LinkedIn co-founder Reed Hoffman has recently co-authored a book with GPT-4, aptly named ‘Impromptu.’ This writer had to resist the temptation of using prompt engineering to generate a provocative introduction while drafting this article on prompt engineering.
Like many others, he also believes that it will not take too long to make this kind of co-piloting a norm, and hence, it is a good time to explore the future with prompt engineering, how it is going to shape the work by eliminating or reducing the relatively low-skill, repetitive mundane day-to-day tasks.
As the initial craze of using ChatGPT and other large language model (LLM) systems to do homework subsides, we must evaluate how augmentation of human and computer intelligence can open new possibilities and deliver hyper-productivity and impact. Human intelligence can be broadly categorized into crystal intelligence and fluid intelligence.
The rote memorization capability and the ability to retrieve random factoids from one’s memory can be described as crystal intelligence. Navigating an unknown city requires one’s ability to identify patterns and juggle various probabilities, which can be associated with fluid intelligence.
These first-generation LLMs currently do not show a lot of fluid intelligence. However, they bring a ton of crystal intelligence to the game. We know that the crystal intelligence of an average human has an upper bound much lower than the one of a decent LLM. There lies a unique opportunity to augment both human and artificial intelligence to gain a many-fold increase in human productivity. Prompt engineering has been advertised as a way to achieve that and has become the hot new role in town.
Before diving into prompt engineering, we want to take a quick detour to see what underlying technological advances fuel this growth area. Even though the academic discipline of artificial intelligence goes back more than half a century, only advances in the last decade in hardware technologies, GPGPUs particularly, opened the door to massively parallel computation capability, a prerequisite for training these machine learning models.
Before the seminal 2017 paper on transformer architecture by Google Brain, various recurrent neural architecture (RNN) and long-short-term-memory (LSTM) were used to train LLMs with modest success. The transformer model led us to the well-known generative pre-trained transfer (GPT) model.
The GPT model is superior to its predecessor in taking less training time and supporting longer context. These two key advances appear crucial for what we are calling today prompt engineering.
The GPT models can be trained on various inputs such as images, text, and computer codes to learn the structure and pattern of the input data and can be used to generate an output of a similar kind. Additionally, the support for longer context and the availability of in-context learning observed in LLMs can be combined to provide prompts to accomplish certain tasks.
In plain terms, prompt engineering is the idea of manipulating textual prompts to the LLMs to get tasks done by them in conjunction with other generative AI models. Currently, almost all the output of these tasks is digital. They range from generating boilerplate computer codes to drafting essays (mostly in English for now) to generating audio and image content. Any good prompt engineering task involves steps like clear problem formulation, breaking down the problem into smaller navigable tasks, adding constraints to the proposed solution, and fine-tuning the responses.
So far, prompts involving not-so-specific requirements often generate broad, generic answers unsuitable in the required context. The prompt engineer’s task is to understand the nuances of LLM’s language processing capability to decompose the prompt in a series of suitable steps and add constraints to make the response useful.
A prompt engineering task usually starts with a stated goal and an initial prompt. As the output gets generated, various tests are performed to check the quality of the output to the desired goal. The more precise the prompts are, the better the output is. This refinement of prompt is the key place where a prompt engineer can truly excel.
Creativity with prompts and the ability to steer the model’s response in the desired direction will be a highly sought-after skill. Prompts also should add guidance on uncertainty quantification on the generated response to avoid fictitious answers. The final goal is to produce a set of clear, precise, and effective prompts that can be used to interact with AI systems, enhancing their utility and ensuring that AI outputs are aligned with user needs and objectives.
These new AI frontiers will have some disruptive impacts on Bangladesh’s work culture and economy. It will open new opportunities for consumers and practitioners and make some old assumptions obsolete. On the consumer end, generative AI models and corresponding prompts to create output from them will launch us into an era of hyper-productivity regarding routine and mundane tasks.
For example, anyone who has compiled month-long sales data into an executive summary and PowerPoint presentation knows the sheer agony of the process. Prompts can be carefully designed to take the raw sales data and perform a series of steps to produce the reports. Similarly, a collection of literature containing the standard operating procedures of a plant can be embedded into an LLM, and precise questions can be asked to find out about the maintenance schedule or shutdown procedure of equipment without hectically searching through a trove of paper documents. Anyone can take existing publicly available court records and compile an instantly searchable reference for exact matches and for finding and citing precedence. AI-augmented graphics will dramatically shorten the iteration loop between the client and the artist.
One common problem with many creative tasks is the ‘blank page’ problem, i.e., where to start. Using prompts can greatly alleviate the problem and generate some starting points for reports, emails, etc. The most exciting thing is that the users do not need a tech-heavy background to rip these benefits. Someone with analytical and critical thinking skills who can decompose a complex problem into a series of simple problems can take full advantage of these tools.
Mundane office work like report generation, analyzing market data, emails, and proposal writing should be outsourced to generative AI with some human supervision. Business leaders should proactively do an audit of the daily activities of their organizations and ruthlessly offload tasks to AI and free up the employees’ time to do creative thinking. This is where prompt engineering will lend itself particularly useful.
On the software development front, there is both good and bad news. Mundane and boilerplate coding tasks are already being outsourced to generative AI. Tools like GitHub Co-Pilot and ChatGPT can generate solutions for simple, routine problems. This is going to have an impact on online marketplaces like Upwork, Fiverr, or Freelancers. Some clients will find it easy and cheaper to use the tools themselves to build websites and forgo the outsourcing step. But then again, developers can also take advantage of these tools and amplify their productivity manifold. Businesses whose revenue model mostly relies on low-skill tech jobs must evaluate their approach in this changing time.
Beyond that, there are huge opportunities to build models catering to local needs. One of the key limitations of the current offerings is the unavailability of the local language for both prompts and output. For a country where English is extremely limited, there is a huge opportunity to offer solutions in Bangla. We should look at China and India to understand how they tackle this challenge, and there should not be any shame in borrowing ideas. While the lack of digitization of records and other documents is a big bottleneck, it is also the path we should invest in.
If we look at Google search trends for prompt engineering, we see almost zero interest till mid-2022 and then a sharp rise in public interest. So, it is still too early to say whether this prompt engineering boom will last or bust. However, it isn’t easy to uninvent something, and prompt engineering in some form will get integrated into our daily lives. Tech companies in the ranks of Google, Microsoft, and some universities are now offering courses on generative AI and how their proprietary prompt interfaces can be used, and first-movers will always have the upper hand.
While Bangladeshi businesses may still shy away from creating a chief AI officer position in their executive leadership, it is high time for them to initiate task forces to explore the application of generative AI in business practice.
Sameeul Bashir Samee is a computational scientist currently working at the National Institute of Health in the United States.