🧠 ai disruption with howie xu

Welcome to another AI Wisdom podcast with Howie Xu. Howie is an AI and data executive, a Stanford guest lecturer. He also posts on these topics and yesterday his Twitter post on OpenAI got 650k views and apparently it's top of Google when you google GPT5. So impressive guy altogether has a huge background. Howie, can you tell me a bit more about yourself, your background and what you're up to nowadays?

Sure Charles. Happy New Year to you and everyone. First, very happy to be here. So my background, I've been infrastructure, software, operating system type of guys for 20 some years. I was one of the early guys at VMware, helped VMware to build a networking stack from scratch and I've been doing software defined network the security. In the last few years I started a AI security company myself acquired by Zscaler and then I moved on to Pallado Network later on. And of course you know, 2024 is going to be an exciting year for me and for everyone in the industry because you know, a lot of new things, right. You know, whether GPT5 is, is coming or you know, other things. You know the genitive AI is like every day you wake up, you know, there's something exciting, there's something new. So, so I can't wait to see, you know, see through the entire year at this point.

Do you feel that the current version of AI, let's call it ChatGPT's algorithm, do you think it can scale to AGI or will we need a modification and a different approach to AI to reach AGI?

Well, you know, depending on your definition of AGI. Right. You know I actually wrote up something about a year ago, probably March ish last year, right? Really depending on your definition of AGI. If your definition of AGI is read a paper, learn it, maybe even read an image and then learn it. Now arguably we are already there and then maybe already there way past that. Of course most people, majority, majority people have not said hey we have reached AGI. But for those people who said we haven't reached AGI, what's their definition? And I'm pretty sure we all have different definitions, so that's number one. Technically there are different views and there's no standard view to begin with, so. So it's going to be a very interesting milestone for all of us to chase because none of us have the same measurement to begin with. And then there is also business non technical agenda behind that. I'm not saying it's an evil agenda, I'm just saying that people have different Agenda. Some people wanted to call it AGI sooner, some people wanted to call it AGI later. If you look at OpenAI investment agreement or whatnot from Microsoft, there is, you know, if you read through, there is actually implications when OpenAI declares its AGI, what sort of things, Microsoft can continue to use it or not use it. So when you mingle all those things together, right, you know, technically we are not on the same page, right? You ask any expert, they have, you know, five different opinions already, let alone you have, you know, thousands of experts on the planet and then you have commercial business agenda, you know, mixed together. So I think it's a mess, but it's a beautiful mess. It's a beautiful mess because, you know, anyone and everyone can articulate a version of it. Now, having said all of that, coming back to your question, right, you know, your question is really, is our, is the transformer architecture, right, Scaling law going to, you know, get to help us to get to the, you know, maybe majority of people, what majority of people are thinking about AGI? I tend to think, yes. I don't think I can articulate, you know, here's the 1, 2, 3, 5 reasons because, you know, even, I mean, let's, let's, let's be honest, right? The GPT3, GPT4 has been working beautifully, but no expert have, is able to explain, you know, why. Right? So how am I going to say hey, hey, before I see the results, you know, this is going to be there. It takes some belief, it takes some conviction. I do have the conviction that a transformer model and then scaling law is going to take us to the next level. When I say next level, I meant, you know, at a minimum order of magnitude better than GPT4, if not even more than that. Right. So I think GPT, you know, five, you know, potentially six. We are going to get there via the current architecture, more or less. And then, you know, with the scaling law, that's sort of how I think, think about it.

In other fields of science, is it common to make discoveries and not understanding how we got there?

That's a very good question. I would say yes and no. So we are very sure, we are very accustomed to, you know, hey, we need the physics, we need the chemistry and then we sort of, I need sure that that's where you're coming from. Right? But do we, I mean, do we humans even understand how we learn things? How we kind of learn things from a baby, grow up, you know, how we, you know, how everything work internally? No, but we, you And I are talking, right? Just because I cannot explain how my brain works. I can say whatever I say is wrong or right by itself, right? So, yeah, I think, you know, it's different. It's very different. You know, it's a very big, it's a very big paradigm shift. You know how we think about this right now? I'm a software guy, I'm a technology guy for decades, you know, I was talking to a friend even yesterday, you know, the way we think about, you know, there are so many things that's a paradigm shift. But even like a software engineering, software testing, software qa, it's a paradigm shift, right? In the last several decades, you build a piece of software, you have a QA plan. This is right or wrong. But with generative AI, with copilots, what is right and what is wrong? How do you even know this is right? Who has the authority to say this is right or wrong? And what do you mean by right? Is that like not having this hallucination or giving you helpful information, helping you to resolve the information? What is the right? So I think, you know, it's a very big paradigm shift. So we need to think the world very differently in this generative AI, in this AI, you know that the new era

for many, what Sam Altman has orchestrated is nothing short of a miracle to me, it seems he reunited a bunch of geniuses in one room, gave them unlimited resources. Do you think this model is reproducible? This, do you think this is eventually what will lead to great breakthroughs in the AI field and in other scientific fields?

Your question is whether the similar model can be used in other fields, right?

Because you founded various startups yourself, right? Some were really, really successful, reached billions of dollars. Is the formula for breakthroughs business or scientific? Is it to reunite a bunch of smart people and give them unlimited resources?

I think my take on this is in theory, yes. In practice it's hard because in theory, when you have this level of the talent concentration and then people have a long term view. Ilya. Right? He has a very long term view. He has his own conviction. He was doing Transformer. He has been doing Transformer, the large language model, big models before there was a result. And even when people laughed at him, laughed at the methodology, GPT2 was being laughed at big time by the industry, by the academics, and he continued that path. So when you have people like them, you know, not everyone is going to work out. But that's a, that's a, that's a model that's going to you know, that's what it takes to get some breakthrough. So I think in theory, yes, in practice it's going to be difficult because you know, when you look back, you know, OpenAI's success, right. It's not just Ilya has the conviction, but also the real talent, concentration around him. Right. It's not just one person, the amount of the money. Right. You know, I think it, you know, it takes a lot to tango in this case and I think in practice it's going to be hard to replicate. So that's Number one. Number two, however, I do think this OpenAI, ChatGPT, generative, AI, whatever you want to sort of categorize it is going to help lots of, lots of the breakthrough in other fields. Anyways, the reason is finally we have. I like the way that Satya of Microsoft, the CEO of the Microsoft put it. He looks at the programming capability, programmers, resources, the digital currency in other fields. In the past we lacked the digital currencies because we don't have enough money, we don't have enough budget, we don't have enough talents. But now with the GPT, with the AI, with the help, you know, we have far more digital currencies now with Additional digital currencies 10x, potentially 100x more than you know, we had before, you know, breakthrough could happen. So I would rather think a breakthrough would happen because of AI, not necessarily because, you know, there are equivalent of the Iliad people sort of being replicated in other fields. That's sort of how I think about it.

But then it seems that our humanity is somewhat limiting this progress as we've seen with the board coupe at OpenAI, isn't that indicative of how us humans are capable of building great things? But in the end ego is bringing us back. How could that endanger the future of AI and even impact humanity by building an AGI that is much too human and capable of destruction?

Yeah, I mean, you touched a very fundamental thing, right? You know, one is AI technology, the other one is human ego. Whenever there are enough humans, there will, there will be enough egos. Right? You know, this is a technology related podcast. I'm going to, I'm not going to sort of get out of the bound, but ultimately you are absolutely right. The defining variable may not be the technology for humans actually it's the human itself and human ourselves. And then ego, how much ego is surfaced, being played out? Maybe the high order bit. And you look at a political world, you look at things globally, you look at it, you know, technology world, you look at why certain companies, certain, you know, administration at a certain management team doing super good job or super bad job or whatnot. Right. You know, if there's a one word that to capture why ultimately it's the one word that you just said, right? Ego. And you know, unfortunately maybe fortunate because we are all humans. Right. That is the higher bit. In the end, everything else may be the distant second. So you brought up a very interesting point. I don't have a good answer. But the reality is there's a technology front. How do we leverage it to the best we can? In the best case, it's a wonderful, wonderful progress, beautiful progress. At the same time, this beautiful progress, GPT4, GPT5 can be used in a way that's not safe, unintended and then could have better consequence. Am I concerned? Sure. But this is no different from other games we have think through for decades or hundreds of years or thousands of years that ultimately we have this crisis. We have those, you know, less beautiful things because of the, you know, human have, humans have ego. Very, very interesting point.

Listen to the pod between Bill Gates and Sam Altman yesterday. And it seems that Sam sees the use cases of GPT first to programming, second, healthcare and third, education. Where would you see the best use cases of AI? The most applicables and potentially the ones that could also get the most venture capital behind them.

Yeah, I think he's, you know, by the way, I didn't listen to that talk yet as sounds like a, you know, good talk, right? You know, programming, healthcare, education, those are the verticals that the technology can be applied. Makes a lot of sense to me. In particular, I'm looking forward to the disruption that AI is going to bring to the table in the education field because education is one area that has not been disrupted much. And we all know that when our next generation, you know, comes to the society, comes to the workplace, it's a totally different ones. And how is the education going to prepare them for that? You know, I'm actually, you know, very much looking forward to seeing what's going to happen and hopefully make some contribution to that too. So that's the sort of the three verticals you just mentioned. It makes a lot of sense I think from the, from the horizontal technology point of view. Most of the energy of the generative AI has been focusing on question and answers, basically right. You know, hey, I give you a lot of documents, you give me the answer, I fetch the most relevant chunk or the paragraphs and then the large language model is going to formulate a Beautiful answer based on that chunk, based on some memory it has already. That's sort of what I have seen in the last one year or so. I think this is still going to be played out because even though in theory we have the fine tuning, the rag, all those technologies, but it still takes time for that to really work, to really shine and then work in the optimal way. So we are not there yet, but I think will take another year or two. I think there is one space that's kind of a wide space today is more than just a question and answers, but deeper reasoning, deeper analysis of the information. Right? Not just finding the interesting chunk and then I formulated the answer, but also correlate lots of different parts of the information and make sense out of it. And then if this information is just within the context window, it might be easier because ChatGPT may already be able to do that, but maybe not because you know, there's all sorts of problem like lost in the middle, all that kind of things, but forget about that. But what if you have gobbles, gobbles of data that's way beyond the context window. Can we leverage the, you know, AI to make sense out of the data? I think, you know, there is a lot of the interesting problem to be

conquered here and you have a son that's currently studying at a lot smart young men there. What would be the application to supercharge his learning, make sure he's motivated and make sure he's accomplishing and executing the work to be ready for this hard job market that's awaiting him.

I wasn't ready to talk about my son, but since you asked that, yes, I do have a son, You know, a, a grade school student. He's still. Right, he's very much into AI actually, you know, he actually wrote a chatbot for his school so that, you know, the pinpoint is it's so hard for people to find the calendar lunch menu. So hey, you know, one natural language away, you get to, you get to sort of know what's the lunch for tomorrow. It turned out to be more painful than people realized to find a lunch menu. So he's very excited about this and what's going to happen to him and his generation. I don't know the answer, but I was talking to some of the AI experts in the last few months. One guy said something that's I think very interesting. So we had a roundtable discussion about the impact of AI for career. Right. That was the thesis of that roundtable and then we had lots of interesting conversation towards the very end when one gentleman rose up and they said, hey, listening to all of the conversations and it sounds like in the future we're not going to have Software Engineer 1, Software Engineer 2 Senior Software Engineer, those titles anymore. What we are going to have is problem solver 1, senior problem solver, principal problem solver. I said, bingo. That is where the future is. I think we still need a problem solver. We still need people who have the judgment. But do we need a software engineer as we understood in the last 20 years, in the next 20 years may not be with all the AI stuff, no one is a programmer anymore and everyone is a programmer with the low code and no code. Of course there's always exception. Someone is going to build OpenAI someone is going to build operating system kernel. That's a small exception. Vast majority of us in the society, do we need to write Java, python, you know, 15 years from now, maybe not. However, everyone becoming a programmer because once we specify things, you know, it will be codified. Now does that mean we don't need people or people with good judgment? No, I mean, probably even more. Right. Because, you know, if nothing else, we have more noise in the world and whoever who can pick the signal out of the noise is going to, you know, be in a better spot. So I want to say, you know, I would agree with that gentleman, you know, basically at the end of that roundtable, you know, his question became a summary. In my view, his question is, are we going to have software engineer 1, 2 senior software engineer or we are just going to have problem solver one, problem solver two and a senior problem solver. And I think that's where the industry is going. So for my own sort of advice to my son and the generation of my son is that, you know, be prepared to be a problem solver more than anything else.

Love that Pain fixer. Love how yeah, you phrased that. And he's already looking for product market fit at such a young age. This is quite inspiring. Where can we find out more information about you, Howie?

About me, I'm on Twitter and then you know, on LinkedIn and you know, and you can find my son's chatbots pantherai.com as well. So you know the. Yeah, so I think the. It's. Yep. Great to chat with you.

My pleasure. That was how we zoo and Charles Cormier for another wisdom podcast.