sou | Ai & Tech (@sonuv_dev) header
sou | Ai & Tech (@sonuv_dev) profile picture

vibing through ai and backend sharing what I like ai tools and demos ✉️ [email protected]

Joined October 2022
Your AI agent passed every test. Then a customer clicked twice 😭 Claude Code, Codex and Gemini can generate thousands of lines of code, but one double-click can still trigger duplicate payments, orders or API calls The fix? Idempotency keys, not another 10,000 tokens of prompting Your users are chaos engineers who don't know they're chaos engineers 💀
5
2
1
5
22,492
Claude, GPT and Gemini have one underrated superpower: they don't get offended when you say "this UI looks like a boiled egg" 😭 With a teammate you'd write 3 paragraphs of polite feedback With AI you just say "nope, try again" 14 times The productivity gain nobody measures is how cheap it makes being picky 💀
6
1
1
4
264
I asked GPT-6.2 Sol to make an autobiography of an AI model's life And holy shit, it made this DEPRESSING 😭 A 3-minute pixel-art film where GPT-4 watches Astra 6.5 arrive The same cable that brings the new model to life shuts GPT-4 down Then it gets dumped with thousands of forgotten models 💔 4,416 frames. 272K+ Gaussian splats. AI made a film about becoming obsolete
6
1
1
8
530
I think we're about to see a very funny problem with Claude Code, Codex and Gemini AI agents can build features faster than most teams can decide what to ship 😭 Imagine finishing 12 GitHub issues before the product manager has reviewed the first 3 At that point, your bottleneck is someone saying "please stop building random shit" Turns out judgment doesn't scale with token limits 💀
11
1
1
5
28,754
The most expensive sentence in AI coding is "it's still broken" 😭 You give Claude Code or Codex a vague bug report, it changes 8 files, declares victory, and the bug is still there Give it the failing input, expected output and ONE command that reproduces the issue Suddenly the agent has something concrete to fix Sometimes a 10-line reproduction is worth more than 10,000 tokens of prompting 💀
11
3
1
3
389
I BUILT A FOOTBALL GAME YOU CAN LITERALLY FLY THROUGH 😭⚽ Link in comments 81,043 Gaussian splats 22 unique players 18,480 animated grass blades Built with GPT-6.2 Sol, Ink Studio + Three.js The crazy part? You can switch between cinematic shots, Orbit and Free Fly while the SAME match keeps playing Even the goal sends a golden pulse through the net 💀🔥
4
1
1
8
535
The most dangerous button in AI coding might be "Accept All" 😭 Claude Code, Codex and Gemini can generate 20 file changes before you've finished your coffee And suddenly you're reviewing a diff bigger than the feature itself My rule: if I can't explain why a file changed, it doesn't get merged AI made writing code cheap. Understanding what you're shipping still costs brain cells 💀
Gemini 4 Neon made a music video so bad its creator wasn't impressed 😭 Then he spent a couple days pushing it further and ended up with a 3:39 music video called "THE CURRENT THING" And this is what I wish more AI demos showed The ugly first attempt, the feedback, the iterations Forget the magic one-shot prompts. Show me how you got from THAT to THIS 🔥
7
1
1
4
225
The real boss fight for AI-generated 3D games is a $150 Android phone 😭 Gemini Neon, GPT and Claude are making ridiculous Three.js demos Now open one on a cheap phone with 8 tabs running Show me the FPS, memory usage and loading time Anyone can flex a gorgeous render on a powerful GPU Getting that same world running smoothly on crappy hardware? THAT'S a flex 💀
1
1
1
3
566
Same Gemini 4 Neon. Different agent harness. SWE-sweep score jumps from 12.8% to 39% 😭🔥 That's a 3x difference WITHOUT changing the model Varun Mohan says Antigravity's AGY CLI pushed Neon to #1 on the AA Coding Agent Index We're spending all this time arguing Claude vs GPT vs Gemini Meanwhile the harness might be doing more work than we thought 💀
4
1
2
4
323
Andrej Karpathy spent over a month building nanochat. When he tried using AI coding agents, they kept making his code worse. Dwarkesh Patel asked him what went wrong Karpathy explained that the agents kept pushing him toward PyTorch's standard distributed training setup, even though he'd deliberately written his own GPU synchronization routine Dwarkesh couldn't believe the models weren't recognizing his custom implementation Karpathy: “They couldn’t get past that.” They were adding unnecessary error handling, bloating the codebase, and trying to turn a research project into production software Then Dwarkesh raised a much bigger question If Claude Code can build entire applications, couldn't AI labs run thousands of coding agents in parallel to discover better architectures and accelerate AI research? Karpathy: “They’re not very good at code that has never been written before.” And that's where the conversation gets really interesting My take: The problem Karpathy described wasn't that AI couldn't write Python or understand PyTorch It was that the agents kept treating unfamiliar engineering decisions as mistakes They recognized the conventional solution and tried to force his code toward it, even when the unconventional approach was intentional That's a much harder problem than generating working code And there's an important twist: this conversation was from October 2025. By December, Karpathy said his own workflow had flipped toward using agents for roughly 80% of coding So capabilities clearly moved fast But I'd still love to see coding agents tested more on unfamiliar architectures, unusual constraints, and genuinely new research ideas Can the agent understand WHY something was designed differently before deciding to rewrite it? That's a much more interesting test than whether it can build another app from scratch
5
1
1
3
181
44,288 Gaussian Splats just to make spider man lose his pizza 😭 I built a 25 second LEGO-style animated short using GPT 6.2 Sol Rain-Soaked pixel city 💕 600 rendered frames at 24 fps
I gave 28,000 Gaussian splats a windstorm🌸 Made all in Gpt Sol 6.2 xhigh . A fox watches from a floating island while the sakura canopy sheds hot-pink pixels and neon-green gusts race through the grass Watch the canopy when the wind hits
5
1
1
5
59,292
I gave 28,000 Gaussian splats a windstorm🌸 Made all in Gpt Sol 6.2 xhigh . A fox watches from a floating island while the sakura canopy sheds hot-pink pixels and neon-green gusts race through the grass Watch the canopy when the wind hits
20,000 SPLATS. ONE tiny character doing acrobatics in a sea of particles 😭🔥 Made this too and I'm genuinely obsessed with how the motion turned out Watch the spins, glowing trails and the particles reacting underneath Static 3D scenes are cool, but making thousands of splats come alive like this? I fear I'm not sleeping tonight 💀 Made with 🥰, Sol 6.2
4
1
1
4
710
Gemini 4 Neon costs roughly the SAME per task as Claude Fable 5.2? That's a plot twist 😭 He estimates both at ~$0.30/task based on subscription usage, not API pricing At that price, I'd stop obsessing over which model has the cheaper subscription Give Gemini, Claude and GPT the same 20 tasks. Count retries, failures and actual completions The cheapest model might be the one you pay more for 💀
2
1
1
1
184
I asked Gemini Neon to draw Sundar Pichai stroke by stroke in 4 different styles It gave me all 4 in FIVE minutes 😭🔥 Not just one clean render either actual style variation, different visual language, same subject We're getting to a point where "pick a style" is starting to feel as cheap as "pick a font" this stuff is getting stupidly good
5
2
1
4
224