1/ You can shrink a language model's KV cache by 200×, in a single forward pass, and it still answers correctly.
At 256k context that's 36 GiB of cache down to ~360 MiB, with no change to the base model.
Here's how we did it 👇
🚨 Holy shxt… The rules of AI image generation just completely changed.
A GPT-based model currently testing in the arena under the bizarre name 'duct-tape' is turning the global AI community upside down. What exactly did they feed this thing?
→ Native-level text generation with zero awkwardness. → Chilling consistency maintained down to the pixel. → Overwhelming illustration quality ready for immediate commercial use.
"Is it just downloading photos from the internet?" That was every tester's first reaction. It's that unbelievable.
A game-changer that instantly crumbles Nano Banana Pro's dominance has arrived. A lot of people might need to start packing their desks again.
The 10 images below were 100% drawn by AI. See for yourself 🧵