feat: periodic or immediate finetuning #128

New Issue

GiteaMirror · 2025-11-11T14:07:12-06:00

GiteaMirror commented

2025-11-11 14:07:12 -06:00

Originally created by @m-c-frank on GitHub (Dec 27, 2023).

Is your feature request related to a problem? Please describe.
i think we can start focusing on aligning the model to ourselves now

Describe the solution you'd like
i think it should be possible to take the interactions you have with the models and use them as training data fairly frequently

Describe alternatives you've considered

to manually do all that with a bunch of scripts or additional cli tools
do it immediately but that might be too energy expensive for now?
outsource the finetuning via some automation pipelines to a cloud provider
do it all locally on a consumer grade machine

Additional context
im working on modeling llms as nodes in a mesh similar to a mycelial mass

my thought was that hyphal networks seem to be much more scalable than neural networks

i mean the largest neural network is probably from a whale somewhere

but if we can combine the largest living organism on earth with the creativity of large language models

then maybe we can scale it much faster

and i want my personal network that can understand me and i need it to be private and local

thats why i think this is probably the easiest way to immediately the exponential growth phase

idea

imagine you have a conversation with some model and you have to iteratively nudge it to the right answer.

we can have a pipeline that uses large language models to identify the core question and the core answer each as one embedding

this way we can teach it to think exactly like us i think

i'd love to see this happen

Originally created by @m-c-frank on GitHub (Dec 27, 2023). **Is your feature request related to a problem? Please describe.** i think we can start focusing on aligning the model to ourselves now **Describe the solution you'd like** i think it should be possible to take the interactions you have with the models and use them as training data fairly frequently **Describe alternatives you've considered** - to manually do all that with a bunch of scripts or additional cli tools - do it immediately but that might be too energy expensive for now? - outsource the finetuning via some automation pipelines to a cloud provider - do it all locally on a consumer grade machine **Additional context** im working on modeling llms as nodes in a mesh similar to a mycelial mass my thought was that hyphal networks seem to be much more scalable than neural networks i mean the largest neural network is probably from a whale somewhere but if we can combine the largest living organism on earth with the creativity of large language models then maybe we can scale it much faster and i want my personal network that can understand me and i need it to be private and local thats why i think this is probably the easiest way to immediately the exponential growth phase ### idea imagine you have a conversation with some model and you have to iteratively nudge it to the right answer. we can have a pipeline that uses large language models to identify the core question and the core answer each as one embedding this way we can teach it to think exactly like us i think i'd love to see this happen

GiteaMirror closed this issue

2025-11-11 14:07:12 -06:00

GiteaMirror commented

2025-11-11 14:07:15 -06:00

@m-c-frank commented on GitHub (Jan 12, 2024):

seems like openai was working on it aswell: https://twitter.com/gdb/status/1745124363362578527

@m-c-frank commented on GitHub (Jan 12, 2024): seems like openai was working on it aswell: https://twitter.com/gdb/status/1745124363362578527