't see what happens afterwards. But in some sense every new technology poses a little miniature sing singularity in that we don't we don't really understand. So uh you know when when uh when um the people on this call were all graduate students while Peter's younger but uh um uh the TCP IP protocols for the internet were rolled out and uh and we sort of said okay the we have uh finished computer networking we can uh use uh move files from any computer to any other computer but we had no idea that social media would be invented and

that it would replace traditional media um and and have all kinds kinds of other large scale social impacts. So I think it's very hard to know how what will be the most effective ways to deploy AI technology. Now some companies are uh we could view what they're doing as exploring trying to find out what will work what will work and what won't work. Um but and and I think we will be engaged in that technological exploration. We need to have our eyes wide open about the limitations of the technology. Um but uh but when uh management says you will all use this technology uh I they're asking for trouble because the truth is that it probably will only really have a strong business case in a fraction of the of aspects and will really require re-engineering the the the whole corporation and and maybe all of its procedures. So, um, uh, the vision that it's somehow a pluggable replacement for a human worker is completely wrong and

is going to get people get companies into a lot of trouble. I think one way of thinking of this too is is that it's it's going to happen sector by sector, right? It's not going to be the raw large language model claude or chat GPT or Gemini that that that transforms the supply chain management industry or manufacturing or or sales. There's going to be a domain expert from that. um from that industry that that knows the pain points and that then becomes you know that works together with an AI expert or becomes conversant in AI and goes through this year-long proc years long process of generating and coming up with new versions of and fine-tuning software that is good for that industry. Um and uh yeah that will take years to to happen. I believe there there's a whole I'll just say one sentence there's a lot of notion out there of magical thinking that AI is magic and in going with that magic is the notion it'll be

instantaneous every sector every problem and it's just not real I I just want to you know um some of the big companies have said we're going to replace our programmers with with u u language based systems and at my company we use we use langu LLM systems for some programming uh techniques a very specific way and it's been very good actually for particular things I could say that in more detail but this has long been the dream when forran came out formula translation the idea was it would get rid of programmers because the you just write down a formula and the program would happen that was the first you know mainstream compiler for for science and it didn't get rid of programmers it just brought lots more programmers on so there's always a goal.

Um, you know, this new technology is going to solve every piece of my problem. I for for you know we I'm in San Francisco whole lot of um we had the um uh uh cruise vehicles and then Whimos up up the and and Zuks a little bit there. Anyway, but I uh started a few years ago asking every Uber driver, have you ever had to clean up vomit in the back seat? and they say, "Oh, yeah. Yeah, I Saturday nights I got to clean up vomit." And people are, you know, saying, "Oh, we'll have these cars that do everything, but they're not going to clean up the vomit." Um, so there's more stuff than than meets the eye in every job. And I think that gets lost.

Okay. So, let's see now and uh what we can do about this mismatch between AI perception, reality. So let's start with Rod because you have been publishing several predictions on AI and robots capabilities and then you regularly check this prediction every year on January 1st. Should we institutionalize this kind of initiative for the whole of AI or is it challenging to make them precise enough so that everybody agrees on the results of the yearly check? We asked this question in the community survey and 59% so you know majority but not really a very high majority of the respondents supported this proposal.

Yeah. So you know I think Gary mentioned uh an instance where someone had made a outrageous prediction which had a real impact uh and then now they're making a different prediction later and they're not held to it. Um, I think we should hold people to it and which is why I did it myself and hold myself to it. But I think there's no there's no there's no motivation for most people to do that. Um, and so uh you know people uh you know unnamed people like Elon Musk for instance um will continue to make wild predictions about their technology um because that's how they keep their valuations of their companies. Um, so I don't think we're going to be able to hold people to it unless we have you, you know, uh, a bunch of grumpy old people like some of us on this call, uh, certainly like me, um, you know, try to hold people to it because, uh, um, there's there's it's not as exciting as as thinking about the

possibilities. It's just not as exciting. I I actually offered Elon Musk a million-dollar bet on one of his things. Um, and got a little press out of it as an attempt to get the press to care about it. Like the Wall Street Journal picked it up and, you know, it was there was some utility in it, I guess. But they didn't it didn't stick. Like getting the press to care is what would really matter. If the press systematically held people like this accountable, it would change the conversation. But the press was interested in for five minutes. Oh, this is cute. This guy offered a million dollar bet. Obviously Musk didn't take the bet and he did I also offered to debate him. He didn't want to engage.

You know, it's not in the interest of uh people like this to actually engage in what is the accuracy of their predictions over time. It would only be if the press pushed it. Maybe if the public really stood up and said we want to understand like are we being lied to and made it in the press's interest, maybe that would happen. Um I think it's great Rod does those predictions. I put 25 predictions out this year about AI for for this year. um for what I thought would happen on my Substack. Um and I think it's good to hold people accountable. Yeah. I I I should say not all um press people are are overhyping.

I think I'll bring someone up that um uh I've got great respect for the articles he's been writing for Kade uh Mets in New York Times. I think they're much more careful than than say the columnists in New York Times about Yeah. the the day-to-day reporters are better than the columnists in terms of keeping to the facts. I think I I was going to say earlier I do think the level of nuance in the press about AI has gone way way up compared to earlier in my in my career. It used to be that either AI was going to destroy the world or AI was going to save save the world. It was all very polarized. And now I see, you know, in the last five years or so, I think there's been many more articles that are more nuanced about look, there's, you know, there's some positive effects, but uh but there's some risks as well, and here's how we need to balance them. I think, you know, the level of discourse around it has been been much more um I

think has been elevated a lot recently, which I think is good. I completely agree with that. And I also think that there's still a long way to go towards um having the press really hold people accountable. Agreed. I don't I don't think we should uh treat u medium articles as press either. Um so, okay. So yes, Gary, just want to go back to the fact that, you know, you propose this bet and then you regularly publish on your Substack, you engage a lot with other people on X, other social media. So to to really, you know, pinpoint the hype, the mismatch and uh debunk it through some factchecking. So, and even in the community survey, 78% of the respondents said that there should be a communitydriven initiative to counter the hype with factchecking claims about AI. So, do you think that this is something feasible or is it would be useful and people would listen and read it? I mean, in principle, maybe. I mean,

I sort of feel like sometimes a oneman army doing that. Um, and I get a lot of push back. I get called a troll and a pest and all of this stuff. Um it would be maybe nicer if there was a larger institutional effort um to do that. I I would support it um you know if people want to do it. I'd be happy to stop doing this if there was a clear firm uh institutional way of doing that. Yeah, there are also initiative from the AI community like the AI index for example that show the current capabilities no and limitations of AI.

So Peter, do you think that these are good and you know good enough ways to educate um all the AI stakeholders about the real AI capabilities of and limitations or should we think about something else? Well, so you know I'm I'm still an active professor. I think everybody else here on the call is an emerit is ameritus now. But I'm I'm still uh I'm still teaching um and a lot of the teaching I'm doing now and and this is not just me. This is I think around across the country it's happening more and more is at the um is about to the broader community now a so I taught I introduced a course last year um that called AI the essentials of AI for life in society it was essent it was essentially an AI literacy course that was aimed at the entire university population people with no computer science background and I think we're at the point where nobody should graduate from from uh you know college from

university with without having at least one course in artificial intelligence and even backing that into um to high school and and um you know and middle schools that that uh it's have being conversant about what AI really is, what it isn't, what the risks are, what what are the myths. um and I I think is is becoming a fundamental uh you know sort of tool and and skill in our society and and we are not um I think we we would be dropping the ball as educators if we have people graduating not being having at least the basic literacy with artificial intelligence. And so um I would love for this to be a required course. I've made all the the materials for this course available by the way.

They're on my website and anybody can look at them. the videos, the assignments that don't require any programming ability um or background um and uh and all the reading assignments. And so um I would encourage people to look at that. We've gotten great feedback from from people on on our campus. We also and you know going to the next level at UT Austin, we have an online masters now in artificial intelligence um at a with at a low price point. It's I think $10,000 um to get a master's degree in artificial intelligence for 10 courses. that's also trying to bring uh allow um people in the workforce to to upskill. Um people can take this while they're you know while they're working do it in evenings or they can do it you know right after uh college and we're getting a lot of good feedback on that as well. I think we're at the point where yeah everybody needs to to have you know some some degree of literacy and then we also need

to really broaden the pipeline of how we can give people more um at least some of the the workforce deeper skills in AI. But yes, I've also been involved in in you know the AI 100 study which is related to the AI index. Think all all these efforts are are you know just in um really essential in people from the inside people who know what AI is educating um rather than us relying on and as we've said before science fiction and and media um and so you know hope I hope this panel plays a role in that as well. So I have another question for you Peter. So you know that tripleAI uh is an association and the conference that covers many areas of AI not just machine learning related areas you know but also symbolic kind of rulebased AI areas. So how is this reflection about the reality of the capabilities and limitation of more modern AI as some people would call it leads to the need for the field to

encourage a diversity of approaches. So yes, LLM, yes, machine learning, yeah, but also other ones because in the community surveys um 80% of the respondents agreed on the need to educate the public including the press, the VCs, about the diversity of AI techniques and research areas that sometimes is overlooked and maybe people don't even know about it. Yeah, this is something I feel really strongly about. think, you know, the the uh the narrative right now, I mean, one one of the reasons that we've gotten to where we are right now with the big boom in in artificial neural networks is that um there was a diversity of of approaches.

There was a time when many people in the field thought neural networks were never going to work and they were sort of on a sideline. Um but the community kept uh there was a part of the community that kept it going. uh the AI world kept a diversity of approaches including neural networks and now they're seeing the the day in the sun. Um, ironically, there are now people, you know, who benefited from that who are saying, "Okay, we won. Now people should only work on neural networks and we shouldn't do anything else." And I think that's a that's a huge mistake. The the what the whole message of this this panel is that there's an overhype that we do not yet have all the tools and algorithms to do everything, you know, to replicate intelligence fully, to do everything we that that we would like to be able to do with artificial intelligence. So we do still need to to encourage and support people who are looking at um ways to

combine symbolic methodologies and probabilistic methodologies with neural networks and support new um new paradigms that could maybe have a hope of building causal world models, you know, and causal models and causal understandings of the world which are still missing. And so um you know I think it's it's a tendency for people to sort of always jump on the bandwagon and do what's you know what everybody else is working on. Um and so you know and that's going to happen. I think we need to actively prop up the the uh the people who have a taste for trying for trying different things. Yeah. I'm gonna have to jump jump off the call. So I just say one last thing. Well first of all it was a wonderful panel. Thank you for including me. Uh and second of all, I completely agree with what Peter said and I think one of the worst consequences of hype is that it has yielded an intellectual monoculture and

we really need to pursue other approaches. Um including I think also neurosymbolic AI. I think neurosymbolic AI is quietly ascendant right now. Um alpha fold is a neurosymbolic system, alpha proof, alpha geometry. Um I think we're going and and the use coding at the back end, right? That's symbolic AI that it um the LLMs uh create for themselves and then run through symbolic interpreters. So I I think we are getting to where neurosymbolic AI is quietly rising but this hype has led to a monoculture that has delayed all of that and that's been a negative consequence and hopefully we'll move past it. Thank you all very much.

Byebye. add and one thing at my company um where we're deploying robots I've counted 13 different AI technologies that we use on the robots you know including neural enhanced stereo but VSLAN geometric path planning some of them 60 years old and that's how real systems get built with lots of different pieces yeah u okay so we are getting to the end we are top of the hour but let me ask the panelists if you you can share with us a final advice about AI perception versus reality. Uh, Tommo, I'm I'm not I guess my my advice is uh uh write your Congress people to keep the National Science Foundation funded because it was during the during the in the 1980s when uh expert systems crashed and uh it was only the National Science Foundation that was funding core AI research and we need that uh that the their they invest in a very broad portfolio of of technologies ies uh always pushing things forward and uh

that's very important for our country. I realize that we have international audience as well uh but but that's an urgent problem here in the United States. Thanks Tom Peter. Yeah, I mean I would totally agree with with with Tom and just you know on top of that um I would say you know if there's I think many people in this audience who are already experts in AI but for those of you who aren't you know there are now resources educate yourself about um about the reality and uh and try to you know to be able to separate out what's um you know what is hype and what is what is fact um and I think the more uh the more conversion and literate people that we have both among the general public and then also among policy makers that will also lead to people understanding the the need for the basic research and for National Science Foundation that that Tom was saying and I I would say I agree with NSF not just for AI or computer science

Brussel it's a it's a it's a disaster for research in general but I would also say to people uh don't follow you know don't be like a 5-year-old playing soccer where you all get around the same ball um because it's going to be really hard to compete there's too many people there go off and do something different and you could have a big impact um um um by by being different. So be different. Thanks. Okay. Thanks. Let me thank you all the panelists for a very insightful discussion about this topic which I think it was innovative also compared to other discussions. Um and I hope it shed some light uh or given some food for thought for the audience on AI perception virtual reality. Uh thanks also for all of those who joined us and attended this panel. The panel has been recorded and will be available on the triplei website and also on the triple aai YouTube channel. So thank you all and uh hope to see you soon. Bye.

Questions about this episode

Who speaks in "AAAI Presidential Panel Discussion: Perceptions vs. Reality"?

Rodney Brooks (Panelist (referenced expert; main speaker on robotics hype)), Gary Marcos (Panelist (critic of hype; media incentives commentator)), Peter Stone (Panelist (AI researcher; education/AI literacy advocate)), Tom Dietrich (Panelist (academic perspective on evaluation and adoption)), Franchesca Rossi (Host).

What is discussed in "AAAI Presidential Panel Discussion: Perceptions vs. Reality"?

In order: 0:00 Rodney Brooks, Gary Marcos, Peter Stone, Franchesca Rossi; 10:00 AI Winters; 20:00 Gary Marcos; 30:00 Chinese Room, Turing Test; 50:00 Neurosymbolic AI, AI Index, TripleAI, UT Austin; 1:00:00 Tom Dietrich.

What is mentioned in this AAAI episode?

19 people, works and ideas are mentioned, including Rodney Brooks, AGI, Gary Marcos, Peter Stone, TripleAI, Tom Dietrich. Each one on this page links to its own page with every episode that mentions it.

How long is "AAAI Presidential Panel Discussion: Perceptions vs. Reality"?

The episode runs 1 h 2 min; the transcript is about 11,531 words in 75 timestamped sections.

More from AAAI

About this transcript: the episode's spoken words as captioned on YouTube, split into timestamped sections. Reclip.Me locates every person, work and idea in the text, keeps the time it is first named, and quotes the words around it verbatim. Last updated January 24, 2026.