@Chris Culp : 我认为人工智能的发展并非一蹴而就，而是经历了多次技术和算法改进后的阶段性发展。目前AI的蓬勃发展，特别是生成式AI的兴起，主要得益于并行计算能力的提升（例如图形处理单元GPU的广泛应用）以及算法的重大改进（例如卷积神经网络CNN和Transformer模型）。然而，未来的发展可能面临训练数据数量和算力瓶颈，算法创新也至关重要。我们需要关注训练数据的质量和代表性，避免模型过度拟合或产生错误的输出。此外，AI的能源消耗和环境影响也是需要考虑的重要因素。总的来说，我认为AI更像是一个由多个‘白痴天才’组成的系统，而非真正的通用人工智能（AGI）。我们需要对人类意识有更深入的理解，才能判断AGI是否可实现。目前，AI更适合作为人类工作者的协作者，而非替代者。在使用AI时，我们应该以解决问题为导向，确保训练数据集的全面性和代表性，并对AI的输出结果进行仔细审核。

我最近与Lycoming学院物理学教授Christopher Kulp进行了一次深入的访谈，探讨了人工智能（AI）的快速发展及其未来走向。Culp教授不仅是一位杰出的物理学家，也是一位备受推崇的科幻小说作家，他的研究领域涵盖非线性动力学、时间序列分析、网络和机器学习等多个方面。

AI的快速发展并非偶然

Culp教授指出，AI的发展并非一蹴而就，而是经历了多次技术和算法改进后的阶段性发展。早期的神经网络算法可以追溯到20世纪40年代和50年代。而如今AI的蓬勃发展，特别是生成式AI的兴起，主要得益于两大因素：

• 并行计算能力的提升: 图形处理单元（GPU）的广泛应用使得大规模并行计算成为可能，显著加快了AI模型的训练速度。
• 算法的重大改进: 卷积神经网络（CNN）的改进提升了图像识别能力，而Transformer模型的出现则彻底改变了自然语言处理领域，催生了强大的大型语言模型（LLM）。

挑战与机遇并存

然而，Culp教授也强调了AI发展中面临的挑战：

• 训练数据和算力瓶颈: 随着模型规模的不断扩大，对训练数据的需求也呈指数级增长。未来，我们可能会面临训练数据数量和算力（特别是能源消耗）的瓶颈。
• 算法创新: 持续的算法创新对于AI的进一步发展至关重要。我们需要探索新的算法，以提高模型的效率、准确性和鲁棒性。
• 训练数据的质量和代表性: 高质量、具有代表性的训练数据是AI模型成功的关键。我们需要关注训练数据的偏差，避免模型过度拟合或产生错误的输出（所谓的“幻觉”）。

AI：协作者而非替代者

Culp教授认为，AI更像是一个由多个“白痴天才”组成的系统，而非真正的通用人工智能（AGI）。 每个模块擅长特定任务，但缺乏整体的理解和意识。 他强调了训练数据集的重要性，指出我们需要了解数据集的局限性，并将其与AI的输出结果结合起来进行分析。 这需要我们对AI的运作机制有更深入的理解，并将其视为人类工作者的协作者，而非替代者。

未来展望：高效、专业化和可解释性

在谈到未来发展趋势时，Culp教授对以下几个方面表示期待：

• 更高效的模型: 降低AI模型的能源消耗和环境影响，这需要在算法和硬件方面取得突破。
• 更专业化的模型: 开发针对特定领域和任务的专业化AI模型，以提高效率和准确性。
• 更强的可解释性: 改进AI模型的可解释性，使我们能够更好地理解AI的决策过程，从而提高信任度和可靠性。

结语

与Culp教授的对话让我对AI的未来发展有了更清晰的认识。AI技术正以前所未有的速度发展，但同时也面临着诸多挑战。 我们需要谨慎地发展和应用AI技术，确保其能够造福人类，而不是带来负面影响。 这需要我们不断探索新的算法，改进训练数据，并重视AI的伦理和社会影响。 最终，AI的成功取决于我们如何将其作为人类智慧的延伸和增强，而非简单的替代。

◉

02:37 AI 的发展并非一蹴而就，而是经历了多次技术和算法改进后的阶段性发展，目前AI的蓬勃发展得益于并行计算能力的提升（例如图形卡）和算法的改进（例如卷积神经网络和Transformer）。

◉

05:54 AI 未来发展可能受限于训练数据的数量和算力，算法创新也至关重要。

◉

08:27 未来两年内，AI 的发展方向可能在于“链式思维”模型的改进，使 AI 的推理过程更加透明，从而更好地与人类协作。

◉

10:52 透明的AI推理过程（例如链式思维模型）能够促进人机协作，推动科研进步。

◉

13:36 AI 可以作为辅助工具，帮助人们克服思维定势，激发创造力，但不能直接复制粘贴其输出结果。

◉

15:57 AI 的应用已经非常广泛，许多人甚至没有意识到自己正在使用 AI。

◉

17:11 当前的 AI 更像是多个“白痴天才”的组合，而非真正的通用人工智能（AGI）。

◉

19:51 AI 的主要问题不在于“幻觉”（错误），而在于训练数据集的质量和代表性。

◉

22:35 使用 AI 应以解决问题为导向，确保训练数据集的全面性和代表性，并将其视为人类工作者的协作者而非替代者。

◉

27:10 未来一年AI领域可能出现高效小型化模型、改进的链式思维模型以及更注重环境影响的模型。

◉

28:40 理想的 AI 工具应该根据具体任务进行定制，而非追求单一万能的 AI。

◉

31:13 AI 可以用于语法检查和文本总结，但仍需人工审核其输出结果。

◉

34:32 AI 能够解决复杂的物理问题，甚至能够对解进行解释，这令人惊讶。

◉

37:01 目前还无法确定通用人工智能（AGI）是否可实现，这需要对人类意识有更深入的理解。

◉

38:30 学习AI入门可以从Andrew Ng的Coursera课程或Darren Geron的《Hands-On Machine Learning》一书开始。

00:00

Forget frequently asked questions. Common sense, common knowledge, or Google. How about advice from a real genius? 95% of people in any profession are good enough to be qualified and licensed. 5% go above and beyond. They become very good at what they do, but only 0.1%. 语法解析

00:16

Richard Jacobs has made it his life's mission to find them for you. He hunts down and interviews geniuses in every field. Sleep science, cancer, stem cells, ketogenic diets, and more. Here come the geniuses. This is the Finding Genius Podcast with Richard Jacobs. 语法解析

00:37

Hello, this is Richard Jacobs with the Finding Genius podcast. My guest today is Chris Culp. He's the John P. Graham Teaching Professor of Physics at Lycoming College. So we're going to talk about machine learning and what's called nonlinear dynamics and how that plays into how machine learning works. So welcome, Chris. Hi, thank you, Richard. Thanks for having me for the show. 语法解析

00:56

Yeah, let's always start out the same way. Tell listeners about you. What's your background and how did you get to where you are? Yeah, absolutely. So as you mentioned, I'm the John P. Graham Teaching Professor of Physics at Lycoming College. I'm also a science fiction author. So I have like, you know, two jobs, if you will. As far as physics goes, I teach physics at all levels. So courses for non-majors all the way up to like advanced physics courses like quantum mechanics. I also teach physics at the University of Michigan. 语法解析

01:22

So teach classes in machine learning and data analysis and things like that. And when it comes to research, my research sort of emphasis are, as you mentioned, in nonlinear dynamics and machine learning and complex systems modeling, which we can talk all about, of course, in the interview. And as a science fiction author, I write both space opera and dystopian science fiction. 语法解析

01:43

I like to try to explore themes of sort of, you know, human identity in sort of societies of high technology or space travel and those kinds of themes. I also write about, you know, AI is usually a part of my novels as well in some form or another. 语法解析

02:03

I'm starting to ask all the people I interview about AI, what has changed in the past couple of years that's allowing AI to suddenly have this renaissance of this explosion? Is it just another layer of weighted factors and now we're getting these emergent properties or… 语法解析

02:22

Why now AI is so salient and so prevalent? What's changed? Yeah, so AI has gone through a lot of different phases over many decades, honestly. So some of the first algorithms for neural networks date back to the 40s, 50s. 语法解析

02:37

And so the field historically has these growth spurts and then these lulls. And a lot of times they correspond to technology. Graphics cards right now are highly parallelizable. They're relatively inexpensive as long as the price hikes don't go up and demand goes up or whatever. But we have really good graphics cards where a lot of these calculations can be done on graphics cards, which are highly what we call parallelizable. 语法解析

03:04

parallelizable, excuse me. And what that means is the card can do a lot of different calculations at one time, as opposed to having to do a calculation, wait for that answer, do another calculation based on that. So we see also, of course, improvements in things like algorithms. Like in 2012, we had improvements in something called convolutional neural networks to improve image recognition. So you might remember back in Facebook and like when all of a sudden you could start tagging your friends. 语法解析

03:31

Or it would basically say, oh, is this, you know, is this your wife in this picture? Like, well, how does it know that? Well, you've been helping to train their AI algorithms, you know, by tagging your friend in your photos. And so, yeah, so we see these like these fits and spurts, right, of based on either technological improvements or algorithm improvements. 语法解析

03:52

and you know this won't be the last one this is just one of a series and don't ask me to speculate on what's coming up next because I who knows right it's impossible to predict that but what's changed though in the is it 语法解析

04:07

The algorithmic change is that, again, there's just a lot more computing power. So you could have like thousand layer neural networks versus 100 layers. And now you're getting this better performance. Yeah, I think both actually. Yeah, I said, like I said, the graphics card's getting better, but also, yeah, the concept of transformers is a new algorithm or newish algorithm. The idea of attention where the algorithm can pay attention to particular parts of tech 语法解析

04:35

or whatever. So I think what we're seeing kind of is this sort of 语法解析

04:39

I'm trying to think of a word, basically this convergence of, in this particular situation, of a lot of different hardware and software developments to really have AI take off. And it's particular, this generative text, these generative images, generative movies. We're really kind of in, I think, a golden era right now of development for, like I said, hardware and software reasons. 语法解析

05:06

A lot of things. And you think it could be like, I mean, do you think AI is going to follow like a Moore's law trajectory or are we going to tap out soon and we'll have to figure out something else? Like, where do you think it's at in this growth curve or what does that look like? It's a really good question. So everybody wants to know that, right? And everybody has an opinion and every person you talk to will probably give you a little bit of a different answer. And, 语法解析

05:26

And so where am I thinking of? I think limits probably coming up in the near future would possibly be training data. So, I mean, right now we don't really know necessarily all the details publicly anyway of how much of the internet, let's say chat GPT-4 is trained on, right? But at some point we're going to run out of internet. So run out of books, right? Run out of things that we have produced. And 语法解析

05:54

Then the question is, you know, like where, what, what happens then? Right. And it could be algorithmic imprints. And so there might be things like, you know, when attention was developed, that was a, that was a big revolutionary idea that allowed, um, you know, large language models to actually come into existence. So once we quote unquote, uh, I don't want to say run out of training data because, you know, new data is being produced all the time on the internet. But, you know, once we have sort of saturated that training data, what's, what's next? 语法解析

06:22

Right. In terms of like how to train our data. Yeah. So I, yeah, it's, it's always hard to say, cause I don't like to be one of those people who's like, yeah, we're, we're running out of like, let's say for example, training data or running out of another, you know, another limitation is power, right. Is power consumption. And can we build power stations fast enough, right. To, to, 语法解析

06:42

to run the data centers, to train these models, which are really power hungry. Hey, what would happen in the U S like you'll, we'll need to build a lot of data facilities. Yeah. You know, and I've heard stories of, you know, for example, certain power companies, like, you know, Microsoft comes into an area and said, Hey, I want to build a data center here. And the power company is like, no, because we cannot get you the power. Right. We just can't. And, and that's why they kind of bought three mile, like part of three mile Island or whatever that deal was to try to like, you know, basically, you know, 语法解析

07:11

generate powers that can generate more complicated models. So I think we actually have some limiting factors in terms of training size, which is probably overcomable, like training size power. But then again, you know, power might be overcomable as well, because we're seeing things like deep seek, for example, that can do things more efficiently than we originally expected to be able to be done. So it's, boy, this conversation, and every time I hear it, 语法解析

07:35

Kind of reminds me of discussions about physics in the early 20th century, at the turn of the 20th century, where you had people like Lord Kelvin writing, physics is over, and people telling their students, don't study physics because there's just details left. And then lo and behold, you know, there's quantum mechanics. And the whole quantum revolution, and then we had chaos theory come up in the middle 20th century. So I'm always hesitant to make these kinds of predictions because I just… 语法解析

07:58

I know the history too well. Okay. That makes sense. Where do you see AI advancing right now? Mostly, it sounds like these LLMs, these integrated, well, these interactive like chats, I guess you could say. Now they're talking about AI agents that actually perform basic tasks and all that. But again, what do you see over the next, since things are moving so fast, over the next two years, what's going to become available for use by, you know, slightly techie interested people? Yeah, that's a good question. I think, 语法解析

08:27

What's gotten me interested right now in particular are some of these chain of thought models, which, you know, roughly speaking, the AI shows its work, right? And its steps and its reasoning. Now, there's been some studies recently put out, very recently put out, that when the AI is showing its work, it's not always being completely honest about all the work. 语法解析

08:46

and it's showing. But that aside, I've been having a lot of fun with my students playing around with these chain of thought ideas. And I can give you a good example. Something we did in the class a few weeks ago. There's a theorem that 语法解析

09:01

we were working with, I was teaching my students about. It's called Stokes' theorem, but that's not really important. If you have any physicists, they'll be nodding their heads. Everyone else will be like, whoa, what's that? But the point is that there's two sides to the equation. There's left side and the right side. And Stokes' theorem gives you two different ways of calculating the same thing, and you have to get the same answer with both of them. 语法解析

09:21

And so we had an AI. So we did a problem on the board and then we had an AI. It was Gemini that we used, solved the problem. And it did it one way and it got the right answer. And so then we asked it to do it the other way and it got two times the right answer. And so then we prompted it in class, like, so why? 语法解析

09:42

shouldn't these be the same? And it responded, oh, you know, you're absolutely correct. These should be the same. And I see my mistake. Let me show you. And it wrote out the mistake and what it identified, it identified the wrong answer as the right one. And that in its chain of thought processing, it fudged a factor of two for the, which was actually the correct answer. So it would match the wrong answer and both sides of the equation. And so it basically recognized, oh, this is not right. 语法解析

10:10

I think the wrong answer is the right one. It didn't know it was wrong, obviously. Right. And so then it just, and then it's chain of thought process. There's like this magical factor of two it's stuck in there. And so that you would get the wrong, the wrong answer to match the wrong answer. I think developments in these kinds of processes is what has me excited, even though this particular case was wrong. Because I, 语法解析

10:32

I think it's going to help us collaborate with AI. If we can start to reliably understand the steps that it takes, then we can actually use this tool alongside of scientists, alongside the AI working together and understanding, you know, how is it coming to this conclusion? And that can, I think, help. 语法解析

10:52

you know, further research. So that's one of the areas I personally find most exciting is these improvements. And with the study I'd mentioned earlier that we talked about how when these chain of thought processes appear on the screen, they're not always completely faithful to what the AI has actually done. I'm actually really excited about that study because now people are looking at this. We're starting to ask questions. How does the AI work? I don't know. How come we don't know an answer, right? 语法解析

11:18

But now we're starting to care, right? We're starting to care. It's like, well, how are these things coming to conclusions that they're 语法解析

11:24

Before we continue, I've been personally funding the Finding Genius Podcast for four and a half years now, which has led to 2,700 plus interviews of clinicians, researchers, scientists, CEOs, and other amazing people who are working to advance science and improve our lives and our world. Even though this podcast gets 100,000 plus downloads a month, we need your help to reach hundreds of thousands more worldwide. Please visit findinggeniuspodcast.com and click on support us. We have three levels of membership from 语法解析

11:52

from $10 to $49 a month, including perks such as the ability to see ahead in our interview calendar and ask questions of upcoming guests, transcripts of podcasts you're interested in, the ability to request specific topics or guests, and more. Visit FindingGeniusPodcast.com and click support us today. Now back to the show. 语法解析

12:11

Oh, that's really good. So what is the reason boom affected every part of AI or only certain parts? It seems like, you know, the LLMs, the large language models have been big time affected. They're answering a lot better. 语法解析

12:23

But what else do you see is improving because of this? My experience is I've seen the biggest improvements with the generative stuff. So the LLMs, the music and video generation, I'm less familiar with those two aspects, but I follow some tech podcasts and they're talking about, oh, these are getting better. They're still like, 语法解析

12:43

great. Right. But you know, that's kind of AI in a nutshell. It's not always great, but it's a, but it's a good, it can be a good collaborator. Right. One of the things that I talk to my students about how to use AI sort of ethically and just, you know, for their own work is to allow it to have a conversation with, so that you can get out of your mental ruts. So I can give you an example with my writing. Okay. As a science fiction writer to make it perfectly clear, I do not write 语法解析

13:11

I do not write AI generated texts. I'm not copying and pasting from AI responses for my books. But you know, as I'm writing and you know, I'm a thousand words into the evening of writing, one gets tired. And so one might be describing, oh, I've got a character blushing. And I'm like, oh my gosh, I cannot say again in this novel, in an early draft that this character is blushing. So I can go over to something like Gemini or ChatGPT and say, 语法解析

13:36

describe someone blushing. Right. And it'll give me a lot of difference. And all of them individually are usually not leaked. Right. But sometimes I can look at that and I can say, oh, I can pull a little bit from that. Or I can read one of them and say, oh, that actually is getting me to think about this now instead. This is how I want to talk about it. So I'm taking the output. Right. What I'm using it is to basically just kick me out of a mental rut. 语法解析

14:01

And I find it, both in my scientific research and in my writing, super useful in that way. Well, I guess if you look at the answer space, you might be the local minimum or local maximum, like wrapped in a small neighborhood of possible answers. And what I think is cool about AI is it has the potential to show you the entire answer 语法解析

14:18

surface of the answer space. And you can see pockets that you never knew existed. Yeah, absolutely. In that regard, it can. But of course, the AI itself can get stuck in suboptimal local minima, if you will. And so that, you know, it gets stuck into like this region. And if you want to explore more, you know, you've got to get that AI to, you know, sort of kick it out of that sort of 语法解析

14:38

sort of stuck as well. So for both the humans and the AI, you have to be worried about, you know, being stuck in that rut, if you will. But I think that's exciting. I think it's an exciting collaborator to work with. Yeah. Good. I don't know. Like, so it seems like, you know, at Google, remember when it came out, everyone said, oh, all the world's knowledge in your hand and, you know, the bar is going to be raised and people are going to be doing all this stuff. And there are some that do that. 语法解析

15:03

But there are a lot that just like watch porn and Netflix or whatever. And, you know, there's amazing technology there, but they don't take advantage of it. And same with AI. I have friends that use it like every day, you know, for all kinds of stuff. Some of them maybe too much. And then a lot of people are like, yeah, I heard about it, but you know, they just don't care. 语法解析

15:19

So I'm guessing that AI will start to be used by a lot of people when you have like simple apps, like an AI app that has an agent in it that the next way the, you know, schedules an appointment or whatever. And then maybe people use it more just as like, again, for these specific tasks. But I don't see a lot of people sitting there talking to it, querying it, trying to figure it out at a prompt better, you know, that stuff. Yeah. Well, you know, the thing is, I think with AI is that 语法解析

15:44

We want to throw a wider net and we're talking about. So all those people that you mentioned who are not querying LLMs right now, they're still involved quite, you know, thoroughly with AI. 语法解析

15:57

Insofar as limited machine learning. When they're on Amazon, that recommender, like, hey, you bought this book, we recommend this book instead, that's AI. If they're streaming Netflix and it recommends new shows, if they get a fraud alert on their phone for a credit card transaction, that's machine learning. That's all done by machine learning, right? Wow. 语法解析

16:18

really what you're what we're seeing with ai it's gonna be ubiquitous you won't even notice it's there and that's good and bad right yeah that makes sense too you um you know i'm sure everyone asks you about agi and what's the likelihood of it and all that um and i guess i see ai is like a series of idiot savants that are strung together to accomplish a task but now they do see you know working more independently and getting a lot more done under one system what do you see 语法解析

16:45

Yeah, that's an excellent question. Again, it's one of those questions where you ask 100 people, you'll get 100 different answers, where you'll have the true believers who do think that something like ChatGD4 is a sentient thing. I'm not in that camp, by the way. And then the other people who are like, this is absolutely not. This is just fancy curve fitting. And I probably, if anything, am more in that camp. But the idea is it sure does look like it's 语法解析

17:11

AGI, right? Like it does, it talks to you and that is by design. It talks to you like a person. It makes you think 语法解析

17:19

It's thinking, right? I don't think it's thinking, you know, I'll give you a good example. Melanie Mitchell of the Santa Fe, she's an external professor at the Santa Fe Institute, has been doing AI for a very long time. And he had a study recently put out where she basically posed the following question to generative AI model. Let me see if I can, I'll probably get the exact wording wrong, but the idea will be correct. 语法解析

17:42

All right. Uh, Jennifer has two sisters. How many sisters does her brother Michael have? Okay. 语法解析

17:50

And so the answer is three because you need to include Jennifer, right? Michael has three sisters. But if you asked like Gemini, we did this in my class. Yes, Gemini is great. It will respond with two, at least as of the time of this recording, right? If your listeners listen to this, who knows what it will say right now. So it's not capable of that kind of response. Even though when we asked it to show its work, it had hints, right? 语法解析

18:17

of understanding, especially we did one of their experimental thinking models. You could see in the experimental thinking model that it was kind of recognizing the fact that Jennifer was actually one of Michael's sister, but it still produced the answer. How do the reasoning modules work? Where's the LLM? Do you know? Yeah, I unfortunately do not know. That's a bit beyond my expertise. I'm not 语法解析

18:40

a ML engineer. I'm mostly, I use it as a tool for my work. I do code up some machine learning, some more basic machine learnings, models, things like random forest support vector machines, a few basic neural nets with libraries and stuff like that. So I do train my own things, but when it comes to LLMs, you know, I have some basic understanding of a few things, but unfortunately not that. 语法解析

19:02

I don't know. What are the holes in AI that you see right now? It's still hallucinating sometimes. It still doesn't or can't share its reasoning. I guess the troublesome thing is that sometimes it kind of tells you what you want to hear, you know, and 语法解析

19:17

It will straight out lie to you with a smile on its face. It will basically, like, this is, you know, it answers with these human tics to make you want to, you know, go along with what it says. But, you know, hallucinations, or basically flat out error, I think hallucinations is a nice term for just error, right? Just wrong. I think it's the 语法解析

19:40

Yeah, yeah, yeah. It should make stuff up, right? Those bother me less than some more fundamental issues that I think need to be tackled. And this is with AI in general. We're talking about 语法解析

19:51

you know, basic neural nets or even like support vector machines all the way up to these LLMs. And that's really getting a handle of what the training set is. What are we feeding? What's the diet we're giving these models? It isn't a good diet, right? And I think that's something that we need to have a serious conversation about. I think there's some people talking about this now of vetting training sets, right? And it's not necessarily where you have a quote unquote approach. 语法解析

20:18

approved or disapproved training set. Like I'm not trying to be authoritarian here, but what I'm trying to say is understanding like what are the limitations of the training set and putting that in the context of what is the output of the AI as opposed to just, you know, dumping a bunch of data into it and then saying, well, you know, that's the, that's the right answer because that's what the AI saw. You just need a little bit of time to integrate it. Then, then you can use it. 语法解析

20:44

Well, also the quality. So, for example, is your training set representative the types of predictions that you're wanting? Oh, what do you do with outliers? And what if the outliers are not uniformly distributed? 语法解析

20:56

high and low, but more high or more low. I mean, there's just a lot to it. Yeah, you're absolutely right. Or think about self-driving cars, right? You know, we're pretty far along with self-driving cars, but there are just rare events, events that are in the long, we call them a long tail, that are just very infrequent that you just don't see enough to, you know, train an algorithm on. 语法解析

21:17

Of course, humans are the same way, right? I mean, how many times do you get the classic driver's ed scenario where the red bouncing ball comes out in front of your car, right? 语法解析

21:27

That doesn't happen very often in the real world. It's an individual driver. And so what do we do as humans? Well, we do, we make the best decision that we can given the circumstance that we're in, right? That's what we do. Sometimes that's as a tragic result. Sometimes it's not. And that's what these machine learning algorithms are going to do. Once it doesn't, if it doesn't have enough of that sort of that element in 语法解析

21:51

in the training set, it's going to start making random decisions. And so what we need, I think, to do when we understand these training sets is where is the sparsity of those training samples? And then, you know, are we speaking to the algorithm? And what does a complete and good quality data set look like? What does it have in it that… 语法解析

22:10

a lopsided or crappy or poisonous data set would have? Yeah, that's a great question. You know, I think that's going to definitely depend on the situation that you're in, right? Those types of things that you're looking at. So I think what we really need to ask first is what is the problem that I'm trying to solve with the machine learning algorithm with the AI? Okay. And opposed to basically just saying, here's a hammer, everything looks like a nail. 语法解析

22:35

We want to start problem first. This is what we do sort of in my laboratory as well. So what is the specific problem that we are trying to solve? And now we're going to train this AI and we train this AI and we need to make sure that this training set has the full spectrum of things that we could possibly see to the best of our ability. You can't do it 100%, right? There's no perfect balance. 语法解析

22:58

batting average here. But you want to take pains to make sure that you have representative samples that are 语法解析

23:05

Samples that are representative of the situations that the AI is going to encounter. Know what are those sparse situations so that when your AI is applied, when your model is applied to that particular case, oh, you know what? We're now using it to look at, I don't know, a million dollar expenditure. I haven't trained my fraud detector on million dollar expenditures yet. Okay. So, 语法解析

23:29

That should be red flagged and then should be sent to a human. It says the AI said this. However, we know that the AI isn't well-trained in this situation. So person, please inspect us. And so that's where really, again, thinking about AI and the machine learning algorithm as a collaborator with a human workforce, as opposed to a replacement. You know, I guess like chess, you know, they play with deep blue or whatever, you know, stockfish, those kinds of engines. 语法解析

23:57

And the best players now use those engines along with their knowledge. They crush everyone else. They do really well. Yeah. And I think, you know, that's going to be true in a lot of knowledge creation fields where, you know, scientists will be working with an AI. Like you just collaborate with an AI, right? Like it's just, 语法解析

24:13

It's not even special. Like at some point, it won't even be special. It'd just be like, yeah, I'm thinking about this new project and I went ahead and started having conversation with Gemini or ChatGPT or, you know, whatever your favorite LLM is or whatever's your appropriate LLM. Making sure that your LLM is tuned to the kinds of problems that you're interested in, right? Because, you know, that will be better. I think the smaller, more focused models will be better in those situations than your 语法解析

24:33

What are some of the terms? I've heard of tuning. What is that? Then there's another term. I forget what that is, but like if you just, let's say, I don't know, you wanted to discuss Elizabethan poetry with the AI. 语法解析

24:47

And you loaded up, I don't know, 500 poems. What would that be? Would that be called tuning it? Or what is that? Can you overdevelop a part of the AI so it skews it to where it can't answer you properly anymore? Yeah, sure. I think that would definitely be possible. I mean, I know in more traditional sort of machine learning topics, some of the things that I teach my students, even simple models like various regression models and things like that, you 语法解析

25:09

you could what's called overfit and so that means is the model parameters are too dialed into the training set think of like um like a kid's connected dot it's like training you had to play baseball but everything they throw to you is like underhanded and it's like a softball pitch so you never get any real experience with tough stuff yeah or i tell my students like if you're studying for a math test and all you do is study the homework problems the professor assigns so 语法解析

25:36

So if the professor gives you a homework problem, like copies of homework problem, puts it on a test, you're going to do really well in that test. But if the professor gives you a slightly different problem and all you've done is memorize the homework problems, then you're not going to perform as well because you have that mental flexibility, right? Because you're too tuned in, you're too dialed into that particular problem set that you've memorized. 语法解析

25:56

Is there a setting that's called that? Or again, is it just overfitting data? What is it to you for an AI to be very flexible? So you want to add a little bit of what they call bias in there. You can penalize certain variations. 语法解析

26:11

At this point, we're probably getting a little technical, but essentially keeping the model that you're using to be less complicated. So, you know, for example, if you might have a curve that looks like a hump and, you know, you could use a polynomial, fourth order polynomials, x to the fourth to fit that plus some other terms, but that might be connecting the data points too closely. So you might say, you know what? 语法解析

26:35

The parabola doesn't connect all of the dots, but it gives me like an average ish of all the dots. It gives me the trend so that now when I see a novel situation I haven't seen before, I'm just going to perform a little bit better because it's just not so focused on the exact training data that you have. 语法解析

26:53

So a little bit of error is a good thing because it gives flexibility. It gives the model flexibility. So what is the, I guess in the next year, are there any major developments that you see coming you're waiting for? Oh, that's a good question. Boy, it's such a fast moving space. A year could be like a decade, you know? 语法解析

27:10

Um, you know, in the short term, I think, you know, I think what's, what's, like I said, what's got me excited is some of this work that's maybe not quite as sexy, but the smaller parameter models that are more efficient that are again, specialized. 语法解析

27:22

to help, you know, researchers in let's say medicine or science or whatever the case might be. Development of more of those has got me interested. The chain of thought for the reasons I talked about earlier has got me interested. I'm also really interested in the more efficient. Now, I'm not an expert in making these things more efficient, but I'd watch and listen to some of this stuff. And that's got me excited too, because one of my concerns is the environmental impact of just burning all this electricity to not just train these models, but 语法解析

27:49

but to query these models too, right? Because that's an expensive process. You're reaching out to a data center, data centers doing some math and sending it back to your computer. And, you know, like streaming, like streaming music and streaming video, these are all electricity, data intensive processes that, you know, have an environmental impact. So I would love to see, and I'm hopeful for more efficiency improvements in that direction. Okay. What's, I don't know, like today, what is making the best AI? It's like what, 语法解析

28:18

But, you know, if you had an AI that assisted you with a given task, what would make it just amazing where it has everything you want? What kind of features or things would it have? Oh, that's a very good question. It's going to depend on the tool that I want to use. So I find myself using three tools. I find myself using Anthropics Cloud, OBI's ChatGPT, and Google's Gemini. 语法解析

28:40

And it really depends on what it is I'm working on. So like if I'm working on fiction, I like to often use Claude. I find that Claude often gives me some really nice alternatives to things I'm trying to say. Or, you know, if I want to, you know, get a character, put a character into a situation, I will maybe chat with Claude a little bit. I'm thinking about this. What are some options here? 语法解析

29:04

And I think it just kind of gives that more literary spin. Again, it's nothing I would ever copy and put into my own work. I don't think that's appropriate, but I think it's just creative in that way. When it comes to code generation, so when I'm producing code, I find myself leaning on Gemini and ChatGPT a little bit more. Gemini purely because it doesn't kick me off. Gemini doesn't tell me, oh, you've reached your limit for today because I'm too cheap to pay for any of these things. 语法解析

29:30

And so find that, you know, one of the, to answer your question, one of them is just don't kick me off. Right. And, but the code I find is pretty good. You know, it's not again, like the fiction, it's not always the case. It's rarely the case. You can just copy and paste from scratch. There's some simple coding things that you could absolutely do for perfectly fine, I think. But some more complicated coding things, especially with like the agent-based modeling I do on complex systems. 语法解析

29:54

You know, you want to have a conversation with it and it takes multiple sort of rounds of querying to get that out. But yeah, I mean, that's kind of what I'm looking for is depending upon the task. I don't know that there will be such a thing as the one perfect 语法解析

30:10

Because I think what we really could stand to work with are things that are specialized for giving topics. Something to help the writer, something to help the coder, something to help the nonlinear dynamics person, right? As opposed to like this one uber AI that you go to and can do everything. Yeah, with stories, what I learned is now what I do is I'll write a whole story and I'll put that into the AI as a prompt. And the AI will clean it up and make it look nice, you know? Like it's… 语法解析

30:37

It's like me, but it puts makeup on me and makes me look good and dresses me up. So I'm not a slob. So that's, that's what I see like the most, the best use for it, at least in terms of writing is so far, you know? Yeah. I usually don't use it for that. Like I said, it's for me, it's mostly that rut, right? Getting out of that, having that conversation, stimulate creativity. I'm very much about stimulating creativity with AI, but 语法解析

30:59

But I do use it to check things like grammar. And again, you have to be careful with it. But, you know, for a quick and dirty, like, hey, is this grammatically correct? You can scan over it. And if it doesn't look suspicious, like, okay, you know, I'll move forward with this and I'll check this with my editor later, right? 语法解析

31:13

It kind of just gets me over that, like just gets me kind of moving forward for sure. It does a pretty good job of summarizing texts, you know, a lot of times. But again, like I tell my students, you know, it's, we're still needing to, to check everything it puts out because, you know, it does produce errors and it's just very important to just, you know, proceed with caution when using it. Right. Yep. True. Yeah. 语法解析

31:36

Other modules that are coming out. So we get reasoning modules like LLMs. What about GANs, the generational adversarial networks? How can those be used to improve an AI? So I have a limited understanding of GANs. This is not something I do a whole lot of. But my understanding is they kind of play against each other. You have sort of like two AIs fighting against each other, if you will. And so, you know, and we see this with like, was it 语法解析

32:01

was alpha zero for the go or how to play go by playing against itself. Right. And I think what you get there is you can sort of more quickly, you can come, you can train these things maybe without giving it so much input. What if you did this? Let's say there's like, we're just making this up. Totally. There's like 10 world renowned players 语法解析

32:20

scientists that specialize in nonlinear dynamics, you being one of them. And I take my AI and I say, here's a problem and I put it in there and I ask it. I say, using the thought patterns and experience and all that of Chris Culp and I don't know, 语法解析

32:37

Bob Smith, critique this and help me find a solution. And then you do the same thing, but you use Joe Jones and Jane Smith and Amy Hex and all that, like three other people to come up with a solution. And you keep kind of rotating through like these top minds and what they would think and answer and say. 语法解析

32:54

and try to approach the problem, you know, nine, 10 different ways with all these permutations and put it together to see what answers you get. Yeah, now you're using it as a collaborator, right? In a sense, and you're having it sort of virtually collaborate. So there's a couple of issues potentially with this. I would say that you would still need to know something about the field that you're working in. You couldn't be a novice, right? Because you have to be able to interpret the outputs to make sure that it is 语法解析

33:16

it's actually pulling together things correctly. Right. The other thing too, is making sure we talk about ethically training these things. You know, did I give company Y permission to, to train like to be me, right. Or at least to be a simulation of me. 语法解析

33:33

And that's something that, you know, artists are really, you know, concerned about right now and rightfully so. So we talk about, you know, thinking in the style of Chris Kulp or thinking in the style of, you know, Einstein or any other scientist, you know, I think there's an issue there we need to be, we need to worry about. But if we put that issue just to the side and just say, suppose we had all the rights, right. And then it's fine. Then again, you still have to be careful because, you know, as we talked about earlier, hallucinations still happen. So you still can't be just pure knobs, right. But again, 语法解析

34:01

With that being said, 语法解析

34:02

you don't necessarily have to be pro expert either. Right. So basically what this is helping do is make these things more accessible. Okay. Is there any, I don't know if you have previews into how some of the AIs are going to work, but is there anything that the AIs are doing that just blows your mind? You're like, I can't believe this, man. This is crazy. Yeah. So I can give you another example from my classroom. We were solving a problem and we were wondering how would AI can do this and messing around with some stuff. And then, 语法解析

34:32

I don't know, sort of off the cuff. I said, you know what, let's just ask it something completely, you know, upper level physics and see what it can do. And 语法解析

34:40

The problem is you have a think of like a disc that's spinning with a pendulum stuck to the end of the disc. And you want to base and we asked the AI to create the equation that describes the motion of that pendulum. And so that's a sophomore, junior level physics problem. OK. And sure enough, it solve it correctly. Yeah. So then I made this offhand comment. I'm like, you know what? 语法解析

35:03

I have a feeling it probably can't interpret the terms of this equation. In other words, like this term is the term that represents the force of gravity. This term represents the spinning of the disc, so on and so forth. And one of my students goes, well, why don't we find out? Let's do it. So we just then said, can you interpret your answer? And what blew me away was it could. It went term by term of that equation and said, you know, this is the torque due to gravity. This is the torque due to the spinning disc. We, my students and I just were looking at the screen with our jaws dropped. 语法解析

35:32

And that was something that I was not prepared for it to be ready to do. I don't know why I wasn't prepared for it. I don't know why I wasn't expecting it, but it just was one of those moments that I, it just blew me away that it was able to, to do that. So yeah, that's that example. It's the, what, 语法解析

35:50

stunned me was the interpretation. The solving, okay, fine, maybe it found that problem in a training set somewhere, so it's seen it before, but the actual interpretation, that was what really, yeah, I was very impressed with. Okay. Well, it's funny, they always say AIs are like black boxes, you know, the inner layers and stuff like that. Yeah. Can 语法解析

36:12

Can they train an AI to say, hey, figure out what's going on inside this other AI? Oh, that's a good question. I don't know. You know, I will be like, I'll be the scientist answer here. I think there are some things that people are, people are definitely, and again, this is a little bit outside my expertise, but people are starting to, you know, answer the question of how do these neural nets make decisions and things like that. It's still really early on in this process. 语法解析

36:34

can we get a trained AI to learn how another AI works? That would be wild. And it might be done. I, unfortunately, I just don't know, but I'm definitely going to look into that because it sounds like something someone's tried, right? That just sounds like something still based on. And I just don't know. Okay. And I know I kind of answered before. I don't know if we got the answer out, but do you think AGI is possible or it's, it's, we're not really, you know, I don't know. I think that, 语法解析

37:01

and I'm going to be like the scientist that kind of waffles, right? I don't know if we understand what consciousness is for humans, right? And so it becomes hard to understand 语法解析

37:11

say, can we then identify, you know, AGI? Will we be able to do that? It's the question too, like, you know, when searching for extraterrestrial life, you know, will we know life when we find it, right, out there in the universe? Will we know a conscious AI when we come across it? And possibly not. You know, I think we need a better understanding of what makes us, us first. And I think asking questions like the one you just asked helps 语法解析

37:38

helps with that, right? I think that there's a place for studying AI to help us understand ourselves. Just like with bipedal robots. Like, you know, you don't really appreciate how hard it is to walk on two legs until you try to make a robot do it, right? And so I think it's the same thing. Will we have AGI one day? Sure. Yeah. But that's pretty far away. I don't think we're anywhere close 语法解析

38:00

close to that. But you know how science works. You know, all of a sudden you think you've got the answer and plop, the next day it's right there in front of you. That's the crazy thing about science. Definitely. Well, 语法解析

38:09

Well, very good. So where can people learn more about AI in a simple way that where can they understand it? And if they want to take a course, what might be a good initial course for them to take? Oh, there's a lot of courses online through YouTube that you can do. I can tell you how I got my start. And some folks might be interested in this, but Andrew Nigg, 语法解析

38:30

as a, he used to have a Coursera course. I think it's still out there. And this was many years ago. But, and so, but a lot of that stuff that, that he covers, it's like this foundationally solid. Like if you want to really get into like the basics of machine learning, then jump into the more complicated stuff. That Coursera course was great. I think it was just called machine learning, introductory machine learning or something like that. I'm sure. NG is his last name. I think it's Kerb. Yeah. And, and, 语法解析

38:58

Yeah. I'm terrible pronouncing names. You know, people say I can't do math. Well, I can't pronounce names and. But you know, he had a course, I'm sure it's been updated, which I got my start with really good. And then there's a book I would also recommend. It's called hands-on machine learning. 语法解析

39:13

And Andrew or no, Darren, G E R O N is the author of that book. And that is a good, I actually use that book when I teach my applied machine learning course. There's a lot of great code examples. There's a really nice GitHub repository for it. And it goes through and just basically, again, it starts you with things like linear regression, logistic regression. These are the 语法解析

39:36

rock solid if you want to understand issues of machine learning and ai you want to understand these things because these things have your cost functions these things you can talk about true positives you know false negatives you could talk about precision and recall and just get a real grounding like a foundational understanding of what's going on before jumping into the more complicated stuff okay 语法解析

39:57

Well, very good. Well, Chris, thanks so much for coming on the podcast and explaining. I know it's still amazing and weird and inexplicable, but thanks for trying. Yeah. Thank you for having me. I really appreciate it. If you like this podcast, please click the link in the description to subscribe and review us on iTunes. You've been listening to the Finding Genius Podcast with Richard Jacobs. 语法解析

40:22

If you like what you hear, be sure to review and subscribe to the Finding Genius Podcast on iTunes or wherever you listen to podcasts. And want to be smarter than everybody else? Become a premium member at FindingGeniusPodcast.com. This podcast is for information only. No advice of any kind is being given. Any action you take or don't take as a result of listening is your sole responsibility. Consult professionals when advice is needed. 语法解析

Edit:2025.04.28

讨论列表 AKP讨论 查看原帖及回帖