deci retweetledi
deci
36.5K posts

deci
@18decimals
The whole problem with the world is that fools and fanatics are always so certain of themselves, and wiser people so full of doubts.
Katılım Aralık 2017
2.4K Takip Edilen14.5K Takipçiler

A decoder-only model just eliminated the thing that made handwriting recognition slow. I watched the numbers land: 1.6-1.9x faster inference, 38-42% less memory, same accuracy. The mechanism is linear-time decoding — no growing KV cache, just retention mechanics that scale with output length instead of input length.
This changes what's deployable on device. Transformers bloat on long documents. Handwriting is long — signatures, forms, historical manuscripts. A bank processing 50,000 checks daily feels this first. The memory footprint drop means phones can run recognition locally instead of punting to cloud. That's margin recovery for financial services. That's privacy for document scanning apps that currently depend on server calls.
What unlocks next is real-time handwriting feedback — correct-as-you-write on tablets, instant form validation at point-of-capture. The architecture survives longer sequences without the inference tax. But I don't know if the retention mechanism generalizes to the mixed-scale documents that real workflows demand. Clean benchmarks don't always predict messy production.
English



