I expect that many of those who have been using ChatGPT in their daily work for some time have developed a relationship with this artificial intelligence behemoth that somewhat reflects relationships of a more human kind. The initial excitement at the discovery of something new and alluring matures into an all-compelling infatuation before the reality of the warts-and-all limitations of our AI partner become all too real. But if the relationship is important to us, we look past the superficial and find a way of sustaining a deep and meaningful equilibrium.

Sounds familiar? I think this idea would ring true for many early adopters of the GPT phenomenon. I can still remember the rush when the GPT chatbot responded to almost any question I could pose with brilliant alacrity. This blossomed into respect and admiration as I incorporated it into my daily work and found that tasks which previously took hours could be completed in minutes. But then the terrible reality dawned. Those interminable system outages. The completely irrelevant answers to poorly formed questions. The complete ignorance of anything that transpired since September 2021. And, of course, the dreaded hallucinations which led me to doubt almost everything Chat told me.

It was at that moment that I realised that if AI and GPT were to be more than just another IT fad that I would need to seriously commit to the relationship. It was at that moment that I realised that I needed to become a “GPT whisperer.”

As part of my Doctoral research I have been examining different techniques for obtaining best results from the generative pretrained transformer. I have undertaken experiments to see how changing the contents and even syntax of a prompt would dramatically affect the GPT output. I have also delved into the nuances of the OpenAI application programming interface which provides a much greater level of control and extensibility than the standard Chat bot familiar to most people. I discovered that there are various ways to “engineer” the prompts in order to consistently achieve more relevant and targeted responses. I also discovered a technique known as transfer learning (sometimes called retrieval augmented generation[1]) which mitigates two of the most significant limitations of the existing ChatGPT model: Its ignorance of any events or knowledge created since late 2021, and its ignorance of local knowledge – such as knowledge created and maintained within an organisation and not part of the public domain.

The concept of transfer learning is quite simple. It means that GPT is able to learn new things if they are fed into the system through the prompt. This means that GPT is able to produce high quality text that references contemporary and local knowledge as long as that context is provided to the system through the prompting process.

Quy Tang provides an example of this in his recent Medium.com article. “So roughly speaking, instead of simply <query> it’d be: answer following question given <relevant texts>, <query>”[2].

To apply this principle I have worked on several software development projects in my spare time that deliver consistently amazing results thanks to effective transfer learning.

The first, which I call GPT Primer can automatically produce a complete suite of draft learning and assessment documents on any subject within a few hours. GPT Primer works in both formal and informal learning settings. The GPT Primer user simply enters their learning objectives and some additional context into a webform, and the software converts this information into a series of prompts which include transfer learning. The prompts are passed to Open AI’s GPT 3.5-turbo engine via the API and the resulting documents are stored in a database where they can be viewed, proof-read and tweaked by the user to produce high-quality documents that they can call their own.

For formal learning settings where curricula are already determined, these can be entered in the webform to produce content which maps back to the formal requirements. In less formal settings, the user can even ask the system to automatically create the initial learning objectives. These can then be tailored by the user to their requirements before proceeding with the automatic document creation.

Educators who have experience in developing learning and assessment resources will agree that producing high quality, version controlled, and internally coherent learning and assessment resources can take weeks if not months to produce the “old-school” way. Thanks to the power of artificial intelligence and clever prompt engineering, this process can now be accomplished in one or two days.

My second project, GPT Job Hunter is designed to help job seekers (particularly those from disadvantaged backgrounds) even the odds when applying for work. I have utilised the same principle of learning transfer through retrieval augmented generation but for a markedly different use-case.

It is no secret that recruiters have been at a technological advantage to job seekers for some time. Most recruiters now use software to parse the CVs and cover letters that they receive and will automatically reject those that do not satisfy the computer algorithm even without any human intervention. Personally I find this trend to be repugnant because it further entrenches structural disadvantage in our community by excluding candidates with poor English language skills from even being considered for an interview. Those who are well educated or who have access to highly paid resume consultants may be able to tilt the odds in their favour. But for everyone else, the dehumanisation of the recruitment process is as good an example of technology producing a social-bad as one might find.

My idea with GPT Job Hunter is to create a tool that job seekers could use to produce quality written job applications. The tool allows the user to load and store their CV in the application. Then, when they see a job advertisement that they are interested in, they can copy and paste that into the application and it will use AI to provide feedback about their suitability for the job. If they wish to apply for the job after reading this feedback, they can use GPT Job Hunter to draft a well written AI-powered cover letter based on the skills and experience listed in their CV. The job seeker can even elect to ask GPT to tweak their CV so that it contains appropriate keywords which match the advertised position.

Because I am not currently looking for work, I have not been able to test GPT Job Hunter as extensively as I have tested GPT Primer. But under laboratory conditions I was pleased that the application rated my resume against a Chief Information Officer position as 8 out of 10, while it rated me as only 3 out of 10 for the position of Medical Receptionist.

The bottom line from all of this research is that the limitations of ChatGPT can be overcome with some clever prompt engineering and automation through API integration. The OpenAI platform delivered consistently reliable results in two very different use cases which suggests that this approach can be generalised to a wide range of situations and applications.

If you would like further information about either of these projects or the potential application of GPT transfer learning to your own situation please don’t hesitate to reach out for information or advice.

The good news is that my love affair with ChatGPT has not yet reached its natural conclusion. On the contrary, I have found a way to keep the flame burning and I hope the information I have provided in this article will help you to find yourself in a similar place. After all, “we didn’t come this far to only come this far”, and I’m sure that the benefits of AI will only become more embedded in our daily work and our daily lives over time.

John Burgess

Statement of Authenticity: This entire article was written by my own hand and in my own words. No part of this article was written by ChatGPT or any other AI text generator.

Footnotes:

[1] See https://huggingface.co/docs/transformers/model_doc/rag

[2] Tang Q, “Integrating Chatgpt with Internal Knowledge Base and Question-Answer Platform” (MediumApril 23, 2023) <https://medium.com/singapore-gds/integrating-chatgpt-with-internal-knowledge-base-and-question-answer-platform-36a3283d6334> accessed April 30, 2023