We are at a pivotal moment in the future of AI.
Our approach at Applied General Intelligence (AGI) has always been to validate our claims internally and only then speak.
We no longer believe we have that luxury given the news over the past week.
A few months ago, we hit a major breakthrough in our approach.
For the past 5 years we’ve worked in stealth on a novel approach that goes beyond LLMs.
We call this new approach a Coherence Maintenance System, or CMX.
We’ve named our prototype system “Arx.”
The breakthrough we achieved came by applying our abstraction of the full system to solve hallucinations while using LLMs for certain sub-tasks.
Applying CMX directly at the problem of solving hallucinations in general intelligence systems was phase 1 in our roadmap. Phase 2 is developing the full system without using LLMs for any sub-task.
While costs have been a huge problem with LLMs, it was not the true defect in the approach.
Hallucinations prevent full scale adoption of AI because outputs can’t be trusted
Hallucinations will always plague LLMs because they’re inherent to their architecture
As we were building our Coherence Maintenance System, Arx, our north star was coherence.
An intelligent system with coherence at its core will not guess.
This presented a challenge we solved by doing two things.
First, deconstructing and fully understanding language.
Linguistics is the core of our foundation.
With that we built our language comprehension engine, grounded in our knowledge base.
As we begin to onboard early users, they will be able to see exactly how Arx chunks and parses every query.
Second, Arx learned how to read so it could then read to learn.
This is the same developmental process every human goes through.
It’s why reading is the first thing any child must do before they can learn other subjects in school.
While this sounds simple, the technological challenge here is extreme.
That is why companies have instead resorted to building LLMs, a convenient statistical approximation of language understanding with all its pros and cons: you live by statistics and you die by statistics.
Arx is fundamentally different – a true breakthrough beyond LLMs..
What’s more is that our approach is by its very design computationally efficient.
Our current cost per query is $.0025.
Cheaper than even the DeepSeek models that have recently come out.
Earlier this year we submitted a small abstraction of our model to Tiger Lab at the MMLU Pro.
We told investors and experts before we ever submitted that we would score the highest of any company on that benchmark.
They ran and validated our result where we scored 82.9%.
This was before the major improvements we’ve made in the past weeks.
We were on the top of the leaderboard for a number of weeks but ultimately we were removed because we are not an LLM.
We agree that we are not and we also believe that LLMs are not the path to achieving AGI.
So while we were disappointed that they could not understand this novel approach we’ve taken, this is not new to us.
In August we raised our seed round.
We are sharing our pitch deck we used for the round and you will see that we were already calling for an approach Beyond LLMs.
Many investors struggled to believe that there was another path forward.
But this round was closed in 2 weeks as many investors immediately understood the approach we were taking and why it would lead to the world’s most efficient, coherent, and reliable intelligence system.
At Applied General Intelligence we believe that winning the AI race is a matter of national interest and it’s one that we are positioned to win.
Not because we need hundreds of billions of dollars to do it but because we’ve taken a novel approach that solves the two biggest problems in AI.
Cost and Hallucinations.
We have purposefully kept our research and IP out of the public domain.
Seeing what has happened over the past week has validated our approach.
We know there will be lots of questions and doubts.
That has been the reality for the past 5 years in building this company.
We said that the world needed to go Beyond LLMs and many are acknowledging that reality now.
We will deliver a giant leap forward that surpasses all the LLMs.
First, we will help solve their hallucinations with our abstraction.
Next, we will make our full model available to take the world from LLMs to CMX and beyond.
Link to our Seed Deck from 2024 here:
https://t.co/cBnQtQYKC9
My short piece, “Last Things,” appears in this issue of Literary Orphans. Big heartfelt thanks to Mike Joyce for giving my story such a wonderful home.
https://t.co/3NsvMNuCyZ
My flash "Sacred and Profane" is up @CraftLiterary today, together with an Author Note. Thank you to EIC @katelyn_keating and Flash Fiction Editor @TommyDeanWriter and team. Thrilled to have this story published with you. ❤️❤️❤️https://t.co/cGq552EJp8
My flash "Nocturne" is up at the brand new @Splonk1 Irish flash fiction journal. Excited to be part of it. Thank you to editor @NualaNiC and the Splonk team. Can't wait to read the issue. https://t.co/H0M87MhPsN
So I wrote a story that knows it's a story and it's got ghosts and symbolism and all sorts of goodies in the dreamilicious @threadcountmag:
https://t.co/k3qf8IXi2Z
If you're having trouble building a magic system for your story, try writing down four bullet points:
*Something magic can do.
*Something magic CAN'T do.
*A price/toll that using magic incurs.
*What it feels like to do magic.
Those four points are often all you need.
https://t.co/gf13vMcWvF Gone Lawn 31 is live! This issue features new work by 23 authors, an excerpt and interview with poet Heidi Seaborn, and cover art by Holly Day.