0:00 Imagine speaking a living, breathing 3D world into existence. Right, just out of thin air. 0:06 Yeah, using nothing but pure math. No software. 0:09 No 3D models. Just lines of code. Just code. 0:13 Welcome to the Deep Dive. Glad you're here with us. Today we are cutting through all the rumors. 0:18 And there are a lot of them. There really are. 0:20 We are looking at the reality of Gemini 4 Neon. 0:24 It's been making huge waves. It has. We will see how it stacks up against the competition. 0:28 Right, because we just got OpenAI's freshly dropped GPT 6 .2. That includes both the Astra and Sol models. Yep. We will also weight against Claude Opus 5 .5. We are going to break down the benchmark data. The raw numbers. Exactly. And we'll explore 0:43 some mind -bending creative projects you can try. Things like math -generated 3D dragons. 0:48 and even fully playable browser games. It is a fascinating time for artificial intelligence. 0:53 It feels like a massive shift. Oh, absolutely. 0:56 We are definitively moving beyond basic text generation. We are stepping right into autonomous world building. Yeah. The capabilities are expanding at a staggering rate. And the barrier to entry is collapsing entirely. It really is. We should 1:09 start by setting the stage with the raw data. 1:11 Good idea. Let's look at the landscape. GPT 6 .2 just dropped on November 17th. Barely any time ago. Right. The industry barely had time to process that release. I mean, people were still testing it. Yeah. But Gemini 4 NEON is the one making massive waves. It kind of stole 1:27 the spotlight. It really did. The benchmark numbers are highly revealing here. They absolutely are. 1:32 Let's look at the Artificial Analysis Intelligence Index. Or the AA index. Right. NEON currently scores a 61 on this index. A 61. Yeah. That puts it slightly ahead of Claude Opus 5 .5. Which is sitting at what, a 59? Exactly. Opus sits at a very respectable 59. And what about GPT 1:51 6 .2? Well, Neon also beats GPT 6 .2 sole. That one scored a 57. A 57, wow. That is a measurable, distinct advantage for Neon. It really is. A four -point lead on that index is actually massive. 2:05 Yeah, people don't always realize how hard those points are to get. Right. It represents a significant leap in underlying reasoning. It does. And Neon also leads in specific coding stress tests. Like which ones? Well, it scored 65 .4 % on cursorbench. 2:20 OK, that's high. Yeah. It also hit 39 % on SWE sweep. Let me pause there. Just for clarity, we should define that for you. SWE sweep. Yeah. 2:29 Go for it. SWE sweep is essentially a test of fixing real world software bugs. Entirely on its own. Right. It is like parachuting an AI into a dark maze. I love that analogy. It has to navigate a massive, messy code base without help. It gets no hints. Exactly. Hitting 39 % 2:48 on that is genuinely wild. It is. It shows strong, autonomous software engineering capabilities. 2:54 Yeah. But, you know, to your point, it is not a complete monopoly. Right, right. Neon doesn't win absolutely everything. No, it doesn't. If Claude Opus is still winning in Terminal Bench, that tells us something. Yeah? What do you think 3:05 it means? It means Neon still struggles with some core systemic commands. That makes sense. 3:10 Terminal Bench is basically testing how well an AI operates a computer's core command line. 3:15 Right, the raw terminal. Claude leads that benchmark with 65 .8%. That's a solid win for Claude. It is. It also takes frontier code with 47 .3%. 3:25 So it's not a total knockout. Right. And GPT 6 .2 sole has its specific victories too. Oh, for sure. It wins chartography. Right. It takes that with 75 .1%. Each model clearly has its own highly distinct strengths. They do. GPT 6 .2 is phenomenal at visualizing data structures. 3:45 Yeah, it's great at that. But overall, NEON leads in the majority of benchmarks we track. It does. 3:50 But you know what I think is the most interesting part? What's that? The thinking level mechanic. 3:55 Oh, yeah. What is truly fascinating to me is how that shifts the paradigm. It changes how we actually use AI. Right. Performance actually scales directly with compute. You literally get what you pay for. Exactly. You can adjust the reasoning power based on the task. It is like 4:12 shifting gears in a high -performance car. That is the perfect analogy. Right. Let's break down the gears. On the low setting, NEON scores 44. 4:21 That is on the artificial analysis index, right? 4:23 Yes. If you push it to medium, it hits 55. OK. 4:27 The big jump. Yeah. High gets you to 60. And the max? X high reaches that peak score of 61. 4:34 Naturally, the cost per task rises accordingly. 4:37 You have to pay for that extra brain power. Right. 4:39 You have to balance your budget against your task complexity. You wouldn't use xHive to write a simple email. No, that would be a waste of money, but you would use it to architect a database. 4:49 Exactly. You buy the reasoning time you actually need. That flexibility is huge for businesses. 4:54 It is, but the enterprise automation results are what really caught my eye. Oh, the auto -backbench numbers? Yeah. Let's look closely at auto -backbench. 5:04 Neon absolutely dominates here with 76%. Wow. 5:08 Gemini 4 Carbon falls closely with 74%. Claude, Opus 5 .5, trails significantly at 68%. That's quite a drop. And GPT -6 Astra is way down at 57%. Whoa. Imagine scaling to a billion queries. 5:24 Right. For enterprise automation, that accuracy difference is completely transformative. It really is. It changes the entire economic calculation for large scale operations. If you are processing a million customer service interactions daily, 5:39 Accuracy matters. A 9 % edge over Claude saves you millions. Easily. It prevents catastrophic downstream errors in your automated pipelines. 5:50 I want to understand the mechanics here. Sure. 5:53 Does bumping up the thinking level just mean it takes longer? It does take longer, but for a highly specific reason. Okay. It is dedicating far more compute steps to complex logic. So it's breaking things down more. Yes, it breaks problems 6:07 down into smaller, solvable components. That makes sense. It tests hypotheses internally before generating the final output token. Oh, wow. It is actively evaluating its own logical pathways in real time. So it literally just spends more 6:22 time pondering before answering? That is precisely it. It is trading processing time for guaranteed accuracy. But raw intelligence doesn't mean much in a vacuum. No, it doesn't. Abstract numbers on a chart can only tell us so much. We need 6:37 to see it in action. Exactly. I want to see that x -high thinking level actually working. Oh, you will. Let's see what happens when it is pushed to its creative limits. Yes. This brings us to the 3D Dragon project. Right. This is where the 6:52 technology gets truly wild. It is mind blowing. 6:55 The developer didn't use Blender or Maya for this project. No traditional 3D software at all. 7:01 None. They didn't import any existing mesh files. 7:03 So how was it built? The Dragon was built entirely through raw procedural code. Just typing? Just typing. It relied exclusively on SDF. Right, SDF. Yeah. Math formula is calculating the exact distance to a 3D surface. That is exactly what 7:21 they are. You aren't plotting points on a traditional 3D grid. Like you would in a normal video game. 7:27 Right. You are writing equations that define empty space versus solid mass. It is incredibly complex for a human to visualize in their head. 7:35 Oh, it's brutal. The prompt given to neon was extraordinarily demanding. It had to act as a master of procedural generation. Yeah. It was given a single 2D reference image of a dragon. 7:48 And it had to recreate that image purely with math. Yes. It had to build the horns, wings, and scales. With just math formulas. It had to logically combine simple primitive shapes. like spheres and cubes. Right. It used custom distance 8:04 functions to create highly organic forms. That is the crazy part. It blended spheres and cylinders together to make muscle tissue. If you have ever tried procedural coding, you know this is hard. 8:14 It is a nightmare. Organic forms are notoriously difficult to create using pure mathematics. Because math wants to be perfect. Exactly. Math inherently wants to be rigid, geometric, and Perfectly symmetrical. 8:26 And nature isn't like that. Nature is messy, asymmetrical, and highly unpredictable. That is the core challenge. But the most genius part of this experiment was the review system. Tell me about that. NEON actually built its own automated 8:39 visual review pipeline. Just on its own? Yeah. 8:42 It rendered the math -generated dragon from five different angles. Is this like setting up five virtual cameras around the model? Exactly. It is stepping back to look at its own sculpture. 8:51 Right. Then it compared those 2D silhouettes to the original reference image. To see where it messed up. Yeah. It autonomously identified incorrect proportions. Wow. It noticed where the wings were too short or distorted. And then 9:02 what? Then it adjusted its own mathematical equations to fix the shapes. That is unbelievable. It repeated this visual feedback loop over multiple rounds. 9:12 I still wrestle with prompt drift myself. No, we all do. I will ask an AI to fix a line of code and it breaks three others. It's a constant game of whack -a -mole. It really is. Human error and prompting is a constant, frustrating struggle. 9:25 For sure. But here, the AI is correcting its own mathematical errors. Completely autonomously. 9:32 It is holding its own hand through the creative process. We all struggle with that drift. But this visual loop is revolutionary. It really is. It is essentially implementing machine vision as a feedback loop. That is so smart. It guides 9:46 the generative text process without you doing anything. Is it actually evaluating its own art? 9:51 Yes. It is running a continuous, autonomous visual comparison loop. So it looks, judges, and fixes. 9:59 It renders the math, critiques the output, and adjusts the underlying code. It acts as both the creator and the meticulous art critic. And it executes all of this without any human intervention. 10:10 Which saves so much time. The final output wasn't just a static image either. What was it? It provided a fully working, interactive 3D web viewer. Just ready to go. The user could rotate and inspect the final model immediately. That is a massive 10:24 leap forward in creative coding. It really is. 10:27 But you know, if neon can use math to generate a single dragon, can it generate an entire interactive universe? That is the next logical question. 10:36 It is one thing to render a static object perfectly. 10:39 It is another to build a world with physics and rules. Well, someone decided to test exactly that hypothesis on X. What did they do? They asked Neon to build a 3D military survival game. 10:50 Okay, a full game. They wanted the entire thing generated from a single prompt. And it had to run entirely in a standard web browser. Yes. 10:59 The technical requirements laid out in the prompt were extensive. I can imagine. They needed a highly detailed voxel art style. Right. It specifically had to use the 3JS framework. Which is pretty standard. Right. 3JS is basically a popular web 11:13 library used for rendering 3D graphics in browsers. 11:16 Right. And a voxel style is like digital building blocks. Exactly. Think of games like Minecraft with blocky pixelated 3D aesthetics. Yeah, that is the exact vibe. OK. The environment it had to generate was vast and detailed. What was in 11:31 it? It included dense forests, flowing rivers and fortified military bases. From one prompt. 11:38 Yes. It needed dynamic lighting to change the mood of the scene. That's computationally heavy. 11:44 It gets worse. It required soft shadows, volumetric fog, and active weather effects. That is an incredibly heavy render load for a browser tab. It is insane. 11:56 Browsers are notorious for having strict memory limits. Oh yeah, they crash all the time. You can't just dump a million textured blocks into Chrome. No, you definitely cannot. It will freeze your computer instantly. But it gets happier. 12:08 The game mechanics were also deeply complex. 12:11 Like what? It needed standard WASD keyboard movement controls. Right, for walking. It required fluid mouse controls for aiming a weapon. Okay. It generated enemy units with distinct patrol and attack behaviors. It had to manage inventory 12:26 systems, scoring, and enemy wave logic? Increasingly difficult enemy waves, yes. Wow. And it had to maintain a perfectly stable 60 frames per second. 12:35 In the browser? It had to do this on a standard desktop browser. That is wild. And it did it without requiring any external 3D assets. So no imported art. Everything was procedurally generated live. The capability on display here is frankly staggering. It is hard to wrap your 12:53 head around. It is handling high -level game design and spatial environment mapping simultaneously. 13:00 Plus, it is managing the underlying engine performance optimization. And it successfully implemented all of those core gameplay mechanics flawlessly. 13:07 It actually worked. The enemies flanked the player. 13:10 The weather obscured vision, the inventory system tracked ammunition perfectly. I need to understand the underlying mechanism here. Okay. How does it stop the browser from crashing with all those voxel details? It uses the 3 .js library incredibly 13:26 efficiently. How so? Instead of rendering every single voxel on the map at all times. Which would kill the memory. Right. It dynamically unloads objects the moment you look away from them. Oh wow. It effectively cheats the browser's memory 13:38 limits. That is clever. It only renders what is in your direct line of sight. Okay. It also streamlines the collision detection algorithms using spatial grids. It builds the performance optimization directly into the procedural code. 13:50 Exactly. It doesn't just write simple code that technically works. It writes good code. It writes robust, engine -level code that runs efficiently under immense stress. Sponsor. Do not write the sponsor read. Just include this exact placeholder 14:04 where the transition will occur. So, getting back to the community experiments. Right. Welcome back to The Deep Dive. We've looked at some massive targeted stress tests today. The big enterprise stuff. Yeah. Now let's look at what everyday 14:18 creators are building. This is my favorite part. 14:20 We are seeing incredible things out in the wild on X. The community experiments really highlight the practical applications. They really do. They also highlight the true hidden cost of using these models. Cost is the ultimate friction point 14:36 in AI development. It really slows people down. 14:39 If a prompt costs $10, you are terrified to experiment. 14:43 You don't want to mess up. Right. You spend hours meticulously planning a single prompt. It stifles creativity. It does. Take a user named NotJazzyOnX. 14:53 Okay, what did they do? They wanted to create a highly detailed sprawling pixel art city. A cool project. They tested both Gemini 4 Neon and Claude Opus 5 .5. To compare them. Exactly. 15:03 They gave both models the exact same architectural prompt. Word for word. They used the absolute highest reasoning levels for both systems. Right. 15:11 Claude Opus 5 .5 took 18 minutes to generate the city code. That's a long generation time. 15:17 It is. That single generation cost around $7 .20. Just for one prompt. Gemini 4 Neon took slightly longer, clocking in at 20 minutes. OK. 15:28 But it only cost around $3 .50. So it took two minutes longer, but cost half as much. Yes. And NEON actually produced a visually superior end result. Really? It handled the intricate building details much better than Claude. That is interesting. 15:42 It understood architectural consistency across the city blocks. It kept things logical. It also generated far more realistic water reflections in the canals. When the cost of generating a 3D world drops like that, AI shifts from being 15:55 a careful planning tool to an imagination engine. 15:59 That is a great way to put it. You aren't just saving money. You are buying the freedom to make mistakes. Exactly. That is a huge win for solo indie developers. Let's look at another striking example from user AI Screening. What did they 16:12 build? They created a complex 3D cycling animation using 3JS. Ooh, that's tough. They compared Neon, Opus 5 .5, and Gemini 4 Carbon side by side. 16:23 Moving characters are incredibly tricky to get right in code. They really are. The inverse kinematics usually look robotic or broken. They do. According to the user's detailed review, Neon vastly outperformed Opus 5 .5. Wow. It came very close to matching 16:38 Carbon's top -tier performance level. That is impressive. Neon particularly excelled at the animated character's body proportions. Getting the scaling right. Exactly. It kept the limbs scaling correctly during the pedal rotation. 16:51 That is hard to do with pure math. It also nailed the intricate mechanical details of the bicycle itself. That attention to spatial proportion and motion is critical. It really shows the reasoning power. There is one more fascinating project 17:04 we should discuss today. Oh, the ship. Yes. User Abhinav Flack built an interactive 3D historical sailing ship. This one is particularly impressive from a prompting perspective. Why is that? He built the entire complex scene with a single 17:20 prompt. Wow. He didn't provide any reference images or examples to the AI. He just described what he wanted in plain text. Just typed it out. 17:28 And the visual output on this one was genuinely stunning. You can see the tiny, accurate details on the ships rigging. The ropes and sails? Yeah. 17:36 You can see the realistic procedural movement of the ocean waves. The water physics looked great. The lighting reflecting off the water was highly dynamic and atmospheric. It looked like a real game environment. The creator actually rated Neon above GPT 6 .2 Sol for this specific 17:51 task. Let's synthesize all of this information for you. Good idea. Let's look at the final capability scorecard based on the data. Based on everything we have seen today, distinct patterns emerge. 18:01 What is the takeaway? GPT 6 .2 Sol remains excellent for massive long context understanding. Right. 18:08 Claude Opus 5 .5 is still a beast for complex systemic coding tasks. But Gemini for Neon is aggressively carving out its own territory. It absolutely is. It is currently dominating in generative 3D modeling. Without question. It is excelling rapidly in game development and 18:23 creative coding. Its visual output from single prompts is largely unmatched right now. And its cost efficiency sits right in the sweet spot. 18:31 It is very affordable. Does this cost advantage change how creators will use AI? It fundamentally changes the entire development cycle for creators. 18:40 How so? Lower costs allow for far more rapid trial and error. You aren't afraid of the bill. 18:46 Exactly. You don't have to stress over perfecting every single prompt. Right. You can iterate, break things, and rebuild them cheaply. Cheaper iterations mean you can experiment and fail much more freely. Exactly. You can explore strange, 18:59 weird ideas without draining your API wallet. 19:03 Let's step back and look at the big picture here. 19:05 It's a big shift. Gemini 4 .0 Neon isn't just generating clever text anymore. No, it's way past that. It is autonomously architecting complex interactive 3D realities from scratch. Right before our eyes. It is building universes from raw underlying mathematical primitives. And it 19:23 is doing it highly efficiently and affordably. 19:25 Which is the crazy part. The traditional barrier between raw imagination and interactive reality is collapsing. Google has officially made the AI race hotter than ever. They really brought the heat. They have definitively shifted the 19:38 battleground. We aren't just fighting over text reasoning anymore. Nope. The new war is over autonomous visual creation and world building. 19:46 I have a specific challenge for you listening right now. Let's hear it. Think of a creative project you previously abandoned in frustration. 19:53 We all have one. Something you dropped because it just seemed too technically difficult. We all have a dusty folder full of those ambitious ideas. Take that exact idea today and dust it off. Bring it back to life. Feed it to Gemini 20:03 4 Neon. Just paste it in. Push the thinking level up to X high. And see what happens. Why not? 20:09 You might be completely shocked by what you can suddenly create. That is a great, actionable challenge to leave everyone with. And will people try it? But I want to leave you with one final lingering thought. OK. If an AI can already generate 20:23 an entire playable game environment from a single prompt in just a matter of minutes. Which is wild. How long until the video games and digital worlds we inhabit are entirely bespoke? No. Generated live, in real time, just for us. using pure math. 20:39 It is a beautifully wild thought to end on. Keep building and keep experimenting. Catch you next time. Thank you for joining us on this deep dive. 20:47 We'll see you next time.
Community discussion