Anthropic Fable Beats OpenAI Astra on Gameplay in Coding Trial

I built the same game with Astra and Fable 5.1... only one was fun — Fireship

Fireship

While OpenAI's Astra generates polished 3D graphics faster, Anthropic's Fable 5.1 builds more engaging software due to superior physics and complex simulation mechanics.

AI-assisted summary of I built the same game with Astra and Fable 5.1... only one was fun — original source. Check the original for full context.

Source published:

Key ideas

  • Nvidia CEO Jensen Huang announced the arrival of AGI following the release of OpenAI's GPT-6 Astra, which was trained on over 100,000 Grace Blackwell GPUs with 400,000 more coming online. (0:00)
  • OpenAI made GPT-6 Astra publicly accessible through its $100 per month GPT Pro subscription tier, promoting it as a major breakthrough for 3D game and asset generation. (0:00)
  • In a direct game-building comparison on a rocket launch simulator prompt, Astra prioritized visual fidelity and slick Three.js rendering within 26 minutes, while Anthropic's Fable 5.1 delivered superior physics simulation and deep gameplay mechanics despite basic UI. (2:02)
  • Astra demonstrated massive leaps in complex 3D modeling—such as generating exploded mechanical watch schematics in minutes—prompting concerns about an impending wave of automated '3D slop.' (3:46)

Detailed notes

## Background Published by Fireship on YouTube on September 09, 2026, the video investigates whether OpenAI's newly released Astra model lives up to the marketing hype. The video description also promotes Mobbin Pro to help coding agents design production-ready UIs that actually convert.

## TL;DR Following Jensen Huang's proclamation of AGI upon the release of OpenAI's GPT-6 Astra, Fireship tests the $100/month model directly against Anthropic's Fable 5.1 by having both generate a custom 3D rocket simulator from scratch. Astra quickly produces a polished Three.js presentation with slick visual assets that appeals to younger players but delivers shallow mechanics, whereas Fable 5.1 takes longer and looks dated while providing deeper physics and authentic simulation systems. Astra also displays astonishing improvements in rendering complex exploded 3D models like mechanical watches, though an autonomous build of a 'Horse Tinder' app demonstrates that minor visual misalignments still separate current models from genuine superintelligence.

## Atmosphere & style The presentation crackles with Fireship's signature sardonic wit, rapid-fire pacing, and dry tech cynicism. Viewers experience a humorous emotional rollercoaster that oscillates between genuine technological awe at massive leaps in 3D generation and comedic relief when subtle flaws puncture the hype. Beneath the fast-cut jokes and existential reflections lies a grounded, pragmatic critique of AI benchmark marketing versus actual end-user utility.

## Bulletin - [00:00](https://www.youtube.com/watch?v=2Xiljy4xzbc&t=0s) Nvidia CEO Jensen Huang announced the arrival of AGI following the release of OpenAI's GPT-6 Astra, which was trained on over 100,000 Grace Blackwell GPUs with 400,000 more coming online. - [00:00](https://www.youtube.com/watch?v=2Xiljy4xzbc&t=0s) OpenAI made GPT-6 Astra publicly accessible through its $100 per month GPT Pro subscription tier, promoting it as a major breakthrough for 3D game and asset generation. - [02:02](https://www.youtube.com/watch?v=2Xiljy4xzbc&t=122s) In a direct game-building comparison on a rocket launch simulator prompt, Astra prioritized visual fidelity and slick Three.js rendering within 26 minutes, while Anthropic's Fable 5.1 delivered superior physics simulation and deep gameplay mechanics despite basic UI. - [03:46](https://www.youtube.com/watch?v=2Xiljy4xzbc&t=226s) Astra demonstrated massive leaps in complex 3D modeling—such as generating exploded mechanical watch schematics in minutes—prompting concerns about an impending wave of automated '3D slop.'

## Key takeaways - [00:00](https://www.youtube.com/watch?v=2Xiljy4xzbc&t=0s) Jensen Huang's declaration of AGI comes with massive commercial interest, as OpenAI's Astra model relies on hundreds of thousands of Nvidia Grace Blackwell chips. - [00:00](https://www.youtube.com/watch?v=2Xiljy4xzbc&t=0s) Access to Astra is gated behind a $100 monthly GPT Pro membership, raising the financial barrier for individual developers testing frontier models. - [00:00](https://www.youtube.com/watch?v=2Xiljy4xzbc&t=0s) Benchmark claims tout Astra's dominance on ARC-AGI over raw Opus, though prior research shows external agent scaffolds had already achieved 99% scores with earlier models. - [00:00](https://www.youtube.com/watch?v=2Xiljy4xzbc&t=0s) Astra targets 3D creative pipelines, capable of outputting Three.js code, Blender assets, and integration-ready pipelines for Unity and Unreal Engine. - [02:02](https://www.youtube.com/watch?v=2Xiljy4xzbc&t=122s) When given identical rocket simulator prompts, Astra finished in 26 minutes with high-polish 3D visuals that suffered from generic aesthetics and shallow gameplay. - [02:02](https://www.youtube.com/watch?v=2Xiljy4xzbc&t=122s) Fable 5.1 took longer and produced outdated Bootstrap-style styling, but delivered authentic orbital mechanics, varied failure states, and deeper physical simulations. - [02:02](https://www.youtube.com/watch?v=2Xiljy4xzbc&t=122s) User testing revealed an age-based split: younger audiences preferred Astra's shiny visual spectacle, while older players preferred Fable's challenging systems and physics. - [03:46](https://www.youtube.com/watch?v=2Xiljy4xzbc&t=226s) Astra's ability to render exploded mechanical assemblies in minutes replaces tasks that previously required hundreds of hours from skilled 3D human artists. - [03:46](https://www.youtube.com/watch?v=2Xiljy4xzbc&t=226s) Lowering the technical barrier to complex 3D graphic generation threatens educational 3D channels by enabling low-effort automated content farms and '3D slop.' - [03:46](https://www.youtube.com/watch?v=2Xiljy4xzbc&t=226s) In an autonomous test building 'Horse Tinder', Astra generated 21 minutes of clean, error-free code, foiled only by a minor UI visual alignment flaw.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=2Xiljy4xzbc&t=0s) Jensen Huang Declares AGI and the Arrival of GPT-6 Astra

Nvidia CEO Jensen Huang took to X to make a staggering declaration: "AGI has arrived. Congratulations, OpenAI team." As Fireship dryly observes, this was an entirely "sobering and totally independent assessment" coming from the exact man who noted in the very same post that OpenAI's brand-new Astra model had been trained on more than 100,000 of his Grace Blackwell GPUs—with another 400,000 scheduled to come online shortly. If Jensen's prediction is accurate, every white-collar worker sitting behind a monitor is now officially obsolete, marking roughly the fifth time that catastrophe has struck in 2026 alone. Now that Astra has rolled out to the general public—or at least anyone willing to fork over $100 for a GPT Pro subscription—it is time to put the marketing claims to the test by pitting Astra directly against Anthropic's Fable 5.1.

Game developers and 3D technical artists were long assumed to be immune to the automation wave, protected by the reality that engaging games demand distinct personality, physical feel, and soul, alongside complex 3D graphic pipelines. Yet GPT-6 Astra has arrived with a reputation as a "turbo AI machine god for 3D games." It can rapidly churn out complete Three.js experiences, construct intricate 3D assets in Blender, and interface with industry workhorses like Unity and Unreal Engine. On the benchmark front, Astra purportedly conquered the ARC-AGI benchmark—a testing ground designed to resist rote memorization and evaluate pure general intelligence—where it crushed Anthropic's Opus sitting at 30%. However, Fireship notes that benchmark marketing obscures the fact that a UC Berkeley research team had already clocked a 99% ARC score months earlier using Opus 4.8 and Fable 5, where the decisive factor was the agent harness rather than raw model weights.

- **GPU Compute Scale:** GPT-6 Astra was trained on over 100,000 Nvidia Grace Blackwell GPUs, with 400,000 additional units slated to launch soon. - **Access Tier:** Deployed to users paying $100 per month for the GPT Pro membership. - **Benchmark Claims:** First public single-model run to top ARC-AGI, compared to Opus's standalone 30% mark, though external scaffold harnesses previously reached 99%.

[02:02](https://www.youtube.com/watch?v=2Xiljy4xzbc&t=122s) The Rocket Launch Showdown: GPT-6 Astra vs. Fable 5.1

Sidestepping synthetic "trust me bro benchmarks," Fireship designed a real-world test by asking his kids what kind of game they wanted to play. The prompt called for a 3D rocket launch simulator where players configure rocket parts and launch their custom craft into orbit. Both GPT-6 Astra and Fable 5.1 were fed the exact same prompt simultaneously. ChatGPT crossed the finish line first, completing the entire application in roughly 26 minutes. Visually, Astra delivered an immediate punch: clean, highly detailed Three.js geometry, polished UI styling, and a striking rendering of the vehicle orbiting a textured Earth. However, the interface shared an unmistakable aesthetic formula with countless other AI creations floating across Twitter, and beneath the graphical sheen, the mechanical gameplay felt shallow and uninspired.

```mermaid flowchart TD accTitle: Astra vs Fable Game Comparison Flow accDescr: Visual breakdown comparing the workflow and results of GPT-6 Astra and Fable 5.1 on the rocket simulator prompt.

subgraph Prompt["Identical User Prompt"] P["'Rocket Launch Simulator: Customize and Orbit'"] end

subgraph AstraFlow["OpenAI Astra (26 min)"] A_UI["Slick Modern UI & Detailed 3D Assets"] A_Play["Basic, Repetitive Gameplay Mechanics"] A_Win["Appealed to 3-Year-Old"] end

subgraph FableFlow["Anthropic Fable 5.1 (Extended Run)"] F_UI["Outdated UI resembling Bootstrap"] F_Play["Complex Physics, Custom Knobs & Explosion VFX"] F_Win["Appealed to 8-Year-Old"] end

P --> AstraFlow P --> FableFlow A_UI --> A_Play --> A_Win F_UI --> F_Play --> F_Win ```

Fable 5.1 took significantly longer to process, finally returning code that looked visually primitive. Its interface resembled something whipped together with "Twitter Bootstrap, which is a library that was written before most of you were born," accompanied by blocky, low-fidelity 3D assets. Yet when it came to actual game design, Fable blew Astra away. It incorporated granular engineering controls, authentic aerodynamic and orbital physics calculations, varied victory and failure scenarios, and punchy explosion sequences whenever a launch went awry. The family trial concluded in a split verdict: the 3-year-old preferred Astra's simple, shiny spectacle, whereas the 8-year-old favored Fable's deep, punishing systems simulation.

[03:46](https://www.youtube.com/watch?v=2Xiljy4xzbc&t=226s) The Looming Era of 3D Slop and the Horse Tinder Test

Beyond mini-games, Astra's raw spatial generation represents a massive leap forward from prior models. Only two months earlier, in July, an attempt using GPT Soul to render an exploded view of a mechanical watch produced nothing more than "a pile of useless garbage." Re-running that identical watch prompt through Astra generated an intricate mechanical schematic that would have traditionally demanded hundreds of billable hours from a skilled 3D modeler. While impressive, this capability also signals an incoming tidal wave of low-effort automated content: the demise of the technical barrier that once protected high-effort educational 3D channels like Jared Owen, opening the floodgates to automated "3D slop" across the internet.

```mermaid flowchart LR accTitle: Mobbin MCP Workflow accDescr: Diagram showing how coding agents use the Mobbin MCP server to build verified UI flows.

subgraph Agent["Coding Agent Environment"] IDE["Agent: Claude / Cursor / Codex"] end

subgraph Data["Mobbin UI Architecture"] MCP["Mobbin MCP Server"] Lib["600,000+ Production UI Screens"] end

subgraph Output["Implementation"] Report["Visual Design Pattern Report"] Code["Target App Implementation with Source Links"] end

IDE -->|"Queries complete flow"| MCP MCP --> Lib Lib -->|"Aggregates reference screens"| Report Report -->|"Adapts verified patterns"| Code ```

To deliver a final verdict on whether Astra is genuine AGI, Fireship subjected the model to the definitive software crucible: building the legendary "Horse Tinder" app without human intervention or bugs. Astra took 21 minutes to output the full codebase, yielding an interface cleaner and more elegant than Fable's previous iteration. After combing through every single line of code without locating an error, Fireship experienced an existential epiphany reminiscent of Antonio Salieri reviewing Mozart's manuscripts: "Here again was the very voice of God." That illusion shattered instantly upon inspecting the visual render: a single carrot asset was misaligned by a handful of pixels, confirming that humanity remains safe from artificial superintelligence for another day.

- **Watch Exploder Benchmark:** Astra rendered a complex, fully exploded 3D mechanical watch assembly in minutes—a task that previously broke earlier GPT models and took human 3D artists hundreds of hours in Blender. - **Horse Tinder Verification:** Completed in 21 minutes of continuous autonomous generation with functional code, undone solely by a pixel-misaligned vegetable icon. - **Sponsor Solution (Mobbin):** Integrates an MCP server into environments like Claude, Cursor, and Codex, allowing AI agents to reference over 600,000 real production app screens to prevent hallucinated, generic "purple gradient vibe slop."

## Notable quotes > [00:00](https://www.youtube.com/watch?v=2Xiljy4xzbc&t=0s) AGI has arrived. Congratulations, OpenAI team. — Jensen Huang

> [00:00](https://www.youtube.com/watch?v=2Xiljy4xzbc&t=0s) sobering and totally independent assessment — Fireship

> [00:00](https://www.youtube.com/watch?v=2Xiljy4xzbc&t=0s) turbo AI machine god for 3D games — Fireship

> [02:02](https://www.youtube.com/watch?v=2Xiljy4xzbc&t=122s) trust me bro benchmarks — Fireship

> [02:02](https://www.youtube.com/watch?v=2Xiljy4xzbc&t=122s) Twitter Bootstrap, which is a library that was written before most of you were born — Fireship

> [03:46](https://www.youtube.com/watch?v=2Xiljy4xzbc&t=226s) a pile of useless garbage — Fireship

> [03:46](https://www.youtube.com/watch?v=2Xiljy4xzbc&t=226s) 3D slop — Fireship

> [03:46](https://www.youtube.com/watch?v=2Xiljy4xzbc&t=226s) Here again was the very voice of God. — Fireship

> [03:46](https://www.youtube.com/watch?v=2Xiljy4xzbc&t=226s) purple gradient vibe slop — Fireship

## Mentioned Anthropic, Antonio Salieri, ARC-AGI, Blender, ChatGPT, Claude, Codex, Cursor, Fable 5, Fable 5.1, Fireship, GPT Pro, GPT Soul, GPT-6 Astra, Grace Blackwell GPUs, Jared Owen, Jensen Huang, Mobbin, Mozart, Nvidia, OpenAI, Opus, Opus 4.8, Three.js, Twitter Bootstrap, UC Berkeley, Unity, Unreal Engine, X

## Video description Get 20% off Mobbin Pro to help your agent design UIs that actually convert - https://mobbin.com/fireship

Now that Astra is actually available to use, let's see if it lives up to the hype.

#coding #programming

Want more Fireship?

🗞️ Newsletter: https://bytes.dev 🧠 Courses: https://fireship.dev

[Source](https://www.youtube.com/watch?v=2Xiljy4xzbc)

中文简报

代码实测：Anthropic Fable 在游戏性上击败 OpenAI Astra 尽管 OpenAI 的 Astra 生成精美 3D 图形的速度更快，但 Anthropic 的 Fable 5.1 凭借更出色的物理效果和复杂的模拟机制，构建出了更具吸引力的软件。 在英伟达 CEO 黄仁勋针对 OpenAI 发布的 GPT-6 Astra 宣告 AGI 已经到来后，Fireship 让这款每月 100 美元的模型与 Anthropic 的 Fable 5.1 展开正面对决，要求两者从零开始生成一个定制的 3D 火箭模拟器。Astra 迅速生成了一个精致的 Three.js 演示画面，其华丽的视觉资产对年轻玩家颇具吸引力，但机制较为浅显；而 Fable 5.1 虽然耗时更长且外观显得过时，却提供了更深度的物理特性和真实的模拟系统。此外，Astra 在渲染机械表等复杂 3D 爆炸图模型方面展现出了惊人的提升，不过在自主构建“马匹版 Tinder”（Horse Tinder）应用时出现的一些微小视觉偏差表明，当前模型与真正的超级智能之间仍有距离。 在 OpenAI 发布 GPT-6 Astra 后，英伟达 CEO 黄仁勋宣布 AGI 已经到来；该模型在超过 10 万块 Grace Blackwell GPU 上完成训练，另有 40 万块即将投入使用。 (0:00) OpenAI 通过每月 100 美元的 GPT Pro 订阅层级向公众开放了 GPT-6 Astra，并将其宣传为 3D 游戏和资产生成领域的重大突破。 (0:00) 在针对火箭发射模拟器提示词的正面对比测试中，Astra 在 26 分钟内优先呈现出高视觉保真度和流畅的 Three.js 渲染，而 Anthropic 的 Fable 5.1 尽管界面简陋，却提供了更出色的物理模拟和深度游戏机制。 (2:02) Astra 在复杂 3D 建模方面展现出巨大飞跃——例如在数分钟内生成机械表的爆炸图示意模型——这也引发了人们对即将掀起一波自动化“3D 垃圾内容（3D slop）”浪潮的担忧。 (3:46)

Reading summary

Testing OpenAI's GPT-6 Astra against Anthropic's Fable 5.1 reveals that despite industry claims of artificial general intelligence, the models diverge sharply between superficial visual polish and authentic gameplay depth. Compute Scale and the ARC-AGI Benchmark (0:00) Nvidia CEO Jensen Huang publicly declared the arrival of AGI following the release of OpenAI's GPT-6 Astra, an assessment contextualized by the disclosure that the model was trained on more than 100,000 Grace Blackwell GPUs, with 400,000 more scheduled to deploy. Astra is accessible to users paying $100 per month for GPT Pro. Marketing highlighted Astra's top score on the ARC-AGI benchmark relative to Anthropic Opus's standalone 30% performance. However, benchmark comparisons obscure that UC Berkeley researchers had already achieved a 99% score using Opus 4.8 and Fable 5, demonstrating that external agent scaffolding often matters more than raw model weights. Rocket Simulator Test: Visuals Versus Gameplay Depth (2:02) A head-to-head test using an identical user prompt for a 3D rocket launch simulator revealed distinct design philosophies between the two models. Astra delivered a functional application in roughly 26 minutes, rendering polished Three.js geometry, modern interface styling, and a textured Earth orbit. Despite the graphical quality, its underlying mechanics were shallow and repetitive, ultimately appealing mainly to a three-year-old tester. Conversely, Fable 5.1 required longer processing times and produced a visually dated interface resembling legacy Twitter Bootstrap alongside low-fidelity 3D assets. However, Fable generated superior systemic mechanics, incorporating authentic aerodynamic and orbital physics calculations, granular user controls, and dynamic failure scenarios with explosion sequences that engaged an eight-year-old tester. 3D Generation Advances and Content Volume Concerns (3:46) Astra demonstrated rapid improvements in complex 3D asset generation compared to predecessor models. Generating an exploded view of a mechanical watch—a task that previously yielded non-functional outputs in GPT models and traditionally required hundreds of hours from human 3D artists—was completed by Astra in minutes. While this efficiency lowers production hurdles, it raises concerns regarding an influx of automated, low-effort 3D content undermining high-effort educational animators. In UI development, agents are pairing with tools like the Mobbin MCP server, which references over 600,000 production screens to mitigate repetitive AI aesthetic patterns. Autonomous Code Execution and Persistent Limits (3:46) To evaluate autonomous software creation without human intervention, Astra was prompted to build a 'Horse Tinder' application. The model operated continuously for 21 minutes to output a functional, syntax-error-free codebase that visually surpassed previous iterations. Despite the flawless underlying code structure, the final visual render displayed a misaligned carrot asset, illustrating that minor visual inconsistencies continue to challenge claims of complete artificial superintelligence. While GPT-6 Astra demonstrates significant engineering progress in automated 3D spatial modeling and rapid code generation, its tendency to prioritize aesthetic formulas over rich functional mechanics indicates that declarations of achieved AGI remain premature.

Source quotes

[0:00] Jensen Huang: AGI has arrived. Congratulations, OpenAI team. [0:00] Fireship: sobering and totally independent assessment [0:00] Fireship: turbo AI machine god for 3D games [2:02] Fireship: trust me bro benchmarks [2:02] Fireship: Twitter Bootstrap, which is a library that was written before most of you were born [3:46] Fireship: a pile of useless garbage [3:46] Fireship: 3D slop [3:46] Fireship: Here again was the very voice of God. [3:46] Fireship: purple gradient vibe slop

Original transcript

[0:00] Yesterday, Nvidia CEO Jensen Huang [0:02] logged into X to make an announcement. [0:04] AGI has arrived. Congratulations, OpenAI [0:07] team. It was a sobering and totally [0:09] independent assessment from a man who in [0:11] the very same post pointed out how [0:13] OpenAI's new Astra model was trained on [0:15] more than 100,000 of his Grace Blackwell [0:18] GPUs, and that another 400,000 of his [0:20] GPUs are coming online soon. Obviously, [0:22] he has nothing to gain here, but is [0:24] Jensen right, or is he wrong about Astra [0:26] being AGI? Because if he's not wrong, [0:29] that means every human being working in [0:30] front of a computer right now is [0:32] officially obsolete. And that's the [0:33] fifth time that's happened this year. [0:35] Well, in today's video, we'll learn the [0:36] truth because yesterday Astra rolled out [0:39] to the common man, or at least the man [0:40] willing to pay $100 for a GPT Pro [0:42] membership. I compared Astra to Fable [0:45] 5.1, and the results were not what I [0:47] expected. It is September 9th, 2026, and [0:49] you're watching The Code Report. There [0:51] was one group of developers who were [0:52] supposed to be safe from AI, game [0:54] developers. Because for a game to be fun [0:56] to play, it needs to have a very [0:58] specific feel and soul. Not to mention, [1:00] the ability to draw and design [1:02] good-looking 3D graphics. Unfortunately, [1:04] though, it looks like the game developer [1:06] and 3D artist jobs are about to become [1:08] just as rare as the Angular JS developer [1:10] jobs. Because GPT-6 Astra has been [1:13] called a turbo AI machine god for 3D [1:15] games. Not only can it slop out amazing [1:18] games in 3JS, but it can also build [1:20] extremely detailed graphics in Blender [1:22] and work with other game engines like [1:24] Unity and Unreal. But what might be even [1:26] more terrifying is that Astra was the [1:28] first model to reach the top of the Arc [1:31] benchmark, which was designed [1:32] specifically as a benchmark that can't [1:34] be memorized, which forces a model to [1:36] demonstrate how well it generalizes [1:38] intelligence. It completely mogs [1:40] Anthropic with Opus sitting at 30%, but [1:42] what they don't tell you is that a team [1:44] at Berkeley already scored 99% on the [1:46] Arc AGI benchmark several months ago. [1:48] They did it with Opus 4.8 and Fable 5, [1:51] but the real difference maker was the [1:53] harness. But today we're not going to [1:54] focus on the trust me bro benchmarks, [1:56] and instead, we'll handcraft some of our [1:58] own slob to compare Fable 5.1 and GPT-6 [2:01] Astra side by side. So, to do that in [2:03] the most organic way possible, I asked [2:05] my kids, "What kind of game do you want [2:07] to play today?" And they requested a [2:08] rocket launch simulator where you [2:10] customize a rocket and launch it into [2:12] orbit. I sent both models the exact same [2:14] prompt at the exact same time, and [2:15] ChatGPT finished first in about 26 [2:18] minutes. When I first launched the [2:19] ChatGPT game, I was incredibly [2:21] impressed. The 3D graphics are highly [2:23] detailed and beautiful with very few [2:25] mistakes, and the UI looks beautiful as [2:27] well. One thing that was kind of [2:28] interesting, though, is that the UI [2:29] looks almost identical to a lot of other [2:32] games that I've seen on Twitter. There's [2:33] a very obvious formula being used here [2:35] that makes it easy to spot ChatGPT [2:37] generated games. And now, as far as the [2:39] actual gameplay goes, I was able to [2:41] launch a rocket into orbit, but I [2:43] wouldn't exactly describe this game as [2:44] fun to play. I think the most impressive [2:46] thing was this graphic of the rocket [2:48] actually orbiting the globe. Now, after [2:50] playing that for a while, I went back to [2:51] check on my Fable prompt, which still [2:53] wasn't done, but eventually finish at [2:55] some point in the far future. And when [2:57] it did, the graphics were embarrassing [2:59] by comparison. The UI looks like it was [3:00] built with Twitter Bootstrap, which is a [3:02] library that was written before most of [3:04] you were born, and the 3D graphics [3:05] aren't nearly as detailed as what Astra [3:07] generated. However, after actually [3:09] playing the game for a few minutes, I [3:11] quickly realized that the gameplay was [3:12] far superior to what Astra had [3:14] generated. For one, the options for [3:16] rocket customization are way more [3:18] complex. Claude provided way more knobs [3:20] to turn and made scientific calculations [3:22] that make this feel more like a legit [3:24] rocket simulator. But then, when it came [3:26] to gameplay, the game was also just more [3:28] fun with a bunch of different [3:29] possibilities for success or failure and [3:31] cooler animations when your rocket [3:33] explodes. But the big test was which [3:35] game do my kids prefer? Well, this [3:37] actually ended in a split decision with [3:39] the 3-year-old preferring the more [3:40] simplistic Astra game with the [3:42] 8-year-old preferring the more complex [3:44] Fable game. It's mostly subjective, but [3:46] one thing's for sure, Astra's ability to [3:48] generate 3D graphics has improved [3:50] dramatically. It just a couple months [3:51] ago back in July, I had GPT Soul [3:54] generate an exploded view of a [3:55] mechanical watch, and the end result was [3:58] a pile of useless garbage. But when I [3:59] re-ran the prompt again yesterday, the [4:01] end result was this beautiful diagram [4:03] that likely would have taken a 3D [4:05] designer hundreds of hours to produce. [4:07] And this actually terrifies me because [4:09] it's going to lead to a new wave of slop [4:10] on YouTube of people copying channels [4:13] like Jared Owen and other people who [4:14] make 3D explainer videos. Videos like [4:17] this used to be what we would call high [4:19] effort and might require hundreds of [4:20] hours of 3D design work in Blender. [4:22] Unfortunately, that moat no longer [4:24] exists and the rise of 3D slop is going [4:26] to be the next big thing that makes the [4:28] internet suck even more. But we still [4:29] haven't answered the question, is Astra [4:31] actually AGI? Well, as you know, the [4:33] only way to find out is to see if it can [4:35] build Horse Tender without making any [4:37] mistakes. It took 21 minutes to run the [4:39] prompt and this was the final result. A [4:41] design that looks surprisingly similar [4:43] to one made by Fable a few months ago. [4:45] However, this design is noticeably [4:46] cleaner and more elegant. And after [4:48] spending many, many hours going through [4:50] every single line of code looking for a [4:51] mistake, I wasn't able to find a single [4:53] one. I felt like I was Antonio Salieri [4:56] reading the flawless music of Mozart. [4:58] >> Here again was the very voice of God. [5:02] >> And then I realized that can only mean [5:04] one thing. Jensen was not wrong. AGI has [5:06] been achieved. Oh wait, actually hold [5:07] on. Wait a minute. This carrot is a few [5:09] pixels off. The false alarm. This is [5:11] just another dumb AI model. If you want [5:13] to build a seriously good UI, you need [5:15] to use Mobbin, the sponsor of today's [5:17] video. It lets your coding agent study [5:19] more than 600,000 real UI screens before [5:22] it designs anything. I've already been [5:23] using Mobbin for over 5 years because it [5:26] shows you complete user flows from [5:27] thousands of applications. And its MCP [5:30] server now puts that entire library [5:32] inside Claude, Code, Cursor, Codex, and [5:34] all the others. So, instead of asking [5:36] for a settings page and getting the same [5:38] purple gradient vibe slop, you start by [5:41] telling your agent to study how the best [5:42] apps build that complete user flow. [5:44] It'll use Mobbin to gather the [5:46] references, build a visual report of the [5:48] patterns they share, then use those [5:50] findings to recreate it and adapt it to [5:52] your app. Every result points back to [5:54] its original screen, so you can check [5:56] the source before trusting a design [5:57] hallucination. Try Mobbin today and get [6:00] a 20% discount at the link below. This [6:02] has been the Code Report. Thanks for [6:04] watching and I will see you in the next [6:05] one.