Me and group of friends are developing open-sourced technology of approximate uploading—sideloading—via LLM with very large prompt. The results are surprisingly good given amount of resources and technology limitations. We hope that it may help with alignment. I also open-sourced and publicly donate my mindfile, so anyone can rune experiments with it.
This is interesting, but I again caution that fine tuning a foundation model is unlikely to result in an emulation which generalizes properly. Same (but worse) for prompting.
Me and group of friends are developing open-sourced technology of approximate uploading—sideloading—via LLM with very large prompt. The results are surprisingly good given amount of resources and technology limitations. We hope that it may help with alignment. I also open-sourced and publicly donate my mindfile, so anyone can rune experiments with it.
This is interesting, but I again caution that fine tuning a foundation model is unlikely to result in an emulation which generalizes properly. Same (but worse) for prompting.