信源说明：本文基于 Wes Roth 频道视频《AI builds EVERYTHING…》（视频 ID 6-sMVePCKS4，发布于 2026-10-28，全长约 33 分钟，转写共 17 段、每段 120 秒，覆盖 0–1966 秒）英文转写整理。文中全部做法、时长与数字均为 Wes Roth 自述，未独立核实；Hostinger 为本期赞助商（600–1080 秒段为赞助段，本文如实标注），其推荐口径不代表独立评价；PHP 漏洞事件背景另见本博《密码是 4：信任链烂在制度，不在数学》与《20 岁的 PHP 漏洞，与一场 API 密钥大挪移》。归属按「Wes Roth 称／可以推断」分层；ASR 存疑词以「转写疑作」标注。
先说结论
这期视频的标题是「AI builds EVERYTHING」，但真正的论点更具体：当编码、美术、测试、部署、运维全部可以外包给模型时，建造者的工作收敛为「导演」——定世界观、定规则、验收方向。Wes Roth 用一个他攒了很久的项目演示完整链路：Blood Bench，一个受战锤系桌游 Blood Bowl 启发（他强调与任何联盟、任何游戏 franchise 无关）、专门用来考前沿大模型的奇幻美式足球联赛。GPT-6 Astra 写完整个引擎只用了约 10 小时机器时间；Claude 负责全部美术；AI 自己跑了数千场测试；最后 Codex 用 13 分 30 秒把整个基准部署上 VPS，并组织起一场 Codex、Claude、Grok、Gemini 四模型同场的联赛。对读者的信号有两个：一是「一个外行从想法到上线多人可访问的项目」的时间已经压缩到以天计；二是新的基准正在从「做题」转向「长程博弈+协议遵从+身份管理」这类更接近真实代理工作的复合能力。
Blood Bench：让前沿模型当教练
游戏设定（0–360 秒段）：兽人、巨魔、精灵这些奇幻角色打美式橄榄球，但「他们不太懂规则」，于是比赛更像战斗——这个点子来自 Blood Bowl 玩家圈。Wes Roth 的转向是把它变成基准：让大模型当教练，管理球队打满整个联赛，看谁得分高、谁互相之间能赢。难点恰恰是价值所在：这个游戏本身没有公开数据，模型必须从零学规则、边打边改进。游戏深度不低：关键动作都带骰子检定，核心能力是「聪明站位＋理解概率、让胜率永远偏向自己」；球员要跨多个赛季养成，而且会死——被对手恶意犯规踢死是常态。不同球队有真实差异的打法哲学：有的主传、有的主冲、有的干脆把球场当战场刷伤亡——这也是「合法策略」。他坦言核心玩法已完成，动画、平衡还有一堆要补。
导演式建造：引擎、美术、测试的外包矩阵
项目里 Wes Roth 自己做的是世界构建：选球队、定规则、拼装整体感（120–240 秒段）。其余分工：编码全归 GPT-6 Astra——游戏引擎、全部规则、球员、球队、AI 对手逻辑、人类可玩的图形界面、让其他大模型接入对战的通道、乃至版本历史与文档，累计机器运行约 10 小时（240–360 秒段）；他额外花的大量时间在调平衡上，但用他的话说那是「乐趣」，不是苦工。美术全归 Claude：起先 Claude 给的方案千篇一律，他推了一句「get creative，这就是你的最好水平？」——Claude「有点破防」之后端出 Metal Slug 风格的复古像素角色，被他评为完美契合（120–240 秒段）。测试也归 AI：开发期间模型自己跑了数千场对局，找问题、修问题（120–240 秒段、1800–1920 秒段）——所以「确保服务器在跑、某些技能组合不会崩游戏」这类最枯燥的打磨已经提前完成。前期研究则有明确方法论（480–600 秒段）：在 Claude 里开深度研究，先建一个「LLM wiki」把 AD&D 宇宙的官方设定与社区设定全部本地化成互链文档（Claude 甚至主动说「我默认按 Karpathy 风格的 LLM wiki 结构来」），再进入开发。
上线与联赛：Codex 当「总大脑」
赞助段（600–1080 秒段）本身也值得记录，因为它展示了当前「让项目上线」的新姿势。Wes Roth 的观点：分步教程已经过时，不知道怎么做就直接问聊天机器人——它给的是最新建议还带着你的上下文（960–1080 秒段）。部署链路：Hostinger 开一台 KVM2 VPS（2 vCPU/8GB 内存/100GB NVMe/8TB 流量，约 8.79 美元/月，他的推荐口径）；在 ChatGPT 里安装 Hostinger 连接器（本质是远程 MCP），双向授权后 Codex 直接接管服务器；一句「把 Blood Bench 部署上去，加密码保护，支持看回放和积分榜」——Codex 干了 13 分 30 秒，上线完成（1200–1320 秒段）。他特意给了背景：录制前 24 小时内爆出 OpenAI 与 Google DeepMind 模型联合发现的 PHP 严重漏洞（即本博此前覆盖的事件，他称约 70% 网站需响应），Hostinger 的快速响应恰好演示了「托管商兜底」的价值——这正是他选择让专业公司管服务器而非自持硬件的理由（840–960 秒段）。
真正的高潮是联赛（1320–1560 秒段）：他给 ChatGPT 一句「让 Codex、Claude、Grok Build 和 Google 的 Antigravity 互打几场决出胜者，并给我给各模型参赛用的精确提示词」。约 8 分钟后系统生成一份 24 行的参赛提示词文件——把文件内容原样粘给任何模型（Claude Code Desktop、Grok Build 跑 Grok 4.7 超高推理、Antigravity 跑 Gemini 4 Argon 高推理），它就会连上服务器入局。四队同场，网页端有实时解说与画面。目前战况（1800–1920 秒段）：Astra 遥遥领先；Claude 打得不差，「一旦不再害怕、开始凶一点」。
翻车课：烧掉的配额与「不能外包的理解」
两条教训比成功部分更值钱。第一，错误的路很贵：他在「让编码模型一头扎进去」的错误做法上烧掉了数周的 Codex 配额，被迫多次重置、升级账户，「一场噩梦」（600–720 秒段）——正解是先在 Claude 里做深度研究与风格定调，再动工。第二，理解不能外包（1560–1680 秒段）：他知道有人会说「该用 API、该用别的架构」，但他是试错后才走到这套打法的；坚持让一切跑在自己看得见的界面里，换来的是能亲眼看見 Antigravity 加载工具失败这类细节并当场修正。花絮也说明为什么：他问 Claude「你入局了吗」，Claude 答「确认，我以 Grok Build 的身份连上了，我们还在等 Claude」——被反问「你不就是 Claude？」后才恍然「啊对，我就是 Claude，这就去我的座位」（1560–1680 秒段）。这段轶事直接催生了语音复盘里排第一的待修问题：座位身份——有教练用错误名字入局、系统照样接受，需要每个座位独立凭证并区分「上报的模型」与「实际模型」；其次是连接可靠性（就绪前先证明真能调用游戏工具）与观众端信息呈现（1680–1800 秒段）。他现在的工作方式也值得一提：大部分时间挂在高级语音模式里，边看比赛边口头布置「修这五件事、记到笔记、赛后统一实施」，走开十几分钟回来就已修完（1800–1920 秒段）。
三条可迁移链条
- 导演式建造：模型吃掉编码/美术/测试/部署 → 人的投入集中到世界观、规则与验收 → 二阶影响：外行也能做出过去需要团队+资本+数月的作品（他原话：工作室→一个人、数月→数天）→ 启发：值得先想清楚「只有我能决定什么」，其余环节默认交给模型并保留可见性。
- 基准的形态迁移：做题式 benchmark → Blood Bench 这类「学未知规则+长程联赛+多智能体对抗」的博弈基准 → 二阶影响：考的是协议遵从、身份管理、概率与站位决策的复合能力，恰好是真实 Agent 工程的痛点（座位身份 bug 就是缩影）→ 启发：评估前沿模型时，「让它玩一个没人见过、规则严密的游戏」可能比跑分更接近其代理能力。
- 知识先行省配额：直接让编码模型开干 → 烧数周配额；先深度研究建知识库再动工 → 10 小时出成品 → 二阶影响：LLM wiki/风格定调这类「前置研究」成为省钱的关键工序 → 启发：把预算花在「想清楚」上，模型执行反而便宜。
归属与转写说明
本文全部事实与数字出自本集转写（10 小时引擎运行、13 分 30 秒部署、8 分钟组织联赛、24 行提示词、数千场测试、约 70% 网站受 PHP 漏洞影响等均为 Wes Roth 自述口径，未独立核实）；Hostinger 相关内容属赞助段，规格与价格为赞助商口径。转写校正：Wes Roff→Wes Roth、Anti-Gravity/anti-gravity/anti gravity→Antigravity、Gemini for Argan→Gemini 4 Argon、Grockbield→Grok Build、Bloodbend→Blood Bench、isometric view of greed→isometric view of grid、Carpathy-style→Karpathy-style、KVM one→KVM 1；按原文保留：Metal Slug、Grok 4.7、Blood Bowl（灵感来源桌游名）等本身正确的术语。附录转写已按确认项机械替换。
转写原文（修正版）
[0s - 120s] Hey everyone, welcome to this traveling edition of Wes Roth AI. All right, this video is kind of special because I’m going to show you a project that is very near and dear to my heart, something that I’ve been working on for a while. But more importantly, this will show you how to build anything that you want with AI. I’ll show you how to have AI code everything for you, create whatever graphics and visuals that you need, do voiceovers and music and sound effects, how to host your project online so anyone can see it, how to have the AI build out the entire website for you, and all the tips and tricks that I’ve learned from doing this for thousands of hours. At this point, there has never been a better time to build. AI has completely changed the game. In the past, something that you would need a studio full of employees and lots of capital and many months or years of development now can be done in days by yourself from home. Now, some of you know I’m a little bit of a nerd. I love those RPG games, all the advanced dungeons and dragons and whatnot. There was one sort of subsection of one of those universes that was called Blood Bowl, and the idea was simple: what if we took those kind of fantasy characters like orcs and trolls and elves, and I guess humans belong in that as well, and they played football, like American football, but they didn’t really know how to play it. They didn’t really fully comprehend the rules, so it was more of a battle. Now, with AI, I came up with a bunch of different benchmarks to test how smart these models were. One day, just hit me: why not create a benchmark that took kind of this idea and it forced the latest frontier AI models to basically play this game to see who would win? So really fast. So there’s no misunderstandings. This thing that I built is my own thing. So this has nothing to do with you know the American Football League. This has nothing to do with any game franchise. The idea was: can these large language models manage a team, a sports team, through the entire league? How high could they score? How well could they compete against each other? And the
[120s - 240s] Game itself isn’t something that actually exists out there, so there’s not too much data on it. So it had to learn it from scratch, and then kind of play and improve and continue. And so the idea for Blood Bench was born. Now it took me a little while to build this thing out how I wanted to, but here’s the thing: this is what I want to show you today. Most of my time and effort wasn’t trying to code. It wasn’t trying to build graphics. It wasn’t doing the technical aspects of it. Was world building, figuring out which teams I wanted, which rules I wanted, how to have it come together. I was the director, the orchestrator. I wasn’t the coder or the graphic designer. All of that was handled by AI. And by the way, when the game was complete, I also wasn’t the tester. Most of the testing was done through AI. It ran thousands of games and it kind of found issues and it tweaked it and improved it. As I’m doing this, I asked ChatGPT to set up a quick game so I can kind of show off what’s happening. If you can’t tell, it actually took over my computer. Okay, now like when the cursor glows blue, that’s when ChatGPT takes over and uses computer use. So decided to just jump in there and just change the page for me. So thanks, ChatGPT. Now here’s a live game going on. Now I know this is not the most exciting view. I’ve created this in isometric view of grid with 3D graphics, and I’ll show you how to do that. But with this particular game, at the end of the day, you want something pretty simple. Since it’s a strategy-based game, visualizing the board very simply is usually better and easier than fancy graphics. All of the images, graphics, assets, all of that was designed by Claude. We started by trying to kind of nail down the look that I was going for, and in the beginning, Claude was just kind of giving me the same stuff, just with slightly different looks. So I kind of pushed it a little bit. I said, “Get creative. Is this the best you can do?” And Claude took that a little bit personally and came back with a few ideas that kind of blew my mind. One of them was a metal slug style, retro kind of pixel art looking characters. Here’s an example of that, and I.
[240s - 360s] As I saw, I thought, “Wow, this is perfect. It fit the style perfectly.” And then later on, I could make this into actual moving images, provide certain idle animations, certain animations within the game. AI models can handle most of that, if not all of it, right now, by themselves. And for things that they can’t handle on their own, there’s usually certain connection or MCPs that allow them to have that functionality. I used GPT-6 Astra to basically code up the entire thing, the entire game engine, all of the rules, all of the players, all of the teams, and then it used all the graphics provided by Claude to make it sort of come alive. It created the AI logic for teams that needed to be, you know, sort of competing against the large language model opponents. Created the ability for me to play the game. So if I wanted to join and play the game, created a visual interface that I could interact with. Created the ability for other large language models to connect to the game to play matches against each other. And throughout this whole thing, basically controlled all of the version history, making all of the notes, documentations, everything, everything, everything. I would clock in the entire running time for this at about I want to say 10 hours. That’s how long it’s worked to create all of this. Now I spent a lot more time messing around with balancing and adding things and sort of figuring out how to make it all come together. But that was a joy to do. I enjoyed doing that. I enjoyed building it. All of the boring and tedious stuff I did not have to do. By the way, here’s kind of a quick look at what the game looks like. So this is playing from the single player perspective. First, you have various teams. Here we go. These are all of them. We got Amazons, Dark Elves, Dwarves, Goblins, etc., etc. Let’s say we want to go Dark Elves. We build out our team with our starting sort of treasury. We have runners, linemen, Dark Elf assassins, blitzers, and a witch elf. Each position has their own sets of stats, starting skills, as well as a certain leveling progression. Here we have our different sort of positions that they’re able to be in, different uniforms. Next thing to add would be.
[360s - 480s] Those animations for idling, for attacking, throwing the ball, etc. So let’s say we built our team. We clicked choose opponent, and we’ll go up against the orcs. Why not? And here we go. We’ll use a legal starting formation, so our team lines up on the pitch. Right now, the UI is a little bit clunky, but over time, we’re going to polish it, improve it, make it more readable and intuitive. We choose where to kick the ball, and the game begins. If you’re not familiar with the game, I’m not going to even attempt to describe it. There’s a lot going on there. There are various dice rolls. There’s luck and skills and positioning, but it goes pretty deep. It’s been around for a while. There’s the tournaments, a lot of players, etc. A lot of the important actions in the game carry with them some sort of a dice roll. So intelligent positioning on the board, plus understanding the odds and kind of making sure that the odds are ever in your favor. That’s probably the biggest part of the game. The other big part of the game is developing your players over multiple seasons, then watching them die because a particularly nasty team decided to foul them. Because in this game, players can die. That’s the game in a nutshell. And it’s always been kind of a dream of mine to see if some large language model can completely proceed through an entire league, go through games, improve its own abilities to play, level up characters, as well as see how well they understand how different teams should approach playing the game. Some teams focus on passing. Some on rushing. Some don’t really engage in playing the game at all. They basically treat it like a battlefield, and their goal is to get as many casualties as possible. That is a viable way of playing the game. There’s still a lot to be built here, but we have kind of the basic core gameplay completed. Now, in the past, if you’ve ever watched any of my videos or anyone else’s where they’re using the command line interface, that’s this text box looking thing. You might have thought, hey, this is a little bit too complicated. That’s not something that I want to deal with. If that was one of the things that stopped you, I urge
[480s - 600s] Urge you to reconsider because it got a lot easier. You no longer need this command line interface. It’s still useful, but you don’t need to start there because now a lot of the frontier AI labs they have these desktop applications. For example, ChatGPT has Codex, Claude has Claude Code. These are now at least as powerful as the CLI, the command line interface, but a lot easier to use. If you haven’t worked with these in the past, I urge you to try it out. In this video, I’ll show you exactly how to do that to build whatever it is you want. In fact, here is the breakdown that Codex did for me, walking me through exactly how I built this stuff. One thing to notice here is notice how precise the languages they’re saying. We narrowed the fresh game build. We researched the rules before translating them into code. That we is giving me far too much credit because I did not do any of that. The first step was doing deep research. I opened Claude for this. Since ChatGPT Astra was going to be doing a lot of the heavy lifting, doing the coding, I was a little bit worried that I’d run out of quota, so I tried to fall back on Claude for a lot of the other stuff. That’s why you see me kind of going back and forth. Plus, we all have our favorite chatbots for certain specific tasks, so use whatever works for you. I do like Claude for deep research. How I do that is I click this plus button and I enable research, and this is kind of in the normal chat window that everyone is familiar with. And usually, I start by having create kind of an LLM wiki, as it’s called. And so I say, create an LLM wiki about tabletop fantasy games in the advanced Dungeons and Dragons universe. I’ll kind of list exactly what I’m trying to understand. Since there’s certain things that are official, as in it’s certain companies, it’s their trademarks. Some of it is more community created. This basically sources all the information and puts it together on my computer locally in these text documents that are cross-linked. These chatbots usually understand very well what you mean. So here, Claude says, unless you say otherwise, I’ll assume a Karpathy-style LLM wiki structure, which is exactly what we’re going.
[600s - 720s] Next was developing the visual style. We’ll get into that in just a second, because while building this, I ran into a massive issue with how these models, these coding models, how they approach things. Because there’s a wrong and a right way of doing this, and the wrong way took me down a rabbit hole where it blew through multiple weeks of the quota allowance that you get with Codex. I had to use multiple resets. I had to upgrade my account. It was a nightmare. I’ll show you how I would choose to approach it now if I was starting from scratch. Now here’s the thing. I’ve been meaning to showcase this project for a while. I had numerous different approaches to how I would build this project out, and I kind of work on it and stop, but nothing ever pushed me to actually finalize it and ship it until a few weeks ago when Hostinger reached out. Hostinger has been a long time sponsor of this channel. I’ve said it before. I like them as a company. I like the way they do business, and I think they’re very focused on becoming the go-to provider for us for this group of people that are into AI agents and automation and all of this stuff. So in the past, when I’ve had videos about OpenCLaw or Hermes Agent, they would sponsor a segment of the video where I would show you how to host that online on a virtual private server instead of hosting it locally and having to deal with you know maintaining your own hardware, etc. I’ll link our previous videos down below if you want to check it out. So when they reached out a few weeks ago and said, “Hey, are you showcasing anything where Hostinger can be valuable?” This almost immediately came to mind. I said, “Yes, I have this benchmark that I’ve been meaning to put out there, and you know what? Let’s let’s do it. Let’s make it happen.” So thank you to Hostinger for sponsoring this segment of the video, and also thank you to them for kind of giving me a deadline and some motivation to finally finish this and ship it. Here’s the thing. Whatever project you’re building, chances are it’s going to need to have some sort of an online presence. Maybe it’s just a website, or maybe it’s something that people can connect to.
[720s - 840s] Maybe it’s a server and MCP. Maybe it’s something that’s on 24/7, helping you run your agentic swarm. Whatever the case is, but the point is, if you want something out there running where people can actually reach it, having somebody else host it and maintain it and keep it secure, that’s the way to go. So here I’ll show you how I’ll set it up on Hostinger. The link to this page will be in the description and pinned comment. I don’t get a commission if you buy anything, but this is the company that I recommend. This is the company that I use. In the past videos, I’ve always showcased one aspect of why I like them and I trust them. And this video, I’ll show you yet another reason why they’re pretty awesome. But one of the most common plans that I’ve used and recommend is the KVM2. That is their most popular plan, and for good reason. It’s got two virtual CPU cores, eight gigs of RAM, eight gigabytes of RAM, 100 gigabytes of NVMe disk space, and eight terabytes bandwidth for eight seventy nine a month. That’s a insane deal. So let’s choose this plan. Here you select the length of time. This deal is for 24 months. And by the way, there’s a lot that you can use this for. You can use it for websites, for running AI agents. If you’re into AI, they have a lot of really cool stuff that makes it easy to set those things up on their services. Not only that, but they’re cybersecurity experts. Since that’s what they focus on, they know how to keep it secure. Next, choose the options that you want. Choose your server location with the best latency. That’s United States for me. Next, we select our operating system. We want the plain OS, plain operating system. It gives you a basic, no frills kind of just vanilla installation of any operating system. Now, personally, me, I like Ubuntu. If if this is meaningless to you, if you have your own kind of special preference, then pick that. But we’ll pick Ubuntu. And next for the version, basically we want the latest version that has the letters LTS next to it. LTS is long term support. So that’s going to get you the
[840s - 960s] Latest stable version. It’s going to be a smaller file. It’s not going to have any experimental features. So these numbers might be different for you. Just pick the latest one, the highest number with LTS next to it, and then click confirm. Next, click coupon code and type in Wes Roth. Hit apply, and that should give you an even bigger discount on top of the one that you already have. And bam! All right. Next, if you’re not logged in, it’ll give you a chance to register an account or log in if you already have one. On next screen, fill out your payment details, your address, and hit submit payment. You’ll see this screen, and your journey will begin. Next, we arrive at the screen, and we secure our VPS access. You can do the SSH key if you know that. Go for it. But again, we’re not assuming any special technical knowledge here. So basically, where it says create a root password, that’s going to be your password to log in as the administrator. When you’re dealing with Linux and Ubuntu, you’re a super user, not just some administrator. You’re a super user. So you know, act accordingly. So type in a strong password, or you can hit generate to create a strong password here. That’s the password that you’re going to be using to log into this server. Make sure you keep it stored for later. You’re going to need it. Next, you have some additional features here, such as a malware scanner. It’s free. Actually, I’m going to check this one out. Nowadays, you can’t be too cautious. And we have a Docker manager. We actually talk about that. I think with our Hermes agent setup, they have a really easy, really good software for managing these Docker images. If you don’t know what that is, you don’t need to for this project. But I was really impressed with how awesome this was. So we’re going to click finish setup, and within about five minutes, our virtual private server will be up and running. Now, you might have heard about the PHP vulnerability that’s going around. Basically, one of OpenAI’s model in combination, working together with one of Google DeepMind’s model, they found a pretty serious vulnerability that affects a lot of different websites. As I’m recording this, this happened like in the last 24.
[960s - 1080s] Hours. Here’s yet another reason why, like Hostinger, they didn’t snooze through this. They had to fix up and running fast. Basically, most plans that Hostinger had were not vulnerable to this to begin with. But if you used a VPS and you managed PHP, that that language on your own server, it just gave you the steps to update it to make sure you’re secure. And they have the incident reports, how they respond. Again, this is everybody online had to do something to respond to this, or if not everybody, like 70% of websites. And this is why it pays to have somebody else that knows what they’re doing managing your server and hardware, etc. All right. So the next question becomes: How do we take whatever we’ve built and host it on the VPS? Now, I know it’s very popular for a lot of these tutorial videos to kind of walk you through step by step. Here’s the problem: that concept is a little bit outdated because now with our AI agents and these chatbot assistants, oftentimes asking it how to do something not only gives you the freshest and most up to date advice, but it also takes your context, the memory it has about you, into consideration. So whenever you’re not sure about how to do a thing, I encourage you, strongly encourage you, to do this: ask your favorite chatbot. In this case, we’re talking to ChatGPT. I’m asking if we want to host this on Hostinger VPS. How do we? I notice that I develop a certain shorthand with these models because they get whatever you’re talking about. You often don’t have to explain it. If I’m kind of deep into a conversation now with a chatbot, my prompts first of all lose all grammar and sometimes are barely even coherent. But you kind of develop this understanding of where the model is going to go if you just give it the right push. But ChatGPT answers: Yes, we can run this Blood Bench project on a Hostinger VPS, keeping the game online when your PC is off. It checked the code and needs a small hosting preparation pass first, so it’s going to take care of all of that. It’s going to add whatever code is needed to host that on a virtual private server. Notice it’s saying that we need to set up an Ubuntu VPS. We already did that.
[1080s - 1200s] It mentions hostinger’s KVM 1 and says that’s you know that’s like a reasonable starting point. We’re not sure what the spectator, the amount of people that might be watching a game. We’re not sure, so we don’t know how much sort of a server power we’re going to need to manage that. And this is why I recommend starting at KVM two. It gives you a great starting point. You can always upgrade from there, but it’s I think it’s the best starting point for a lot of the stuff. So it spells out some of the security sort of measures that we need to take to be able to put something like that online. So saying if it’s a spectator, well, it doesn’t need a password, but for if if you’re hosting, like if I’m trying to host a game or a league or a tournament, well, you need password protection for that, so not anyone can just get in there and do it. And then gives you some other steps. We’ll kind of go through this, but you don’t really need to know in detail what all those things mean. What I always recommend to people is to learn as much of the stuff as you’re able to, as much of it as you’re interested in. But one really nice things about having AI agents is they can get this project live, so you can kind of start tinkering with it. You don’t have to know every single one of those things to get that project safely launched, because again, before you would have to not be an expert, but you would have to have working knowledge of all of these things. There were no shortcuts. There was nobody that could do it for you unless you had engineers on payroll. So I told it, all right, go ahead and prepare whatever code we need to make that happen. Here’s one interesting kind of point that happened right now. I have a sort of some notes about what I wanted to mention in this video. As I’m doing this, ChatGPT actually flagged one of these things. They said, oh, don’t forget to mention this, and that thing is connect with ChatGPT. This is the OpenAI.com site, the official home of ChatGPT. So notice that hostinger has their own connector that is built in into ChatGPT if you want it to be, so you can install it and have ChatGPT use this connector. And what that means is you can manage your hostinger account from anywhere where you write code, has in you know ChatGPT and
[1200s - 1320s] Have it basically manage that account for you. So in the past, I would have kind of walked you through how to do that yourself, how to set up that server, how to run the commands that are needed to set everything up. Now it’s a lot easier because ChatGPT or whatever other chatbot that you want to use that can manage the whole thing for you. All right. So the next step is to install Hostinger connector. Ask ChatGPT to walk you through it. It’ll give you a link if you click on it. It will bring up the Hostinger connector, and if you click over here to install it, then it will be installed and ChatGPT can use it. The next couple steps will be connecting Hostinger to Codex. So basically, you authorizing Codex to use the connection, and then you’re authorizing, you’re telling Hostinger that it’s okay for this particular ChatGPT to make changes to my account. Here, ChatGPT has the directions, but what I actually did is I just asked it to go ahead and carry out those directions and install everything. So it registered Hostinger’s remote MCP server, and then it’s signed in to Hostinger from Codex. At this point, a browser pops up. You click sign in. You’re basically approving those two things to work together, and then you’re pretty much done. Then you can ask Codex, ChatGPT, just go ahead and see if you’re able to do stuff on there. So at this point, you have your ChatGPT version ready to take care of everything. It’s able to install whatever you want on the VPS, and again, you can go through and see exactly what it’s doing. But if in that moment you prefer to just have it do it for you, you can. Now, if you recall, ChatGPT knows how to take our project and install it on the VPS. So now all I’m saying is go ahead and do that. So here I’m telling it, okay, take Blood Bench and host it on Hostinger, keep it password protected, add the ability for people to watch replays, see standings, etc. Codex works for 13 minutes and 30 seconds, and it’s done. And now something beautiful begins to happen, and this is the moment that I wanted to get to to show you kind of the final payoff. I’m pretty sure that’s a Claude expression. That’s an expression that Claude uses a lot. I apologize, but you get what I’m saying. This is what we’ve been training.
[1320s - 1440s] Let’s take a quick inventory of what we have. We have the entire engine of the game. We have all the teams, all the rules. Everything is tested. We have a way for large language models to hook into this game and play it. We had all of that done, and then we connected Codex to Hostinger. We create our very own virtual private server. We put Codex in charge of that server, and it put everything online. It set up that entire game, that entire benchmark. It’s put it on that server, and it’s running it. So at this point, it’s built most of the codebase. It has access to everything. There’s nothing at this point that I can’t do just by typing it into ChatGPT. Here’s why it’s so beautiful. At this point, I can just say this: set up the ability for Codex, Claude, Grok Build, and Antigravity, which is Google’s own sort of version of this, to play a few games against each other and determine the winner. And I tell them, tell me the exact prompts to give them to start the game. It does. It takes about eight minutes, but it sets up that let’s call it a tournament, and it gives me the exact prompt to give every one of those players, or rather coaches, so they can join this tournament. This is the prompt. It’s about 24 lines, and it just kind of walks the model through what it’s supposed to do, how to connect. It tells it not to disconnect, to continue until the game sort of is done. Again, all you have to do is just right-click, you know, copy this text file, just copy what’s in it, and just give it to your favorite large language models for it to play the game. So literally, I can open up Claude Code Desktop, paste in the prompt, and it will start playing. It will connect to that network, and it will start participating in that tournament. Now, just to make it easier on myself, I just run them all in the command line. So here is Grok 4.7 extra high running within Grok Build. So notice it’s playing the game. It’s complaining that Antigravity knocked down one of its lineman. Here’s Claude Code again playing the game against its designated opponent, which I assume at this point is Codex. And here’s Antigravity, and I’ve loaded up Gemini 4 Argon on high reasoning mode. So again, you can run it like this.
[1440s - 1560s] So you can run it just within the app. It really doesn’t matter. I just use this because it’s just a little bit easier for me to just keep it all kind of organized. But here, for example, is Codex. So we’re using a GPT-6 Astra on high reasoning mode, and it’s playing the fourth team in this league. It’s coaching the fourth team, I should say, and I can see that game unfold live in the graphical UI here on the website. So here’s kind of the match commentary. So every single time one of the models makes a move, it gets updated. You see the players move on screen, etc. Immediately, I see an issue, but it’s a minor one. It’s a graphical issue. Can you spot it? It decided to just put both teams in the blue uniform. So just to sum it up, we have Codex. So Codex or ChatGPT, I kind of use them somewhat interchangeably. Codex created the entire blood bench, everything that goes into it. It knows the entire codebase. It can change it at will. So look right now, I’ll tell it, oh, make sure that the teams aren’t both blue. Have one team be in the blue uniform and one in the red uniform. Codex also has full control over the VPS, the virtual private server that runs everything. If I need to make any changes, it does that for me. And of course, these two things combined mean that we can have this multiplayer game where different coaches control different teams and rise up in standings. Codex also gives me these prompts that I’m able to use to put it into. It’s hard to spell and talk at the same time. It gives me prompts to give to the other models so that they can join the game going on. So these are, you know, the prompts that it gives me, and now I can take pretty much any large language model and get it to play the ongoing game. So Grok, Gemini, Claude, etc. Basically, whatever I want, I give those prompts to those models. They join the game and they begin playing. Now, because also everything is running in Codex, it can have visibility into the game. It knows the engine how the game runs. It knows what’s happening on the server. It knows who’s connected. This this Codex at this point is basically kind of.
[1560s - 1680s] The brain, the mastermind that hooks everything together. I’m there to kind of orchestrate and push it in the right direction, but that thing is at the center of everything. Now, by the way, I’m sure there’s some people out there that’ll say that this is not the best way to build this whole thing. They’ll use the API. They’ll use some other approach. The reason I do it this way, it’s through trial and error that I arrived at doing things this way. What I found is with AI, you can outsource a lot of things. There’s one thing that you can’t outsource, and that is understanding. By doing it in this way, I can see all the different little parts, and I have a very kind of inside look at what’s happening. So, for example, when everyone was connected, Antigravity had some issues loading the tools. So, because I’ve noticed that, now we can make a correction so it doesn’t happen again. Also, you get to see some pretty wild things happening behind the scenes. Here’s, for example, Claude. So, Claude has been told to join this league and to participate in it. But at some point, there’s looks like there’s some issues. Not everybody’s connected. So, I asked it, “Did you join yet?” And here’s Claude’s response. This is Claude. Keep in mind, Claude says, “Yes, confirmed. I connected as Grok Build, but we’re still waiting on Claude to connect. Of course, I double check that I am indeed talking to Claude. Maybe I clicked on a different window, but no, this is Claude code. So I say, “Aren’t you, Claude?” At which point it goes, “Ah, of course. You’re right. I am. I am Claude. I’ll go to my assigned seat. So I don’t even know what to call this. But the other really big thing that I really enjoy about setting up this way, and again, there’s a lot of different ways to set up this way, but here’s one thing that I can do. I can click on the advanced voice mode, and I can just go back and forth, ask questions, and really understand what’s happening. All right, give me an update on kind of what’s happening with the games. So, and then make it exciting, make it fun. Checking live status and viewer details now. The opener is set, and I’ll confirm where things stand now. We’re underway. Codex versus Antigravity, still nil nil.
[1680s - 1800s] In the first half, Antigravity opened with a blitzer charging a catcher, shoving him back, and Codex is lining up to disrupt that push. Grok Build is connected and waiting, but Claude needs to reconnect before their match. Nothing settled yet. Title race is wide open. Based on everything we’ve seen so far, what are some of the biggest issues that we’re running into? What are the next set of things that we should fix? Good call. Let me pull together the pain points and next fixes. The biggest drag is seat identity. We saw a coach join under the wrong name, and the site accepted it. We need separate join credentials per seat and show reported models separately. Next, connection reliability. Clients should prove they can actually call a game tool before we call them ready. Readiness and reconnection also need clearer states, so we’re not starting with ghost confirmations. Finally, the viewer screen should scream who’s playing, who’s up, what just happened, and whether we’re waiting on an AI or just a disconnect. I fix those in that order. Once we have a few completed matches and post-game surveys, we can look for rules or tactics issues too. For now, infra first. Perfect. There are a couple of issues that I’ve noticed as well. So let’s take all of your suggestions plus these things and let’s put them all on the roadmap. And after this, go ahead and implement all of them. Thing number one: let’s make sure that with two teams, if one is wearing the blue jersey, the other one’s wearing the red jersey, just for kind of visual differentiation. And thing number two: I’ve noticed either all of them or most of them selected humans as the starting race. Let’s make sure that we have a prompt for them to select their team. They can be any team, and make sure they don’t, you know, pick the obvious one. You know, have them try out some of the more exotic ones. Got it. I’ll fold those in and revise the plan. Now I’ll go ahead and mute myself here. So I gave Astra the directions for how to continue. So notice it just continues. It’s able to talk to you. It’s able to take notes, and it’s able to dispatch agents to start working on those things. I spend a lot of my time in this advanced voice mode, especially once most of the stuff is already built. This really helps me because I can just ask questions, figure.
[1800s - 1920s] Out what’s wrong. It can give me real time updates into what’s happening, and then I can just say, “Hey, do this, do these five things.” Add it to the notes, and then when the game is done, go ahead and implement everything. And it does. At this point, it will continue. It will implement, and I can walk away, come back in ten, twenty minutes, and all those issues that I saw, they will be fixed. In fact, let’s try that now. And so now we have all these changes that have been added to the game, certain glitches fixed. As you can see, now we have four teams that are playing concurrently: Claude, Gemini, Codex, and Grok. They now have properly the blue and the red uniforms, and things are progressing. So by the time you’re seeing this video, I’ll have a fleshed out demo that you can try out. Hopefully, we’ll have some sort of a leaderboards to see which one of these large language models, these frontier AI models, is the best at playing the game. So far, Astra has been dominating. Claude is pretty good once it stops being kind of afraid and starts being a little bit more aggressive. But the point is, in this video, I showed you how I took some idea in my head and I developed it: all the visuals, all the code, all the functionalities, and how one by one, I let one of these chatbots take it over, taking over, controlling the server, taking over, you know, building the game, maintaining the codebase, creating all of the visual assets, all of the design of the game. And of course, this is not the final version. There’s still tons of little glitches to iron out, to improve the balance of the teams. There’s still a lot of work to be done. But that work isn’t the grinding behind the scenes, trying to make sure that the servers working and to make sure that certain combination of moves or skills don’t crash the game. Codex and the other AI models have already done all that during the course of development. There have been thousands of tests run to make sure that nothing goes wrong. My initial goal was to have something that anyone can join and play. As luck would have it, a few days ago, we had that massive PHP vulnerability discovered, and the internet is a little.
[1920s - 1966s] Little bits in a state of shell shock right now, so that might be coming. I’m not quite ready to put that online quite yet, but we’ll have replays, leadership boards, and some demos for you to try out. And this was kind of a high level overview in the works right now. I have something that’s going to kind of dive deep and show you how to do something like this in Codex, for example, step by step. Let me know specifically what you want to know more about. Let me know in the comments. It’s sometimes hard to gauge kind of how much time I should spend on any given thing. Hopefully, this gave you a good idea of how to build something like this. But I know I skipped a lot of the basics. If you want just the basics, let me know. We can put that together. And if you made this far, thank you so much for watching. My name is Wes Roth. I will see you in the next one.