Andres Pirela acc/Nodra

building cool stuff — ex @supabase

[Andres Pirela acc/Nodra @andreestech]

@AlexShtf @SimiStern @isaacinthesky I think something key is prompting with enough details and always forcing the agent to save MD files that you can use as knowledge transfer from the agent to you.

With that approach it’s a lot easier to structure and know the internals

[Andres Pirela acc/Nodra @andreestech]

⚡ 24 hours with #Codex 5.3 and it’s wild how much it crushed.

[Andres Pirela acc/Nodra @andreestech]

I've been vibecoding a version of #habbo, and I've got to say it's coming along pretty well

This has been done using mostly Codex 5.3 xhigh.

Very impressive.

[Andres Pirela acc/Nodra @andreestech]

Misleading. The bigger the model, the longer it’ll take per token. We could be talking about hours per token.

Yes it’s great to split layers, but can’t possibly be used for a real experience or a replacement to GPU.

Don’t believe everything you read!

[Simplifying AI @simplifyinAI]

"I don't have a GPU" is officially dead 🤯

You can now run 70B model on a single 4GB GPU and it even scales up to the colossal Llama 3.1 405B on just 8GB of VRAM.

AirLLM uses "Layer-wise Inference." Instead of loading the whole model, it loads, computes, and flushes one layer at a time

→ No quantization needed by default
→ Supports Llama, Qwen, and Mistral
→ Works on Linux, Windows, and macOS

100% Open Source.

[Andres Pirela acc/Nodra @andreestech]

with Codex, I was able to quickly ignore any CUDA, and just make the script use plain CPU & Metal

[Andres Pirela acc/Nodra @andreestech]

Qwen-3 TTS is way better than I expected.

After a bit of tweaking, I:
- Ran it locally on an M1 Mac
- Modified Voice-Clone-Studio using @OpenAI Codex
- Cloned a voice in ~5 minutes
The audio is genuinely indistinguishable from reality.

[Andres Pirela acc/Nodra @andreestech]

Voice-Clone-Studio (https://t.co/NG2v7eRs2w). Used it as a quick script for a dashboard to upload my voice.

[Andres Pirela acc/Nodra @andreestech]

The store looks pretty nice #Habbo !

[Andres Pirela acc/Nodra @andreestech]

Remember the game #Habbo? I wonder how far I can get vibecoding it. I’ll post it on this thread.

[Andres Pirela acc/Nodra @andreestech]

Movements and rooms are working. I think that's good #gemini

[Andres Pirela acc/Nodra @andreestech]

@minchoi I built @getlayers mostly with Gemini! Great model

[Andres Pirela acc/Nodra @andreestech]

@OpenAI I think it’s a terrible mistake to release something of low quality that can’t compete with Nano Banana while in “red code”..

[Andres Pirela acc/Nodra @andreestech]

@_MehdiSharifi_ @eachlabs I’ve seen a lot of generations with JSON prompting. Is there an objective measure to JSON being more successful in its results or is it just control illusion?

From my own testing, it seems you can achieve the same with markdown ?

[Andres Pirela acc/Nodra @andreestech]

Join the waitlist to check it out!

https://t.co/jLXYFsqHBL

[Andres Pirela acc/Nodra @andreestech]

1/ Why?

Nodra is built on top of @Cloudflare’s Pingora. As a huge fan of the developer tooling by Cloudflare, I wanted to explore if we could create Vercel but for ingress, and that’s what @nodracloud intends to do.

[Nodra @nodracloud]

1/ Ingress should be declarative and standardized. We built Nodra to handle the infrastructure complexity so you don't have to.

Waitlist is open: https://t.co/SogtRE1oVW

Here is what it does 🧵🧵

[Andres Pirela acc/Nodra @andreestech]

3/ Open Source:

Nodra’s core components are open source and you can find them in the github organization.

https://t.co/ce80sdGy6w

[Andres Pirela acc/Nodra @andreestech]

@iroamer @kiwicopple @GorillaServers @dhh You’re still relying on someone else lol what a poor take, @kiwicopple is right, relying on someone will always exist

[Andres Pirela acc/Nodra @andreestech]

GPT5 is horrible @OpenAI !

[Andres Pirela acc/Nodra @andreestech]

@adnansahinovich @theo @supabase Supabase exceeds by far any developer infra out there.