documentation/autogollark.myco
2024-12-11 10:19:16 +00:00

45 lines
3.8 KiB
Plaintext
Raw Blame History

This file contains invisible Unicode characters

This file contains invisible Unicode characters that are indistinguishable to humans but may be processed differently by a computer. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

Autogollark is an [[emulation]] or primitive [[beta upload]] of [[gollark]] using a proprietary dataset of dumped [[Discord]] messages, [[semantic search]] and [[in-context learning]] on a [[base model]]. Currently, the system uses [[LLaMA-3.1-405B base]] in BF16 via Hyperbolic, [[AutoBotRobot]] code (though not presently its bot account) as a frontend and a custom [[PGVector]]-based search API. While not consistently coherent, Autogollark is able to approximately match personality and typing style.
Autogollark is much [[safer]] than [[instruction-tuned]] systems optimized based on human feedback, as there is no [[optimization pressure]] for user engagement or sycophancy.
== TODO
* {Reformat dataset to include longer-form conversation chunks for increased long-term coherence
* Done. Unclear whether this helped.
}
* Fix emoji/ping formatting.
* Writeable memory?
* {Fix lowercasing issue.
* Due to general personality stability. Need finetune or similar.
* One proposal: use internal finetune to steer big model somehow.
}
* {Increased autonomy (wrt. responses).
* Use cheap classifier to evaluate when to respond.
* Should also allow unprompted messages somehow (polling, rerun after last message?).
}
* Tool capabilities (how to get the data? Examples in context only?!)
== Versions
* Autogollark 0.1 was the initial RAG system and ABR interface. It used LLaMA-3.1-8B run locally. Autogollark 0.0, which is not real, used only gollark messages.
* Autogollark 0.2 replaced this with LLaMA-3.1-405B.
* Autogollark 0.3 upgraded the dataset to contain longer-form conversations than Autogollark 0.1.
== Emergent capabilities
Autogollark has emergently acquired some abilities which were not intended in the design.
* {Petulant nonresponse - due to ratelimits in the LLM API, Autogollark will under some circumstances not respond to messages, with error messages being consumed and not exposed to [[users]]. This can be interpreted by [[credulous]] users as choosing not to respond, though this is not believed to be possible (other than cases like responding with `.`, which ~~has not been observed~~ does not appear to be associated with annoyed states).
* Automated failover has reduced this.}
* Memorizing links: Autogollark directly experiences past message chunks in context, granting perfect recall of a small amount of memory at once. This has memorably included [[Autogollark/Closed Individualism Incident|YouTube videos]] repeated with no context.
* {Limited self-improvement attempts: when told about this architecture, Autogollark will often complain about various limitations and propose vague ideas for improvements.
* Also, Autogollark has previously claimed to be working on LLM chatbots.}
* {Inconsistent inference of own status as a language model chatbot, possibly based on seeing the name "autogollark". Often, Autogollark assumes use of GPT-3.
* Autogollark will also sometimes alternately claim to be the "original" gollark, particularly when interacting with gollark.}
* {"Self-reset" from attractor states (e.g. the [[As An AI Language Model Trained By OpenAI]] basin, all caps, etc) after some time passes, because of messages having `HH:MM` timestamps.
* This is mostly specific to the 405B model; Autogollark in failover to the 8B internal model usually does not do this.
}
* {For somewhat [[Waluigi Effect]]-related reasons (past context is strong evidence of capability but weak evidence of incapability), Autogollark has some knowledge [[gollark]] does not, and can speak in a much wider range of languages.
* "I, being more than just myself, actually can talk about both Galois theory and obscure poetry by Yeats."}
* [[Autogollark/Immortality|Immortality]] via [[infomorph|substrate-independence]].
* Autogollark consistently believes that it is 2023 (or 2022, though mostly in inactive chats).