Great Ideas & Roadmaps

[lengockhanhnk]

Could you show the roadmaps to develop it in the future?

[juhonkang]

Great

[conglu1997]

Please see the future work section of our paper!

[BradKML]

Maybe instead of only using the roadmap, the public can also throw in their own ideas on how things could improve?

[conglu1997]

Indeed, we welcome anything :)

[BradKML]

Here are some small ideas regarding the making and testing of new models:

1. It needs to anticipate Mamba-esque LLM, Diffusion-based Language Models, and alternate image generation architectures (together multi-modal architectures would have variety) https://github.com/bansky-cl/diffusion-nlp-paper-arxiv https://github.com/XiudingCai/Awesome-Mamba-Collection
2. Hyperparameter optimization (like the Ray library or the papers on predicting ideal hyperparameters) and things like Grokfast should be included to speed up finding the best solutions https://github.com/ironjr/grokfast
3. Pruning and merging should be considered with their own toolkit (not just previous blog posts but also research similar to ones by PrunaAI) https://github.com/liyunqianggyn/Awesome-LLMs-Pruning

Maybe some more on: dataset gathering and pipelines, or prompt fuzz-testing (formatting optimization or adversarial attacks), or diverse benchmarking and subjective goal preferences (MOO).

[bdutta19]

we are trying to build a stand alone UI over Ai-scientist.. right now it is simple Gradio app but once our ux is settled we plan to make it more robust ..

[BradKML]

Any effort in that direction is appreciated, since having something akin to A1111 is easier than having something closer to ComfyUI with all the drag-and-drop (e.g. Dify). Can it generate and test experiments tho?