With the onset of tools like ChatGPT, Stable Diffusion, Dall-E, Midjourney and more, and increasing amount of people are being exposed to the power of large language models (LLMs) and transformers. These tools are incredibly powerful and open the door to new possibilities, previously beyond the reach of our technical capabilities. Generative models have transformed users into sorcerers: accepting spells via a chat box dialogue (or, discord, in Midjourny’s case) and producing text and art from seemingly thin air.
Users are currently expected to interact with these models through prompts; as such, the act of generating text or art via these prompts has been dubbed, prompt engineering. It’s easy to consider prompt engineering as the future – as the application of these models expand and invade all our daily workflows, much of what we do will likely be tied to writing a well tailored prompt to get the oracles to provide us answers. We fail, however, to recognize that variations of “prompt engineering” define our present, and our past. If you’ve ever used a search engine like Google, you are well aware that your ability to get the right search result is directly correlated to your ability to ask the right question. This idea stretches beyond the internet – in life, getting the right answer is directly tied to asking the right set of questions. Prompt engineering is only the latest evolution in a phenomena that has existed since humans could communicate. If you want ChatGPT to provide you with the right flavor of text, you need to give it a well tailored prompt (i.e., ask the right question, in the right way).
While powerful, the premise that the generative power of these models will make all work obsolete is fundamentally incorrect. Few shot learners like the GPT models will be able to excel at certain tasks that require little expertise, but they lack what a human operator has by nature: accountability. If your company leverages a GPT model to automatically handle user chats, and the chatbot someday decides to use an expletive or an offensive phrase, who do you point the finger at and blame? You’ll apologize profusely to the customer, probably provide some sort of discount, but you have no recourse to ensure such an incident never happens again because these models are black boxes. You can’t fire the model, or give it a poor performance review; you could yell at the proprietor serving you the model, but the model is a black box even to its creators, and they can make no promises. While you likely wouldn’t hire a human employee that would use such language with a customer, this just an extreme of an end-user interaction gone awry; human operators will make mistakes too, but you can teach them and help them correct their mistakes in a way that you can’t with machine learning models.
LLMs, like many deep neural network architectures, are beyond the realm of explainability – that makes them a business liability, especially for client facing applications. You typically would like to minimize the opportunities for agents that can cause irreparable harm to your brand to interface with end users. The immediate solution here is that a human operator could vet the responses that a GPT model spits out before surfacing it to a user – in this way a single operator becomes more efficient by being able to handle more customers, since they are not required to come up with custom responses for each customer and context. This is the future of work. The augmentation of human intuition with the generative power of these models will make the individual much more efficient, and allow them to scale their work and expand the radius of their impact.
The symbiotic relationship between human and machine is one that is still in its infancy. We use machines, but not in the most efficient way. Models like ChatGPT and its successors will increase our capabilities, but they can’t replace us. The conversation will be different once (if?) we ever reach Artificial General Intelligence (AGI), but for now, these tools are best understood as allowing individuals to move up the value chain.
There are, of course, dangers associated with these models as well. Any powerful tool can be as harmful as it is useful. The ability to generate cohesive short and long form text from a well formed prompt means social engineering gets easier for malicious actors. The spread of misinformation will explode, and it will become increasingly difficult to tell the difference between what’s been produced by a human, and what hasn’t. Misinformation without the aid of generative models was already difficult to control, but the problem will likely become intractable as these models see increased adoption. The lines are now forever blurred, and the Turing test is no longer a sufficient measure of a machine’s intelligence.
We should be cautiously optimistic about new technology, and be aware of the the need to develop that technology, but also invest in educating users and the general public on how the world around them is evolving. Explainability is an area that will need better UX as we consider wider applications of generative models. The tools at our disposal will continue to grow more powerful, but so to must our understanding of them. We can’t move up the value chain if we can’t use our tools effectively, and we can’t use our tools effectively if we don’t understand them.