게임 및 콘텐츠 제작에서 AI의 미래

### 게임, 코딩, 그리고 그 너머의 AI의 미래
* 🎮 구글 리서치는 신경망을 사용하여 고전 게임 둠의 실시간 버전을 개발하여 **기존의 코딩 없이도 비디오 게임을 제작**할 수 있는 혁신적인 변화를 선보였습니다.
* 🤖 신경망은 **게임을 렌더링할 뿐만 아니라 로직도 관리**하지만, 몇 가지 결함이 발견되어 이 기술의 초기 단계를 암시합니다.
* 🌍 이러한 발전은 비디오 게임과 기타 콘텐츠가 개인화되고 온디맨드 방식으로 생성되어 **사용자가 설명하는 것만으로 자신만의 게임 세계**를 만들 수 있는 미래를 제시합니다.
* 💻 커서 AI는 사용자가 자연어를 사용하여 애플리케이션을 구축하고 코딩을 대중화하며 **잠재적으로 수십억 명의 새로운 개발자**를 양성할 수 있다는 점에서 주목받고 있습니다.
* 📈 주요 벤처캐피털인 a16z는 여러 프로그래밍 언어에서 **90% 이상의 정확도로 코딩 작업을 지원하는 AI의 성장 추세**를 강조하며 **Cursor에 투자**했습니다.
* 🤖 보스턴 다이내믹스는 인상적인 신체 능력을 보여주는 **아틀라스 2 로봇**을 소개하며 이러한 휴머노이드 로봇이 가정에서 보편화될 수 있는 미래를 암시했습니다.
* ☀ Reflect Orbital은 우주 반사판의 별자리를 사용하여 태양광을 활용하는 것을 목표로 하지만, 이 개념은 아직 **이론적인 개념**으로 유익할 수도 있고 해로울 수도 있습니다.
* 📝 메타는 비디오 게임과 자율주행차에 큰 영향을 미칠 수 있는 인간 중심 비전 작업을 위해 설계된 모델 모음인 **사피엔스**를 공개했으며, AI 규제에 대한 논의는 계속 진행 중입니다.

Google의 신경망은 실시간으로 비디오 게임을 제작하여 콘텐츠 제작의 혁신적인 변화를 예고합니다.

there's been so much news already this week that I had to create a video immediately. Let's get right into it. The first one is truly mind-blowing. I have been talking about the future of video games, the future of content in general, and now we really see a glimpse into what that future can be. Google research just released a paper that details their method for creating the game Doom in real time using a neural net. That means that this game was not pre-programmed. I cannot emphasize that enough. This game was not written in traditional coding languages. It is being created on the fly by a neural net. Let's watch a little bit more of this video because I really can't even believe what I'm seeing. So Doom is a game that came out, I believe, in the '90s, one of the most popular games to port to every device possible. It's been ported even to calculators, and now we have a neural net powering it. Now what you're seeing here is truly the future of video games, the menu, health.

the logic for how the game works. The reloading, absolutely everything is being rendered in real time using a neural net, and you can actually see a little bit of issues here and there. So, obviously, the zombie guy is kind of disappearing and then coming back. Some of these numbers in the bottom right are changing without really needing to change, so there are definitely some issues. But to look at what has been accomplished really blows my mind. So, in the future, we are going to have video games delivered to an audience of one, just like we're going to have TV shows and movies delivered to an audience of one. That means that a neural net is going to be able to create any content we want simply by describing it, and it's going to be consistent. It's going to have all of the logic we need, if it's a video game, to actually play it, and we're basically going to be able to design any world that we want to play with. Now, taking this a step further into the future, a lot of people have discussed

Sora by Open AI and other neural network techniques for generating video as World simulators. And if you're not convinced that these models can be World simulators in the future, True World simulators, I don't know what's going to convince you. Interestingly, Elon Musk replied to this video saying Tesla can do something similar with real-world video. That is fascinating. I'm not exactly sure what he's referencing, but as soon as there's more information, I will report it to you. I suspect that with all of the real-world information that Tesla has collected from all of the vehicles on the road, they essentially can reproduce the real world using a neural net. So, it's not actually there, but they're reproducing it anyways. That is a world simulator if I've ever heard of one. So, the teams at Google research, Tel Aviv University, and Google Deep Mind have released a paper, released more videos. And if you want me to do a full breakdown of this, let me know in the comments. I'm probably going to do it, but

Yeah, let me know. Anyways, next, if you haven't been hearing about Cursor AI lately, it has been all over Twitter and the internet in general. People are realizing how powerful Cursor AI is and building incredible things without actually needing to know how to code. If you're not familiar with Cursor, it is basically a code editor that is AI native. You go in there, you simply describe in natural language what you want to build, and it builds it for you. You can edit the code using natural language, you can iterate on the code, you can test the code, all using natural language paired with AI. This should be very convincing that in the short and medium term, we are going to have billions of developers. And then, you know, I've said this a million times, after that, we won't need any developers. But in the meantime, it's going to be a lot of fun because everyone is going to be able to create apps whenever they want, regardless of skill set. I've seen examples where entire iOS applications have been built. I saw a video of an

8-year-old girl builds a chat application to chat with Harry Potter, and again all of this in natural language. And now a16z, one of the most prestigious VC firms in Silicon Valley, has made a huge investment in cursor, the first line. LLMs are getting good at writing code, and that is a massive understatement. All of the major AI models can now perform basic programming tasks reliably with greater than 90% accuracy. They are starting to tackle more complicated real-world tasks, and they can do it in more than 80 languages. However, writing code isn't just about writing new code. The majority of a developer's time is spent maintaining, debugging, and tweaking code, and to do that reliably requires actually understanding the code and the intent of the system. This is something that large language models to date have not been great at, maybe with the exception of the project AER AIDR, which I've reviewed on this channel. Being able to pass in a massive code base into the context window of a large language model has been a limiting

factor on the capabilities of what those large language models can do with code. But as context windows increase, as RAG solutions get better, and as we figure out more ways to map code bases and really just compress code bases using large language models to iterate on large code bases becomes more and more of a reality. We believe that, therefore, the interface between programmers and AI models will soon become one of the most important pieces of the dev stack. And we're thrilled to announce our Series A investment in Cursor, the leading company working on this problem. Cursor is a fork of VS Code. So, VS Code by Microsoft is an open-source coding editor. It's the one I use. It is extremely popular open source, as I mentioned. And now Cursor forked it and made it AI native, and it's heavily customized for AI-assisted programming. It works with all the latest LLMs and supports the full VS Code plug-in ecosystem. What makes it special is the features designed to integrate AI into developer workflows, including next action prediction, natural language edits, chatting with your codebase, and a bunch

of new ones to come. So, if you haven't used cursor, regardless of if you're a programmer or not, you need to check it out because you can build anything you want. Now, we are beginning down the path of having billions of programmers, and I cannot wait. Next, Boston Dynamics released a new video of their brand new Atlas 2, I believe it's called, their new robot, and it's pretty impressive. So, let's take a look. It's only 22 seconds, but let's see. So, we have it jumping on the ground and now doing push-ups seemingly easily. This is a beautiful form factor, much more consumer polished than their previous versions. And there we go, it gets back up, and that is a very short demo video of what this robot can do. So, very, very cool, and yeah, these humanoid robots are getting better and better, and I cannot wait to see them in every household. Next, in what looks like complete science fiction, and it's not actually artificial intelligence, but it's so cool, I have to cover it. There is a new company called Reflect Orbital.

that essentially can point the Sun or a reflection of light from the Sun at any location on Earth using a constellation of in-space reflectors. And yeah, this is absolute science fiction. So, this is a demo video, and as the community note says, in case this isn't obvious, this is a simulated video. No such satellites have been launched yet, and the practical effect of on-demand light from orbital mirrors remains unproven, but it's still cool to think about. Let's take a look at the spot. Yeah, watch this. Bring it over to us here. Isn't that crazy? No way. All right, so the demo video looks really cool. It is not real. It is just a simulated video, according to the community notes, and it could be cool for the future, but I can also see it being used as a weapon. So, this is the website Reflect Orbital Sunlight After Dark. Reflect Orbital is selling sunlight using a constellation of in-space reflectors. So, let's take a look at how they propose doing it. The sun is a huge fusion reactor that supports all life on

Earth. The Sun's light contains 24 trillion times more energy than humanity uses today. It is an unending source of energy. Reflect captures that sunlight so they have this in-space reflectors that are circling the globe and brings it down to Earth using mirrors and space powering a solar future and lighting our world. So there are a number of use cases that I can think of. Obviously, powering solar arrays is probably the most obvious one. When it's nighttime, your solar arrays are done. They're not being powered obviously because there's no sunlight. But if you can have sunlight focused on a specific position, focused on the solar array directly, those solar arrays can be running at 100% efficiency 24 hours a day. So applications are closed limited availability delivery begins Q4 2025. So they have this kind of cool demo thing on their website. And yeah, you could check it out. It's definitely far away still. And we'll see if it actually manifests into something real and practical. But it's cool to think about nonetheless. Next, meta.

presents sapiens: a foundation for human Vision models. So, we present sapiens, a family of models for four fundamental human-centric Vision tasks: 2D pose estimation, body part segmentation, depth estimation, and surface normal prediction. And so, we're seeing all of those right here: we have the pose, the segmentation, the depth, and the normal. It natively supports 1K high-resolution inference and is extremely easy to adapt for individual tasks by simply fine-tuning models pre-trained on over 300 million in-the-wild human images. So, obviously, this is good for video games, this is good for the future of content in general, this is good for autonomous vehicles. There are a number of different applications for this, and I'm so happy that Meta continues to release awesome open-source projects to the world. Next, a new letter from former OpenAI employees who have been advocating for whistleblower protections in AI on SB 1047 legislation. Now, if you're not familiar with that, that is the California legislation to essentially restrict AI. I'm not going to get too

deep into it, but it's California legislation for AI. Sam Alman, our former boss, has repeatedly called for AI regulation. Now, when actual regulation is on the table, he opposes it. So, let's read a little bit about it. So, dear Governor Nome, Senate President Pro Tempore Maguire, and Assembly Speaker Rivas. Open AI and other AI companies are racing to build artificial general intelligence. Along the way, they may create systems that pose a risk of critical harm to society, such as unprecedented cyber attacks or assisting in the creation of biological weapons. If they succeed entirely, artificial general intelligence will be the most powerful technology ever invented. And they go on to talk about being extremely concerned with the accelerating progress of artificial intelligence towards AGI and specifically why Open AI has not given the public any confidence. So, in the absence of whistleblower protections, Open AI demanded we sign away our rights to ever criticize the company. Despite touting cautious and gradual deployment practices, GPT-4 was deployed prematurely in India in direct violation.

of OpenAI's internal safety procedures. More famously, OpenAI provided technology to Bing's chatbot, which then threatened and attempted to manipulate users. They've had major security breaches, and prominent safety researchers have left the company. So, another mix-up within OpenAI, and I don't know what to think at this point. It does seem like there's a lot of churn and a lot of drama going on inside OpenAI, but I guess that's what happens when you're building the future. In a recent video, I talked about how OpenAI has shown the government GPT-5, according to an article from The Information. So, we'll see how all of this transpires, but right now, OpenAI definitely makes me a little bit nervous. They kind of seem like they're talking out of both sides of their mouth. On the one side, they're saying they're being cautious and they're all for regulation, but at the same time, they are opposing regulation, which, by the way, I don't necessarily agree with that California bill. And then

At the same time, they're also showing the federal government some of their top-secret coming projects. So, what do you think about this? Do you trust OpenAI? Let me know in the comments. Next in news that I'm very excited about: Microsoft will release the controversial Windows Recall AI search feature to testers in October. Now, I was extremely disappointed when Microsoft decided to delay AI Recall. And if you don't remember, the Recall feature basically takes snapshots of your Windows environment all the time, and you can search through everything you've done on that computer using natural language, exactly like you would do with ChatGPT, except now it has all of the context of everything you've done on the computer. Now, it is opt-in, so if you're not into it, if you think there are security and privacy risks there, you are totally welcome to just not opt into this feature. And that's why they delayed it, because I believe a few researchers found out that everything was being stored in plain text or unencrypted. Now, I believe everything was being stored locally, but that's usually not enough.

they have to encrypt it in some way to truly be secure. But apparently, they fixed a lot of these issues, and now it's coming soon. This is a future I'm very excited about. I am less concerned with the privacy and security issues, and I just can't wait to use it. And as soon as it comes out, you know I'm going to use it. Now, here's the thing: AI co-pilot plus PCs have come out. They're pretty awesome, but for the most part, a lot of the features that were promised with them have not come out yet. And this is one of them that promises to be a real game-changer, at least for me. The AI hopefully is going to be powered locally by the NPUs and the other chips in these AI PCs, and that is a great use case for on-device AI. Next, Salesforce is continuing to get into the AI game. They have released Einstein SDR and Einstein Sales Coach, two new autonomous AI sales agents to scale your sales team. So, doing sales with AI is definitely a bit scary, but if

you're a sales team, and you're looking to scale up and automate a lot of the stuff that is super manual. Then this might be a good solution. So here it shows a video of the agent being created. Then you can customize the agent, telling it when you want it to actually act and assigning it leads. This is, you know, less than $10,000, so maybe low-quality leads. And then here's an example. So we have an outbound cold email by Einstein. Simon replies, How much does your solution cost? Einstein automatically replies. So again, all of this has been done by kind of entry-level SDRs to date, but now it can be done by AI. But I think if I see an email from an AI, I'm probably going to ignore it. I already ignore all cold emails for the most part, but now definitely from AI, I'm going to ignore it. And the signal-to-noise ratio is going to be off, so I'm actually not sure how I feel about this. I've run sales teams in the past, so I think from that

perspective. I would have appreciated this, but as now on the flip side of it, as a consumer, I probably don't want a bunch of emails coming in and just a bunch of noise in my inbox much more so than I already have. All right, thanks to Sean Ralston for pointing this out. Apparently, there are two new models in lmc.org engine test and little engine test, and these are both from Google. These are both test models from Google, and he was kind enough to do the marble question, and it passed it perfectly. So, I don't know much more about these models. These might be the new models that were just released by Google. Here is Logan Kilpatrick from Google saying, Today we are rolling out three experimental models: a new smaller variant Gemini 1.5 flash AB, a stronger Gemini 1.5 pro model, and a significantly improved Gemini 1.5 flash model. You can test these all out on AI Studio. I've tested them. They still haven't performed all that well, being honest, but according

to chatbot Arena. It performs really well, basically just below Chat GPT-4 latest. I just haven't been able to get it to perform all that well, but I'm willing to give it another try definitely next. Elon Musk posted a video of their AI super cluster in action, and it looks incredible. Let's take a look. So he said video of inside of Cortex today, the giant new AI training super cluster being built at Tesla HQ in Austin to solve real-world AI. And so I guess by real-world AI, he probably is referring to Tesla autonomous vehicles, but maybe not. Maybe this is actually Grok 3. I'm not sure, but nonetheless, this is just a massive super cluster, and the energy consumption must be tremendous. Still very cool, and I'm all for another company producing Frontier models and increasing the competition. Now I just want to quickly touch on this next story, but I already covered it in depth in a previous video, which I'll link to in the description below. OpenAI shows Strawberry AI to the feds, the federal

government and uses it to develop Orion. So, the gist is strawberry qar. The thing that we've been talking about and hearing rumors about for a while now is actually being used to train their next Frontier Model, GPT-5, code-named Project Orion. The way it does that is by creating really high-quality synthetic data, and that is one of the limitations that AI companies are facing right now. There just isn't enough new high-quality training data. So, there's really two solutions to that problem. One is you do a lot more with the data you have, and that's something that Sam Altman has talked about in previous interviews. And two, you create high-quality synthetic data. So, using one model to create data for another model to train on. And we haven't really seen a Frontier Model being trained on purely synthetic or majority synthetic data to date, so it'll be interesting to see if this actually works. And as I mentioned, they showed it to the feds. So, again, as I mentioned earlier in the video, they're kind of

talking out of both sides of their mouth right now. OpenAI is they're showing this model to the federal government. They are keeping the weights private. They are keeping the code private. Are not releasing it in open source in any way. And I just don't agree with this approach. I would much rather everything be out in the open. And Jimmy Apples all the way back in November 2023 says let's conquer the cosmos mood curious Jimmy. So Jimmy already posted about Orion almost a year ago. So if you're looking for real OpenAI leaks Jimmy Apples he's the one to follow next. The company named Cerebrus Systems introduced Cerebrus Inference and they have currently taken the top spot for inference speed. Llama 3.1 70b at 450 tokens per second. 20 times faster than GPUs. 60 cents per million tokens. A fifth of the price of hyperscalers. Full 16bit precision. And they were actually able to reach 1,800 tokens per second for the AP model which

is just insane to think about. So yeah, if you want to try it out, they have their own custom chips, they're running their own inference, and it's super fast. So definitely check them out. So that's it for the news today. Let me know what you think of all the stories in the comments below. If you like this video, please consider giving a like and subscribe, and I'll see you in the next one.

Doom Powered Entirely by AI, Cursor AI, Meta SAPIEN, OpenAI Drama, Project Orion

Artificial Intelligence (AI), Open Source, Generative Art, AI Art, Futurism, ChatGPT, Large Language Models (LLM), Machine Learning, Technology, Coding, Tutorials, AI News, and more

** Exclusive Pinecone Partner **

Matthew Berman

비디오 게임과 콘텐츠 제작의 미래

Google의 연구는 신경망이 기존의 코딩 방식에 도전하여 게임 둠을 실시간으로 제작할 수 있는 능력을 입증했습니다.

이 기술은 비디오 게임과 기타 콘텐츠를 필요에 따라 개인에게 제공할 수 있는 미래를 암시하며 잠재적으로 콘텐츠 제작에 혁명을 일으킬 수 있습니다.

또한, 사용자가 자연어를 사용하여 코드를 작성하고 편집할 수 있는 AI 네이티브 코드 편집 기능으로 새로운 개발자와 앱 제작이 급증할 수 있는 Cursor AI가 주목받고 있습니다.

프로그래밍 및 혁신 기술에 대한 AI의 영향

8세 소녀가 해리포터와 대화할 수 있는 채팅 애플리케이션을 개발하여 AI 지원 프로그래밍에 중점을 둔 회사 Cursor에 상당한 투자를 이끌어냈습니다.

현재 주요 AI 모델은 코딩 작업을 안정적으로 수행하고 있으며 코드를 이해하고 유지 관리하는 능력을 향상시키고 있습니다.

VS Code의 포크인 Cursor는 개발자 워크플로에 AI 통합을 맞춤화하고 있으며, Reflect Orbital은 아직 시뮬레이션 단계에 있지만 우주 반사경을 사용하여 온디맨드 햇빛을 제공할 것을 제안하고 있습니다.

Reflect는 우주 반사경을 사용하여 태양 에너지를 포착하고 지구로 전달하여 연중무휴 24시간 태양열 어레이에 전력을 공급하는 것을 목표로 하고 있습니다.

Meta는 비디오 게임, 콘텐츠 제작, 자율 주행 차량에 응용할 수 있는 인간 중심 비전 작업용 모델 제품군인 Sapiens를 출시했습니다.

전 OpenAI 직원들은 인공 지능 개발 경쟁과 회사의 투명성 부족에 대한 우려를 표명하며 AI 내부 고발자 보호를 옹호하고 있습니다.

OpenAI의 내부 드라마와 마이크로소프트의 AI 리콜 기능

OpenAI는 대규모 보안 침해와 내부 혼란을 겪으면서 안전 절차에 대한 우려와 규제에 대한 상반된 메시지가 제기되었습니다.

이 회사는 연방 정부에 일급 기밀 프로젝트를 공개하여 불확실성을 더했습니다.

한편, Microsoft는 자연어 검색을 위해 사용자 환경의 스냅샷을 찍어 이전의 보안 문제를 해결하는 논란의 여지가 있는 Windows 리콜 AI 검색 기능을 출시할 예정입니다.

영업 분야의 AI 자동화: 양날의 검

콜드 이메일에 대한 응답과 같은 수동 영업 업무를 자동화하는 데 1만 달러 미만의 비용으로 AI가 사용되고 있습니다.

이렇게 하면 리드 할당과 후속 조치를 간소화할 수 있지만, 필자는 AI가 생성한 이메일의 효과에 대해 회의적이며, 받은 편지함의 소음을 가중시킬 수 있다고 우려합니다.

Google은 새로운 실험용 AI 모델을 출시했으며, Elon Musk는 자율주행차 개발을 위해 Tesla에 대규모 AI 훈련 슈퍼클러스터를 구축하고 있습니다.

OpenAI의 프로젝트 오리온과 Cerebrus Systems의 추론

OpenAI는 고품질 훈련 데이터의 부족에 대응하기 위해 합성 데이터를 사용하여 차세대 프론티어 모델인 GPT-5(코드명 Project Orion)를 훈련하고 있습니다.

이 회사의 접근 방식은 모델의 가중치와 코드를 비공개로 유지하여 오픈 소스가 아니라는 비판을 불러일으켰습니다.

한편, Cerebrus Systems는 초당 최대 1,800개의 토큰을 처리하는 맞춤형 칩을 통해 속도와 비용 면에서 선두를 달리는 빠른 추론 솔루션을 선보였습니다.