Page 20 - AI Vol 1: Foundations of AI
P. 20
likely to be eliminated completely. One of the engineering can further refine their outputs,
current areas of focus in LLM development is especially in complex or specialized tasks. For
on making the models better able to tell when a example, a sophisticated combination of Chain
user request requires a factually accurate answer of Thought and Multi-shot prompting was
as opposed to when the user is seeking creativity demonstrated in a study to enhance GPT-4’s
or problem solving from the model and allowing performance, enabling it to outperform medical
the model to respond accordingly. Still, for the domain-specific models in medical exams. This
foreseeable future, users should expect the need finding suggests that while everyday use of LLMs
to check the accuracy of information provided might not require intricate prompting techniques,
by LLMs to continue, particularly for high there is a significant potential to unlock additional
stakes applications. capabilities through careful and innovative
prompting strategies.
8. PROMPT ENGINEERING
Prompt engineering is a technique that involves 9. THE FUTURE: AGI/ASI AND AI AGENTS
strategically crafting inputs (prompts) to elicit Artificial general intelligence (“AGI”) and
more effective or specific responses from LLMs. A artificial super intelligence (“ASI”) refer generally
prompt is essentially the initial text input given to to AI systems capable of reasoning at the level
an LLM, which guides its subsequent output. The of an average human or at the level beyond even
art of prompt engineering lies in understanding expert humans, respectively. These systems are
how different styles or structures of prompts can only theoretical, and even how AGI and SGI are
influence the model’s response. defined are the subject of much debate. That is,
how will we know that an AI system truly has
Two notable strategies in prompt engineering general intelligence on par with or even surpassing
are Chain of Thought and Multi-shot prompting. humans? Also, does it matter if these systems
Chain of Thought prompting involves structuring truly “understand” and are “intelligent” in the
a query in a way that leads the model through a same way humans are, or is it sufficient that the
step-by-step reasoning process, often helping it systems simulate such capabilities to a convincing
to produce more accurate and logically coherent degree, as discussed above. While these systems
responses. Multi-shot prompting involves are years, if not decades away, assuming they are
providing multiple examples or iterations of a possible at all, it is important to understand that
query to guide the model more precisely towards it is the explicit mission of many of the leading
the desired response. There are many other AI companies to develop AGI and eventually SGI
strategies, and this is an active area of research as systems. Moreover, understanding and adapting
more techniques are discovered to get LLMs to to current AI technologies will better prepare
produce better results. agencies for if or when AGI and SGI systems
become widely available.
While LLMs like GPT-4 are trained to respond
effectively to natural language requests, prompt
20 | VOLUME 1 FOUNDATIONS OF AI | LOZANOSMITH.COM