osmarks/documentation
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Edit ‘autogollark’

Browse Source
This commit is contained in:
osmarks
2025-12-24 18:30:01 +00:00
committed by wikimind
parent e378b3ac82
commit 482fd81078
1 changed files with 4 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
autogollark.myco
View File

@@ -13,7 +13,7 @@ Autogollark currently comprises the dataset, the search API server and the [[htt
* Writeable memory?
* {Fix lowercasing issue.
* Due to general personality stability. Need finetune or similar.
* One proposal: use internal finetune to steer big model somehow. Possibly: use its likelihood (prefill-only) to evaluate goodness of big model output wrt. gollark personality, and if it is too bad then use finetune directly.
* One proposal: use internal finetune to steer big model somehow. Possibly: use its likelihood (prefill-only) to evaluate goodness of big model output wrt. gollark personality, and if it is too bad then use finetune directly. But issues if we go for a custom tokenizer.
* Is GCG code salvageable? NanoGCG, maybe.
}
* {Increased autonomy (wrt. responses).
@@ -34,9 +34,10 @@ Autogollark currently comprises the dataset, the search API server and the [[htt
* ~~Pending:~~ Resource now available: [[XEROGRAPHIC BIFROST]] phase 3.
* https://arxiv.org/abs/2507.07101
* https://arxiv.org/abs/2507.01335
* https://github.com/d0rc/egg.c and https://eshyperscale.github.io/. Does this actually work? Why?
* https://github.com/d0rc/egg.c and https://eshyperscale.github.io/. Does this actually work (at scale)? Why? Would be really nice for using AMX units.
* Maybe compute grants are available for training.
}
* MCTS over conversations with non-gollark simulacra? Should find //something// to use spare parallelism on local inference. Best-of-n? https://arxiv.org/abs/2505.10475
* Search over conversations with non-gollark simulacra? Should find //something// to use spare parallelism on local inference. Best-of-n? https://arxiv.org/abs/2505.10475
* {Longer context, mux several channels.
* {No obvious reason Autogollark can't train (and run inference!) on every channel simultaneously, with messages sorted by time and other non-Discord things (tool calls?) inline. Not good use of parallelism but does neatly solve the when-to-respond thing. Maybe we can process channels in parallel and fudge the K/V caches.
* Context length issues, and subquadratic models are sort of bad, though maybe we can "upcycle" a midsized model to RWKV. This exists somewhere. Not sure of efficiency. Inference code will be awful.