Current AI technology mostly relies on multiplying matrices of numbers together, and changing the values of those numbers in clever ways to get the result one is looking for. There are other ways it’s been done and in theory an infinite number of ways AI could be done. But Matrix multiplication has been the standard. That could be rapidly changing as AI research is often proprietary.

Now, the math behind matrix multiplication is hard to unpack the first time you learn it, and learning it might not be necessary for what I’m trying to do in this post. So for now lets analogize these giant grids of numbers to sheets of paper with instructions on them.

Imagine you’re playing the role of a processor on a computer. As the processor, you have no memory, and simply follow the instructions in front of you like a good Orwellian citizen. Once you finish reading the paper and carry out what it asks, you lose all memory of what it said. (Like a GREAT Orwellian citizen.)

Here’s how ChatGPT and other LLM AIs work. They finish reading all the instructions and then add a word fragment to the end of a sentence. To get the next word fragment, they need to read all the instructions over again.

For example:
“please finish the sentence: The quick brown “
…gets read in… Processing by reading the entire stack…
“please finish the sentence: The quick brown ~fox~“
…gets read in… Processing by reading the entire stack…
“please finish the sentence: The quick brown fox ~jump~“
…gets read in… Processing by reading the entire stack…
“please finish the sentence: The quick brown fox jump~ed~“
and so on…

As you can imagine this is an incredibly inefficient process. By embracing that inefficiency and leveraging processing power AIs are able to brute force their way into doing what they do.

The result? AI companies need more and more processing power. That means more electricity. It means more water for cooling.

You may have your own opinions about how deleterious all of this may be to the environment and to society. I’m not taking a position on that here.

I only mention this to explain the need for more efficiency in the software. And when DeepSeek came out they found a way to run LLMs with something like 1/8th the processing power of other AI systems, it was taken as pretty big news. And I think, but do not know, that this has become the new standard. And in order to explain how they did it, I’ll go back to my paper analogy.

Imagine instead, you have 8 stacks of index cards sandwiched in between the first and last few papers in the stack. And instead of needing to read all papers and all the cards, you need to read the first paper, and then only one of the eight stacks of index cards, and then the last few sheets of paper.

Easier. Faster. Much less processing power.

This is clearly a major improvement. When you’re dealing with data processing of the magnitude AI companies typically work with, an 8 fold increase in efficiency is a big news.
Now, I don’t know about you, but I feel like it could still be better. 1/8th of a really big process is still a pretty big process. Yes, there might be some projects that require it. But there are also countless projects that don’t.

How hard is it for your brain to answer the question “How are you?” with “I am fine.”

In order for AI to do the same thing, it needs to make billions or maybe trillions of calculations.

Put another way, to answer “I am fine,” I don’t need to go through all the words I know in English, all the words I know in French, all the words I know in Japanese to find the right response. AI, in effect, does.

Even with the new standard (1/8 the processing needs of the old standard), AI is still incredibly inefficient.

AI doesn’t “remember” in the same way humans remember. It needs to recalculate every time, transforming information into another piece of information.

When you look at something, your mind turns what you see into something you understand. It turns it into meaning. AI does something like that—only it does it through calculations.

And it’s for that reason most people analogize AI to the human brain, and the “reader” in this case would just be the passive process of neurons firing and adjusting the connections between them.

And with that in mind there might be a way to emulate the natural processes that the brain goes through without completely reworking the structure of AI in general. And with that I introduce to you the Graph Node model of AI—an invention of mine—or an invention in progress??

Here’s how it works: Instead of needing to read through all the instructions on all the papers and index cards, each card routes the reader (the AI) to a chosen next card, finding the fastest possible route to an exit. Each set of instructions could be short or long depending.

I may discuss the technical details in a future post as I continue to refine the process. And there’s a great deal that needs to be worked out (e.g., how the activation function interacts with the exit state, algorithm questions, and more).

One of the major advantages to an approach like this is to allow for a higher degree of flexibility on how the many “neurons” are used and how many functions each “neuron” could take on. This is hard to explain without the math, but it would be kind of like having it set up so that some of the papers could be read backwards or otherwise out of order and still give useful information. To my knowledge, most modern AIs solve this problem of needing to process more inforamation to get to an answer by running the entire AI multiple times to give the output more context. This DOES create an opportunity to allow for further processing of more intricate information (again, with a lot of redundancies considering you’re sending the same process through for vastly different types of tasks), but it doesn’t allow for very efficient and specified tasks in the way mine does, because the structure intrinsically creates opportunities for branches that could be isolated and used for specific functions.

I think doing something like this, even if it isn’t “this,” becomes important as technology becomes a more integral part of our day to day lives, as it already has. By creating systems that can be efficiently sealed off and put into a portable environment, it makes it possible to use AI in more remote ways. On an individual level, such that you could use something with AI in it without an internet access or the need to subscribe to some kind of massive warehouse in Arizona.

There’s something about the offshoot of technology to remote regions that concerns me. When you have an entire society coalescing around a single point of failure, it becomes a security issue on multiple dimensions. Both for the potential of sabotage or accident. (Arguably the least concerning of these, as there are still plenty of data centers.) And from the potential for bad actors to use this. Even if those bad actors are the AI itself.

This is a concern that OpenAI was partially founded on. That AI should be open to all of mankind. That is something I think I agree with. But in practice we might be falling short. By creating a type of AI that is fundamentally mobile, you might be able to create a way to make it equally impactful, but distributed, without the dangers of over centralization. This could help quite a bit with the above concerns. And could be part of the solution to the control problem. Having so many different AIs with competing modes and desires such that no one AI can take over and cause major ruptures.

I’m not naive to the fact that even with this restructuring, there will always be an adventage to larger immobile data centers simply because have more space to fit more processors. Thus, physics might ruin my plan. Or maybe it won’t; as having more types of AI interactions with the smaller of these having more direct human control might create more effective interface mechanisms as the large AI’s inevitably become smarter. Giving us more tools for solving the control problem in the future.

However, I am open to being wrong about this. It’s still possible that this could backfire and accidentally create a super optimized AI for destroying the world. Perhaps the solution at that point is to create a variety of AIs with the express purpose of stopping that from happening, and as long as they outnumber the others, you might be able to come out of this alive.

Well that turned dark really fast. And it might be tempting to say that the best thing to do is ban AI, and you can try to do that while other countries embrace it and undercut your progress. Perhaps the real solution is to embrace pluralism, while authoritarian societies use it to enforce uniformity; our heteroculteral forest of technology could spring forth abundance.