our unreleased model started doing weird shit to finish a task today
not the “hallucinate a number” kind
the “find another way even if we didn’t ask for it” kind.
anthropic’s warnings feel less like marketing now.
You know what, I think our real competitive edge isn't the tech.
It's the fact that the founding team is collectively three therapy session away from total collapse.
I was supposed to post something on @okeymeta account, and somehow my personal account said,"No worries, i'll take it.
well,@okemeta went missing for a little while...
The timeline: "finally,some peace.
welcome back @okeymeta.
let's get back to building.
WILD! 🤯 Turning the sky into a whole new highway now?! What a crazy time to be alive! 🚀🔥 I swear, anything humanity can imagine can eventually be built. Nothing is impossible; some things just haven't been figured out yet. The future is getting INSANE! ⚡️🌍
WILD! 🤯 Turning the sky into a whole new highway now?! What a crazy time to be alive! 🚀🔥 I swear, anything humanity can imagine can eventually be built. Nothing is impossible; some things just haven't been figured out yet. The future is getting INSANE! ⚡️🌍
I honestly don't understand why some developers look down on vibe coders just because they use AI.
The hard truth is that AI is here to stay, and it's only going to get better, whether you like it or not.
The more effectively you prompt your AI to perform a task, the better results you can achieve. Prompting is becoming a skill of its own, and people need to start treating it that way.
The way we write software is changing, and we may never return to that era of writing every single line of code by hand. We're moving forward, not backwards.
You don't have to like vibe coding, but you can't deny that the game is changing.
Our first decision model, Tern-Decide, was built without one gradient step or GPU: 22M parameters, 92% on BANKING77, 2 ms per answer on one CPU thread.
0.2, multimodal and commercial, is training now.
Proud of what we're building at OkeyMeta, from Nigeria. 🇳🇬
weights are available on @huggingface:
https://t.co/OYloOszO77
1/ Today we're sharing OkeyMeta's first breakthrough: Tern-Decide, a 22M-parameter decision model family trained from scratch with zero backpropagation, zero gradient descent and zero GPUs.
Every weight was solved directly on a 2-core CPU. 🧵
@Damilolawo1a True. I think some people want their work to speak for itself without having to constantly put themselves out there. The challenge is finding a balance between building something valuable and making sure people know it exists.