will to code

Today I wanted to make something with gpui then I realized that I didn't understand it enough So I built a to-do list app Super basic super simple you can just look at the codebase and understand what gpui is all about I am sooooo bullish on this framework GitHub Also discovered this amazing library. Its the shadcn for gpui. gpui-components

Cloudflare is down again lol Someone should really build this: One click infra migration

#friday

New deepseek model gpt-5 and gemini-3.0 level performance while being 10x cheaper It cost them 600k upto a mill to go from 3.1 to 3.2 which rivals frontier models that were trained from scratch for 100s of million It's over And it will be even more over when deepseek trains a new foundation model. The current one is Deepseek V3 from late 2024.

Another cool model It's a GUI agent that can use a phone and follow instructions accurately 4B params but still outperforms larger SOTA models

AI is humanity's great project This is interesting from deepseek. They make one agent generate mathematical proofs and another agent grade it's reasoning. They completely ignore the answer. My mind is blown. AGI is solvable only through the game theory we introduce in to agentic systems

I'm so pissed off I can no longer continue watching this How is it fair to put a system that sees millions of moves ahead against a human whose brain has very limited compute Limit the search depth of these systems to human level and they'll crumble Humans are yet to be defeated We are the master species 🧠 It's so disappointing to see the general consensus among everyone being "computers have defeated humans" I despise this whole narrative and will never accept it because it's built on lies and distracts from real AGI benchmarks and goals.

In programming we often say be efficient don't repeat yourself but does nature share that same sentiment

TL;DR: Nature is efficient when it needs to be but it fundamentally cares more about survival than elegance survival comes first, efficiency only matters when it helps survival

ML is so beautiful man 🫠 It's where math and cs meet together and become art

I just finished understanding how the image encoder works 😅 Ya this project ain't getting finished today

#friday

SuperTonic a TTS model that can run on an e-reader Huggingface | Website

I ran qwen-3 0.6B model which is only 500mb on my laptop and it generated this website for a restaurant we are at the precipice of something

Insane price to quality ratio

There's so much stuff coming out right now its hard to keep up $0.0025/image 0.84s/image ⚡️ 6B parameter https://huggingface.co/Tongyi-MAI/Z-Image-Turbo

So far African tech bros were making a very decent living doing the usual programming jobs. Now AI is one shotting them and we will all be forced into doing the real hard technical jobs that AI isn't good at Net benefit for the continent everything is e/acc everything is adaptation

Great ads worked because they were rare and hard to make If everyone can make great ads then people will just get desensitized to it Ads already have such a low conversion rate. The reason that huge companies like coca cola show ads is to subconsciously influence your decision making. They are maintaining a brand not trying to grow it. So just because you can make an amazing high quality ad, it doesn't mean you'll get users. I'm very skeptical on media generation. I'm willing to change my mind but only when I see it's utility in the future. Let the dopamine rush die out and let's see people use it then.

Today's project was an interesting experiment When Copilot first launched, LLMs were relatively weak and slow. We were forced to train custom, small-parameter code completion models, and that became the industry standard. But why are we still doing that? General purpose llms have become powerful and we now have inference providers that deliver 2000+ tok/sec So, I decided to test a new approach. I used gpt-oss-120b on Cerebras to build a tab-autocomplete extension for VSCode. Its as fast as copilot and sometimes faster (with reasoning enabled lol) The major advantage here is flexibility. For instance, I was able to get the model to treat comments as prompts and actionable instructions, rather than wasting tokens (like copilot) auto-completing them. The sky is the limit here. This approach could eventually replace chat interfaces, have a auto-complete.md, and make agentic decisions. Much more to explore in the future! 🤟 [Code]

#friday is here