How Is Prompt Engineering Evolving? First Prompt Engineering Conference 2023 (Track 2)

TL;DR

Prompt engineering is evolving through the strategic design of prompts that help queries flow more easily through language models. Stefan Pan connects this evolution to constructal theory, a framework based on a law proposed in 1996, and examines information flow in encoder-decoder architectures, token sampling and generative AI. Read on for the talk's definitions, examples and theoretical framework.

Transcript

all right as folks continue to come into track two I again am your host and moderator today West Shields I am joined by Dan clear of promp Hub and we are GNA kick things off with a wonderful talk from Stefan Pan um he is a uh CTO software engineer it architect um who focuses really on delivering generative AI Solutions throughout the software indus... Read More

Key Insights

• "today is about the evolution of prompt engineering and before deling into the topic a few words about me and my motivation regarding this talk" (0:38)
• "prompt engineering is a accting edge discipline that focuses on harnessing the full potential of language models by strategically designing input instructions prompts" (2:16)
• "queries flow through the model and they will flow more or less easily depending on how well that have been crafted" (5:35)
• "you probably know that llms are considered zero short reasoners" (8:50)

Questions & Answers

Q: What is prompt engineering?

Prompt engineering is described as a discipline focused on harnessing the full potential of language models by strategically designing input instructions, or prompts. The talk highlights language models and prompt design as its two main subjects.

Q: Why is prompt engineering evolving so quickly?

The speaker describes prompt engineering as a relatively recent domain that is evolving at an unprecedented rate. He also points to exponential growth in machine learning and artificial intelligence publications on arXiv.

Q: What is constructal theory?

Constructal theory is presented as a theoretical framework based on the constructal law. The speaker says it comes from mechanical engineering and non-equilibrium thermodynamics.

Q: What does the constructal law state?

The law states that a finite-size flow system must evolve its configuration to persist over time. This evolution gives easier access to the currents flowing through the system.

Q: When was the constructal law proposed?

The constructal law was proposed in 1996 by Duke University professor Adrian Bejan. The transcript also says he received the Benjamin Franklin Medal in 2018 for his discovery.

Q: How is prompt engineering a flow system?

Queries flow through a model and move more or less easily depending on how well they have been crafted. The talk uses this idea to connect prompt engineering with constructal theory.

Q: How does information flow through an encoder-decoder architecture?

The encoder compresses the input sequence and is presented as the more straightforward, low-resistivity flow. The decoder is described as more complicated and therefore a higher-resistivity flow.

Q: How are token probabilities explained as flows?

During generative model sampling, token probabilities are treated as flows. Rare, low-probability words encounter more resistance, while common, high-probability words are associated with lower resistance.

Summary & Key Takeaways

• Prompt engineering can be copyrighted like software code, ensuring originality and protection.

• Patenting prompt engineering involves demonstrating novelty, non-obviousness, and utility, considering the US Supreme Court's Alice decision.

• Balancing copyright and patent rights is crucial in prompt engineering to protect intellectual property.