LLMの先へ:AI CEOの実験 ― Skyfall AI 共同創業者 Sam Pasupalak 氏

#425 Beyond LLMs: The AI CEO Experiment — with Sam Pasupalak, Co-Founder & CEO, Skyfall AI

Making Data Simple1時間3分
LLMの先へ:AI CEOの実験 ― Skyfall AI 共同創業者 Sam Pasupalak 氏

LLMの先へ:AI CEOの実験 ― Skyfall AI 共同創業者 Sam Pasupalak 氏

Making Data Simple

0:001:03:04

要約

Skyfall AI の共同創業者兼CEO、Sam Pasupalak 氏が、LLMの限界と「ワールドモデル」による次世代AIの構想を語った回。長期計画、動的環境への適応、推論効率の3点をLLMの弱点とし、ワールドモデルで計画しLLMで実行する分業を提案する。同社は小規模ソフトウェア企業を買収し、AI CEOに運営させる実験を進めているという。人間関係や採用、重要な承認は人間に残すと述べた。

この回のポイント

  • ●Pasupalak 氏は、LLMは長期計画、変化する環境への対応、推論効率の3点が弱いと述べ、自律的な企業運営にはワールドモデルと継続学習が必要だと主張した。
  • ●ワールドモデルは「状態と行動のペアから次の状態を予測する」仕組みで、潜在空間で複数の未来をシミュレーションし、最適なものを選ぶという説明だった。
  • ●同社は従業員10人未満の小規模企業を買収し、最新のワールドモデル技術で運営を自動化している。結果は数か月以内に発表予定とのこと。
  • ●初期結果ではLLMより最大100倍規模で安く速いと述べたが、詳細は近日公開予定の発表まで明かされなかった。
  • ●人間関係の構築や採用、大型の資金調達・商談、確率的で判断の難しい意思決定には人間が必要だと述べ、決定論的な業務ほどAIに任せられるとした。
  • ●同氏は、ゼロから事業を起こすAI創業者は現状不可能で、既存事業の運営の方が容易だと語った。

章立て

  1. Sam氏の経歴とMaluba

    ウォータールー大学出身で、Maluba創業とMicrosoftによる買収、深層学習による自然言語理解の取り組みを紹介する。

  2. Skyfallの狙いとLLMの限界

    ポストLLM時代を見据え、長期計画・動的環境・推論効率という3つの課題を挙げる。

  3. ワールドモデルとは何か

    LLMをパターンマッチャーに例え、状態と行動のペアから次の状態を予測するワールドモデルを説明する。潜在空間での計算により大幅に安くなると述べる。

  4. 企業買収とAI CEOの実験

    小規模企業を買収してAIに運営させる計画と、人間に残す役割や承認の範囲を議論する。

  5. 文化・自律度・仕事の未来

    AI主導企業の文化、決定論的か確率的かによる自律度の線引き、新卒エンジニアや人文系の将来を語る。

  6. データセンター投資と最大の賭け

    ワールドモデルがLLMの推論コストを下げるという見方と、エンタープライズ超知能の構想を語る。

  7. ラピッドファイア質問

    過大評価されている点やAIに任せたくない領域などを聞くコーナー。

解説記事

Making Data Simple の第425回に、Skyfall AI の共同創業者兼CEOである Sam Pasupalak 氏が登場した。同氏は、LLM(大規模言語モデル)の次の段階として「ワールドモデル」を掲げ、AIがCEOとして実際の企業を運営する実験について語った。以下は番組内の発言に基づく整理で、数値や見通しは本人の主張である。

経歴:AI冬の時代から深層学習の最前線へ

Pasupalak 氏はウォータールー大学でコンピュータサイエンスを学び、2011年に卒業した。当時はAIが注目されない時期で、モバイルやクラウドが主流だったと振り返る。同氏はYoshua Bengio氏やRichard Sutton氏と仕事をしたと述べている。創業したMalubaは機械読解や対話システムの研究所に成長し、2017年にMicrosoftに買収された。Transformer以前に深層学習で自然言語理解に取り組み、『ハリー・ポッターと賢者の石』を読み込ませて質問応答させるデモが当時の目玉だったという。

LLMの限界と「ワールドモデル」

同氏が目指すのは、AI CEOが自律的な従業員と自律的なソフトウェアを束ねる「完全に自律した企業」である。CEOの仕事には3つの難題があるとし、LLMはいずれも苦手だと述べた。

  1. 長期計画:CEOは数週間から数年先を計画するが、LLMは次トークン予測の構造上、せいぜい10〜100ステップ先までだという。
  2. 動的環境:世界は常に変化するが、LLMはウェブ上のデータで事前学習されており、再学習は高コストである。
  3. 推論効率:人間の脳は1日あたり約500ワット時だが、複雑なマルチモーダルなLLM呼び出し1回で1000ワット時を超えると主張した。

これに対し、同社はワールドモデルと継続学習に注力する。ワールドモデルとは、状態と行動のペアから次の状態を予測する仕組みである。映画『マイノリティ・リポート』のように複数の未来をシミュレーションし、最適なものを選んで行動に移すイメージだと説明した。LLMは「サッカーの試合を全部見て記事も読んだが、自分ではプレーしたことのないアナリスト」だという比喩も用いた。計算は高精細なピクセルではなく潜在空間(抽象化された表現空間)で行うため、LLMより少なくとも100倍安く速いという初期結果があると述べたが、詳細は近く予定する発表まで明かさなかった。

LLMとの関係:計画は世界モデル、実行はLLM

同氏はLLMが消えるとは考えていない。ワールドモデルで高レベルの計画と未来の選択を行い、選んだ道をLLMが実行する分業を描く。この方式ならLLMが完走まで無駄にトークンを消費する場面が減り、推論コストが下がってLLMの用途がむしろ広がるという見立てだ。同氏は、巨大データセンター投資への疑問や、大手ラボがLLMに巨額を投じているため方向転換しにくいという見方も示した。これは本人の意見として語られている。

AI CEOの実験:小さな会社を買って運営する

Skyfallはすでに複数の小規模ソフトウェア企業を買収し、ワールドモデル技術で運営の自動化を進めているという。対象は従業員10人未満の会社で、今後は買収額を100万ドル未満から1〜10億円規模、さらに上へと広げる構想だ。きっかけは、経営シミュレーションゲーム「RollerCoaster Tycoon」でLLMが全く運営できず、ワールドモデルでは有望な結果が出たことだった。結果は数か月以内に発表するとしている。

同氏は「ほぼ自律的」と表現し、顧客データや支払いに関する許可など、人間が承認する部分は残ると認めた。一方で、データやERPの処理は人間よりAIが優れると述べた。人間に残す領域としては、採用候補者の口説き落とし、従業員の動機づけ、大型の資金調達や企業間商談といった対人関係を挙げた。自律度の線引きは判断が決定論的か確率的かで決め、確率的な判断には人間の承認を置くという考え方である。

仕事と社会への見方

同氏は、運営業務の「下働き」をAIが担うことで、人間は創造的な活動に向かうと楽観的に語り、ルネサンスと産業革命が合わさったような時代になると述べた。コーディングエージェントが成功したのは論理的で決定論的な領域だからであり、新卒のエンジニアには厳しい状況になる一方、基礎を知るシニアは生産性が高まると見ている。進路としては、人文・社会科学が重要になるという見方だ。最終的な目標は、誰もが起業できるよう、製品開発から営業・マーケティングまでのAPIとモデルを提供することだという。ただし、ゼロから事業を立ち上げるAI創業者は現時点では不可能と明言した。

まとめ

本回の主張は、LLMの限界をコスト、長期計画、適応性の観点から整理し、ワールドモデルとの併用で補うという構図だった。数字は同社の内部結果で、詳細は未公開のため、近日予定されるリリースと実験結果の公開が検証の鍵になる。日本の読者にとっては、AIが業務実行だけでなく経営判断にどこまで関与できるか、人間の承認をどこに置くかという設計論が参考になるだろう。また、AIセラピーや人生相談へのAI利用に慎重な姿勢など、技術と人間関係の線引きについての示唆も多かった。

文字起こし(英語・自動生成)

Let's go. You're listening to Making Data Simple, where we make the world of data effortless, relevant, and yes, even fun. Hey, podcasters. Al here. Hello. Hope everybody is really well, really well. I am here with Sam Pasupalak from Skyfall, CEO of Skyfall AI. So it's like Skyfall. See, that's my Adele. Did that sound like Adele to you? Pretty close, right? Not so much. Nailed versus an adult. Guyfall is based in Toronto and San Francisco and is developing AI world models, which stimulate the physical world. So Sam is a serial killer. No, no, he's not. Sorry. He's a serial AI founder. He started, is it Maluba? Is that how you pronounce it? It's the last company, yes. Yes, he started Maluba while studying computer science at the University of Waterloo. Obviously, he is an overachiever.

It grew into a deep learning research lab focused on machine comprehension and conversational systems, and then Microsoft acquired it in 2017. Today, he's the co-founder of Skyfall AI, which builds AI world models and is testing whether an AI system can run a real-world business as a CEO. This is crazy stuff we're talking about here today, so I'm intently curious. he works to make the kind of capability available to all small businesses so that's the end goal here hey sam thank you for being here i know there's some trouble sam was trying to find the right spot to our location to do this podcast there are people around him and i think we found the right one right thank you thank you so much al by the way for having me in the podcast very excited to talk about skyfall and what we're up to oh i'm super excited man i couldn't be couldn't be more excited to be honest with you so speaking of excitement why don't you give us a brief introduction about yourself maybe a little bio so uh i basically graduated from waterloo

on 2011 it's a it's a tech school based out of canada we do a lot i remember just remember lots of all-nighters coding up all night that's what that's what they used us used to make us uh do all the time which is it's very intense it's extremely intense and back in the day i was very interested in AI this is like let's call it like 2008-2009 and we used to take the introductory courses in AI and I built this whole robot using the Lego framework and integrated with speech recognition natural language understanding and then this robot could go around the hallways and and you could it could do very simple tasks like go to like drop a paper for the for the one of the props, carry like a little quote thing around the hallway. And you could talk to it in natural language and stuff. So that's how it got started. When I graduated back in 2011, I was very interested to create universal Turing machines.

And this was in the midst of like the AI winter, by the way. You know what was hot back then, 15 years ago? It was like mobile. It was B2B staff. It was cloud computing. it was even social to some extent uh with twitter which came up a few years ago and stuff and and facebook and instagram came off the same year 2011 actually so that was what we thought back then and no one really cared about ai uh there was a few other labs like deep mind and us and just a few others that were doing pure ai research so back in the day we worked with like i worked with like Joshua Bengio, who's the godfather of deep learning, and Richard Sutton, who's the godfather of reinforcement learning. So this was way before they were Turing Award winners. And we created, like, the state-of-the-art algorithms in natural language processing using deep learning. So I think back in the day, it was called natural language understanding. It was not called LLMs. This is way before Transformers.

I was there. I remember. and uh yeah i remember in my last company one of the coolest products that we had was you know we're we're created a one of the world's first uh personal assistant voice-based personal assistant kind of like cd but way better because cd was ontological based back in the day so basically rule-based structures and stuff which they had created this huge graph and we were using cutting-edge deep learning models for natural language understanding. And I remember one of the coolest demos we had was, if you think about, like, the Harry Potter book. This is 10 years ago. The Harry Potter and the Philosopher's Stone. So you could use our algorithms, upload the book, and you can do any question answering on that. And that was a massive deal back in the day. So this is basically any question answering about the philosophers. her stone book without without the whole building the machine learning pipeline the same way it used

to be uh so yeah again that was very cutting edge using deep learning uh and uh yeah microsoft basically ended up acquiring my last company it became microsoft research canada google acquired deep mind just a couple years before that my team was very much involved in like integration with bing and integration with the first version of gpt1 and gpt2 and so on and so forth uh because OpenAI and Microsoft used to be in very close collaboration, if you remember, for the longest time. So that's kind of the background story of me and what I've been up to over the last decade or so. Wow, that's a hell of an introduction. Look, you don't look that old, man. That's a compliment. Just be lifting weights or something, keeping good care of yourself. I don't know. I just try to stay healthy, you know, just eat the healthiest food I can. I bike to work every day. and, yeah, try to keep in shape. You have to, man. This is a stressful job. That's what I'm talking about. You've got to do what you've got. I got you. So are you in downtown San Fran?

Is that where you're at? Yes, I'm in downtown San Francisco and financial district right now. Yes, been there many times. That's a great place. And speaking of eating healthy, great place to eat. I don't know if it's always healthy, but it's always good one way or another. Every day I bike from Presidio all the way to downtown and go back. So it's like 35, 40 minutes each way. That's awesome. To the hills. So it's, yeah. You're getting your exercise in is what you're saying. Yeah, that's what I'm trying to say. Since there's no time to work, so commute is the only time to work. Yeah, that's the way to do it. You've got to fit it in, man. It's like work-life integration. That's what it is. That's what I always preach, and you're doing it. You're doing it. So what the heck is Skyfall AI then? I mean, you must be a James Bond fan or something. I won't make you listen to Adele again. But Skyfall AI, how did you come up with it? No, so I think, of course, I think James Bond, especially the Daniel Craig movies are amazing. But I think the name Skyfall, I think why we call the company Skyfall is because we do think that we're very much inspired by this whole agents concept, you know, and the AI agents going and doing the work and saving the world.

You know, in Skyfall, James Bond saves the world. We do think that some of the stuff we're going to do is save the world as well. So I think that's the inspiration behind the name. In terms of the company, we're building pretty much for the post-LLM era. So let's talk about that. Let's deep dive into that. So if you think about large language models, which have been predominantly used by most people around the world right now, it's kind of like they're very good in information retrieval tasks. So they're very good in like, you know, they've been trained on the entire web, on Reddit, Wikipedia, all the different data sources around the web. They've been pre-trained on that. And you can do any question answering on that. Right. So I think, but they're not good in certain things, especially if you want to create, let's call it an autonomous business, which is what we want to create. We want to create a completely autonomous business, which is led by an AIC. Right. So what does a completely autonomous business consist of? It consists of autonomous employees, which is one part of it,

and the second part of it is autonomous software. So software that runs by itself and all these employees that coordinate amongst each other for getting the tasks done. So if you think about an AI CEO who's going to direct these autonomous employees and the autonomous software, the job is extremely hard. If I talk about my personal experience as the CEO of a company, I would say I'd categorize the problems I solve into three high-level buckets. The first problem, I would say, is long-horizon planning. So I always have to plan hundreds of thousands of steps ahead for everything I want to do. Let's say I have to do a fundraise in the upcoming few months, or I have to build a product, or do the research, or have a go-to-market plan, do the marketing. There's so many things going on in the company, right? So we have to like always do long horizon planning in weeks and months and years, in fact, as a CEO. This is something that LLMs are not capable of. Like they plan like 10 to 100 steps ahead predominantly because of the architecture by which LLMs are being built, which is the next token based architecture.

So that's number one. Number two, quickly, I'm going to highlight is dealing with dynamic environments. So again, as the CEO of a company, you have to make a lot of decisions every single day. It's based on external factors, based on internal factors within the company. The world is always changing and continuously evolving. And I think this is again where LLMs come short because LLMs are again pre-trained on the data in the web. And the model weights and everything don't change continuously on a minute-by-minute basis or an hour-by-hour basis as the world is adapting and changing. It's impossible to go and retrain the LLMs again because it's too expensive. You just can't do it. So I think this is the second constraint that LLMs are severely lacking on. The third constraint is what we call the inference efficiency.

So in inference efficiency, what I mean is, let's talk about the human brain. The human brain is able to process information in a very efficient way. You know we consume information. And it's around 20 watt hours is what the human brain processes information. So for the whole day, it's around 480 watt hours. Let's call it 500. And a single LLM call, on the other hand, can process like, let's say, a multimodal complex call that you're going to make through an LLM. It's more than a thousand watt hours. And that's why you can see all these crazy data center rollouts across the United States that's happening right now. So in order to tackle these problems of building this AICU to run an autonomous organization, we're focused primarily on two areas. One is world modeling and second is continual learning. So world modeling at the highest level, the concept is as human beings, we try to imagine the future rollout. We try to dream what the future is going to be, and according to the different simulations that we have,

we choose the simulation that best fits into what we want it to be, and then take a policy or an action in the real world. That's kind of the highest level concept of a world model. Okay, okay. Let me break it down. I got it, man. You're throwing a lot at us. I want to make sure that it's all understood. Let me back up and summarize kind of a little bit of what you said that I think is profound, because I have read this about what you said. And that is you said that we're at the beginning of the end of LLMs as the dominant AI paradigm. That's pretty aggressive. They're not going away. They're useful. But for more complex problems that require an understanding of how actions change the real world, LLMs are not going to get us there. That's what you said. What you said is that's where world models come in, and that will be the next major leap. And these are models that will reason about states, actions, and possible futures, hence Skyfall AI. So a couple of questions.

Do you really think LLMs are going to take a back seat? I think, Al, LLMs are going to be here for a long, long time. but they won't be the only predominant architecture that's going to be used by everyone around the world because LLMs are going to be good for the execution layer, right? For like once you have the high-level plan of what you want to do, LLMs are excellent at execution. So if you want to make LLMs even more pronounced and used by users around the world, like if you remember, I don't know if you have heard of the Jevons Paradox. Yes. If you want to increase the efficiency, so you have to increase the efficiency of the LLMs. And how do you do that? You do that by using something like a world model, like how we imagine the world to be, is you sell world model as a dreaming, as a service. And you use the world model to kind of like imagine the different future rollouts. And then once you select a future rollout that you think is the best paradigm,

then you use the LLM to carry out the action. That's how you do it. Let me pause. Let me jump in there, though, again. you got to tell me what the world model is. How does that differ from a large language model? I mean, break that down a little bit so you don't just gloss over it. Yeah. So from, like, let's say the audience perspective, in a large language model, what happens is, again, you're pre-training the data on the entire web, right? You're pre-training the entire large – that's why it's called a large language model. So it's pre-training on different sources, Reddit, Twitter, Wikipedia, whatever you call it, entire web language, right? It's written on all the text sources. And if you think about it, that's why it formulates. It's like a very large-scale patent matcher. Think about it like that. It's like an extremely large-scale patent matcher. And it's like a very smart sports analyst. A sports analyst that has, let's call it soccer for the sake of it. A soccer analyst that has watched every single game of soccer on the web,

It read everything about soccer on the web, but has never played a game of soccer in the real world. That's what a LLM is, like in a nutshell. Does that make sense to you, Al? I think it does. So that's the LLM piece. That's the LLM, yes. That's the LLM piece. Let's talk about the world model piece. Again, just for the audience, if you think about what a world model is, it's like given a state and an action pair, what's the next state? That's kind of what a world model is at the highest level, right? So it's a state action pair predictor, right? So if you think about, like, I don't know if you remember the Minority Report movie by... I do, yes. So if you remember, they had, like, 20 different simulations of what the future is going to look like, right? So then you select one simulation or one future which best fits into the roadmap you want it to be. You select that future and you go out there and do that. Now, the other thing is like, since LLMs are based on next word prediction, it's very sequential in nature.

You just go one token after the other So you just carry out a whole action end to end with that whereas the world model what happens is like you try to like imagine the different simulations you get different options of what what the different futures are going to be and you roll and you do the latent space rollout so what that means is like basically try to imagine in like uh let's say you have pixels let's say you have pixels in terms of like visualizations and stuff instead of doing the high fidelity pixels and doing reasoning on top of that, which will be much, so much more expensive, you do it in the latent space. And the latent space abstractions is way cheaper and way more cost, way more cost effective than doing a full, full blown LLM rollout. And which will take like, you know, according to the, according to the preliminary results we have with our world models, for example, compared to LLMs, we are building something that's at least 100x cheaper, 100x faster, and 160x cheaper, more cost-efficient than the same calls in LLM.

I'm not going to disclose more about this because, for example, we have our upcoming launch coming out in a few weeks, but that's kind of the highest, higher-level breakdown of like… But does that mean you can choose multiple scenarios, multiple future scenarios? You can choose the one that you want? Yes, yes. You can have different multiple scenarios, and you choose whatever you want, and then you go down that path. And you can do it way faster because you think in the abstract, in the latent space. You don't think in the token space just like LLMs are doing, and you do it very sequentially just like LLMs. Yeah, you're going to pardon my ignorance here, but if you're looking at a world model, do you build it just like you would build an LLM? Is it like separate? Are you building it like Claude would build? You build it from scratch. We're building it from scratch. Yeah, absolutely. So that's got to be pretty pricey, though. You've got to have GPUs and everything else like that. So that must come from your venture funding. Is that how it works or what? If you're not venture funded, I don't know how to do a business. Well, that must be a lot of venture funding. That's the – yeah, I don't know how you would do it either.

It's all very GPU intensive, especially the world modeling algorithms. It is, in fact, yes, yes. Let's just put it at that. The answer is yes. It's expensive in your building. This is on par with Opus or anything else? Well, so far internally we are trying to build the models to showcase that they're much more efficient than LLMs in terms of how cost-effective they are and how fast they are, how fast the rollouts can be. but yes if you want to build something that generalizes in scales like opus or astra or whatever you call it then it's then you have to do the fundraising like that to get there yes all right all right let me okay i think i got it so it's like i don't know how it's a foundational model in its own context is what i'm trying to say but are you you're also scraping the internet then? So you're going to have everything that like... No. So again, LLMs are scraping the internet because

they're doing the reasoning in the tech space, right? For example, the world modeling we are doing, the first domain for example is going to be design. We're doing the world modeling, the thinking on the visual space. So it's very different. So yes, you have to collect data from the internet or whatever different sources you have, but it's not going to be similar to how LLMs collect the data and process the data. All right. I have more questions there, but I want to get to the next question. I don't know how to jump into this question, but to jump into it. And that is, the way I ask the question will hopefully just domino a lot of different explanations. And that is, I hear you guys plan to buy a small software business. Yes. And you want to let an AI CEO run it. Now, two questions. One is, seriously? That's the first question. And second question is, how does the world model tie into a CEO? What's the connection?

Yes. Yes. Well, again, so I think a bunch of this will come up in the upcoming press releases that we're going to release. So I won't reveal much, but I will say that. Well, that's no fun, dude. Nobody listens to this podcast anyway. You can just reveal it all right here. Nobody will hear it, man. No, I do think that what I can reveal is we have bought these companies, and what I can reveal is that we are using our latest world modeling techniques to automate these companies. That's all I can reveal. Now, are these successful companies? So you bought some successful small companies. You replaced them with your AI CEO, and you're going to step back and let it go. Can you say that? Comment on it. If I say this, Al. Just nod your head yes. Say it like this. It's up to you to take the interpretation, which you want. But, yeah, I do think the goal, I'll tell you the goal. The goal is definitely, Al,

is that these companies should run autonomously, completely autonomously. You think about, like, the go-to market, the sales, the customer service, the, you know, marketing, all the parts of the organization. and we are buying these companies, keeping this into perspective. And, yeah, we want to – the experiment results will reveal itself in, let's say, within the next couple of months of what has happened with these companies. All right, so I got you. And, look, some of this has happened today. But what I hear you saying is the difference here is there's no human in the loop. No, so some of this has happened today. I'd say some of this has happened in like specialized verticals. Like you can think about like, okay, I'm trying to do marketing automation, or I'm trying to do sales automation, or I'm trying to do customer service automation. Yes, there are lots of companies who do that. But to build, like to buy an entire company and to run the entire company end-to-end,

that has never been done, at least to my knowledge. I've never seen that in my, like I'm on Twitter and my team. No, no, I agree with you. But let's just be clear, no human in the loop then. It's completely autonomous. At this point of time, mostly autonomous. Let's put it at that. Yes. In my opinion, human has to still give like the permissions and stuff with respect to customer data and such. So those things you still need human in the loop. like and payments for the admin stuff let's call it the admin stuff but not the not uh carrying out the actions and taking the decisions those kind of decision making power obviously the agents is what we believe the agents are much better at and you know why because humans will not be able to crunch the numbers and go through the databases and the erp systems and whatever is inside the company better than AI. It's impossible. Just think about it from that perspective. Well, let me ask you this, just to make sure we're talking the same language. What is an AI CEO in your definition? So if you think about majority of my job, I think majority, at least 50% of my

job is operational tasks. At least 50% of my job is operational tasks. And I talked about the three different parts of my job. One is like high level decision making under uncertainty. The second one would be, let's call it like long horizon planning, all the time we have to plan and such. These are some of the roles which we think an AI CEO should be able to do way better than any human being because they have all the data, external factors, internal factors, all of that within the company and outside the company because it's all data driven right but as what we think the ai ceo cannot do is let's say human relationships like motivating the employees right like these are the things like making sure like uh so i think the inter-human relationships within the company i don't think the ai ceo can do that in my personal opinion uh maybe maybe maybe what do you need

into relationships for inside the company if they're all AI. It's just bots talking with bots. Yeah, absolutely. But right now, we're doing it for smaller companies, right? But we are not automating Microsoft. That's not ready yet, right? So maybe a few years down the line, that's ready. But we have to start. We have to be very pragmatic and start with, like, be very practical with what we can do right now. So the companies we're trying to acquire right now are, like, less than 10 people. It's not like, oh, we're buying, like, 1,000 companies. company. The point is, I think for everybody listening, it's coming. Are we going to put everybody out of a job? Is that what's going on here? Is that your belief? I don't think so. I think it's going to make entrepreneurship even more democratized. I think that's what the goal of my company is. I think what we want to do is we want to give everyone the tools so that anyone can start their own company from scratch. I think that's why we're buying these companies. It's for the data, training all the world models with all this data that they have. Historical data also So they have, there's so much historical data that we are using to create these world models.

Again, you remember what the concept of a world model is. It's like given a state and action, what's the next state, right? So you're trying to do these simulations across millions and millions of entries across these companies and stuff, and use that to make your models better and better over time. And once you're able to do that, then at a smaller scale. So it all started out when last year we're trying to, I don't know if you know the game Roller Coaster Tycoon. It used to be very famous 20 years ago. It's like a business simulator game. You basically operate a theme park and you try to increase the revenues of the theme park. You try to organize these crazy rides and then make sure your customers don't end up vomiting and stuff if the rides are too steep. Yeah, I'm telling you, it's like a crazy. No, I think I remember that actually. Yeah, yeah, yeah. And then we started experimenting with RollerCoaster Raccoon, and we saw that LLMs really can't run the game at all. And we were like, okay, so let's start experimenting with world models. And we saw a lot of good results, promising results from world models.

We were like, so this year we're like, you know what? We should actually buy companies. We should actually, like, not do it in, like, a game world, but in the real world. So we started buying companies in the last few months, and the results are phenomenal, which you're going to know, hopefully, in a couple months. So now we're trying to buy companies for less than a million. Hopefully next year we're going to try to buy companies 1 to 10 million and then 10 to 100 million and move up the ladder. I think that's how we do that. In terms of overall revenue? In terms of overall revenue. Market share, yeah, yeah. Or market cap, whatever. All right, I got it. You can automate almost all of the operational side, but you said, like, human relationships are different. Where exactly is the line? between the two i think human relationships for example like let's say i'm trying to hire someone right now like like a high profile researcher let's say they're very much in high demand sounds like this is personal you're trying to hire a high profile researcher keep going

let's say i'm trying to hire because yeah i'm trying to give a real life everything i tell you is like real like from my own experience so uh um so let's say i'm trying to hire this person I have to go, I have to wine and dine them, I have to like talk about different scenarios, create like sense of urgency, you know, like there's so many factors involved in this, in this situation, which an AI can never be able to do. Now, what the AI can do, I'll tell you that it's, again, operational. It's like going through the list of, like, all the research papers that we're going through, figuring out who the right candidates are going to be, figuring out what the messaging is going to be to these candidates. Those are the parts, again, and sending them the emails off. Like, those are the parts which, again, AI can do. It doesn't do very well with LRMs, but using world modeling and stuff, obviously, we've shown a lot of, like, internal progress in that, yes, you can do a lot of these things autonomously but i think with uh but the interpersonal

relationship let's say an employee is not working very well uh and such in order to motivate them and such okay what's going on is there something wrong in your personal life can i do something for you those kind of those kind of things it's impossible for ai to replicate in my opinion in my opinion at least maybe i'm a bit old school but you know it's like yeah maybe you know with the younger generation and stuff, they're so hooked on to, like, let's call it, like, some of these tools and stuff, and they get dating advice from these AI tools, and some of them even date an AI avatar, and so maybe it's slightly different for a different audience and stuff, but I do think that, from my perspective, I think the inter-human relationships cannot be replaced with AI, at least not as of yet. All right, so let me ask you this. I got one question and I'm going to take a pause. And that is, it's kind of a variation of the last one, but I'm trying to close it in. That's why I asked these two.

So, Sam, if you're not allowed to hire another human outside of the relationships you've already described, and you say you're not allowed to hire another human, you've got to run a company entirely with AI, what breaks first? What do you think breaks first? And you can't, we've already talked about the relationships, but what else do you think would break? So is your question that if we are running a company completely autonomously without any employees, what happens? Is that what you're trying to say? Yes. Well, what breaks first? I mean, if something's going to go bad, we've already talked about relationships and hiring stuff. What do you think, what is your anticipation that could break or will break? to carry out like the entire operational piece you can have the ai do majority of it but for the high level approvals and stuff you do still need the humans and for that level of human judgment like you still need the humans in the loop i do think for the smallest part but again you're doing you're like doing 99 of the work let's call it using ai and then you have humans

like approving things like that's what human proving what give me i mean like the ai can approve things too so give me an example of something you wouldn't want the ai to approve you don't trust it i definitely would not trust ai with like making the hiring like like hiring the person it's impossible for me to to imagine like you talking to a bot and like like let's say an ai and then let's say the ai i'm the ai ceo and a research candidate a star researcher is talking to this AI CEO and this AI CEO has convinced this researcher completely to join the company. I cannot see that happening. But maybe there are people who are going to be convinced today. But I don't think so. I think it's a lot of like relationship building and stuff. Even like, let's say the fundraising and stuff. Like for smaller scale fundraising, sure, you can have like, you know, you can just like have an AI avatar based on like metrics and everything, do the fundraising and everything. But for bigger fundraising stuff, or even to close enterprise kind of deals, let's say,

you need human from both sides. I don't think that I cannot imagine like a AI, let's say, salesperson from my company talking to, let's say, let's say someone in IBM, let's say, and like an AI IBM sales representative and both of them negotiate with each other and figure out what the right deal contract size is going to be I don know I don imagine that It possible though but I get your point It the human interaction is still needed I get it So let me pause for a second. I have got the great, beautiful Catherine Main here who is also serves as the producer. She comes up with some great questions and I wanted you to give you the floor a little bit, Kate, to be able to ask any questions you may have. Sam, this is fascinating and I have your website pulled up and you've got a great blog that's the hierarchical planner that outperforms LLMs while running it 55 times faster at just a fraction of the size. And this is from January. So it's really incredible. One thing that comes up as we talk about this, though, is what happens to

culture in an AI-led company? How does that work? What happens? So I think the whole purpose of us doing something like this of is trying to again democratize the tool so that anyone can be an entrepreneur right i think if you think about a mission statement of what we're trying to do that means you will have so much more time in your personal life to pursue your passions or interests or whatever you want to do whether it's like you know you think about like you want to design like a spacecraft or something crazy like that or you want to like read up about like philosophy and come up with new philosophical ideas and such right so we think we think this will enable an age of abundance in my personal opinion and you'll have much more time to spend with family and friends and and stuff like that because all the operational work is taken care by ai so that's i think in a personal because i think it will free up so much time is what i'm trying to say katherine now what happens within the organization is basically uh there'll be

let's call it humans who are going to be involved only in like the high level decision making and stuff and that's how we imagine humans should not be doing all these let's call it grunt tasks i i personally do a lot of grunt tasks every day even though i'm the ceo of a startup we have 25 employees i do so much grunt every single day and it's mindless and it's and it's operational in nature it's and in more than 50 of my job is that uh even though we have so many ai automation tools are used across so many uh so many different uh parts of the business so i think humans will not be doing a lot of these operational tasks is what what we imagine the future to be and they should be doing more creative tasks they should be doing more tasks which will be like yeah i think this should be like the renaissance plus industrial revolution combined is my personal opinion is what the future is going to end up being because in the renaissance if you remember like a lot of

different creative ideas across so many different disciplines were were done over this uh over the 14th and 15th century i think something like this would uh building the ai ceo and a completely the autonomous organization will also help in moving the human race in that direction. Wow. That is an amazing vision. I love that. It's so optimistic and positive about what potential is unleashed by doing this, right? That's amazing. Can you talk to me a little bit more about the autonomy part of it? Because that's where I think that we get tripped up. People have, at different skill levels, people should have different levels of autonomy. Don't you agree, right? Because you shouldn't give 100% autonomy to me, who I may be very unskilled in doing something. You wouldn't want me to be completely autonomous. So let's revisit the idea of how you dial in autonomy and guardrails and how these decisions, what makes a human level

decision? What makes a decision okay for the AI to make? Does that question make sense? I kind of rambled there. Does that make sense what I'm asking? What you're trying to get at, Catherine, is basically uh which i which i kind of like alluded to before is like anything that's like that's like very deterministic and and uh and very uh what you call it yeah it's it has to be very deterministic in nature yeah i think deterministic is the right word and operational in nature you make the ai kind of do that anything that's a bit probabilistic which is that's why ai CEOs so much harder to build an AI software engineer, by the way. AI software engineers are all logical. It's very logical and deterministic in nature. That's why coding agents have taken off and they've done a phenomenal job, I think. And by the way, I think, like, let's call it, sorry to divulge a little bit, but I think the entry-level software engineering jobs, it's going to be so, so tough, like, for people because it's going to be extremely tough because now it's turned into basically natural language.

And you need a senior person in software engineering, for example, to kind of guide the alarms, run the experiments, and see how the experiment results are going to be. So if you're entry-level software engineer, it's going to be very, very difficult in my personal opinion. I saw a report from a few months ago. I don't know if that holds true yet or not, right now or not. But basically, majority of the Stanford graduating class, like four months ago, still were looking for jobs. and they are the most sought out software engineers out there. This was like in May or something like that. And they were still looking for jobs. I mean, that's crazy in my personal opinion. So, but what it creates is like for the senior researchers who have been through the whole, who know the basics and who know the LLM paradigm, they can be so much more productive than ever before. I think that's what's happening from at least the engineering space is what I'm trying to see internally. But bringing it back to the autonomy, the deterministic,

basically how deterministic is the decision that needs to be made is what should inform the level of autonomy that is given. Yes. If it's much more probabilistic and such, if it's very high on the probabilistic side, then you probably need a human to kind of approve and stuff. If it is very much deterministic, then you let the AI run the whole thing end to end because they're so much faster and much more efficient. So what are the jobs that our new undergrads should be going to do if it's not software engineering and we're going to have an AI CEO running the three of us? Al and I are just going to be the figureheads on this podcast and we're going to have some AI CEOs running us. I think you know what's going to happen, in my personal opinion. Like liberal arts, social sciences, those are the things that I think will take off and will have a profound impact in society. I've always said that I mean the arts are coming back will end up coming back Absolutely and I did a little stint in London School of Economics and I studied

political science after I sold my company to Microsoft and I was like why do you do that because I really wanted to because when you're stuck in computer science and tech for so long you need kind of like at least for me I needed to explore a different side of the world so I do think that's what's going to That's what humans are going to move towards. And Renaissance was mostly about like science and liberal arts and combination of that. Like, I think that's where humans are going to be. The creative side is what humans should move towards. And whether it's like, in my personal opinion, whether it's like sports or philosophy or history or whatever the humans are interested in, And that's where the energy should be divulged in, not in doing the operational, I'll call it grunt tasks that we do. Like, okay, let's move to this CRM or let's go to these databases, move the data from here, data migration projects from here to there. That's very grunt and manual in my personal opinion.

And I do think that AI is way better than humans in these kind of tasks. But Sam, let me ask you, would you want to work for an AI CEO? would i like to work for a ceo i mean it depends i i if it's if it's a small organization and stuff then we're already i shouldn't say that but you'll see the results in like six to eight weeks when we do the press release like of what happened with all the acquisitions that we're trying to do but for larger organizations i think i will have a ai ceo twin that can that can run all the operations for me. And then I can kind of like take the decisions for the harder decisions. Let's call it the probabilistic decisions. And on the, yeah, I think the deterministic decisions should be taken by my AI CEO twin. And I think the probabilistic decisions, I have to kind of give approval. I get you. I've got a lot of friends that have the AI or the CEO twins.

and it's working out well. I got one gentleman that created his AI twin, CEO twin. He said, I want to mix somewhere between Charlie Munger. I think it was just Warren Buffett and Charlie Munger. I want them to challenge me in the context of those individuals, blah, blah, blah, and put this all this stuff. It's pretty cool, I have to say. I mean, they have really long debates, and he learns a lot in terms of doing it. But, hey, before I pause again, Kate, any other questions you want to ask? I want to go through a rapid fire before we end. No, I'm just so excited that I'm going to ask you to come back because I have questions about the technology that I don't think that you can answer right now because we've got to see it come out and got to have everything released. But the idea of the – how did you say it? It was the action, right? The state action pairs. The state action pairs. Yeah. I want to understand that more as it comes into planning. Do you have any – can you expand just a little bit more on the technology behind that and what that means?

Or do we have to wait until our next session after you release? Yeah, I think we're doing a major release in a few weeks, the first major release in a few weeks. So we'd have to wait for that. And then the second major release would be in November. So we'd have to wait for that, Catherine. So just a few more weeks. I think so. I guess I was like, oh, should I really push and ask because I'm reading all of this on your blog and your website from January, and there's some really great questions because the idea of making it so much faster, and as I'm conceptualizing it right now, is almost getting ahead of the inferencing that you would do with a large language model, right? So you're getting ahead of it, but what you're anticipating is different than, like you said, he would use a large language model and that's just been trained on the web so so katherine i can just say this which is like what i can say is like you use something like a world model for high level planning and you use the and the and then you decide the different futures like okay

these are the different futures you can see and you select a future and you let the lm run it like you just run it to the end that's how that's how we imagine it because otherwise you're wasting so much tokens, so many tokens in LLM. Because when LLM does an execution, it has to complete the whole thing end-to-end. You don't have too many options when you run an LLM. That's what I'm trying to get at. Yes, so I think world models have a huge inherent advantage for that, and that will democratize LLMs even further, is what I'm trying to say. Yes, I can't wait to see it. It's going to get the cost way lower than what it is. Like people are building all these data centers and stuff, and the costs are going to plummet in terms of like how much they're spending on inference because of world models. That means LLMs can do much more is what I'm trying to say. One way or another, Sam, I agree with you with the technology advancing the way it is. It feels to me, and look, I'm not no genius, but that it's irresponsible in how much money they are spending on these data centers at this point in time.

because I think it's going to change. I think it's going to change. And some of those commitments, some of it's just commitments to spend on a data center, and I don't know that they'll ever come true. So if the stock price is going up just based on commitments, I think people need to make sure they're doing their research. I also think that because the big labs have invested so much money in LLMs, now because their whole R&D, their whole product, everything's like based on the like the you know the prompt prompting and the co-pilot kind of experience then you have to have the data centers also built out because you have to have the projections of the crazy revenues that you're going to potentially make so they can't they can't back down from that right so i think that's why something like this can be disruptive at the same time also you will also have much more use cases of llms than ever before i do think that's going to happen but at the same time, it won't be as expensive as it was until this period of time. You can't just ask him directly because you know he can't say.

You've got to do something indirect. You've got to say, like, if you were to use the technology for this, what would that look like? You know what I mean? So then he's kind of like, oh, wait a second. That's how you've got to do it. Anyway, all right, awesome. First of all, I want to make sure we get this out. What's the biggest bet you're making with Skyfall? Kind of like your mission statement. We started with that before we turned on the mic, but what's your biggest bet? So we want to reimagine a world where there's going to be enterprise super intelligence. That's the vision of this company. And we want to democratize the tools so that anyone in the planet can be an entrepreneur. Anyone in the planet can be an entrepreneur. So what that means is that suppose as an entrepreneur, you have to build a product, design a product. build a product. Then you have to figure out go to market. Then you have to figure out sales. Then you have to figure out marketing. All the whole operations end to end. We want to expose these models and APIs so that anyone can build their own business from scratch.

Currently, let me give you a preview. I don't know. You guys are very good at asking questions. Let me just tell you that. It's like I want to keep that. I want to keep these answers for the product lot but I'll reveal a little bit right now so currently it is very very difficult for AI to let's call it build a business from scratch based on the experiments we're doing it is easier to get an existing business and run a business via AI but to be a let's call it an AI founder it is impossible right now at least right now so that's what we want to move towards based on the experiments we have done over the last year or so on this. Yeah, that's what I can reveal at least at this moment. I can see, Sam, how you'd be a good fundraiser. I mean, you've got that. Seriously, that's a compliment too. I mean, you've got the knowledge, you've got the passion, you've got energy,

you've got the answers to almost all our questions. So good for you, man. I think that's going to go out of the way. If you think about it Al like you have to if you look at Sam Altman okay Sam Altman the other Sam the way more successful Sam Yes I don know He is going down He going to gonna go bankrupt with all these data centers and you gonna go straight up so keep going uh so um no so i think i think one of i heard from many uh from many um founders and stuff i was in like in a uh ceo retreat a couple days ago and they met sam multiple times i think one of the most important things that you need as an entrepreneur is to sell your vision to both the internal employees and make them very excited about your vision and external investors and i think open air would not be what it is if sam often wasn't raising like the hundred something 130 billion whatever there is in the last last round like you if you want to change the world in a profound way you have to be an amazing fundraiser at this moment of time in history

So kudos to him for doing that. And, yeah. How big is your company, by the way? I mean, at this point. We only have 25 employees right now. How many AI employees do you have? How many AI employees in our company? Yeah. We're a research lab. So I think, well, they do have the AI agents that run and do the experiments for them. I haven't asked them for a count, but let's call it the humans. We have 22 or 25. I'm sure they have at least 10x the number of agents at the minimum, the number of agents that are running experiments within the company. All right, where can folks reach you and where can folks reach your company? Yeah, we can be found on Twitter. SkyfallAI is the Twitter handle. That's the best way to reach us. Or you can go on our website, skyfall.ai, as well.

All right, fantastic. Hey, I got one more thing. I want to do some rapid fire with you. Will you do that with me? Kind of fun. We'll just have fun. It's like 15 questions, but it's rapid fire. Just some questions that I've been listening to you here. So when I say rapid fire, you just give an answer. And it doesn't have to be perfect. You know, we understand your answer can change. It's like somebody asking you what your favorite song is. Today it might be X. Tomorrow it can change. But I just want to see where your head's at. All right? Fair? You don't have to answer it if you don't want to, but they're all easy questions. I promise you. Sure. All right? All right. What job would you personally hate to see AI take away? A sports star, like a sports player, like a Cristiano Ronaldo or Messi or something like that. I'd love to see humans still playing sports, and I love the FIFA World Cup. And, yeah. Sounds good. Best business book you've ever read? Best business book I've ever read? I think back in the day I read, I'd say the Steve Jobs autobiography was one of the favorite.

I don't know if it counts as business or not, but that inspired me. I think Elon Musk's biography as well was well written. It was pretty good. I read both of those, by the way. Fantastic. What is the most overrated AI company? Right now? Yeah. There's a lot of them. I don't want to be in trouble because then I'm going to piss off some of the potential investors for me, so I can't reveal. All right. There's a lot of them. I'll get that. There's a lot of them. Okay, I can answer this question in a different way. Okay, there's a lot of companies right now that are raising funds based on just like who they are. But they are, it's like, let's say you're a researcher in DeepMind or OpenAI. Even if you're a researcher in DeepMind or OpenAI for like six months or something like that. Then you go out, you leave the company and you start, you leave the parent org and you start your own startup.

I think those kind of companies will not fan out in the long term because I manage researchers. I've worked with researchers every day for 15 years now. I don't know if an AI researcher is the best CEO or co-founder of a startup because you have to deal with a lot of tenacity. You have to deal with a multitude of problems. Like, I can't do their job. Like, I can't be a great researcher. Don't get me wrong. I can never be as good in research as they are. Otherwise, I'd be doing research, right? But at the same time, I think the founder skill set is very different than a researcher skill set. So I think what people are missing over here is like confusing the two between founder and researcher. All right. Most underrated AI company. Most underrated AI. You should be able to answer that because that's giving somebody a compliment, actually. Yeah. Most underrated AI company. Maybe IBM. Maybe IBM. Oh, that's a good answer. That's a good answer. There's a lot of people going to be saying, good answer.

Go ahead. Go ahead. Yeah, no, no, you elaborate. If you're going to talk about IBM, I'll let you do it. No, because I remember, like, I was very inspired. When I was doing Maluba back in the day, we were doing cutting-edge NLP with deep learning. IBM's Watson was, like, crazy, like, insane compared to what we were trying to build. And I was just shocked at the scale at which you guys were building it. This was, like, 10, 12 years ago. Let's call it that. Right, right. So I think IBM always comes up with crazy products, and you just have to, like, yeah, but I do think it's very underrated, some of the research that you guys are trying to do. I hope it gets more. I think you guys need more PR and something like that. Yeah, like we don't like to tout our own stuff. I don't know why that's a little bit of a culture. A lot of the AI stuff right now is about PR. It's like you have to be good in tech, but you have to also be a great storyteller. That's what Sam Hortman, right? We were talking about him before. Great storyteller. so i think maybe someone yeah pr is what's missing

no no no yeah i'm asking you the questions you're giving me the answers that's what this is about i like it this is fun we hopefully keep going just a little bit longer what's the best llm out there right now you think astra is pretty strong uh and obviously claude opus 5.5 is not bad either but Astra is what my researchers tell me. They love Astra and they're using it to the max. We pay all these subscriptions to all these companies every month. Sounds good. Sounds good. Would you let AI manage your money? Would I let AI manage my money? To be honest, yes. Because I know my bankers will be very unhappy to hear this. But if you think about what have they done in the last few years, it's basically following the S&P 500 or something like that. So I think at the end of the day, yes, it's higher risk. That's what the algorithm is. Yes, I agree. But I do think it ends up being close to the S&P 500.

So I'm like, okay, you know what? Maybe algorithms should run this, definitely. I do think so, yes. Well, let me ask you this then. Would you let AI choose your spouse? there's a lot of there's a lot of apps out there before you answer that i think it's happening today would you let it so so let me try to answer this giving a personal anecdote oh okay so so i have tried to use this uh in the past let's call it uh and uh the matching the back in the day it wasn't great at all to be very honest with you i think i think uh most of the that's why i still stick to my point about like human relationships i think that some uh referral from a person is way better for dating in my personal opinion like from someone you know uh at least it has worked out for me better than like then then going on these apps and and

matching happens with apps obviously so there's obviously cases where it has succeeded but for the vast majority of my friends who've tried the apps out, it hasn't really worked for them. And maybe there's a small percentage it works for, but I just don't know them. That's good. That's a good answer. So just a few more. We're almost done. The pain is almost over. What's the most? It's amazing. There's no pain. This is like, it's amazing. The pain is going back to work and doing all those operational tasks. Same for me. I'm just delaying it. That's what I'm trying to do here. I got people texting me right and left. I'm like, forget that. This is fun. What's the most useless thing people use AI for? What's the most useless thing people use AI for? That's a hard question. Let me think. AI is pretty useful in a lot of things, in my personal opinion. Like, maybe I'm only very picky. Okay, I will say this. So I think, yeah, okay, this is a good one. um i think people are using ai as like uh i know some people who are using chat gpt and

and gemini as such as as they're like um how do you call it as a personal assistant it's it's really good let's call it the opposite would be to get advice as a psychotherapist or something like that i think that's where i think it's not good at all and it always starts to agree with you and it always starts to make you feel good and i think it's it's doing as a therapist i don't think ai is doing a great job yet because i think the human touch is definitely missing the and how humans interpret the other person's emotions that's completely missing uh ai just tends to agree with you so i think the ai therapist is something where i think they're not there yet what's your take on that i think you're i think you're on to something there if you're asking me, hey, I'm the host, man. Sam, come on. No, I would say that, look, I think AI is very useful, and it's good to bounce ideas off of, but I think people are using it too much for, like, just making every freaking decision and just going out there and thinking that it's going to give the

right answer all the time. You've got to use your brain. You're so collaborating with, yeah, yes. And you know what, Al, the problem is? When you're making your personal life decisions based on AI, Think about what the training data for that is. Reddit. People go and rant on Reddit about their stuff and rant on Twitter. And it's trained on that. And you're making a decision based on what the training data for, pre-training data for Reddit and Twitter, that's pretty bad life. That's exactly right. But most people don't get that. You're exactly right. So I think personal life decisions for, oh, should I, like, let's say, break up with this person or what should I do with ABC in my decision-making in my work? These things you should not use AI for. I don't think so. I think that's like, yeah, for me, it's like a no-no, big no-no. Well, I also find it troubling that people are having relationships with AI. Relationship advice. Relationships. I mean, relationships, not even advice. Relationships, period.

That scares the hell out of me. I think we're old school, Al. I think the new generation is different than us. Let's just put it at that. Yeah, but I think we all have to adapt to the process. Maybe I'm old school, but I still think going out, meeting a companion, you know, trying to, you know, take somebody out to eat, have a conversation. That's it can be hard. I mean, because you're trying to find somebody that you're compatible with, whatever. The easiest thing to do is to get on AI and then have it tell you you're the best thing since life spreads. You know, it's like you're great. It's easy. But that's trouble. Send them an emoji of a hamburger or send them an emoji of flowers, and then that's the date you have. You know, that's what. So what's the biggest AI hype right now? I still believe that LLMs are much more overhyped than what it should be. Of course, it's extremely useful, but LLMs are much more overhyped. The other thing that is overhyped right now, in my personal opinion, is this whole scaling back of AI.

And let me tell you why. I think what's happening is these big companies like Anthropic and OpenAI and such, they're not going to meet their revenue targets. And it's getting much more expensive in terms of the data center usage and all that stuff. So what's happening is they're like, you know what, in order to and they're all going to go IPO very soon. They're going to go IPO very, very soon in the next few months. So the way they're doing it is, in my opinion, again, this is my opinion. Doesn't mean it's a reality or truth, but it's my opinion. I have to say my opinion. so uh so so you don't have to i'm glad you do keep going so what i think is there's like a high level agreement amongst like the leaders of these companies of like let's all agree on a ai slow down so as to speak uh because they're not meeting their revenue targets that's what's going that's what's going on right now and they frame it is like oh yeah it's gonna kill the world the world is in danger, blah, blah, blah. I think that's all complete, in my opinion, complete

both. That's my honest opinion. Because they're not going to meet the revenue targets and not going to hit profitability, that's what's going on. Alright, two more questions and then I'm going to let you go. If you had to bet your net worth on one technology for the next 20 years, what would it be? Just one technology. World modeling, 100%. That's going to be the next big thing. That's the next thing. just like you know in 2020 how scaling laws made llm's a reality you know with scaling laws basically it's like okay put more data uh and and and uh bigger algorithms lots of processing power and stuff then it's going to scale the algorithms are going to scale same similar a similar paradigm is going to happen with world models over the next 12 to 18 months you could already see it but The Fairfields startup got acquired by AMD like two days ago, I think, for $8 billion. It's just a research lab. So people are already seeing that world models have huge potential. So that's going to be the next big paradigm after LLMs.

All right, here, I'll finish with the final question. I can keep going, by the way. You're fun to talk to, dude. You're very knowledgeable, and you answer the questions, except some of the ones that are confidential. But we're going to bring you back, and we'll get the technology here in a second. After the launch. All right, after the launch. if you weren't building an ai company what would you be doing i would love i'd have loved to be like uh you know this this comes from the heart i'd have loved to do something in like sports like i'd have loved to be an athlete until you're a sports guy no so so i do think i'd have like a lot of people i know personally uh tell me that i have the attitude of like you know like i have to win no and give my 100 effort no matter what and i just never had the exposure when i was growing up in India because it's a very different mindset over there. Oh, you have to study and work hard and do whatever to get good grades and stuff. So my dad really pushed me hard on that. And because of variety of factors, I'm very appreciative. Well, it served you well, it looks like right so far. Anyway, keep going. But I'd love to do something in sports. But

yeah, maybe that's for another lifetime. All right. Fantastic. Hey, Sam, you're a good man. Thank you for being here. It's been a pleasure talking to you. It's been fantastic. thank you Al, thank you Catherine, appreciate it hey listeners, you know what I'm going to say hit us on almartintoxidated at gmail.com we'd like to hear from you, we'll get to people on that you want to hear about until next time, rate us wherever you are and we'll see you on the podcast, later

番組の概要欄(原文)

Send us Fan Mail What happens when an AI isn’t just helping run the business, but is running it? Serial AI founder Sam Pasupalak joins Making Data Simple to talk about moving past LLMs to AI “world models,” which simulate how the physical world works. Sam sold his first company, Maluuba, to Microsoft. Now, at Skyfall AI, he’s testing whether an AI system can act as CEO and run a real business with no humans in the loop. Tune in to hear how world models work, what happens to company culture wh...

出典・この記事について

番組
Making Data Simple(IBM Big Data & Analytics Hub)
元の回
#425 Beyond LLMs: The AI CEO Experiment — with Sam Pasupalak, Co-Founder & CEO, Skyfall AI
配信日
要約作成日
公式サイト
www.ibmbigdatahub.com/tag/7039

この要約・解説記事は、元の音声を文字起こしした内容をもとに AI(Claude)が日本語で作成したものです。 固有名詞や数値に誤りを含む場合があるため、正確な内容は元の音声でご確認ください。制作方法について

X でシェアApple Podcasts で聴く

関連エピソード