Introduction to ChatGPT
If you haven't tried ChatGPT and are a bit lost about where to start, the following will help you. ChatGPT is software that allows you to use artificial intelligence; it's a type of chatbot developed by OpenAI, capable of engaging in conversational dialogue through natural language processing. It can perform a variety of tasks including answering questions, composing essays, summarising documents, developing software, and recognising images.
Background and Evolution of GPT Models
Let's understand, in simple terms, more about GPT-4 and other models, (my apologies to the clever people who worked on it).
Machine learning (ML) has been around since the late 1940s. The recent development of being able to program Artificial Intelligence (AI) through natural language has occurred as a result of combined breakthroughs in mathematics, the internet, and computer hardware, in particular, memory and processing speed.
The mathematicians and engineers designed and built clever code called a Generative Pre-Trained transformer (GPT). You can think of this as a next-word predictor, like the one you use when you text. Once the algorithm (GPT) has been built, it is trained. This requires a vast amount of data typically taken from the web and books. The GPT processes this data and works out the relationship between words so it becomes a next-word prediction model. This takes a lot of computing power to do. Once complete, the model is not perfect; it needs fine-tuning, which is done by asking it questions, seeing its response, and giving it feedback.
Once the process is complete, it can be used either as a closed model, like GPT-4, where you have very little control over the way it works via its weightings, or as an open-source model like LLAMA from Meta, where you can adjust the way it works by changing the weightings. Some models are small, like Mistral’s 7 billion parameter model which can run on a laptop, whereas GPT-4's 400 billion parameter model requires Graphic Processing Units. These are very powerful arrays of chips and, at the time of writing, are expensive and highly sought after.
Overview of GPT-4 and why you need it and not the free version
There are several versions of ChatGPT you can use. The free version is GPT-3.5. The paid version is GPT-4. I’ll describe some of the differences:
GPT-4 has a significantly enhanced capability to understand and generate more text. This is because its model has 400 billion parameters*, whereas GPT-3.5 has 175 billion parameters.
GPT-4's context window is approximately 25,000 words (32,000 tokens), while GPT-3.5 has a window of 6,500 to 7,000 words (8,192 tokens).
This is important; the larger context window allows for more extended interactions, enhanced linguistic finesse, programming power, image and graphics understanding, and a reduction in inappropriate or biased responses. However, while GPT-4 offers more depth and complexity in its responses, it is slower than GPT-3.5.
GPT-4 has a limit of 25 to 50 messages every 3 hours; GPT-3.5 is unlimited. From experience, it is quite feasible to hit this cap, and you have to use 3.5 or have a coffee break before proceeding further with GPT-4.
GPT-4 is multimodal; it can do more than generate text; it can code, read documents, examine pictures, generate pictures, write computer code, and a lot more.
If you want to use the best model with all the functionality and are serious about using ChatGPT, then you need GPT-4. At $20.00 a month, you will soon find, as I have, it’s good value. I do not get paid to write this by OpenAI.
Practical Applications of GPT-4 in Knowledge Work
Many of us are employed in what we call knowledge work, which involves many separate tasks. You could describe these as writing, talking, analysing, explaining, presenting, and persuading. Information comes to us, we process it and in turn, add value to it before passing it on to others via conversation, emails, texts, spreadsheets, graphs, and presentations.
All of these tasks have one thing in common; they come under the heading of communication. To improve our communication, we are learning new skills and adapting old ones to make sense of the world and convey messages to others.
When you use GPT-4, you realise that, in its multimodal form, it can integrate into what you do. It has tools and features which can be used in various situations to help you. GPT-4 is the Swiss army knife of knowledge work. The tools that it has when used properly, can not only speed up the way you do things but can also improve what you do. And if you are doing things faster and better, this is a “good” thing. In a market-based economy, it makes you more efficient and therefore less prone to obsolescence.
How GPT-4 helps
When you consider all the skills we use at work, you realise the development of GPT-4 has not been a haphazard process. Its multimodal functionality helps us undertake what we do. Let's have a look at some of these features:
Writing: Through the chat or prompt interface, you can communicate with ChatGPT and ask it to do anything related to text. It will write an email, a thank you note, a memorandum, a letter; this list is endless. It will write in a particular tone or style or enter a conversation with you.
Information retrieval: GPT-4 has a training database drawn from a mass of information from the web. You can ask it for facts, information, and opinions on just about any topic. The model that it currently works on has a cut-off point of April 2023. If you want reliable information after this date, you must ensure that GPT-4 is enabled to browse the web.
Browsing: GPT-4 does this through Bing. When you converse with GPT-4, it can search the web for more up-to-date information and use that in its response. With the browse function enabled, GPT-4 will decide when to browse. I’ve found it helpful to ask it to browse, particularly when you want events post-April 2023 to be included in your conversation.
Data analysis: GPT-4 has a powerful data analysis capability, which works with both numeric and text data. For example, you can upload a spreadsheet and ask it to analyse trends or upload text-based customer feedback and ask it to comment on sentiment. With trial and error, GPT-4 will help you gain insights quickly, easily, and cheaply.
Vision: You can upload pictures to GPT-4 and work with these. For example, you can upload a handwritten note and get it transcribed, and then rewrite the contents using GPT-4. You could upload a picture of a report, for example, a technical medical paper or diagnosis, and ask GPT-4 to explain the contents to a layman, and it will do it.
DALL·E: This is Chat GPT's picture-generating model. You can ask GPT-4 to generate a picture by giving it instructions in the prompt window; it will then generate a picture that mirrors your request. You influence the output by altering your prompt.
Writing computer code: GPT-4 enables non-programmers to write executable code, it not only generates code snippets that run directly but also offers features like code cleanup, code changes, and insights on coding practices. It can write in languages such as Python and HTML, making coding accessible to individuals with varying levels of technical expertise. Furthermore, GPT-4's autonomous code-writing capabilities are closely linked to the data analysis function in that it writes code to undertake the data analysis tasks; prompting the model leads to code writing and execution and as a non-programmer, this is how I use it.
Custom GPTs: These represent an advanced yet user-friendly feature that allows you to create tailored AI models based on specific prompt configurations. Imagine the process of interacting with GPT-4 by crafting detailed prompts to guide its responses or actions. These carefully constructed prompts can be transformed into custom GPTs, enabling you to automate repetitive tasks or specialised queries without the need for continuous prompting or retraining of the base GPT-4 model. These custom models can be saved for personal use, shared with select individuals via a direct link, or made publicly available through the GPT Store. The GPT Store functions similarly to an App Store, offering a platform where anyone can discover and utilise custom GPTs created by others. I find the custom GPTs very helpful as they let me use GPT-4 in a way that would normally require developer skills.
Accessing GPT-4
There are several ways to access GPT-4; the best way depends on what you need it for:
For individuals seeking a straightforward and user-friendly way to interact with GPT-4, a personal subscription to ChatGPT is the best option. It provides easy access to the AI's conversational capabilities without the need for technical knowledge or development skills. This subscription is well-suited for a wide range of personal use cases, from learning and entertainment to productivity and creative assistance.
For business users, there is either API access, which is direct access to GPT-4 through OpenAI's API, (encrypted key), this is suitable for developers and businesses looking to integrate AI capabilities into their applications or services. Or Team GPT which provides collaborative features and more confidentiality (you can stop OpenAI training on your data). It's designed for enterprise use, requiring a more expensive business subscription due to a minimum of two seats.
You can also use OpenAI’s sandbox, which is another environment primarily for development where you can adjust the parameters and see how various prompts behave.
Finally, you may be using GPT-4 without knowing it through third-party websites or applications integrating the API into their offerings. This method can vary in user experience and functionality, depending on how the third party has utilised the GPT-4 capabilities.
The World of Work and GPT-4’s Role
In the knowledge economy, our work involves reading, writing, and arithmetic (if I can use that as a description of coding, analysis, and strategy), and the faster that we can obtain information and convert it into something valuable and then communicate with others, the more efficient we become and the more valuable we are.
When you consider GPT-4, it has all the things that we do built into it. It can write, browse, analyse, study pictures, make pictures, write code, and enter dialogue, it can save your prompts and store them for you; it is multifunctional.
When you see this, you realise GPT-4 and similar general models are massive steps forward in our ability to deal with data and convert this quickly and intelligently into something usable and helpful. Because these models are the worst models that we're going to use and they are getting better and better very quickly, they will soon be on par with, or better than, human experts in a given field. Where your work is not at an expert level, then it is very likely that these models will replace what you do because they are cheap. This means economic disruption for many. The only way to take advantage of these changes is to embrace the use of these models in the work you do. The best way to do this is to see them as helpful graduate trainees who have a very good basic understanding of how the world works, but they do not currently have in-depth expertise that has been gained through experience. By working with them, much of the day-to-day work can be delegated out, and you can concentrate on using your experience to take your work to a higher level.
Summary
Models like GPT-4 have changed things. In white-collar industries, we are paid for our knowledge and ability to use information, process it, and complete all sorts of tasks. I’ve called this communication.
At first sight, GPT-4 is a chatbot, but once you explore all the things it can do, there is a whole suite of tools like the proverbial Swiss army knife. There is no training manual because how you use it depends on what you want it to do, your curiosity, and your willingness to accept change. Because this computing power is cheap and effective, the workplace will change; if your job can be done by GPT-4 or its offspring, your bargaining power is likely to decline.
The rate of change will be much faster than, say, web usage as the infrastructure and incentives are in place, and the advancements are speeding up, not slowing down. You should consider bringing these models into your work to future-proof what you do.
Use them as helpful assistants, treat them in the same way you would a co-worker who you are instructing, and you will be amazed at what they can do for you. This is truly where you can learn. Using GPT-4 is a very practical skill that is largely gained from practice.
You can get the basics in a few hours but “intuition” about its capability needs a lot more time and experimentation.
By working together, you can see what you can do that it can’t, this is where you are adding value and it’s where the future is. Critical thinking, emotional intelligence, and complex problem-solving are increasingly valuable in a GPT-4 augmented workplace. This is both a great change but also a massive opportunity to do new things.
*A note on parameters:
Think of the parameters in a GPT model as the internal parts of an engine that you can adjust to optimise performance. These parameters are akin to the nuts, bolts, timing belts, spark plugs, and fuel injectors in a car's engine.
Parameters as Engine Components:
Adjustable Settings: Just like you can tweak the timing of the spark plug or the air-to-fuel ratio to get the best performance out of an engine, in a GPT model, the parameters (which are essentially weights in the neural network) are adjusted during the training process to make the model more accurate. These adjustments are based on the data the model is trained on, similar to how you might tune an engine based on the type of fuel it will use or the conditions it will operate in.
Complexity and Performance: The number of parameters in a GPT model is similar to the complexity of an engine. An engine with more components (like a V8 versus a V4) can be more powerful if tuned correctly, but it's also more complex to manage. Similarly, a GPT model with more parameters can process language more effectively, capturing nuance and complexity in text, but requires more data and computational power to train and tune.
Optimisation for Efficiency: In the same way that you might use a turbocharger to increase the efficiency and output of an engine without making it bigger, optimisation techniques in machine learning adjust these parameters to improve the model's performance without necessarily increasing its size. This involves fine-tuning the model's parameters to ensure that it can understand and generate language as accurately and coherently as possible.
Fine-Tuning for Specific Tasks:
Imagine you have a basic engine setup that you can then fine-tune for different purposes, like racing, towing, or fuel efficiency. Similarly, once a GPT model has been pre-trained on a large dataset to understand language generally, it can be fine-tuned by making smaller adjustments to its parameters for specific tasks, such as writing in a particular style, translating languages, or answering questions. This fine-tuning process adjusts the model's internal settings to specialise in the task at hand, much like adjusting an engine to optimise it for a specific type of driving.
In the workshop of machine learning, parameters are the internal components of the GPT "engine" that determine how it processes and generates language. By adjusting these parameters through training and fine-tuning, similar to how a mechanic would tune an engine, we can optimise the model's performance for specific tasks or improve its overall efficiency and accuracy. Just like with engines, the skill lies in knowing how to adjust these parts correctly, balancing power, efficiency, and usability to meet the desired outcome.