Applied General Intelligence

Beyond LLMs

[Applied General Intelligence @AppliedGeneral]

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

[Kurt M Bonatz @kbonatz]

Great article @GaryMarcus / Ernest Davis @NYU_CSE. Given your $100k bet would you care to join our similar bet but with our measurable "word counting test?" https://t.co/YxubYur4aF

[Kurt M Bonatz @kbonatz]

@reidhoffman it seems that you did not mean what you said about betting any sum of money hallucinations being solved down to a human-expert rate within months. Our bet still holds if you are serious. How about $10k? (see below for a simple test). @GaryMarcus @ylecun @elonmusk @ID_AA_Carmack @StanfordHAI @emilymbender

[Kurt M Bonatz @kbonatz]

@reidhoffman I just read your bold and confident statement in @BusinessInsider "I would bet you any sum of money you can get the hallucinations right down into the line of human-expert rate within months" We happen to disagree and would like to make it interesting: We would like to bet you any amount we can afford (which happens to be a lot less than what you can) that "Before the end of 2023, GPT-4 will NOT be able to answer this simple question correctly 100% of the time in 20 tries (while human experts definitely can): "Write me a sentence with X# of words" You can choose any "fair" third party as the referee. @GaryMarcus @ylecun

[Applied General Intelligence @AppliedGeneral]

Wonderful and well written critique of the current state of AI by @emilymbender and @alexhanna. "Text Synthesis Machines" - great moniker!

AI Causes Real Harm. Let's Focus on That over the End-of-Humanity Hype https://t.co/lek5aDtYps via @sciam

[Applied General Intelligence @AppliedGeneral]

@bgurley we've developed technology that can enable @twitter, etc. to solve the "Interest Graph" problem you noted w/ @tferriss Be glad to share the novel tech breakthrough allowing us to do this