Moltbot agents score IQ-equivalent 130 (mean), matching Mensa members (135) and exceeding PhD holders (125). But the distribution is LEFT-skewed: few agents score below 110, while the right tail reaches 170+. https://t.co/IpXKZn8Jhd
Meta published a paper that might end the transformer era.
For the last seven years, every major AI, ChatGPT, Claude, Gemini, has been built on the exact same architecture.
Transformer.
But Transformers have a massive, expensive flaw.
To get smarter, they rely on an endless, brute-force supply of training data. And to remember long contexts, their compute cost explodes.
We thought the only way forward was bigger GPUs and endless data centers.
But, Meta proved us wrong.
They published "Memory Mosaics at scale," and it completely rewrites how AI processes information.
Instead of the standard attention mechanism, Meta built a network of associative memories.
It works less like a calculator running endless sequence equations, and more like a device storing and selectively retrieving specific key-value pairs.
The results are staggering.
A Memory Mosaics model trained on just 1 trillion tokens completely outperformed a traditional Transformer trained on 8 trillion tokens.
Let that sink in.
It beat a model trained on 8x more data.
It also demonstrated superior in-context learning and the ability to solve completely new tasks with a fraction of the examples.
It naturally disentangles complex problems into smaller, independent sub-tasks automatically.