← Back to Main Page

Top

Last Updated: November, 2023

Introduction

Artificial Intelligence (AI) is a branch of computer science that aims to create machines that mimic human intelligence. This didn’t originally mean creating machines that can think for themselves, but rather, creating machines that can be programmed to think or learn in a way that appears intelligent. However, there seems to be very little, or potentially absolutely no difference at all between “actual intelligence” (whatever that means) and “artificial intelligence” - more on this later.

The language I’m using could be seen as anthropomorphising, however, these models are reflections of some ways the human brain works and learns, and I believe the analogies using human-specific language to be well-accurate enough.

Why Does It Work?

The “intelligence” and “reasoning” abilities of any given AI system are not binary - AI is not either intelligent or not intelligent - these reasoning abilities exist on a sliding scale, similarly to humans. There is only so much reasoning a particular model can perform, and only so much the model can “remember” (or juggle in its “mind”) at any given time. An AI system’s intelligence and working memory size seem to be the constraints on AI ability, however, this is still very high-level, and still hasn’t really answered the question. Funnily enough, we are once again asking "what is intelligence?"

Technically, all these models do is predict the next word (actually “token” instead of “word”, but that’s not too important a distinction at the moment). That said, viewing these models as “only text prediction and nothing more” is majorly missing the point. These Natural Language Processing (NLP) models (like ChatGPT) have been trained on such a large amount of diverse text that they have learned and understood the inherent underlying logic and patterns of human language as presented through text. Language is the vehicle in which we humans use to convey our thoughts, reasoning, feelings, ideas, aspirations, and emotions - these are emissions and admissions of our own intelligence. Hence, intelligence is a fundamental requirement in creating the thoughts we have and the language we use.

Clearly, it seems that by instructing an AI to learn how to effectively model language, it must also learn how to effectively model intelligence itself to be able to excel at modelling language. Stemming from “only” an advanced text prediction algorithm, there is a veritable emergent “ghost-in-the-machine” with a surprisingly organic and authentic display of robust artificial intelligence.

What's On Offer

There's a few conversational LLMs available, with varying usability, function, intended purpose, and ability. Here are a few of the more well-known ones, with some added polemic on my part. This list is in no way comprehensive, and there's a lot more to be said about open-source LLMs, but I digress:

The First Pitfall

At some point in the future, I’ll collate more information on open-source LLMs. However, for the time being, I’ll mainly be writing about usage of ChatGPT and GPT, specifically GPT4.

The first pitfall would be not evaluating ChatGPT-Plus, the USD$20/month subscription to ChatGPT. The free version of ChatGPT (ChatGPT-3.5) is a golf buggy – impressive, if before you could only walk or run. The paid version of ChatGPT (ChatGPT-4) is a Ferrari. This is for two reasons:

  1. ChatGPT-Free is based only on GPT-3.5, a fine model for understanding text and performing basic intelligent tasks on texts. ChatGPT-Plus, however, is based on GPT-4, a much more recent, and much, much more powerful model. Not only can it understand text and language, but it excels at reasoning, problem solving, intuition, and creativity - it’s much more able to do what you might need it to do. I generally put it this way:

    GPT-3.5 could write a single python function at about a 65% success rate – if you work with it to fix the errors, you could probably generally get it to successfully create a python function.

    GPT-4 could write a whole python program with about a 95% success rate. If errors are encountered, simply copying and pasting them back into the chat window will generally bump that up to 100%.

  2. ChatGPT-Plus allows access to the hugely expanded User Interface. Aside from normal text output (called “inference”), it can also autonomously:

    • write and run python code
    • browse the internet
    • perform proficient computer vision, including recognising and reading text in images
    • upload and download files to/from the chat window
    • generate visuals and art using Dalle-3

For example, one could simply drag and drop a text file like this one (a dataset from the Australian Bureau of Statistics detailing employment trends across Australian states, by time and by company size) into ChatGPT4, and ask for some interpretation to be done (in this case, calculate the difference between Victoria and the Australian mean) and for some graphs made. It writes some python while I step away to make a coffee, returning coffee in hand to this:

A line graph titled 'Difference in Employment Trends: Victoria (VIC) vs Australia (AUS)' displaying three lines representing different sizes of companies by number of employees: 0-19, 20-199, and 200+ employees. The horizontal axis marks the dates from early 2020 to late 2023, while the vertical axis shows the difference in employment index with values ranging from approximately -2 to +6. The lines for all company sizes show fluctuations over time, with notable peaks and troughs, indicating variances in employment trends between Victoria and Australia as a whole.

I'll eventually come back to this page to write more, but for now I'll leave it here. Be sure to check out part 2.