Sabitlenmiş Tweet
Tried my first fully "Local AI" vibe coding with @unity experiment tonight. DeepSeek V4 Flash (@MiaAI_lab recipe for 2x sparks) + MCP for Unity + Box3d for Unity with @pidotdev as the harness.
English
Iman Mostafavi
1.4K posts
Iman Mostafavi
@imost
Previously co-founded 50-person game studio (100M+ downloads) | Currently AI product engineering | Local AI (2x sparks)
San Jose, CA Katılım Mart 2009
201 Takip Edilen946 Takipçiler
@gregisenberg This is the correct take for an entrepreneur. If you’re not skewing optimistic, you’re ngmi.
English
There's panic right now that AI is eating indie software. I see it differently.
For 10+ years you could spin up a few hundred auto-generated pages, rank them, and let Google send you signups for free.
Now the answer people used to click your site for sits right inside ChatGPT and Google's AI overviews.
If your whole business ran on that one faucet, the water's turning off, and it feels like the end of the world.
And I've built a few of these projects, so I feel that darkness too.
But let's look at the data. Has it really gotten dark/worse for software founders and indie devs??
Let's look a Stripe Data...
New companies on Stripe Atlas are up 130% year over year. Companies are hitting $10M ARR within three months of launch at double last year's rate. One in five charges its first customer inside the first month, up from 8% in 2020. Fastest business formation anyone's ever recorded!!
And the solopreneurs, the exact people saying they're getting crushed, are winning the most.
63% of new companies on Atlas are solo founded, and AI native solo startups pull 2.3x the revenue by month 24. This category was basically non existent at the scale we're seeing rn like 5 years ago.
TLDR;
So my thinking it's a channel shift wearing a collapse costume. The people hurting share one setup: they sell inexpensive tools to other indie hackers, and/or their traffic comes from SEO.
That's the most exposed spot in the whole market, because your customers can rebuild your $20 SaaS in a weekend, and your traffic source is the exact thing the models replaced. Kinda feels like 2 cannons pointed at one little boat.
Call me an optimist, but here's my take....
Building software is a bigger game than it's ever been.
More companies, forming faster, making money sooner, run by fewer people, than any point in history.
@levelsio@levelsio
I'm seeing a trend here of declining revenue and traffic with indiehackers On my own projects too Maybe big VC products too but I wouldn't know cause they don't share revenue To me it seems clear BigAI is cannibalizing everything that used to be apps Not bad btw, just times are changing and we have to adapt
English
@BosonJoe @RyanFogleOP12 Just tried to submit a benchmark for your GLM 5.2 recipe but got this. Let me know if there's anything I need to change on my side.
English
First contributor to howtospark!
Thanks @RyanFogleOP12
Check out the MCP server to join up 🤝
Ryan Fogle@RyanFogleOP12
It is here let’s go! @nvidia @NVIDIAAI #LocalAI #NVIDIA #AI #DGXSpark
English
My current favorite 2 Spark recipe is by @MiaAI_lab but I can’t seem to replicate the ~66 tok/sec single stream speed. I get something in the low 50 range.
Is there a standard benchmark to check if there is something misconfigured on my cluster setup?
github.com/MiaAI-Lab/Deep…
English
@u1tra_instinct Really looking forward to trying this when @BosonJoe is finished cooking.
English
Joe has been doing some phenomenal work pushing glm5.2 on Dual DGX-sparks to now from 14 initially . 24.7 toks/sec is amazing for such smart model with large parameter. He is helping everyone who has two sparks to run usable local intelligence at home.
He is not going to be satisfied until this goal is achieved please follow him if you support his work
Joe Muller@BosonJoe
GLM 5.2 on 2 @NVIDIAAI DGX Sparks @ 24.7 tok/sec 🚀 Next step is to claw back some quality while keeping the speed steady We've come a long way. When I stared it was 4 tok/sec 🐌
English
@Launch1Labs Congrats! May not matter much since the GX10 has pretty good cooling, but they will run slightly cooler (1.4 C at load) if you flip it onto the other side (power button down). Might not be as aesthetic though.
English
Second GB10 secured.
This is what a future development team looks like:
Two NVIDIA GB10-powered systems running large AI models locally.
Mac Studio + Mac minis orchestrating agents.
Raspberry Pis serving as always-on operators for monitoring, automation, and network tasks.
UniFi using VLAN segmentation to isolate agent workloads and secure access across the system.
The org chart is starting to look a lot more like a network diagram.
English
@EMostaque I’ve had it make my MBP just hard power off. Only app that does this.
English
@Daniel_Farinax Have you switched over to the recently released Unity CLI?
English
The Unity game being developed by Grok Build in a loop is coming along really well. The quality jump from ThreeJS to Unity is mindblowing.
Playable free demo coming soon 🔥
Dan@Daniel_Farinax
Do you remember the first game I made with Three.js and Grok (via web) by copying and pasting the entire output from the web chat as a single HTML file? This was back in March 2025 when a few of us on X started the Three.js Vibe Coding trend. I’m rebuilding the game with Unity now, and with the powerful Grok 4.5. Crossing fingers 🤞
English
@JensenHuang @nvidia A DGX Spark (or two) belongs on every developer’s desk to run open models. Thanks for this incredible device.
English
For my first post, I’m sharing a letter @NVIDIA signed on why open models matter.
AI will transform every industry, power every company, and be built by every country.
Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty.
The world needs both frontier closed models and frontier open models.
images.nvidia.com/pdf/Open-Weigh…
English
Nice try Claude prompt-suggestion, but I tell you when we're done.
English
@BosonJoe @0xgilbert Hey Joe is there any reason you went with FP8 over NVPF4?
English
@0xgilbert Here's my recipe for dual sparks:
howtospark.com/recipes/laguna…
English
Super excited about this—a 118B model that punches above its weight in coding tasks! Stronger than Deepseek 4 Pro Max while also cheaper and faster. Win + Win = ?
UsePod.ai@UsePodAI
Laguna S 2.1 is live on UsePod! Poolside launched Laguna S 2.1--an open-weight Mixture-of-Experts coding model that beats rivals 10x its size. It's live today on UsePod. Enjoy 118B total parameters with only 8B active per token, 1M token context window, and frontier coding performance: 78.5% on SWE-Bench Multilingual (topping the published leaderboard), 70.2% on Terminal-Bench 2.1, and 40.4% on DeepSWE v1.1. Built specifically for agentic software engineering and long-horizon coding tasks, with max thinking mode that lets the model set its own test-time compute budget. The efficiency is remarkable--8B active parameters punching at the level of models with 80B+ active parameters, small enough to run on a single NVIDIA DGX Spark, yet capable enough to compete with the largest coding models. Open-weight under OpenMDW-1.1 license, designed for complex workflows, interactive development, and multi-step software engineering. Available now through UsePod's marketplace at 20% market pricing.
English
Added more balls and a few fun props (generated locally with TRELLIS.2) to the pit, this time with my agent driving it through the new @unity CLI.
It's still experimental, but it seems to work really well already. Feels faster like it has more control over the Editor compared to both the official Unity MCP and the most starred open source alternative, MCP for Unity.
If you're a Unity dev, just point your agent to the docs URL and it should get you going.
Unity CLI
docs.unity.com/en-us/unity-cl…
TRELLIS.2
github.com/microsoft/TREL…
English
@pbteja1998 What’s the reason you won’t allow in app purchases? You don’t want to give Apple a cut, or something else?
English
Got rejected for the 2nd time for SiteGPT.ai iOS app - not sure what to do about it...
I don't want to add payments inside the iOS app... but don't know if there is anyway around this...
English
@chongdashu Are you also using the official Unity MCP or just Unity CLI directly controlled by your agent?
English
I gave the Unity CLI a try - and it's GOOD!
Connected with Grok 4.5 via Grok Build - it handled the Editor by itself.
Took the 2d platformer template and told Grok to give it a visual uplift, while keeping the game mechanics.
Spriterrific for the character animations and Nano Banana 2 lite for the static sprites.
I think this is a huge unlock for gamedevs, especially those who want to use Unity!
Unity@unity
We’re excited to introduce the Unity CLI, our new terminal-native path for connecting coding agents, CI pipelines, and custom tooling to Unity. Unity MCP remains fully supported for your existing setups but to help with the transition, the CLI includes an "MCP Mode," allowing you to leverage existing MCP-based agents without starting from scratch. We want to remove barriers to innovation, so we are making these tools free! Both the CLI and MCP will be free, with no concurrency limits on the MCP server.
English
@MiaAI_lab @u1tra_instinct How many tok/sec single stream are you up to with GLM 5.2 on 3 sparks now? I need to start budgeting for spark #3
English
The DeepSeek V4 Flash recipe by @MiaAI_lab they’re adapting for this article happens to also be the one I’ve had the best experience with on my sparks.
Miguel Salinas@Vercantez
English
I gave Laguna S 2.1 another shot since they updated the weights today. Here's "Flappy Laguna" in 1 prompt (0 retries) on Laguna S 2.1 FP8 Dual DGX Spark.
Recipe @bosonjoe howtospark.com/recipes/laguna…
The bird turned out a little funny looking, but otherwise it's a functional game.
English