直播预告 | 滑铁卢大学邓云天教授：0.6B何以超越32B？神经编译技术解析
Merged summary
TL;DR - PaperWeekly announced a talk on Program-as-Weights, a “neural compilation” approach that turns a language-described function into a reusable local neural program. The announcement claims a compiled 0.6B-parameter model can outperform direct calls to a 32B model for targeted tasks.
- Functions are compiled once rather than invoking a large language model for every input.
- Compiled programs run locally like ordinary Python functions.
- Demonstrations include word games, language-controlled 3D characters, and website assistants.
- The broader aim is to make large models builders of reusable tools instead of repeated problem solvers.
Sources (1)
直播预告 | 滑铁卢大学邓云天教授：0.6B何以超越32B？神经编译技术解析
TL;DR - PaperWeekly announced a talk on Program-as-Weights, a “neural compilation” approach that turns a language-described function into a reusable local neural program. The announcement claims a compiled 0.6B-parameter model can outperform direct calls to a 32B model for targeted tasks.
- Functions are compiled once rather than invoking a large language model for every input.
- Compiled programs run locally like ordinary Python functions.
- Demonstrations include word games, language-controlled 3D characters, and website assistants.
- The broader aim is to make large models builders of reusable tools instead of repeated problem solvers.