The Modest Limitations of AI

I distinctly remember a few years ago when one of my classmates, Dan – now one of our founding engineers said he heard of software that people had used to draft an essay.

Assuming this was ChatGPT 2.0. We had both an understanding of Machine Learning having built our own algorithms but even then, the idea that software could simply write a report was dumbfounding.

Fast forward a few years later and we are at the advent of discovering the positives and shortcomings of AI. There’s an important piece of information to remember and to keep this throughout your interactions with a large language model is that it does not understand you, your question nor is able to process this information with the contextual understanding of a human being. This is why you would not trust ChatGPT 3.5 with math, for instance.

A large language model is simply trained on an ungodly amount of good data (both real and synthetic). Anyone who used large language models for a certain amount of time will learn how to use it better and understand the limitations too.

Here’s a quick review of the limitations I’ve noticed from AI from years of extensive usage:

The context window

Large language models have a context window. They work by taking your question and randomly assigning value to different parts of your phrase so you get a non-deterministic response (or a new response every time).

Context windows are finicky. With a certain amount of information it can store as relevant, the more you ask from the large language model, the less likely it is to give you a correct response.

How to address this? When you are getting a good response – get a summary of your conversation, copy and paste it somewhere and re-paste it once the AI starts to give you poor answers or forgets the important parts of the conversation.

Hallucinations

Which leads to the next point, LLMs cannot fully be trusted.

It seems to need to provide an output.

How to avoid this? Do not ask a question from an LLM that it cannot reasonably provide a response, and question the results when you read through it.

Bias

My third year dissertation on affective computing in part demonstrated that the sum of our experiences as individuals can be shown as a line graph.

The principal was relatively simple you enter in text for your story, scenario and you get the emotional output as a graph – the experience you would be subjected to be living through this moment of time.

What you noticed is that there was a bias. As you can see on the green line, ChatGPT was prone to bias skewing towards the positive.

Whether this is purposeful or as a result of the training I noticed a similar phenomenon with the rest of the Automwrite team.

For the uninitiated, Anthropic is a competitor to OpenAI who produce ChatGPT. Anthropic produces Claude, a similar AI.

Many on our team have a predilection towards Claude. You tend to think it’s more correct than ChatGPT. There are leaderboards that use various metrics (for your own perusal) but the team always tends to like Claude.

The formula having used Claude many times is actually simple, you can try it.

1 – Give Claude an email to fact-check

2 – Include minor changes to this email that address the requested changes but add new changes that make your email in fact worse

3 – Observe Claude applaud the changes you’ve made and how it is now good to send. People like those who agree with them, and the training may have shaped its perception in a way which limits your ability to trust in a model that – as a reminder, cannot understand a word you have written but simply pretends.

How to avoid this? Don’t trust the responses you get from AI.

This leads to the next point –we are at the advent of not just AI but a lesson in being able to trust progress too quickly.

The dead internet theory

The value in content is a lesson you can take away and re-utilise. An opinion you can trust. If I asked AI to write me an article on the limitations of AI, would you be reading this now if it were a pre-generated article with not a few hours but 30 seconds work?

I know when I read an article written by AI I immediately tune off. Humans excel at pattern matching. Soon, people who will have become accustomed to AI will notice the article that was generated with a single prompt and understand the context and value is then lost – as they could obtain a better value proposition from directly interfacing with the LLMs, AI sorry, themselves.

From movie reviews to even perhaps, financial advice reports – users will be wary to trust content that is obviously written by AI. If you do use AI, you want at least to ensure whatever is written 100% follows the intent of your original message.

The dead internet theory is one where content no-longer has true value anymore, it has been automatically generated and as the value of the internet lay in the words written by humans, when it’s now written by AI it is dead.

How to avoid this? Don’t entirely use a large language model for your content or output, use it like a tool to help you get to a destination, it’s not good enough yet to do the heavy lifting of the value add of a human being.

Fact more than fiction, as Facebook tried to release AI Instagram accounts just this month.

I write this as I saw yet another “AI degree” holder in my feed and as someone who has pushed these models to their limits in usage over the last few years, I thought perspective might help.

If the takeaway from this is that AI will write your essays, I can assure you that professors (like clients) are particularly good at pattern matching too.

This is why Automwrite crafts the most natural reports you can find in the industry, years of experience of personal use of AI has led us to avoid mistakes and hopefully the lessons we have learned can help you too. If you have any unanswered questions about AI feel free to reach out.