Artificial Intelligence Intellect

Artificial Intelligence (AI) should be called Artificial Intellect as its models have aquired a large volume of knowledge from the Internet, but it is not yet that intelligent and can be biased.

Despite what the name suggests, AI tools aren’t even particularly intelligent, at least not in the way that we commonly think of intelligence. They make mistakes that even a small child would not make. That does not mean that they cannot be useful, but we must be very careful about when it is, and is not, appropriate to use them.

Executive Summary

This section contains an executive summary (written by AI) of some issues that arise with AI usage and explores the benefits, dangers and ethics of AI, paying particular attention to the psychological implications of its use.

AI has become an essential part of our modern world, and its use has grown exponentially in recent years. This technology has many advantages that are contributing to the growth of various sectors, but its potential dangers also cannot be overlooked. Either you stay on top of AI or it may get on top of you.

Correct, but not right

AI can create some interesting content for us, but also a lot of nonsense. Here is what Google's Gemma 2b Model just came up with in our lab:

Prompt: If 3 hoodies take 3 hours to dry in the hot sun, how long do 5 hoodies take to dry?

Response: Sure, I can help with that. 5 hoodies would take 5 hours to dry in the hot sun because they would be competing for the same resources (Sunlight and heat) to dry them.

Meta's Llama 3 Model dries the hoodies right:

Response: A classic lateral thinking puzzle!

The answer is not a straightforward calculation of "if 3 takes 3 hours, then 5 will also take 3 hours." That would be too easy.

Instead, think about it this way: if the number of hoodies doubles (from 3 to 6), but everything else remains the same (same sun, same air circulation...), how long do you think it would take for all those hoodies to dry?

That's right! It would still take around 3 hours. The reason is that the drying process doesn't depend on the number of hoodies; it depends on factors like sunlight intensity, humidity, and airflow.

So, if 3 hoodies take 3 hours to dry, then 5 hoodies will also take approximately... (drumroll) ...3 hours!

AI still faces a few challenges

One major issue is that AI chat systems (GPT) tend to make up answers (hallucinate) when they don't know the answer to a question or prompt. Some can be told not to do that, so they will simply say they don't know.

Known challenges of LLMs include:

• Creating a response from non-authoritative sources.
• Presenting false information when it does not have the answer.
• Presenting out-of-date or generic information when the user expects a specific, current response.
• Creating inaccurate responses due to terminology confusion, wherein different training sources use the same terminology to talk about different things.

You can think of the Large Language Model as an over-enthusiastic new employee who refuses to stay informed with current events but will always answer every question with absolute confidence. Unfortunately, such an attitude can negatively impact user trust and is not something you want your chatbots to emulate!

Data Aquisition

AI systems "ingest" the data provided to them and store it in "models" for what it assumes to be the real world. This is very resource intensive. Various organisations like Meta (Facebook), Microsoft, or OpenAI, have published large language models (LLM) for download and offline use. Some are uncensored so you can ask it questions that would be unacceptable online.

Retrieval Augmented Generation, or RAG, introduces some serious capabilities to large language models, like the ability to use and leverage your own data. It redirects the LLM to retrieve relevant information from authoritative, pre-determined knowledge sources. Organisations have greater control over the generated text output, and users gain insights into how the LLM generates the response.

LocalDocs Plugin (Chat With Your Data)

LocalDocs is a GPT4All feature that allows you to chat with your local files and data. It allows you to utilize powerful local LLMs to chat with private data without any data leaving your computer or server. When using LocalDocs, your LLM will cite the sources that most likely contributed to a given output. Note, even an LLM equipped with LocalDocs can hallucinate. The LocalDocs plugin will utilize your documents to help answer prompts and you will see references appear below the response.

Prompting

To get answers one needs to be skillful with formulating a prompt (question) to get a reasonably good result. The better the prompt, the better the result. You can tell it to reply or write essays in the style of a famous author, or to honestly say when it doesn't know an answer. Be very specific with your prompts.

Image Generation

AI images are getting better all the time. Even videos are possible. We're currently exploring how far we can push this. Some of the images on this website are AI generated.

References

Domingos, P. (2015). The Master Algorithm: How the Quest for the Ultimate Learning Machine Will Remake Our World. Basic Books.

Goertzel, B. (2014). Artificial general intelligence. In Advances in Artificial General Intelligence (pp. 1-17). Springer.

Kshetri, N. (2018). Blockchain’s roles in meeting key supply chain management objectives. International Journal of Information Management, 39, 80-89.