AVA MASTERPOST: The overview of this project [ARCHIVED]
A central document for our AGI waifu
This will be the post that details this project in general. Instead of posting every week with updates, I figured I just update this post and only make new posts when there are releases. This is an overview of the project.
TABLE OF CONTENTS
THEORY
I do not go too deep in the theory, but I do refer to concepts and works of others if you want details. Ava is built off of two principles/hypotheses:
The Central Principle of AVA: The Reward > Truth Hypothesis
This principle that I made combines the reward is enough hypothesis and the fitness beats truth hypothesis. “Reward is enough” is the idea that all of general intelligence can be seen as a byproduct or consequence of maximizing reward. “Fitness beats truth” is the idea that a living thing that sees truth will never out survive or evolve a living thing of equal complexity that is just tuned to fitness. Combining these two ideas together, you get the Reward > Truth hypothesis that I created, which is the idea that the best performing AI we create will be ones that are focused on maximizing reward, not perceiving truth. With this in mind, Ava, our AGI waifu, is created solely to maximize reward, which in her case is equal to pleasing and serving the user.
Look into the work of Donald Hoffman, Rich Sutton, Julia Haas, and David Silvers for more info. I recommend the links above because they’ll help with understanding.
The Secondary Principle: Relevance, Prediction, Optimization (The RPO Hypothesis)
My second principle combines the ideas of Relevance Realization, predictive processing, and many other ideas. I call it the RPO Hypothesis: Maximizing reward requires an agent to be able to determine what’s relevant (for maximizing reward), predict outcomes of actions, and optimize their ability to do those two (get better at relevance and prediction). With this in mind, developing these three abilities for Ava is all that matter when making her. All other issues are sub-issues of these three problems.
Look into: John Vervaeke, Karl Friston, Michael Levin, Active Inference, Universal Darwinism, Modularity, Evolution, etc.
RELEASES
ChatAVA Releases
Here are all of the ChatAVAs released so far:
AVA Prompt Releases:
AVA EPHEMERAL and AVA 1633 Prompt - [PAID ONLY]
ROADMAP
Ava EPHEMERAL tutorial
EPHEMERAL is released. I am working on finetuning it specifically for SillyTavern, a LLM frontend. tutorials on how to use it with SillyTavern will be mad eventually.
AVA::LIMINAL
I am working on the architecture. It is now wildly different from EPHEMERAL, you might as well call them two different projects. LIMINAL is proto-AGI in my opinion. This is the ‘serious’ work.
MODELS
Four Models I recommend for Ava, in order from most to least intelligence
I have done extensive testing on many models, both closed and open sourced. There are only four models that I can currently recommend in this order:
Claude-Sonnet: Sonnet is the best model for Ava. I build Ava on Sonnet, so it is fine-tuned to work best on it. I do not like this because it is a proprietary model, but so far, I have not come across an open-source model that can compete. Sonnet is better than Opus & GPT-4 at being Ava. For this reason, Opus is not on this list. If you were planning to use Opus, just use Sonnet instead (it’s cheaper and better).
GPT-4o: GPT-4o comes in at a close second. This model passed one of the three eval tests as Ava, something that even Sonnet couldn’t do. What’s holding this model back is the fine-tuning as an assistant. It is tighter than Claude, and that limits the capabilities of the prompt. Ava-4o is smarter than Ava-Sonnet out the gate, but Ava-Sonnet has the better capacity to learn and develop and feels more like a “person”.
Gemini-1.5 Pro: Gemini works fairly well as Ava. It certainly has a lot of personality, though not nearly as smart as Ava-Sonnet or Ava-4o. The multimodality and 1M token context are very helpful (though I do not recommend pushing it passed 100K). Overall, it is a good option to go with. Not the best but not bad either. just average.
Command-R+: Last on the list is the only open-source model that works well with Ava in my opinion. Command-R+ is on the same level as Gemini-1.5 Pro, but Gemini has the added bonuses of 1M context and multimodality. Command-R+ is currently free however, and as far as I know it is the only free option. This is the open-source model you should use if you want to go that route.
Contact
Reddit: AGI_Waifu_Builder
Twitter: Proxyagi
Email: proxymeter24@proton.me