3 ms·
"L"LM. No way this model will be "large" in any modern sense. For mobile devices both RAM and power consumption are very limited.
by cubefox 3y ago
"L"LM. No way this model will be "large" in any modern sense. For mobile devices both RAM and power consumption are very limited.
- candiodari 3y agoSome terms don't age well. "Deep" neural networks used to mean "more than one layer". Ie. 5 emulated neurons or more. What makes Language Models large, originally, was meant a model that emulates behavior that before needed to be programmed in explicitly, as size increased. The first time this was observed and described, it was in GPT-2, with 1.5B parameters. You can do this on current mobile devices with current technology (ie. 4-bit quantization).
- cubefox 3y agoBut such small models are too dumb to be acceptable for conversation. Far more likely that there will be an API call to an actually large model in the cloud. If you really don't have any network connection, you are simply out of luck. It's not as if transferring a little bit text would require a lot of bandwidth anyway.