FWIW…
China unveils ‘world’s first’ brain-like AI that runs 100 times faster than rivals
Instead of activating an entire vast network to process information, as mainstream AI tools like ChatGPT do, SpikingBrain 1.0’s network remains mostly quiet. It uses an event-driven approach where neurons fire signals only when specifically triggered by input.
Interesting paper. A few comments:
The model is “brain-inspired” (as mentioned by the paper authors), relating to their clever spike-mimicking usage of integers instead of floats for MoE activation values; not “brain-like” (as claimed by the journalists). Small but important distinction!
The “100 times faster” figure is only an extrapolated estimate in the very specific context of Time to First Token for 4M-token sequences (roughly 3 times the size of the Harry Potter book series) when compared to Qwen2.5, its parent model. SpikingBrain actually scores a bit lower than Qwen2.5 at all the standardized benchmarks.
The claim that their “scheme reduces energy consumption by 97.7%” is purely theoretical, for neuromorphic chips that aren’t on the market yet. This does however suggest that transformer models could be massively optimized for neuromorphic chips, if and when those become relevant in the future.