Tuning ChatGPT: Finding the Right Frequency

The ChatGPT 4 model has a staggering 1.76 trillion parameters. We can think of parameters as internal settings or dials within the model that refine how input (tokens) is processed to generate output. Alternatively, one can liken the parameters to representing the diverse opinions or voices of 1.76 trillion people. The prompts you provide to ChatGPT determine which subset of these opinions it draws upon to generate a response.

If you don’t provide models like ChatGPT with the right prompts or training, it's uncertain whose 'opinion' you are tapping into. Certainly, you can instruct ChatGPT to adopt the role of a copywriter to enhance the output. However, will it reflect the insights of skilled copywriters who truly understand your brand?

Jordan Wilson l, host of the Everyday AI Podcast, often emphasizes the critical importance of effective prompting and highlights the drawbacks of evaluating ChatGPT's output based on zero-shot prompts (no training). He argues that "garbage in, garbage out"—high-quality input is essential to fully leverage models like ChatGPT. Jordan illustrates this with an example where ChatGPT generates advertising slogans for Nike, comparing outputs from zero-shot prompts with those from trained models and human copywriters.

Using a zero-shot prompt is like asking a billion people, who might only have a passing familiarity with Nike and slogan creation, to unanimously decide on a slogan. The result? Predictably average. With his expertise in prompting and a deep understanding of Nike's brand from past collaborations, Jordan fine-tuned ChatGPT with approximately 4,500 words (or around 30 pages of text, yes you read that correctly). In my analogue this process is like refining the model's focus to the perspectives to say just a thousand people whose insights align closely with Nike's branding ethos. The difference was clear: while the untuned model produced a generic slogan, the optimized ChatGPT's output rivalled those crafted presumably by a highly skilled human copywriter (based on popular opinion).

Zero-shot prompting can be unpredictable, akin to soliciting opinions from a random group of people. Much like in life, it’s about asking the right questions to the right group.

As we're all beginning to understand, getting effective output from ChatGPT requires more than just inputting a simple prompt; expecting spectacular results from minimal input is unrealistic. To achieve success, you must provide the model with precise and relevant information. As Jordan demonstrated, tuning the parameters to resonate with the right "group of people" involves both time and expertise.

Just as you wouldn’t expect exceptional results from a new copywriter or design agency without providing them with a comprehensive design brief—including information about the company, its purpose, values, brand identity, and guidelines—you should have similar expectations when working with ChatGPT. To get the best outcomes, it's essential to supply it with detailed context and specific instructions.

#AI #ChatGPT #MachineLearning #TechnologyInnovation #DataScience #Advertising