7 Things I wish I knew before using ChatGPT

The past year has seen great hype surrounding AI and Chat GPT. There's no denying that the advancements made in this field are remarkable. However, the hype around AI can sometimes be likened to a gathering at a local pub where old-school friends bragg about potential and success. Still, the reality is more mundane and less revolutionary than the stories suggest.

?

So, what is it then??

?

Chat GPT is a Large Language Model (LLM); essentially, it is a powerful computer that uses vast amounts of text data to mimic human interaction. Here, the word 'mimic' is critical. ChatGPT identifies patterns in text through complex algorithms and understands conversational language, predicting responses. The accuracy of these models depends on the quality and quantity of training data and the calibre of the algorithms used for training. They improve over time with human feedback and iteration but are still fallible and rely heavily on human input. Despite the hype about its 'intelligence', they are not independent of human intervention.

To leverage its potential without falling into the trap of complete scepticism or utter confidence in these tools, there are some worthy things to consider:?

?

1. There are diverse models?

ChatGPT's free version uses GPT-3.5 and can complete simple tasks. However, it has some limitations, such as sometimes giving false responses. It's reasonably accurate for summarising and drafting text but not as great for more complicated tasks like creating worksheets.

Claude 2 is an upgraded version of Chat GPT 3.5 that allows you to upload documents for more contextual interactions. However, the free version has usage limitations. Claude 2 is especially good at creating summaries of long documents and teaching resources, making it the best free model available for general tasks.

Microsoft's Copilot uses Chat GPT-4 under the hood. This versatile LLM can generate images and tasks from uploaded pictures. It's free, but chats are limited to 30 responses, and the context window is only 2000 characters. While its output is less reliable than Claude's, it's worth giving it a go.

Google Bard is a language model that has access to up-to-date internet information. However, its quality of answers is not as good as other models, and it lags behind in accuracy. It has recently undergone an upgrade, but it is recommended to use it through Google Workspace for more functionalities.?

?

2. If you want the best, you have to pay

The most advanced LLM is ChatGPT 4 turbo. It offers extensive capabilities. Suppose you are serious about exploring the potential impact of LLMs in your area of expertise. In that case, getting a subscription with a limit of forty messages every three hours is worth it. With suitable prompts, the outputs are way superior to the other models.?

ChatGPT 4 gives you access to the GPT Store. This new feature provides access to specialised mini-chatbots for tailored responses in specific fields and for different tasks.

Microsoft 365 CoPilot is a new tool that offers exciting functionality inside Office products. Its 'reference a file' feature is a standout utility. However, it has glitches and is yet to deliver on its promises, being less than a week old at the time of writing.?

?

3. Context and Prompting are key factors

?These models don't understand language the same way humans do. They use complex algorithms to generate responses based on the context provided rather than retrieving responses from a source. This means they can develop plausible but fictitious responses. If you're not an expert in a field, you could be misled into thinking the responses are accurate, leading to misinformation.

The effectiveness of LLMs heavily depends on the context and specificity of your prompts. AI can't guess what you want. Your responses will be better if you tell them what you want and are specific. It would help if you detailed the content type, target audience, language tone, output format, etc. Also, these models seem to hallucinate less when they are provided with the source of information from which they will produce their outputs.

?

4. Manage your expectations?

Expect outputs to be drafts needing further refinement. The better the model, the higher the draft quality. It is a helpful assistant; the buck stops with you.?

5. Consider inherent biases and privacy concerns

LLMs can be biased because they learn from the data they are trained on. Even with built-in safeguards, their results could lead to harmful stereotypes, prejudices, or discrimination against certain groups. Additionally, some models handle private data differently. For example, Open AI uses 'interaction data', i.e. your chats for training, whereas Claude 2 promises not to use your data for model training. If you're concerned about privacy, consider anonymising your data or avoiding entering sensitive information altogether.

?

6. LLMs are not good with numbers

It is counterintuitive, but these models are generally unreliable for math or factual data unless they are contextually bound to a reliable source like a textbook. For example, when you upload the information you want the LLM to use. Also, ChatGPT 4 has a feature called 'code interpreter', specifically designed for number crunching. It is much more reliable.?

?

7. LLMs should be off-limits for students

These tools best suit teachers who want to create resources, write emails, and automate simple administrative tasks. They can also help turn a blank page into a helpful draft. However, they are unreliable for assessing student work as they cannot make accurate inferences. This is especially true for the free models. Teachers should be aware that students can use these tools to complete coursework that requires writing. Even though experienced teachers can often recognise plagiarism, there is no sure way to confirm it.

I recommend against promoting their use among students. Writing and critical thinking skills risk being underdeveloped if students have access to these mod

Despite the hype, there is still a lot of potential for ChatGPT and other AI-powered tools for educational purposes. However, it's vital to approach them with a balanced perspective and understand their limitations. We need to keep a critical eye to ensure they are helpful aids, not a hindrance. As LLMs are still in their early stages, let's explore their possibilities with an open mind while remembering they are only a means to fulfilling our educational goals.

?

Thanks for reading,

?

Francisco?