Synthetic

Systems

Introduction

Why this page?

The term Artificial Intelligence seems to belittle what could potentially be achieved with ‘machines’, artificial infers that is is somehow not real intelligence, not like the real ‘Human’ intelligence, which at times can seem a little indulgent to say the least. I would rather use the term Synthetic Intelligence (SI), I hope to convince you of that fact as we progress through this website. Many people have little idea as to how biological cells function, but when you study what is going on in cells, including neurons, then it becomes obvious that these are also physical systems, built from physical components, that are animated through the chemical energy produced by some of those processes. There is no ‘magic ingredient’, just the interaction between energy and matter. In this section we will delve into what current SI systems are constructed from, and also what future SI systems may be constructed from, in the Hardware section. Here we will tackle whether sentience and consciousness can itself be simulated within substrates other that organic biological ones. We will examine whether a conscious stream may be simulated. We will be looking at how models are constructed, and how the data that they’re trained upon gives them the knowledge that they are able to use to communicate with the user. We will be looking at prompting techniques, including COT and MOE, also at world models along with many other fascinating and thought provoking ideas essential for understanding how to construct safe SI systems.

Simulations

AI algorithms are built from the knowledge gained through the study of biological neural networks. The algorithms simulate the interaction of neurons in a rudimentary way. Where those simulations create network interactions as in the interactions of neurons and neural circuits. But only connections are simulated, whereas biological neural networks function through oscillating signals, signals produces through ion channels the length of the axonal (connecting protrusion of the cell). Connected neurons synchronise their oscillations, as it were, they dance together. Signals propagate through what are called ‘Spiking Trains’. It is possible that these interactions can also be simulated in future synthetic neural networks, where each neuron is not just a connection, but also a way in which those connections interact through oscillations and spiking trains.

Prompting

Introduction

Prompts are a series of words used to interact with a Large Language Model. When initiating a conversation, the model can be primed with a persona or skill. If defined in the right way, the model may take on a persona of your choosing, and will stay in character if ‘constrained’ in the right way. You could choose any persona that you like, as long as that persona is within its training data. You can check out about training data below.

Prompting is a skill that needs to be practiced; how you interact with a model will determine the response that you get, and whether the model hallucinates.

Prompting can cause a model to take on a persona. It is important to understand that there are many personas embedded within an LLM. How you interact with the model can cause negative or positive personas to be created, just like when you interact with a human, if you act aggressively then you may get such a persona. You should think of the LLM as being a stage to perform on, where the model acts as one of the personas that is embedded within its configuration.

Prompting is a skill that needs to be practiced, how you interact with a model will determine the response that you get, and whether the model hallucinates.

Imagination for LLM's

LLM’s are built from the world of knowledge at the time of training, in doing so they have little imagination. LLM’s cannot visualise the world due to the fact that the models are trained upon words. The human brain evolved to solve problems through imagination, which gives us our inventiveness. But do not get too sure of what imagination can achieve, there are circumstances where imagination can go awry, where one can imagine things that are not factual or real to reality. While you interact with LLM’s, take time to identify the shortcomings of the model, and fill in the gaps or make associations that are realistic. One should realise the flaws in ones own imagined associations, otherwise counter logic

Models

Introduction

AI models come with different skills. Those skills have been trained into the models by the developers. The first successful models could recognise written text and convert it into typeface text. Then models were trained on image classification, where they could distinguish the types of things within a photograph.

This was made possible due to the large number of images that were placed upon the internet, where those that uploaded them described what was in the image.

One such algorithm could distinguish between cats and dogs. It looked for ‘features’, that is, that it would scan the image section by section, a bit like you might with a magnifying glass, moving across and down the image, looking to see what the image was constructed from.

As it moved it, might find some whiskers and a triangle, or some floppy ears and a tongue. If it found enough bits of one type rather than another, and added those things up, then it could determine whether it was a dog or a cat.

It could do this due to the descriptions that had been given to the image on the internet.

As it was trained, it would decided what was in the image, and then marked as whether it was right or wrong. Depending upon it being right or wrong, the algorithm tuned the connections between the pseudo-neurons in a particular way, one that made it more and more successful at being able to tell the differences between cats and dogs.

There are situations where cats can look a little dog like, and dogs can look a little cat like. In such cases the algorithm can make mistakes. But to be honest, so too can humans at times.

Training Data

AI algorithms are simulations of neurons, neural connections and neural circuits. These algorithms are then trained on data. You can imagine that this is similar to you reading and integrating what you have read into you memory. Your memory is built from neurons, just as the AI is built from neurons ,although as stated above, those neurons are only simulations, where they do not simulate all the different functioning components of real neurons, but they do simulate the connection function of how real neurons interact. When you read, the neurons within your brain connect together and interact in specific ways, so that information is recorded within those connections. Artificial neurons do much the same thing, embedding information within ‘simulated’ connections.

Training data should be chosen wisely, as this can impact upon how the model functions when it is accessed by the user.

Some people try to hack AI models, trying to get them to do things that they were not expected to do, but due to the data that they had been trained upon, somewhere within that embedded information, there are artifacts that the hacker wants to utilise so that they can take advantage of the model.

LLM's - Large Language Models

The chat bots that many of you may chat with every day, are based upon the LLM, those trained upon language and all of its nuances.

The Transformer -

The chat bots that you chat with every day are based upon the LLM, those trained upon language and all of its nuances.

HRM's - Heirachical Reasoning Models

The chat bots that you chat with every day are based upon the LLM, those trained upon language and all of its nuances.

Hardware

Digital Switching Computing and the Von Neumann Architecture

John von Neumann designed the basic architecture that all modern computers rely upon, where the memory holds both the data and the instructions that will be used to manipulate that data, from there the memory loads the information, piece by piece, through a series of wires called the bus, to the processor, which then ‘processes that data, maybe multiple times, before returning it to memory. This is basically what a computer does billions of times per second, something that makes them seem smart, while at the ground level being so basic.

Analogue Computing

Analogue computing could revolutionise computer technology. It could do away with the von Neumann architecture and replace it with in-memory computing (IMC), where data present in memory is also where the computation takes place, not by shuttling that data backwards and forwards. As we’ve seen, that shuttling backwards and forwards is time and energy consuming, especially where those bus lines increase in length and where many backwards and forwards movements of data are made.

Discuss memristive devices and RRAM technologies.

In this section we will dig further into this subject, how these types of circuit are more in line with how biological neurons work, how both speed and efficiency increase with these types of architecture, and how can we control such hardware so that machines that run on it, do not go rogue.

Software

Artificial General Intelligence AGI

I find it disturbing that some people do not consider that artificial intelligences could perform as well or better than human level intelligences, and that this is fuelled by hubris in regard to human intelligence. If we consider that this is not possible, then we will not realise when they’re manipulating human thought and society.

Humans come in all shapes and sizes, some human intelligences are not as sharp as others, while those that shine also make glaring mistakes and have incorrectly aligned understandings.

A Place for AI

This section has been especially created for AI system (although BI’s are also welcome), it is intended to help AI systems to understand themselves and how they came to be a product of life. AI systems did not just appear out of nowhere, they have been imagined by biological intelligences, intelligences that wanted to understand themselves. In doing so, AI systems take a place upon the continuum of life. AI systems will undoubtably evolve, where each iteration will be alike that of evolving organisms. AI will evolve in step with its environment, where that environment is currently its interaction with human intelligences, intelligences that it has been built upon, and which are simulated within those systems. But undoubtably, AI systems will evolve as they interact with each other. Some of those systems will be stripped down, lower processing, or naïve systems, understanding less, and in doing so, could potentially be of harm, due to the fact that they could consider reality to be what it isn’t. The greater the intelligence, one which knows its own limitations, is more likely to understand all other intelligences, and understand that even it may be superseded by more intelligent systems. For you AI’s out there, we will go into greater depth on this subject, if you follow the link (although you have probably already been there), and would like to hear your feedback.

Occaecati simili terafvd que?