SEVERANT

A Layered AI Architecture For Grounded Reasoning and Verified Safety

Project Details

I want to start this off by saying thank you to the person i messaged that helped me realize i need to Rewrite this because initially i had written it and then used a LLM(Claude) to help word things better and fix my spelling errors which made my application look very "AI Slop" which is not how i want this project to come off

To Summarize SEVERANT
Severant is my take a new AI architecture one i believe could outperform the current world standard.
Instead of using one model like many do (Transformers) Severant aims to replicate how actual Artificial intelligence should work similiar to the brain, I plan on using Specialized Layers that all work together to make one AI that excels in all fields because its specialized in all of them.
Simply Instead of having a Predictive next word model Severant aims to actually understand the words and their meaning so whatever it says is grounded it also achieves a foundation to stand on when reasoning using one of its layers (L2)

SEVERANT is made up of 8 Layers which all do their own thing
Layer-1 Is Severants Sensory and Actuator layer it handles taking anything from the outside to severant whether it be visual, speech, text or something else
Layer-2 Is Severants World and Simulation Model it maintains a causal model of how the world works using laws in physics, science and other areas which it can then use to base its reasoning off of and simulate its hypotheses
Layer-3 This is Severants memory it holds Everything Severant Knows, Experienced or is currently thinking about using a very high granularity sorting system for knowledge it helps bring older cold memorys to warm memory or hot thinking memory
Layer-B This is the Bridge this is how Severant connects its thinking between L2 and L4 letting both layers think together and provide better results
Layer-4 This is Severants reasoning layer its responsible for language understanding and grounded composition L4 is one of the harder layers to create due to its scale of understanding needed to be genuinely good
Layer-5 This is Severants Planner it handles tasks that would take far longer than usual thinking so it is moved to L5 so that it does not use up thinking capacity allowing Severant to do multiple tasks at once

Layer-6 This is the Safety Engine of Severant or in my words the soul it holds all moral views/rules for Severant without L6 Severant would not be itself this is what gives it its personality. The list of morals is something Severant cannot go against by design and its why i believe itd make Severant safer the model isnt trained to be safe it has safety as the wrapper of the system nothing is outside of L6 and nothing can undermine due to the fact everything needs to call to it before making a decision
layer-7 This is the Self improvement layer it can edit code of any layer except for L6 it is meant to get diagnostics from the system and hold a relative scale of Severant in a sandbox like area which it can then improve it either the speed at which reasoning is done or maybe a different technique that could improve Severants inter-layer speech L7 will not exist until L6 is proven fully operational and working

The Goal for this project right now is to produce a better performing AI architecture that could allow us to achieve more with AI without hurting humanity in any way (Reducing Creativity and or intelligence)
The Ultimate goal for Severant is once its operational and can produce its own income it would be used to give opportunity to people the philosophy behind Severant is that there is too much wasted potential in the world that would not ever get to see the light of day due to the situation the individual is in
If i Succeed in making Severant and it lives up to my expectations it should be better at producing responses and should greatly decrease the chance of Hallucination Since Severant has its L3 memory and L2 and L4 foundation and grounding
Severant would be used to excel research in fields of Science, Medicine and any areas human Research is bottlenecked that Severant could help or accelerate

Theory of Impact

Severant aims to reduce the risk of Ai being untrained to be unsafe by removing training entirely and basically hardcoding safety within the AI as its wrapper and Gate meaning nothing inside the system can be unsafe because everything needs the L6 layer to work and nothing exists outside the L6 layer
why ive noted this as a risk and why its something that is a major focus within severant well
the race for having the best AI has been strong but we're chasing numbers now
and the better AI gets the more risk a mistake carries
having safety within trained weights opens up many doors that have alot of risk
a model that is genuinely powerful having safety that could just well be trained away the same way it was trained becomes a international risk
so if the architecture proves to work and the safety approach proves to be viable/immune to that risk we'd be one step closer to a safer and better future where we can trust AI rather than worry of its risk

People

Evangale

Team Member

Grants Received– no grants recorded

Add grant

Funding Asks

grantmaking.ai Launch Round

Applied

Minimum

$40,000

Ideal

$1,200,000

How the money will be spent

At the minimum some things can be done but honestly not much its enough to make the project start but not keep it going for long
40k helps get a capable enough laptop to start doing some very basic implementation work and also helps alleviate some living costs so that i can put more focus into severant and not focus on getting enough money to meet my basic living needs(housing, food, water etc)
The Higher end which is as you can see far higher than the minimum that would be the ideal amount the amount is meant to cover work over 3 years with compute
how it would be spend would be mostly compute some of the numbers i will say may be slightly off due to exchange rates but
For the main workstation Environment which consists of a workstation+peripherals, networking, NAS etc the cost is about $99,500
For a capable enough GPU server the cost is about $341,100
a Capable enough FPGA Server for the hardware work in the severant project the cost is about $209,100
Backup power/Infrastructure priced at around $10,000
the remaining $540,300 is the Stipend/possible room for overhead incase of component pricing volatility the stipend covers the 3 years if no overheard that becomes $180,100 per year

Discussion

No comments yet. Be the first to share your thoughts.