What comes to mind when you hear the words “Thinking machine”? I picture a highly intelligent system that reasons logically, almost like a human! But is an LLM a Thinking Machine?
It is a hard question, but personally I think the answer is no. A system that was originally built to translate text from one language to another, and that suddenly turned out to imitate human thinking without understanding it, cannot be a TM.
And that was the lead-in to some global changes in AION!
I counted that I have been working on the project for about half a year. In that time I ran more than 100 architecture experiments, trained the AI for more than 20 hours and… the results seemed negligible to me. Almost the whole previous chain, the one I told you about with such enthusiasm in earlier articles, turned out to be fundamentally wrong. A system this huge is incredibly hard to maintain. I asked myself: “If the chain does not work, does it make sense to keep improving a path that may have been wrong from the start?” And by then I had done an enormous amount of work without ever reaching the cognitive core… I built a temporary test bench with only a couple of modules left, and the chain now looks like this:
LS → CoS ⇄ ES → LS
I had to change the cognitive core, and the way experience is stored.
CoS is no longer a separate “thinker”!
Before, I had no idea how exactly to make the cognitive space work: build it as a core with a single neural network, or maybe create a very complex adaptive mechanism… Or maybe forget it, I am very tired, the AI will somehow manage and synthesize a finished system for me? But no. In practice, if you do not understand how a specific module works, no AI will solve such a complex task for you. You need certain skills…
An idea came up that could work on paper. Let me introduce my next new mechanism: CL (Cell Logical)! The inspiration for it was ordinary biology. What are all living organisms made of? Right, cells! But what happens if you take a single cell out, will it be a living, whole organism?
One cell is good, thousands are better!
Digging deeper and deeper into human consciousness, I may have understood what I was missing. And it was right on the surface. I looked at cognition as logical operations, but as soon as I looked toward simple neurons… As you know, the neurons in our brains are connected to each other and pass information between them, but they are more complex than we usually say. Did you know that some neurons can store information about a person’s image? But okay, let’s stay on topic. The cells in my system think over an answer before saying it or passing it to other neurons. Something like CoT in an LLM, but working through numbers and faster :) During this “thinking” a cell can pass intermediate calculations to the others, which in turn can decide whether to use them or not. This way a task can be “thought through” from different sides, with better results than if we had one big cell, or just one big thinker.
On the tested tasks this system showed very good results compared with my other experiments! And the most interesting part: it works. ES saves experience and loads it into the cells, and they use it to improve their reasoning. Of course, at this stage I cannot say that a cell noticed its mistake and its behavior changed right away: it takes more unambiguous mistakes before a cell can formulate experience.
In the long run such a system could be incredibly useful for long research projects (which is what I am building this project for), the kind an LLM physically cannot sustain because of its limited context window. AION, in turn, using its past experience of working on a project (which may even count as knowledge, more experiments are needed), could improve its logical reasoning at the same time and work on the project more efficiently, holding an incredibly large amount of data. One of the interesting mechanisms is move reduction. After working, the cells consolidate experience to shorten their reasoning and derive more general, working rules that can be used on bigger tasks. Roughly speaking, the first time the system works on a math problem that takes 100 steps. After consolidation the steps drop to 60, the third time to 20, and so on down to the minimum number of steps that can solve the task. Of course, all of this needs testing, and some of what I describe may turn out to be architecturally impossible, but I will do everything I can to make it real. Most importantly, if my next experiments prove that my system really does take over experience while working, that will be an incredibly strong result. Machines that learn as they go — isn’t that amazing?
So, let me show you the first results of my work! They are modest, of course, but the main thing is that there is progress!
I used to have a cell that ran on a GRU, but as soon as I switched to my own architecture… the results are incredible!
- How much it has to learn in advance. Old neural network: 10 minutes of training. New cell: 0, it learns right from your three examples.
- How long a task takes. Old neural network: milliseconds. New cell: from 6 to 200 milliseconds (numeric formulas need about 20 seconds per rule).
Besides that, on some tasks I showed that ES works:
And probably the most interesting part for you: how does the system talk right now?
— What color is the sound of rain?
The sound of rain is simply fairy-tale and pink.
— Can you get tired of resting?
Yes, of course. But I prefer an active rest. I plan to go to the beaches and enjoy nature.
A conversation with memory:
— My name is Daniil.
— I live in Grodno and I love coffee.
— What is my name?
Hi, your name is Daniil.
— And where do I live?
You live in Grodno and you love coffee.
The quality of these answers is about 2.5/5.
I think this is a good note to end the article on. I hope that by the end of October or November I can show you better results than these…