the prompt engineer profession died before it had time to appear

the prompt engineer profession died before it had time to appear

Researchers are sure that the profession of prompt engineer will not receive much attention and will not become in demand. The point is that language models formulate content creation queries better and more accurately than humans. Therefore, a specially trained neural network can become a prompt engineer.

Rick Battle and Tej Gollapudi, who work at VMware, said that language models often react unpredictably to requests. For example, a very detailed and detailed query can give a high-quality result, but non-obvious hints significantly improve the situation. For example, if you ask the language model to explain its reasoning step by step, it solves mathematical and logical tasks better.

The researchers noticed that even asking for a step-by-step explanation of actions was not always helpful. The quality of the answer is affected by the question, the set of input data and other parameters. Different combinations give different results. It is noted that language models understand written language and answer questions with text, because of this people anthropomorphize neural networks. All this leads to the fact that a person tries to explain the request in the way that he would pass the task to another person.

This approach does not work well. For example, Rick Battle talked about the method of composing prompts using a language model. Most of such queries are illogical to a person, but give a better result than the method of selection by trial and error. In one case, the query generated by the neural network began with a reference to Star Trek, and this significantly improved the generated response.

Battle claims that neural networks are a set of mathematical models and algorithms, and people cannot know exactly how they work and which of the teams have a greater influence on the neural network. Therefore, even users who call themselves experts in prompting simply find successful combinations, but cannot immediately compose good requests.

As another example of the expediency of using a neural network to generate prompts, a recent experiment by the Intel team is cited. The company’s engineers have developed a tool for Stable Diffusion, which receives the user’s request and then transforms it into the optimal one for the neural network. As a result, the images generated using the transformed prompts are more detailed and aesthetically pleasing. To transform requests, Intel developers trained a special neural network.

Image generated by human query (left) and transformed query (right)

Researchers believe that the profession of prompt engineer has already lost its relevance, without having time to become in demand and popular. Current language models can generate more correct and efficient queries than humans.

Related posts