Podcast interview with Dean Leitersdorf, co-founder and CEO of Decart, about Mirage, a real-time video-to-video generative AI model that restyles a live video feed frame-by-frame in about 40 milliseconds. After a live demo of Mirage transforming the host's feed into an interactive game-style world, the discussion covers how the model achieves real-time speed, Decart's "live stream diffusion" architecture, and solving the error accumulation problem in real-time video generation. Gaming applications are explored, such as restyling existing titles (e.g., GTA-V with a winter filter) and texturing vibe-coded prototypes, alongside Dean's view of real-time generative models as a new creative medium. The episode closes with a debate on AGI—how close it is, what counts as progress, the post-AGI society, and why AI brings back the age of the generalist. # Channel Presentation Observations ## Format & Presentation Style - **Interview/podcast format**: Hosted conversation with a guest (Dean, co-founder/CEO of Decart), including live screen-sharing demos of AI products ("Let me share my screen... this is real me in the interview screen"). - **Hosted by "Dan"**: The guest addresses the host as Dan Shipper in the closing outro, and the host is described as running an influencer-based business ("Thank god for me because that's my business model"). - **Includes a promotional outro**: A stylized subscribe/like appeal referencing the show name "AI and I": *"smash that like button and subscribe to AI and I... a journey into the future with Dan Shipper as the captain of the spaceship."* ## Intended Audience - **AI-interested, technically curious general audience**: The host frequently summarizes for listeners ("for people who are listening..."), asks for definitions of technical terms ("Can you define differentiable for people?"), and uses accessible analogies. - **Viewers familiar with the AI ecosystem**: References to Genie 3, Dream Booth, Claude Code, Suno, and "vibe coding" are used without extensive background explanation. ## Claimed Scope (Self-Description) - The outro explicitly frames the channel as covering **ChatGPT and AI knowledge**: *"pure, unadulterated knowledge bombs about Chat GPT"* — an explicit creator claim about channel content. - The host states his business model is **influencer-based content**: *"that's my business model"* when discussing influencer-driven social media. ## Demonstrated Presentation Patterns - **Long-form technical-philosophical discussion**: The conversation moves between hands-on product demos and abstract topics (AGI timelines, ancient Greek democracy, organizational theory). - **Host positions himself as a friendly skeptic/interrogator**: Offers counterarguments ("I'll take the opposite on that") and asks probing questions ("How did you make this?"). - **Personal familiarity with guests**: "you're one of my favorite people to talk to in AI," "It's been a while" — suggesting a returning-guest, relationship-driven interview style. ## Notes on Claims vs. Demonstration - Guest statements (e.g., "Mirage is the only real time video to video model in the world," "$100 million at a $3 billion valuation") are **guest claims about their company**, not channel self-description, and are not independently verified in the material. - No curated metadata was supplied; all observations derive from the transcript. # Delivery Observation: "This AI Makes a Video Game World in 40 Milliseconds" ## Overview Two-speaker interview conversation: **Dan Shipper** (host) and **Dean** (guest, co-founder/CEO of Decart). The delivery is informal, demo-driven, and highly conversational, with frequent interruptions, overlapping enthusiasm, and a late stylistic break in the outro. --- ## Tone - **Enthusiastic and casual throughout.** Both speakers use filler-heavy, unpolished speech: "Uh," "um," "you know," "like," "okay." - **Mutual excitement during demos.** Reactions escalate quickly: - Host: > "Dude, this is so cool." / "this is this is so amazing." - Guest: > "It's it's it's exciting times." - **Playful teasing.** The host needles the guest about naming: > "very Jenzy of you" (after "Dulu"), and the guest jokes back: > "We We all have to be We all have to be a bit delusional, okay?" ## Vocabulary - **Heavy AI/tech jargon used casually:** "vibe code," "next token prediction," "autoregressive," "diffusion," "PTX," "distribution." - **Recurring informal intensifiers:** "insanely" (> "went insanely viral"), "mindblowing" (> "it is just mindblowing"), "completely," "exactly." - **Guest's coined/branded terms:** "LSD model, a live stream diffusion model" — with a self-aware joke: > "it's a very technical term." ## Recurring Phrases - **"Exactly"** — used repeatedly by both speakers as agreement markers (e.g., > "Exactly. Spot on." / > "Exactly. And so basically what you're saying is..."). - **Guest's repeated "Let me let me"** construction: > "Let me let me show you one cool thing." / > "Let me let me pull this up." / > "Okay, let me let me let me show you one cool thing." - **"Okay"** as a frequent pivot/reset word for both speakers. - **Guest repeats "very" for emphasis:** > "very very very efficient" / > "very very very optimized assembly code" / > "humanity's future is very very vivid." ## Teaching Habits - **Learning by demo first, explanation second.** The guest shows the product before explaining: > "Let me share my screen. And as you can see, see this is this is real me..." - **Host actively translates for the audio audience:** > "for people who are listening, uh basically what's happening is..." / > "So just for people listening again, basically you're in your portal world..." - **Analogy-heavy explanation.** Both speakers rely on extended analogies: - Guest: > "some problems are just a big hay stack with a single needle" vs. haystacks "full of needles." - Guest on model degradation: > "your entire screen just becomes reds or blue or green." - Host: > "you need a skeleton to have a... your body work" (bones vs. muscles/tendons analogy); "it's kind of like a jellyfish." - Guest on curriculum learning: > "It's sort of like... the way a a child learns to run." - **Host requests definitions:** > "Can you define differentiable for for people?" — prompting the guest to teach in plain terms (> "think of a differentiable or continuous function something that it's it's okay if if you're close by"). - **Scaffolding via questions.** Host restates and checks understanding: > "Okay, let me see if I understand. So basically what you're saying is..." — then the guest confirms: > "Exactly. Spot on." - **Concrete numbers and specifics:** latency figures (> "40 millisecond delay... The next version... is going to be 16 milliseconds"), team sizes (> "The cart is roughly 70 people now"), and company milestones. ## Communication Patterns - **Stammered repetition for emphasis** — both speakers, e.g., > "So So what you're doing" / "it's it's it's the dark abyss of AI" / "I I I completely agree." - **Back-and-forth co-construction of ideas.** The philosophy passages build jointly: host offers the countable/uncountable infinity framing, guest responds with the needle/haystack analogy; host later calls back: > "this is a really good book end to the podcast because this goes right back to... one needle in a... hay stack." - **Friendly disagreement.** Host pushes back on AGI timelines: > "I'll take the opposite on that. I don't think in 12 to 18 months..." — disagreement stays civil and idea-focused. - **Interruptions for clarification:** > "Wait, let me stop you there for a second." - **Direct audience appeals.** Guest: > "if if you're listening in and you want to be an alpha beta tester, shoot me an email directly." - **Hyperbolic claims delivered confidently:** > "There's only two interesting problems in the world right now... Nothing else matters. Really, nothing else matters." ## Outlier Moment - The final outro shifts to an exaggerated, scripted promotional register alien to the rest of the conversation: > "you absolutely positively have to smash that like button and subscribe" ... "Dan, I'm absolutely hopelessly in love with you." This is an **isolated stylistic moment**, not a pattern of the interview delivery. ## Uncertainty Notes - Attribution of specific verbal tics to one speaker vs. the other is occasionally uncertain because the transcript's speaker markers are inconsistent in places; e.g., the bones/muscles analogy appears within a block where speaker turns are ambiguous. - Some repetitions may reflect transcript disfluencies rather than deliberate delivery emphasis; the pattern is real (it recurs), but its performative intent cannot be confirmed from the transcript alone. # Observation Record: "This AI Makes a Video Game World in 40 Milliseconds" ## Speakers - **Dan Shipper** — host of the podcast ("AI and I") - **Dean (Dan? identified as guest)** — co-founder and CEO of Decart (transcribed variously as "Daycart" / "the cart"), makers of Mirage. *Note: transcript identifies the guest as "Dean"; his company name is inconsistently transcribed.* > Speaker attribution below relies on conversational flow in the transcript; some `\>\>` markers are ambiguous. --- ## Product Claims: Mirage (Decart) - **Dean:** Mirage is "the only real time video-to-video model in the world right now" — takes any video stream (camera, conversation, game) and changes it via a prompt. - **Dean:** Current version has a **40 millisecond delay** between input frame and output frame; the next version will be **16 milliseconds**. - **Dean:** Achieved by writing highly optimized GPU code in **PTX** ("the layer below CUDA" — GPU assembly for Nvidia GPUs). - **Dean:** Mirage is an **autoregressive** model rather than bidirectional — a diffusion model combined with an autoregressive transformer, which he calls an **LSD model** ("live stream diffusion model"). It generates frame by frame, "like training a video model just on next frame prediction." - **Dean:** The model takes two input streams: the video to edit, and what it already generated, to predict the next frame. Longer context acts as "the model's memory" (e.g., restoring a black shirt when someone re-enters frame) and is critical for seeing motion, not just static images. - **Dean:** The hardest problem was the **error accumulation / repetition problem** — after 100–200 frames the output would degrade to a static color. Solving this took ~six months of research and "several thousands of experiments." It required unlocking "roughly seven or eight things," each **binary**: "until you unlock it nothing works... you never know how close you are." - **Dean:** Development used a stepping-stone strategy: **Oasis** (real-time Minecraft generation, launched November, went "insanely viral") worked on a single small distribution; they then progressively widened distributions (street footage, etc.). - **Dean:** Use cases include a Minecraft mod in alpha ("change Minecraft to be Barbie land"), texturing vibe-coded games, game engine integration, and filtering existing games (a GTA 5 winter demo). ## Philosophy of AI vs. Classical Computing - **Dean's differentiability framing:** AI is good at **continuous/differentiable** tasks (approximate image generation is fine); CPUs are good at exact, non-differentiable tasks ("add five and five, always gives 10"). AI "would suck at" reversing or implementing a hash function. - **Dan Shipper's framing:** Two ways of seeing the world — a **countably infinite set** with one right answer (classical search/programming) vs. an **uncountable infinity** of solutions where every solution is "kind of meaningful." His view: you start with the uncountable infinity, then solve within it with countable methods. He flags, "I know I'm probably sounding very galaxy [brained]." - **Dean agrees** with the needle-in-haystack analogy: some problems have one exact needle; others have "an entire area full of needles." - **Dean:** Criticizes Google's **Genie 3** demo claim that AI can replace everything including game engines. His position: keep the game engine for deterministic state (e.g., "exactly 71 gold coins," the pickaxe in a chest from five years ago) and let AI handle the creative/approximate layer. - **Dan Shipper's skeleton analogy:** deterministic code is bones; AI is muscles/tendons providing flexible motion within rigid limits; without bones "you're kind of like a jellyfish." Dean endorses the analogy ("I love the jellyfish analogy"). - **Dan Shipper's emergence analogy:** classical game code is like atoms obeying equations; Mirage's texture layer is like emergent conscious experience/qualia. ## Product Claims: "Dulu" App (Decart) - **Dean:** A consumer app coming "in the next few days" — upload a photo, get hundreds of instant variations; video camera features coming ("turn my friend into Elsa," add a monkey that can high-five back). - **Dean:** Not initially a social network; users post to Instagram first. He reports internal interest in **trends** (e.g., "turn your hair into grapes") and notes friends/family engaged with each other's edits — an implicit return to friend-based rather than influencer-based social dynamics. - **Dan Shipper (challenge):** Cites prior burnouts — Apple Animojis, Dream Booth-style Stable Diffusion avatar trends. **Dean's response:** (a) the Dream Booth paper's last author recently joined Decart; (b) quality has jumped enormously in two years; (c) images are a gateway; real-time video and audio are the genuinely new thing. - **Dean:** Decart runs alpha/beta testing; invites listeners to email to become testers. He predicts that in two or three years "almost all pixels and audio waves" will be AI-generated or AI-modified, so people must build tolerance and understanding now. ## AGI and Economic Predictions - **Dean:** "Only two interesting problems in the world right now": the race to AGI and the race for consumer dominance. Splits AGI into a **Terminator/singularity stage** ("a few years away") and **economic AGI** (machines outperforming humans at economic work), which he says "could easily start creeping up on us in **12 to 18 months**." - **Dean:** Believes this will be "humanity's best golden age," citing a theory that **democracy arose in Greece** because technological progress gave people time to think rather than work fields 14–15 hours a day; AI-provided free time could analogously enable a creativity renaissance. - **Dean (revised claim):** Possibly in 12–18 months AI enables companies "way bigger than anything we saw so far" — noting zero trillion-dollar companies in 2017 vs. ~ten today — and that "there's a very good chance" the **entire stock market doubles in value** in that window. - **Dan Shipper (direct disagreement):** "I'll take the opposite on that. I don't think in 12 to 18 months... we'll have seen that happen." His reasoning: - AI is better/faster/cheaper than his lawyer on the right prompt, but **"given the right prompt is an incredibly difficult question."** - Self-sufficient AI depends on **continuous learning** (weight updates), which he believes context engineering alone won't deliver. - Child-development analogy for AI autonomy; his preferred AGI definition: "when it is economically profitable to leave your AI on all the time" — several years away and very domain-specific (e.g., some coding first, but not PTX-level work). - He concedes a possible 12–18-month **stock market bubble** but not sustained true value. - **Dean (partial concession + example):** Agrees AI won't replace coders yet; describes using Claude Code with Flutter over two weeks to contribute to UI despite knowing no CSS/HTML/JavaScript — his value was "common sense" intuition (e.g., spotting a data-flow bug). His dad, retired dean of medicine, informally serves as "senior adviser for common sense" at the company. ## Athens, Generalists, and Organizations - **Dan Shipper's historical claim:** Athenian democracy required **generalists** (statesman, juror, warrior); it "stopped working" when Athens became an empire — the equivalent of a trillion-dollar corporation — because empires need **specialists** (e.g., you don't send a farmer to lead the Sicilian expedition). Specialization then persisted in Western society. - **Dan Shipper:** AI, as "a thousand specialists in your pocket," lets people stay generalists longer — he cites his own ~15-person team of almost all generalists. His hope: a return to a "golden age of Athens type vibe." - **Dean:** Endorses this; says almost everyone at Decart (~70 people) is a generalist, and the company remains "completely flat." - **Dean's organizational theory:** He spent ~two years studying why organizations fail. Claims Google (180,000 people) and big tech generally "don't really work" — can't get the best out of smart people. Key bottleneck: **information flow** and limits on creative slots — "humanity has never been able to construct an organization that gets a thousand creative people and lets them all be creative at once." Any organization enables only ~20–30 creative generalists (startup teams, or Roman legion generals over 100,000 non-creative foot soldiers). Non-creative jobs are easy to define in "three bullet points"; creative roles are hard to define and risk-laden. - **Dan Shipper's diagnosis of big-company failure:** (1) information flow problems, and (2) employees optimizing for promotion over doing the right thing. **Dean explicitly says** he has "no idea how to solve the second one." - **Dan Shipper's explore/exploit framing:** creative work has an **explore** phase (figuring out what to do, frame of puzzle unset) and an **exploit** phase (execution of defined puzzles). AI is currently good at well-defined problems — but "surprisingly few problems you can define that are valuable" — so humans can spend more time exploring. He ties this back to the opening needle-in-haystack/uncountable-infinity discussion. - **Dean:** Agrees — "AI is going to leave the creative stuff to us" — and adds that AI can read everything in a thousand-person organization (every email, every call), overcoming communication bottlenecks that previously meant "not a single person knew everything that was going on." ## Preserved Tensions and Qualifications - **Unresolved disagreement:** Dean's 12–18-month economic AGI / market-doubling prediction vs. Shipper's explicit rebuttal; the two do not converge — Dean shifts to a softened productivity claim, and Shipper allows only a possible bubble. - **Tentative claims flagged by speakers:** Shipper twice qualifies his own analogies ("I'm probably not going to get it exactly right," "sounding very galaxy"); Dean hedges repeatedly ("it's going to do something. I don't know exactly what"). - **Self-acknowledged limits:** Dean concedes he cannot solve the promotion-incentive organizational problem; Claude is useless at PTX-level work ("it's the dark abyss of AI that no one wants to touch"). - **Aspirational vs. demonstrated:** Many product claims (16ms version, Minecraft mod, Dulu app features) were stated as upcoming, not shipped, at recording time. # 视频来源观察 ## 视频 - **标题：** “This AI Makes a Video Game World in 40 Milliseconds” - **视频 ID：** 01a0f15a-db02-705a-8d5c-476ec97c94b0 ## 说话者和角色 - **Dan Shipper — 主持人/创作者**（音频最后确认了姓名；他介绍了“AI and I”节目，并接受了结尾的致敬）。 - **嘉宾（Dean，Daycart/Decart 的联合创始人兼 CEO）** — 主持人称他为“Dean”，后来又提到了“Dan”，嘉宾在结尾处回答了“Thanks so much, Dan”；嘉宾表示他是 Daycart（转录文本中的拼法）的联合创始人兼 CEO，并给出了邮箱地址 “dan@decart.ai”（在转录文本中呈现为 “dan atthequart.ai”）。录音中的说话者姓名/拼写不确定。 - **旁白/合成结尾声音** — 最后一段（“smash that like button…”）看起来是自动生成的宣传性结尾，并非人类说话者的自然话语。 ## 说话者逐字记录的归属内容（经证实的事实） - **嘉宾的陈述：** 共同创立了 Daycart/Decart；筹集了 **$100 million，估值为 $3 billion**（由主持人陈述）；推出了 **Mirage**（实时视频到视频模型）、**Oasis**（实时 Minecraft 模型，于 11 月发布）、即将推出的 **Minecraft mod**（发布时处于 alpha 测试阶段）、即将推出的应用 **“Dulu”**；Mirage 的当前版本延迟为 40 毫秒，下一版本为 16 毫秒；构建了一个 “live stream diffusion (LSD)” 自回归模型；为 Nvidia GPU 编写了 PTX 汇编代码；公司规模约为 70 人；Dream Booth 的最后一位作者加入了旧金山办公室（之前在 Snap，再之前在 Google）。 - **主持人的陈述：** 运营着一个由约 15 人组成的团队（“Every”），团队成员几乎都是多面手；不同意 12–18 个月经济型 AGI 的时间表；提出了基于儿童发展类比和“永不关闭 AI”的 AGI 定义。 ## 引用的第三方声明/来源 - **Google Genie 3** — 嘉宾将其描述为一款演示的世界模型，声称可以用 AI 生成的世界完全替换游戏引擎（据称“未产品化”）。 - **OpenAI / GPT-2 / GPT-3.5 / ChatGPT / Veo** — 被引用作为上下文（循环、聊天机器人、视频模型）。 - **Yann LeCun** — 主持人将复合误差问题归因于他（未直接引用）。 - **Dream Booth** — 被引用为之前 viral 式图像编辑趋势；由新员工共同研究。 - **Anthropic 的 Claude / Claude Code** — 嘉宾描述个人使用（Flutter UI 代码），并明确表示它*没有*用于 PTX 汇编工作。 - **雅典民主、罗马军团、福特工厂、苹果万亿美元市值里程碑（2018）** — 用于论证的历史/事实参考。 ## 宣传片段（已标记） - **嘉宾推广：** Mirage 的现场演示，即将推出的应用 **Dulu**（免费图片编辑、即将推出的视频/相机功能），Minecraft mod alpha 版，以及征集 alpha/beta 测试人员的邮箱呼吁。 - **结尾的合成结尾** — 明确要求点赞/订阅的促销信息，最后以“Dan, I'm absolutely hopelessly in love with you”结尾；这似乎是自动生成的，不代表主持人本人。 ## 值得记录的主张不确定性 - **公司名称拼写** 在文本中不一致（“Daycart”、“theart”、“thequart”、“every”）；嘉宾的名字呈现为 “Dean” 但邮箱使用了 “Dan” — 身份细节明确保留为不确定状态。 - **融资和估值数据** 仅由主持人在介绍中陈述。 - Genie 3 的功能是嘉宾对 Google 演示的第三方描述，并非直接来自 Google。 ## 评论区/元数据 - **转录文本中未提供任何精选元数据**；无法从所提供的内容中确定评论区的可用性。 Let me let me show you one cool thing. Then we can dive in right into how this works. And then you get like tissue holder then you can turn it into a gun \>\> and every once in a while if you do this then it like shoots out something. How did you make this? \>\> That's a really fun question. So So what you're doing with what you see with mirage is it gets a live input stream and creates a live output stream. It just generates them frame by frame and not the entire video at once. So it's kind of like training a video model just on next frame prediction and not next token prediction. You just have to predict the next frame each time. \>\> Dean, welcome to the show. \>\> Thanks so much for having me. It's been it's been a while. Been a few months. \>\> Thanks for coming. It has been a while. uh you're one of my favorite people to talk to in AI. You're in this like really interesting intersection of like doing incredible new stuff at the frontier, but you're also a big philosophy nerd. Um and uh so there's there's we're just going to talk about a lot of good stuff. So for people who don't know, you are the uh co-founder and CEO of Daycart. Um you can describe for us what Daycart is in a second, but you do awesome um real-time generative video uh models. Uh you just raised $100 million at a $3 billion valuation. Congratulations. Um, yeah, welcome to the show. \>\> So, so excited to be here. You know, I read the the the newspaper every uh when it gets to my to my inbox and always fun to have these conversations. Yes, we have so many exciting things going on right now. Like, these have been incredible past few weeks and we have the tons of launches for the weeks uh for the upcoming weeks, which uh you know, we can get into to to a few of them here, but it's it's uh exciting times. We're creating a completely new way for just people to interact with AI, just to have fun with AI. \>\> So, you just just to give people a sense of um what you do, you just launched this thing called Mirage, which is kind of crazy. Can you just show us Mirage? \>\> Yes. Let me let me pull this up. But we just launched Mirage. It was a few weeks ago, like three weeks ago. Uh by the time this will probably up, it'll probably be four or five weeks uh since it was launched. Mirage is the only real time video tovide model in the world right now. And what that means is that you can take any any video stream that you have, whether it's this conversation or your camera or a game even, and just put it in through the model and change it just with a prompt. So, what we can I I'll bring this up right now so I can show it. Let me share my screen. And as you can see, see this is this is real me in the, you know, in the interview screen. And and you have the Mirage version of me. So, the current Mirage version of me is in the Versail Palace. Or we can do some Blocky kind of thing. Okay. Or we can do an anime version. Uh \>\> this is this is so amazing. I literally just saw this for the first time like right before we started recording and it was like a oh my god moment for me. Um for people who are listening, uh basically what's happening is you're on a web page and uh you have your camera going. So I I see an image of Dean on my screen. And then when you started Mirage, it just in real time transforms you into effectively a Pixar character, something that looks a little bit like a Pixar character, but different themes and a different setting. So right now you're on Jingle Bells, now you're a wizard, you've got like lights coming out of your hands. Um it's it's crazy. And uh I like I feel like \>\> Lego. \>\> So how did you make this? I'm made of Lego. Look at this. \>\> Okay, let me let me Okay, let me let me show you one cool thing. Then we can dive in right into how this works. But I love this is one this one of the the prompts I love. If you put it in the portal world and then you get like like physical objects like a blocky thing, okay, like this is like, you know, tissue holder. Uh then you can turn it into a gun. \>\> And every once in a while if you do this then it like shoots out something. Wow. I It's uh Yeah, like like one one of the things I'm dying for the next version to have is so I can be like this and be like, "Okay, like and then like a huge portal pops up or something." Like imagine a real Rick and Morty gun. \>\> Dude, this is so cool. \>\> That's cool. \>\> That's actually cool. \>\> Um so yeah, just for people listening again, basically you're in your portal world portal world. You picked up a tissue box and the tissue box became a gun. And if you move the gun in the right way or you move the tissue box in the right way, it kind of shoots a little bit. Um, so it sounds like I mean right now the possibilities are like oh real time realtime video games uh that you can just you can just play. Uh is that is that what we're like where you're going with this? \>\> So I think the possibilities with this what this kind of tech does is it actually enables a different way to interact with AI video. What did AI video do so far? It was a great piece of tech that gave us cat videos on Tik Tok. Okay, that's that's we got out of AI video. Now, for the first time, computers can actually respond to us as we're, you know, moving things, changing things. They can in real time change it. So, it can be used anything from for gaming. We're going to have a mod coming out for Minecraft. Uh we have an alpha testing right now. It'll be coming out in the next few weeks. We can you're playing Minecraft and you just through through the game you you tell the mod, hey, change change Minecraft to be Barbie land and everything becomes Barbie. So you can take it you can take it to gaming. We're also integrating with lots of the game engines so that as you're developing a game, you can actually see this or you can vibe code a game end to end. So imagine vibe coding with with an LLM a game and doing all the texturing with Mirage. So So that's that's one category. all the real time interactive gaming categories. Huge. \>\> Wait, let me stop you there for a second. So, are you saying that you said it's video to video? So, when you're vibe coding a game, how does that work? Is it it's generating video that you're playing or is it also able to output like the 3D mesh that you might import into Minecraft? So, one of the one of the things that people have been going crazy over, you can just vibe code a game in like, you know, low poly, just bunch of blocks moving around and and turn that into into just have Mirage do the the coloring of that. \>\> Oh, let's see if I can find an example for this game. \>\> Okay, let me see if I understand. So basically what you're saying is um it's really hard to make a AAA blockbuster shooter game, let's say. Um but it's not that hard to make a a 3D game, especially if you're vibe coding a 3D game that's uh just a firsterson shooter where the uh the gun is a like a rectangle or like a a cube. \>\> Exactly. And so basically what you're saying is you make that game and then when when your game engine renders that really basic game, Mirage takes it and turns it into a video that has all the textures and all that kind of stuff. \>\> Exactly. Spot on. And that's uh people have been playing with this online and it is just mindblowing to see what people came up with that you can actually you can just uh some some people just use you know outputs from Unreal Engine and pass them through Mirage. Uh, others actually just, you know, vibe coded games and added Mirage on top. I can I can send a few examples and uh, you're editing this video, we can put it in. Um, but it actually just lets you build a game very quickly with an LLM. Just half an hour like, oh, add this tree here, add that there. It doesn't even have to build a tree. It just builds a blob that kind of looks like a tree. And Mirage just comes and completely changes it. So, it's going to do something. I don't know exactly what, but it's going to do something big to gaming. Uh either taking existing games, you know, where we were here at the office. We're playing GTA a few days ago and we passed GTA through a winter filter and then boom, all of GTA 5 just turned into a winter version. And there there's no winter in regular GTA 5. So, sure, you're going to be able to get a bunch of varieties of existing games, but you're also going to be able to create a game end to end just fully AI based. Just vibe code it and texture it. One of the things this is making me think of, and I'm sorry, we're already going to the philosophy stuff. Um, but I'm not actually sorry. One of the things that it's making me think of is there's this really interesting thing happening where the underlying game is still works with classical code and then the what Mirage is doing is sort of doing this emergent layer of texture in the same way that you know maybe you can break down the world into you know atoms and molecules that that work according to like specific equations that we can um we can figure out. But then our conscious experience of the world is this emergent textured full of qualia type thing. And you kind of need like or at least the way that we understand the world is is it has both or at least the way I understand it. And it seems like previously we had to make that um all games had to be able to reduce down to that deterministic code and it was very very difficult and timeconuming to to set that world up and now uh you have this sort of emergent layer that you can just vibe on top of the basic I don't know Newtonian or classical world that you can build with programming. Is that kind of what it is? I love I love that you're going into this. You got me thinking about this. Um there's some stuff that regular coding is great for and other stuff that AI is great for. What's AI great for? AI is great for doing stuff that's differentiable that's continuous. Okay, fine. Drawing a picture that's kind of continuous. It doesn't matter if I get exactly your face or something close. It's fine. What's AI definitely not good at? Um reversing a hash function or something. Okay. Or even just implementing a hash function. Anything that that is very that is not differentiable, AI would suck at. \>\> Can you define differentiable for for people? I know you're a uh you know, you're you're like a math genius. Uh and so I think I might have an idea of what uh what you're saying here, but can you define can you define it for us? \>\> Definitely. So think of a differentiable or continuous function something that it's it's okay if if you're close by. Okay. Like it's fine if I if I'm creating your image. It's fine if I get the exact image, but it's also fine, you know, if I'm missing a few things. If your shirt's a bit off, if the if the wall behind you is slightly different, if your glasses are slightly bigger, I don't have to exactly get what you're talking about. And and that's that's the quantities of that of a differentiable or a continuous function. If if you're not talking about a if you have a function that's not differentiable, not continuous, it can be it it's very critical that you get it exactly right. Now, regular CPUs are just great at doing stuff that needs to be exact. Regular CPUs in our computers, if we tell them add five and five, they will always give us 10 and they know exactly how to do these these deterministic operations. AI can do the nondeterministic operations where we don't need stuff to be exact. I totally that this this is something I've been playing with. A way that I've been formulating it is um and I'm I'm going to probably not get it exactly right, but we can find it together. There's two ways of seeing the world. One is as this sort of uh countably infinite set where there's one right answer and you can guarantee yourself that you're going to get to the right answer if you just keep going through like counting through the set. And that's the sort of classical way of of seeing the world. That's a sort of classical search function for the right answer for a um a program. Another way of seeing the world is there's this it's an uncountable infinity of different solutions. And so you're not ever guaranteed to get to the right solution, but every solution is like kind of meaningful. It's kind of right or it's clo you can get further or closer to it. Um, and both of those are really valuable, but you need to, at least my view of reality is you need to start with that kind of um, uncountable infinity where everything is full of meaning until you get to kind of like the right zone and that's when you can start to solve with the countable infinities and like count out the right solutions. I know I'm probably sounding very galaxy to people listening, but I we'll we'll unpack this in a way, but I'm curious what that makes you think of. \>\> Oh, but I I definitely agree with everything that you're saying and I think that we need both. So, we both need some problems are just a big hay stack with a single needle. You need to find that exact needle to solve it. And other hay stacks have an entire area that's full of needles and and they're not all the same needle and maybe one needle is better than the rest, but they're all an okayish needle. And then you just need to find the rough area where you have that that that bunch of needles concentrated. \>\> Now, let's take an example. So, you know, Google released uh Genie 3 recently or you know, they showed a demo of it. they they didn't productize it where they claimed okay well it's it's a world model can generate uh as you go you just you put in your keys and your mouse you move them around and it generates frames for you okay that gene gene was trying to claim well let's replace everything let's just have an AI create a completely new world for us and we don't need any of the of the old software like game engines that existed beforehand AI can just do everything what we're talking about with Mirage is the combination you'll have you have the game engine doing the deterministic Oh, I need to remember that you have exactly 71 uh gold coins in your pouch. Uh, and I need to remember that, you know, you took this pickaxe and you put it in that chest 5 years ago. And if you go back to that chest, it'll still have the same pickaxe you put it there five years ago. Have the game engine do that. But the the more creative parts where it doesn't matter if you specifically hit the exact needle, but you get something right, have the eye do that. For example, texturing it or making it look different. I think a good another good an analogy that might unpack this for people is um you need a skeleton to have a uh uh your body work and your bones define like those are the hard limits on what you can do. But you have joints and you have muscles and you have tendons that your your bones can pivot around and that allow you this infinite flexibility within this rigid system that guarantees certain things. And if you have bones, you can stand up. Um, and if you have muscles and senus and tendons, you can move around in all of these different ways that um just bones wouldn't let you do. Uh, and for a long time, we've only been able to build computer systems that that use bones. And now we have muscles and tendons that like that that allow us to uh to move around in all these different new ways, which is really cool. But you need the bones because if you're just tendons and muscles without bones, you're kind of like a jellyfish. Like it's just very hard to pin you down. You can do certain things, but it's it's hard. \>\> I love the jellyfish analogy. By by the way, if if I'm understanding you're right, you're saying that when we build Terminator, it's not just going to be metal robots, but they'll also have, you know, uh they might have skin and boat surround skin surrounding them. But yeah, it's um \>\> uh but no, I I I completely agree and and this is really what we're getting at that we're enabling this from a technical perspective. We're creating video in real time that can touch everything from gaming but also to you know this Zoom call or this is this is a Riverside call Riverside instead of having you know just you seeing uh we could be seeing each other we could be seeing something completely different um but it's it's touching the ability for people to take stuff that's in in their imagination and then add it to an existing real world. So we'll have this this core that's the same. The world's the same. physics is the same. You can take your imagination and apply it to to what you're seeing. How did you make this? You ask a really fun question. So to crack realtime video, we had to crack two things. On the one hand, we have to write lots of very optimized GPU code. We sat and wrote lots of assembly for for GPUs for Nvidia GPUs. It's called PTX. And we had the things that's below CUDA. So like lots of people know CUDA and that's like the Nvidia software stack that lets you really write good stuff for for GPUs. We had to write in PTX which is like the layer below CUDA. It's the actual assembly that gets written on the GPU. So you had to do that to actually make it very very very efficient. \>\> And did you vibe code this with Claude or how did you do it? \>\> Oh no no no. I mean Claude's amazing. I love Claude code by the way. I use Cloud Code a lot. uh get good job and anthropic but no it like it has no idea what it's doing on those levels of the stack. It's it's it's the dark abyss of AI that no one wants to touch including themselves. Now so yeah on one hand we have to write like really lots of assembly code for GPUs to get this to be efficient. You know the Mirage the current version that you saw is 40 millisecond delay. That's 0.04 04 seconds between when a frame comes in and when the model spits it out. The next version of Mirage is going to be 16 milliseconds delay 16. \>\> Wow. \>\> Okay. And to do that you really have to write very very very optimized assembly code for GPUs. Now on the other hand you have to completely build a different kind of model. So all the video models we know today are what's called birectional and we had to build an auto reggressive one. What does it mean? All the video models today, if you go to, you know, I Google VO or a bunch of these, you put in a prompt and then generate spits out like a 5-second clip and and it does a lot of processing. So, it thinks for like a minute and then you get your 5-second clip. What we had to do here is if we're generating a 5second clip or an hourong clip, doesn't matter. We're not generating the entire clip at once, but we're generating frame by frame. So, so what you're doing with what you see with mirage is it gets a live input stream and creates a live output stream. It just generates them frame by frame and not the entire video at once. That to do that it's kind of a combination between a video model and an LLM. So, you know how LLM just generate the next token? \>\> Yeah. \>\> So, it's kind of like training a video model just on next frame prediction and not next token prediction. You just have to predict the next frame each time. \>\> Okay. Okay. So, what you're saying with the autogressive stuff is um what you're doing is feeding in, let's say you're you're halfway through the video, you're feeding in the the the the frames of video that you've already generated to the model to produce the next one. I assume you're also feeding in like a source frame of the video that you want to generate. Yeah. And that's how it works. \>\> Exactly. So, you feed it in two two video streams. one is just the video stream that it needs to edit and two what it already generated and it needs to predict the next frame. \>\> And what is the um advantage of doing it that way versus just putting in uh the the two frames the last two frames uh you know the last video generated frame and the the last uh uh video input and doing that next frame versus like a bunch of frames. What is the advantage of that? \>\> Nice. So you you really need the the longer context you have, the more you know of what happened because if if for example, you know, I was wearing a black shirt in in in this stream and now I leave the camera and I come back, how do you know like to to put on the black shirt again, not to change it? So it's it's kind of like the model's memory to be able to see what was in the past. And also it's very critical for the model to to see motion because it's not just images. It's uh if if if for example you use the model to add a portal gun, it needs to be able to see me doing this action in order to be like, okay, now I need to create a portal. So it's very valuable for the model to see motion and not just static images. \>\> I see. And is it like a diffusion model or how like what kind of model is it? \>\> Yes, it's a diffusion. We call it an LSD model, a live stream diffusion model. \>\> I have no idea why you're laughing. It's it's a very technical term. Uh so these I see that these these these LSD models \>\> it's a live stream diffusion model. So it's a combination of a diffusion model with an auto reggressive transformer which which really just lets you generate predict the next frame uh every time that you're generating it. Now the hard part about this, do you remember like early LLMs like even like GPT2 2.5 days or even when GP3.5 came out they'd get stuck in loops. Do you remember those days? \>\> Yes, I do. \>\> Ever happened to you? So, so, so you know for listeners sometimes like in the early days of of LM you talk to it and it'd be great for like you know a few a few back and forths and they'll just start repeating the exact same thing. you'll just say, you know, repeat get stuck on a word or on a sentence and keep on saying that exact same thing over and over and over again. That same problem that LLBs dealt with a few years ago comes back when you try to do auto reggressive video models. What happens is the auto reggressive like the model's great for the first few seconds. Uh, Mirage, their biggest challenge was we could easily get Mirage to be great for 2 3 seconds, but then slowly starts to degrade and it gets stuck in this loop until it just gets stuck on a single color and your entire screen just becomes reds or blue or green and and solving that repetition problem was the hardest thing about Mirage. Now, if you use Mirage, it can do an infinite stream. You can keep on going with it for hours on end. Is that the um is that from the yan lacun problem with LLMs which is that errors start to compound and so the longer the more Yeah. So how did you solve that? \>\> So it's the exact same problem. It's the error accumulation problem but every frame you're generating you're going further and further from the distribution that you actually saw. You get slightly more error as as you keep on going until after like 100 200 frames you just lose everything and you're a static color. So the there's solving that took you know six months of research. We we we had an easy version of Mirage that worked six months ago just could hold for like two seconds. And so so going back you said you know your question was how do we build this? So on the one hand you have to do lots of GPU code to actually get this to run fast. On the other hand we had to solve lots of very tough research engineering problems on the model layer. It required I think at this point we ran several thousands of experiments until we really got it right. \>\> And is it is it one thing that you unlocked where it just it fixed everything or is it a bunch of little incremental things that you fixed here and there? \>\> You know what you talk about is a great it's a great point about AI. It was roughly seven or eight things that we had to unlock. And the annoying part is it's binary. If you un if you until you unlock it nothing works. And then when you unlock it, suddenly it's all better. And that's that's an annoying part about AI research that you never know how close you are. You have no idea if you're going to solve this problem next week or two weeks from now or two months from now or two years from now. \>\> But how did you know, for example, on the second thing that you added, it's still not working, but how did you know that it was better and you should keep that improvement? So, we broke this down to lots of different smaller problems. First, be able to do this just with images or be able to do this just with with certain types of conditions. As you know, for example, with Oasis was Oasis was the last thing that we launched uh that went insanely viral back in November. It could generate Minecraft. \>\> Remind people what Oasis is. \>\> Yes, it's it's a model that could generate Minecraft in real time. Went insanely viral. Everyone played with it. Was very fun for us. We just came out of stealth and that could that was a real-time video model that worked just for Minecraft. The reason that was actually an interesting step in the way towards Mirage if you try to create a video model that is great just for one single distribution. So it it just does Minecraft. It doesn't know how to do everything in the world. It is much easier for the model to learn that. So for example that that's one of the the problems that we used as as a stepping stone. we we would take larger and larger distributions. So you begin with Minecraft which is a very small distribution. It's all roughly the same and then you do okay let's do um just like street walk take like a bunch of just video of like streets uh across the world and let's be able to simulate that until you grow and you add more and more complex distributions and then each time you see like okay can I crack the next distribution? Can I actually get it to work with with this additional challenges? It's sort of like I don't know the way a a child learns to run. It's like first you you sit up and then you pull yourself up and then you like there's there's there's a bunch of steps, right? Um \>\> they have to use more and more muscles as as they uh as the challenges get get harder and harder. Yeah. \>\> Um but you know, all this takes me back to why are we even doing this? So, I'm a huge believer that there's some some entertainment, fun, creativity, experience that we're going to unlock in the next year that will completely change everyone's lives. And that that's what that's um that's what we're here for. And the way the way I like think about is this. What's the internet? What was the what is the old internet before before Chat GBT? The old internet before chat GBT 3 years ago, it used to be four things. It was knowledge. It was creativity. Like knowledge is everything Google oriented. Creativity is Facebook and Tik Tok and Instagram, but also Netflix and Roblox. Everything you do when you're not doing your homework, shopping and messaging. That's the internet. That's the That's the entire internet. There's nothing else. \>\> You're forgetting porn. \>\> Just these four categories. \>\> That's the four categories of the internet. But continue. That is an interesting point. Anyway, the internet used to be four things. Uh, not five for anyone wondering. Now, the first thing, knowledge. Chat bots have just obliterated knowledge. In the new internet, anything we used to do knowledge oriented, which whether it was search or docs or anything knowledge sharing or knowledge creating or maps, it all is going into the chatbots. Doesn't matter if OpenAI wins or Gemini wins or Xi wins. Someone's going to win that. But chatbots are replacing knowledge. What happens to creativity? Is there going to be a completely new experience that lets us be creative, lets us have fun? that's different than than what we already know so far. And and the way I like for you to get the easiest way possible is if you go up to a 12-year-old today, a random 12-year-old, either in Kansas or in Germany or in Japan, and you ask them, "What's Chad GPT?" What are they going to tell you? \>\> A chatbot. \>\> Yeah. What what what's what do they do with it? \>\> I do my homework with it. \>\> Exactly. Every 12year-old on the planet will tell you the exact same thing. Oh, Chad. Oh, yeah. That's the AI that does my homework for me. Okay, it helps me with my homework. And then you ask them a follow-up question. What AI do you use when you're not doing your homework? And they have no idea. Okay. They stare at you like, I don't use an AI when I'm not doing my homework. I'm on Fortnite or Minecraft or Tik Tok. What do you mean an AI when I'm not doing my homework? And my entire belief is AI is changing everything. It really, it really is. And it can't be that AI doesn't create completely new experiences. They just get to have when we're not doing our homework. \>\> Yeah, I I mean I definitely resonate with this. I think it's really clear that every new technology creates new content formats, creates new art forms. And we maybe have cracked some of that. You know, you can see like Sununo for example is like gesturing in that direction. But there's probably or definitely a bunch of things that are on the horizon that we haven't quite figured out yet because for example, video games used to be really hard to make. Now everyone can make a video game. Um that's a really interesting thing. What does that look like? And it's probably not going to look like traditional AAA games. Um we'll we'll evolve it into something else that has for the for the medium that we've we've just invented. \>\> Exactly. I think my question for you is how do you think about the strategy of inventing a new medium like that? So, um you're obviously starting with the foundation model. Uh an interesting question for me is for example, OpenAI started with the foundation model, but it didn't really work until they created CHBT. It didn't really work until there was a form factor that was consumer friendly. Are you all working on that? How's what is your strategy for taking this foundation model and turning it into something that people actually use to make stuff? So, right now there's no app outside, but by the time this gets uh this gets published, there will be an app um because it's coming out in the next few days. And I completely agree that this this great tech has been, you know, there's this term called AI experts that it's people who know how to use AI, know how to prompt these systems and everything. You have to turn everyone into what's called an AI expert. the a random kid on the street or, you know, a grandma living somewhere should be able to be an AI expert and just and and what we need to do is really find ways to put this in hands of people, put it in an app, put it in something that's comfortable for them to use and and let them access this great tech that's been built over the past few years in in audio and video. I think it's a combination of of what's going to create these new experiences and really put this in the hands of people. I think we'll do a lot of things. We're going to publish lots of different experiences through this app and it it remains a mystery to see which ones will take off and which ones won't. And it's kind of an experiment we're going to be running together. Like we're all going to get we're all as humanity going to understand. Okay. What what is fun about these experiences that we can do? \>\> Can you show us the app? I actually can't on this computer. It's a new computer. So, let's do this. I can explain it here. So, the the app's called Dulu, okay? It's uh \>\> very Jenzy of you. \>\> Come on. We We all have to be We all have to be a bit delusional, okay? It's it's it's what we're building. We're building the one place you go to when you're not doing your homework, which is the Lulu. Okay? Now, what you're going to get in the Lulu right now, if you go ahead and download it, you're going to be able to just upload a picture of yourself either from your camera roll or from, you know, take take a photo and and the app itself will just create hundreds of variations of it instantly. So, you can see yourself eating pizza on the Eiffel Tower on top of Mars or you can see yourself um it's it's actually all the the gender stuff is really fun when it turns a guy into a girl and then vice versa. so we can see you in a dress. Um or you can use it to express your emotions. You can be one of the things you can do is you can be angry at you know friend or something and you take a selfie and you tell the eye okay turn me make me angry and you know you'll suddenly get like devil horns or something and like smoke coming out of your ears. So what you'll be able to do is there is you can take any any photo that you have and just express yourself in a way that you just couldn't before. Right now, if you download it, you're going to get all these uh these imaging uh tools. You're also going to get over the next few days the same thing with video. And it's going to be added as a camera. So, you open up your camera and you just point it at something and you you point at your friend, you're like, "Okay, turn my friend into Elsa and suddenly your friend becomes all blonde and with a blue with a blue dress." Or you can point it at your friend and say, "Hey, um turn like add a little monkey on their shoulder." And you'll have a little monkey sitting right here. And they can high-five the monkey. and the high five the monkey high fives has been back. So all that's in the Lulu and there are going to be so many releases coming out over the next few weeks. Every week we're going to have a release for the Lulu with with new capabilities that we're adding in constantly on the video and the audio side. \>\> And is it like a social network or are you posting this to Instagram or Facebook or whatever? \>\> First stage you're going to be posting this to Instagram and whatever. But what we really found interesting is the is is is trends when when someone finds what someone here at the office playing around with the Lulu and it turned their hair into grapes. And that was actually super cool and they wanted to share that with someone else so that you could also turn your hair into grapes cuz other people in the office were like asking hey how'd you do this? Can can I do this as well? And so I think one of the key parts of that we already find that is very interesting in in this category is how do you have this place you can discover new ideas for what the AI can do with uh together with you and and try these out yourself instantly and see if you like them. See what see what they make you feel. So, I'm not I don't think it's going to be a social network in the sense that we got used to, but it's going to be something completely new. Like adding in a feed there, I'm not sure it's it's it's the thing people will want or that will want, but we will want something else. We'll want the ability to find new trends, to find new cool stuff that's happening. Uh, another thing that we found was really interesting. You know how social media today is not really social? Like it used to be all social graph and friend based. You used to be able to do things on Facebook, Instagram with your friends. Now it's all influencer based. You you see something cool is you see trends that are happening, but you don't it's they're not our friends. \>\> Thank god for my for me because that's my business model. But I'm Exactly. \>\> I I do miss those days. \>\> But there used to be days when we had, you know, like our actual friends on social media. And we found that this was actually really cool with Dulu. People actually cared what their friends thought about the the edits they were making. So I could make like a hundred different variations of myself and be really cool to see if if if my friends or my family I'd send it over to them and and they usually would like the ones that I hated. And that would be like that would trigger like a huge spark inside of you like, "Holy this is they're they're seeing the world in in a way that's different than I am." And it's it's fun to do that. \>\> So I think this is very cool. It reminds me though of some things that have already come out. So easy one is Animojis from Apple. Um there was also like a whole trend maybe like two years ago of uh you know I can't remember what it was called. It was like dream booth or something like that. Basically um stable diffusion generations of people you know head shot basically it you know it's you in a forest or whatever. So there's there have been some waves of this before. Um, and those have all kind of uh burned like gone viral and then burned out. So, what is your how do you think about that and how do you think about this as being different? \>\> So, by the way, it's it's um you know, you bring up Dream Booth. We have one of the experts on the subject, the the person who was the last author on the on the Dream Booth paper fear. He just joined us uh a month and a half ago uh in our into our SF office. He used to be at Snap at Google before that. Um, and you know, I was having this conversation with him as well. They're like, "Okay, what's uh what what's what's happening here now?" The tech has come so far over the past 2 years since Dream Booth came out that back then it used to be a great PC. Now, it actually works. Like you look at the picture like, "Oh, wow. That really is me and doing something insane." And it doesn't take any time to do that. the AI is creative enough also to to come up with new scenes. So, I think there's a there's been a huge huge jump on on the quality of these things that actually really gives you a different experience. But B, I agree that images are just a gateway to to what's enabled with video and audio. current version to Lulu. I think we've all seen the these image editing apps. And in here when we're just making it much easier for for active people to access completely for free that and that creates a different dynamic, but it's still just a precursor to to the things that are being added right now, which is video and audio. When you do this with with video and audio, when you do it in real time, when I just take out my camera and point it at something and and you know, ask the the trash can that's here next to me like to turn into an elephant and it becomes a little elephanty trash can just walks around the room. That is something completely new that we never had before. If I ask if I ask the AI to to add a little uh turtle here on my on my shoulder and the turtle starts crawling around my shoulder and like rests down here and goes to sleep. These are things we never could really do. And I think it's very interesting to see what people will do with it. And how are you how are you testing this internally? It seems like um the idea is probably kids are going to be using this a lot and so do you have alpha or beta testers who are in that age category that you're you're doing this with or how do you how do you make something that you think kids are going to love if you're not kids yourself. So, first of all, we have a bunch of alpha and beta testers. And that's the thing that, you know, just excites me the most talking to to all these people. By the way, if if you're listening in and you want to be an alpha beta tester, shoot me an email directly, dean at theart, dan atthequart.ai. Um, would love to be able to show your tech before it gets released because that's the way that we're exploring this together. I mean, we're creating great tech. We're using it for ourselves and it's it's fun for ourselves but but exper like having people experiment with it and understand together what is what this is going to enable humanity to do. We're all just building this together and and we we'd love we'd love to have everyone's input on this. I think one of the critical things from our perspective in two or three years sure the world is going to have almost all its pixels and and audio waves be AI generated or AI modified and we have to put this tech into people's hands right now so that people also build the tolerance and understand how to use this and and what this means for them in their lives. What are your AGI timelines? Ooh, I think there are only two interesting problems in the world right now. Only two. Nothing else matters. Really, nothing else matters. One is the race to AGI. And two is the race for consumer dominance. That's it for building the biggest consumer company in the world. Those are the only two races that matter. Nothing else really matters. Everything else is fun. everything else, you know, is cute and and and maybe worthwhile, but but not not not actually interesting. I think with AGI, we have to split into two. There's a Terminator stage where, you know, we get to singularity and it's smarter than all of us and and we have, you know, it's a good question, what happens to humanity then? And the second question is when do you get economic AGI? when just machines are able to do any economic job better than all of us or better than the vast majority of humanity that just makes no more sense to do any kind of virtual job virtual work anymore. The second one, the latter, it's probably a few years away. The former the the former is a few years away. The latter though, the economic AGI could easily start creeping up on us in 12 to 18 months. \>\> Wow. that will we'll slowly start seeing AI just just be economically better than a lot of us and at creating economic value and that will will have dramatic impact on society. Now, I'm a huge believer that this is going to be humanity's best best golden age, best time possible. We're we're going into a new world. Sure, AI is getting better. It's getting stronger. It's getting smarter, but humanity's future is very very vivid. And we have to we we have to understand how we how we build this and at the same time just build the best world possible for humans. Okay, let's uh let's take that one at a time. I think the the the big important thing I wanted to talk about is uh if I understand you correctly, you think that uh most economically valuable work will be able to be done by AI in the next 12 to 18 months. \>\> Yes. Another way to say that is um it will be less and less appealing to hire humans to do work and we'll be hiring AI to do work or or what is what do you think are the implications of that? What are you really saying when you say that? I think that it's very clear at this point that much of the work whether it's lawyers or or you know accountants or or lots of jobs that are very very virtual oriented a lot of that work can be done really well by AI and it's getting better and better and and there's no reason to assume that it won't continue to get better. Now it does mean that a lot of us will have to adapt just like we all adapted when the internet came out or when previous generations of technology came out. And it also means that humanity will get a lot more time to to to think to be creative and to explore what's what's in our hearts and not just what's in our minds. What what I like going back to a lot, why did democracy first start in in Greece? I I was actually very intrigued by this. Democracy first started in Greece because people had the time to think. People didn't have to work the fields 24/7. Before that, they had to be agriculture was very, very hard. You would be in the field every day for 14, 15 hours and you would just go to sleep and go back to work the next day. As technology got better and people started getting some time to be able to to take off and not work all the time, the first thing that we got was that the ancient Greeks became philosophers and they started thinking, okay, why are we even on earth? Why what are we doing here? Uh what's the right way to govern people? And they created all these systems that we still use today. And we were able to get that only because humanity had more free time on its hands. With more and more economic value coming out of AI, we will have more time on our hands to be creative and to use our imagination. And that's something I'm very excited for. That that's that's something that we have to give an avenue for people to be able to use their added their their more added time to be more creative to to actually do stuff that makes them feel something. So, there's a lot here and I want to go into the the ancient Greece stuff because I have a whole I have a whole take on that too and I think there there's a lot of overlaps actually. \>\> I would love to hear what you think on that. But yes, there's so many parts here. \>\> But I want to start with the just going back to knowledge work. You could pretty much be done by AI because I just uh I will I'll take the opposite on that. I don't think in 12 to 18 months if if you come back on the on the podcast that we'll we'll have seen that happen. And there's a couple there's a couple reasons I have, but maybe I'm wrong. Um, the first thing is I think you're right. Like let's take the lawyer example. I think that you're right that um if I think about the things that I I asked my lawyer to do um given the right prompt, the AI is going to be better and faster and cheaper than my lawyer. However, given the right prompt is like an incredibly difficult question. \>\> Yes. And um I don't think we're 12 to 18 months away from uh self-sufficient AI that's able to know what prompt to give it at itself at the right time. And I don't think that that is a uh that's just a point that we pass. I think that's very spiky depending on the domain and the context and all that kind of stuff. So the way that I think about AGI for example is um I actually think about it in terms of child development. When a child is first born, they are not independent at all from their mother. Uh, as they as they get a little bit older, you can leave them alone for a minute, 5 minutes, 10 minutes at a time until when they're 20 or 20 years old or whatever, 18, theoretically, they're adults, they can do whatever they want. Um and I think you can see uh AI following that same trajectory where if you take coding for example uh it started with tab completes it's like one step \>\> right it was just tab completes it was like and this was just two years ago \>\> exactly um uh but if you uh but now it's like cloud code is like 10 or 15 minutes but only sometimes so like you can't use cloud code on your deep down assembly stuff, right? But I think I I think uh a better definition of AGI is when it is economically profitable to leave your AI on all the time. It's always working. It's always doing something. Um however even in in that scenario which I think is pretty far away like several years away at least um because it well let me just say like I think that's a good definition because it will require a lot of things that we don't currently have. So for example continuous learning I think the only way to get AI that knows how to prompt itself in the right situation at the right time is if it's good at continuous learning. And I don't think that um better context engineering is actually is actually going to fix that. I think that's going to be maybe part of the solution, but you have to be able to update your weights um is my my belief. Um so so that's why I think I don't think it's 12 to I don't think it's 12 to 18 months away. I think it's I think it's much farther and um there it will be very very domain specific when it is totally AGI totally independent. So maybe coding is first, but that'll still be only some kinds of coding. Like it's still not going to be for your your CUDA improvements necessarily or I don't PTX improvements. \>\> Right. Right. \>\> So So that that's kind of that's my AGI take. Um what do you think? \>\> So I love this because it you said when does it become economically viable, profitable to just keep your AI running all the time? Okay. And clearly, by the way, we're at that point with computers. Keeping my just my my laptop on all the time, okay, is is something that generate like the the I didn't even think about how much electricity it costs me to keep my laptop running because I just know the value I'm getting out of it is insanely higher than the I know how much is my electrical bill for my Mac cost a month. 20 bucks maybe. Um, so it's it's definitely economically viable to keep your uh your Mac on uh 247 and we're probably going to get there with AI as well that like I any any price I I'll pay for actually running this AI will just make itself back instantly. And you gave like the cloud code example. I completely agree it's not going to replace coders. It's because I've been using I've been using cloud code over the past uh two weeks to to code some stuff for our for our app. So I've been using Flutter. I have no idea how UI works. I literally never wrote any CSS or HTML or JavaScript or any of these things. Okay. Uh I'm in the through through OMI like write assembly for GPU codes. I love that. No idea how UI works. But I was with a team working on the on the UI of the app and I was able to actually contribute a lot to that project via cloud code without knowing anything about the syntax of these languages. And what the contributions were were that I had intuition of like okay okay okay okay you got to be doing something wrong because there's no way you're storing the image here but also somewhere else and and so you must have sent it through something else in between and having the intuition of okay what's the actual challenging part you had to do of of moving the data around or whatever and just letting pod code do the actual typing that was something that personally I've been experiencing over the past two weeks and it's been really great Um, and it's been very valuable. So, I think that maybe a different way to phrase is it there's a chance that in 12 to 18 months, AI lets us be so productive that we're able to create companies that are just way bigger than anything we saw so far. And I'll give a concrete example of this. Today we have 10 trillion dollar companies in the world. Give or take depending on exactly when you check and how Tesla's stock does. But we have we have 10 companies in the world that are past a trillion dollars in market cap. Back in 2017, we had zero. We got our first one in 2018 when Apple crossed the line. And that's that's what 2018 that's seven years ago. Okay. Seven years ago, we crossed the line of first having a first trillion dollar company. Now we have 10 of these. It's a lot a lot of it is due to just technology making us that more efficient so that we could actually make a lot more money and a lot more bring a lot more value to the world. And there's I think there's a very good chance in the next 12 to 18 months AI just creates so much value in humanity that the entire stock market doubles in value. \>\> That's an interesting one. So now I want to go back to Athens because I think this this actually this dovtales with with the point you're making. Um, so I'd probably say yes. I don't think 12 to 18 months. I think it's that's way too quick. I think it it will take many many years for it to like uh even let's say you you vented this this AGI. Let's whatever our definition of AGI is, but you never want to turn the AI off. We meant it tomorrow. I think in 12 to 18 months, maybe there'll be a bubble in the stock market, but it wouldn't be like a thing that would \>\> true value \>\> maintain. Yeah. Okay. \>\> Um I think it'll take a lot longer. But I als but I do grant your point like uh it will allow us to build much bigger companies than we could we could before. I also think it will do something else um which is allow us to build many more smaller companies that accomplish much more than uh than you could as a small as a small company. Uh and and and and I want to talk about why. So, back to your Athens point, I actually don't know why democracy arose in Athens. I like I remember like Solon's laws or whatever from my history of ancient Greek class, but I can't remember why he why he did it. Um, but but the thing that I love about Athens is uh is is it's a society of general generalists, right? Um, democracy sort of requires that uh direct democracy requires that where you can be a statesman, you have to be a uh a lawyer, you are a prosecutor, you are a juror, you are a warrior, \>\> you're everything. You need to be good at everything. \>\> You have to be good at everything. And so the question is why did that stop working? And the reason it stopped working is because Athens became an empire. It became the equivalent of a trillion dollar corporation. And uh what's interesting about \>\> I love that \>\> what's interesting about being an empire is you actually need specialists because you can't send some farmer to be the general on your Sicilian expedition. Like you want someone who's got a lot of experience. Uh and so specialization requires uh sorry specialization allows for collaboration over uh across larger and larger organizations of people. Um, but then you lose the sort of the sort of generalist thing. You you incentivize uh being a specialist, which I think has basically just just continued in Western society over the last like couple thousand years. Um, and there's a lot of good things from it. But also, if you're like me and you, we love being generalists. It's fun. Um, \>\> Exactly. And I think what's interesting about AI is because you have this uh this thing in your pocket that's like a thousand specialists in your pocket. Um it allows you to uh do a lot more for longer as a generalist. You don't have to specialize as much. And so for example, if we take every, we've got 15 people, almost everybody inside of every is a generalist and is doing multiple jobs. The the like lines between jobs starts to blur. \>\> Incredible. Um, and they can do that because they have AI. And so my hope is uh maybe in a similar way to yours, we get back to uh uh we we we start moving back toward more generalists uh and this sort of like golden age of Athens type vibe because people can get more done uh uh with with AI. They they have all the specialists and they can coordinate uh across more people without having to specialize. I love this theory that we're going towards a world which is going to be more more we'll actually have space for generalists. I think it's going to happen because two reasons. One is exactly what you're saying. If we if AI gives us tools to jump into fields we're not experts in but but the AI has you know it has seen that experience and we can just through the AI actually manifest uh our our general abilities. Yes. Exactly. like you know like what I was doing this these past two weeks with cloud code and Flutter. I have no idea how to write iOS or Android apps. But I do have common sense. What I have is common sense and and I can use that common sense combined with what Claude knows about how to write code to to create this thing. Uh you know my my dad he's retired. He used to be dean of medicine at at one of the Israel universities. Um he's now now he's He's not officially at the cart, but he does he does come with us to he he he hops by the office like twice a week and and he likes being called the senior adviser for common sense. Okay? And he's like, "Look, I'm not a lawyer. I'm not a coder. I'm not a business person, okay? But I have common sense. That's what I have. That's what I can contribute to this to the situation." And so I think that's that's really huge point that they like what you're saying is yeah we'll actually be able to have generalists that can act as if they were specialists in each field. So that's that's really valuable. Something else that I think about a lot in terms of AI. So every is 15 people. The cart is roughly 70 people now. And I think you've done a tremendous job with having 15 people that almost all of them are generalists. that is typically incredibly hard to do but and and you know it it it speaks for itself because you're actually able to create insane stuff that no one else can. Also here at the cart almost everyone at the company is is a generalist and an independent thinker and they just go ahead and they do stuff without being told. Like we're we're still we're still we're 70 people and we're still completely flat. For many years I spent around two years of my life doing this. I was thinking a lot about what organizations work and what don't. Now here lots of organizations like you said like the most trillion dollar companies today just don't work. Okay. Google has one 180,000 people. They don't really work. It it doesn't we can't get the best out of them. They they they're smart people. They're talented people. But Google has a hard time getting the best out of them. But it's not just Google. It's every big tech. And and this got me thinking, what's the biggest bottleneck to actually running a company or a group of people and telling them, hey, you all have to do the same, like go build that one huge thing together. What is the bottleneck? And what do you think? \>\> Um, so I would say there's one thing is just uh there's a real information flow problem. So flowing information from the bottom to the top and from the top to the bottom which AI solves. And then the other thing that I think is is really important is humans uh have different motivations depending on the circumstances that they're in. And inside of a big company they tend to be more most interested in the thing that will get them promoted rather than like just doing the right thing. Um and that's for a lot of different reasons. Uh it's partly due to information flow, but that's that's the that's I think a big thing, \>\> right? So I have no idea how to solve the second one. \>\> Yeah. \>\> Uh and I agree that that's you know, fine. Okay. But even if we did have tons of people who who only really wanted to actually build insane stuff, I'm not sure we can fit a thousand of them in the same company and and still actually get things done because of the first uh constraint that you have, the the information flow constraint. and and you see the organizations that function incredibly well and those that don't. That's really the the key difference. So, so with an organization that's up to 100 people, sure, information will flow and it'll be fine because you can get everyone in the same room and they can all talk to each other and just move information that way. There's also really big organizations that did work really well. The ancient Roman legions. Okay, you can have a 100,000 people with swords and shields and you'll tell them, "Okay, go conquer that hill." and they'd go conquer that hill. Now, the reason that those organizations worked was because not a lot of information really needed to flow. You could just tell them, "Look, we're on this hill. They're on that hill. March and stab." Okay, that's the information that you needed to get to your 100,000 people on the field. And what I tried to do over a few years and then the the conclusion I got to was organizations that work versus ones that don't, it's how many generalists do you have inside the or how many slots do you have for generalists. If you have only 20 30 generalists in the organization, that's great. They'll think of crazy stuff and they'll go ahead and they'll uh they'll pull it off. But if you have a thousand generalists, they will get stuck on communicating with each other and understanding, wait, can I do this? Can I not do this? Is it okay that I do this? And and that's where humanity has never been able to construct an organization that gets a thousand creative people and lets them all be creative at once. So any organization you're in can only enable maybe 20 30 people to actually be creative. Whether it's a 20 30 person startup or a Roman legion fine, he had 20 creative generals and 100,000 foot soldiers who didn't have to be creative at all. \>\> Really interesting. And why is that? \>\> It's a good question why that is. But I think that when your job doesn't require creativity, it's rather easy to define. Okay, your job is to go ahead and you're working a Ford factory. Your job is to pick up this tire and put it on the cart. Okay. Um, and it doesn't matter if you do it exceptionally well or you do it average. If you do it average or exceptionally well, maybe it's 50% difference. Um, but if you have to be creative, like, okay, go build a completely new experience for people or build some breakthrough technology, that is a role that you need to be able to to take risks. You need to be able to do stuff that that's much more complicated. It's hard to define. And for people the way that's we we have information bottlenecks in that we can't communicate too much. We can't I can't take everything that's in my mind just you know telepathically give it over send it over to you. And so we have to be able to dumb things down and explain stuff in very very simple terms like okay your just job is just to do this and I can explain in three bullet points. And so when you need when something's creative it's in its early stages not that easy to define. \>\> I think um I I agree with this. I think the the way that I would describe it is in any kind of work, there's a there's at least there's two phases and they're kind of fractal within each other, but like we can just for now say there's two phases. There's the explore and then there's the exploit phase. \>\> In the explore phase, you're trying to figure out like what do I even want to do? In the exploit phase, it's like it's just an execution problem. You're just solving puzzles that you like the the frame of the puzzle is set. And um we currently use a lot of humans for the exploit thing and we may be able to use a higher a higher ratio of humans for explore now that a lot of the AI is going to be able to do exploit because AI is currently very good at problems you can define but there are surprisingly few problems you can define that are valuable um compared to all the all the problems in the world. Um, and actually I think this goes right back to this is a really good book end to the podcast because this goes right back to the the first thing we talked about which is um one needle in a in a in a in a hay stack uh or this sort of like infinite uh stack of this infinite force that you can go down and every every path you can go down is meaningful and you kind of have to uh you kind of have to explore that infinity and there's no right answer. Um and so I think we can we can in a world where AI is quite good at finding the needle in the in the in the haststack or something like that. Uh we can spend a lot more time exploring and AI is also a good tool for exploring. But it it ultimately comes down to uh those I guess those two phases of creative work maybe is a is a way to say what you're saying \>\> exactly and AI AI is going to leave the creative stuff to us. that's going to do the stuff that is is well defined and that's great and not only that because it overcomes lots of communication bottlenecks. It can actually see everything we're doing. You know it can in a thousand person organization beforehand before AI no one not a single person knew everything that was going on in a thousand person organization. No one read every email and heard every phone call. Now we actually can and so it can help us communicate better and help us each be creative and let us explore until we find uh something that's valuable to to go ahead and exploit and then can do the exploitation on its own. I think that's a great way to sum up what I think will happen over the next few years that AI will do the stuff that's more well defined and we'll leave it leave it to humans to be creative. Dean, this is a great place to leave it. This is an amazing conversation. Thank you so much. Thanks so much, Dan. This was this was amazing and we should do it more often. \>\> I would love that. \>\> Oh my gosh, folks, you absolutely positively have to smash that like button and subscribe to AI and I. Why? Because this show is the epitome of awesomeness. It's like finding a treasure chest in your backyard, but instead of gold, it's filled with pure, unadulterated knowledge bombs about Chat GPT. Every episode is a roller coaster of emotions, insights, and laughter that will leave you on the edge of your seat, craving for more. It's not just a show, it's a journey into the future with Dan Shipper as the captain of the spaceship. So, do yourself a favor, hit like, smash subscribe, and strap in for the ride of your life. And now without any further ado, let me just say Dan, I'm absolutely hopelessly in love with you.