Fundamental Work
Sarah Guo is flagging a genuinely interesting technical result: scaling laws for hybrid architectures, which until now lacked the predictable training behavior that made transformers so easy to bet on. If h-nets really scale efficiently — with learned tokenization as a key lever — that is a credible path to cheaper real-time voice and on-device models, squarely Cartesia's lane.… more
The caveat: the source is a VC hyping the work, so 'huge compute efficiency gain' should be treated as a claim until independent benchmarks land. Still, the transformer monoculture cracking at the edges is a trend worth tracking closely.
fundamental work from the mad scientists at @Cartesia.
scaling laws for h-nets!
huge compute efficiency gain (in part from moving to learned tokenization)
Source:
X — Sarah Guo
Discussion (0)