100 million users in 60 days! ChatGPT broke records with that number by some margin (Although eventually broken by Threads by Meta). A cited 97% of business owners think ChatGPT will help their business. This means in 2023, Artificial Intelligence became an agenda item in boardrooms around the globe.
However, while the general public only started paying attention to AI this year, those working in machine learning labs know it's been around for decades.
ChatGPT is an LLM (Large Language Model). LLM is only one of the many categories where AI has made astounding progress. Other AI disciplines, including computer vision, predictive models, and autonomous driving — to mention a few — have achieved noteworthy milestones.
The current focus, however, is on NLPs and ChatGPT specifically. LLMs were invented to tackle the seemingly insurmountable obstacle of communicating with computers using "natural language." Natural Language refers to how a language has naturally developed and evolved over time through groups of humans speaking and using it. This is in contrast to using a logically developed programming language to communicate with computers.
ChatGPT captured headlines for the following reasons:
- NLP (Natural Language Processing) was considered the most challenging area of AI. Computer scientists expected it to be the last area of AI to make meaningful advancements. Shockingly, NLP leapfrogged all predictions and became the first area to see rapid adoption.
- It is the first AI tool that the general public has gotten to play with frictionlessly. While not open source, it is very much open access.
- It works with a high degree of accuracy across almost all use cases. Many of the parts of ChatGPT that are not open access include plug-in abilities, which means its uses and utility are near universal.
The frustrating thing about a black swan event like ChatGPT is that now every man and his dog has an opinion about AI. These opinions are shared with the rest of the world through social media.
Unfortunately, this means the amount of uninformed nonsense espoused on LinkedIn and Twitter is staggering. Wild click-bait headlines abound: "Use ChatGPT to write emails," "Let AI do your copywriting," "10 ways marketers can use AI", or "Everyone should use ChatGPT for XYZ."
To understand why some of these claims are misguided or just blatantly wrong, we must look behind the curtain—learning how AI tools like ChatGPT work, enables you to make informed decisions on where and how to implement AI.
A Basic Explanation of LLMs
In a nutshell, language models absorb a library of text and predict words or sequences of words with probabilistic distributions. Breaking that down, absorb here means the AI is training on x to mirror output; a library of text is made up of books, blogs, websites, social media interactions, etc. Sequences of words mean phrases, idioms, sentences, related terms, and concepts. Lastly, probabilistic distributions refer to how likely a word or sequence is to occur.
For a clear example, when you say, "Tom likes to eat …", the probability of the next word being "pizza" would be higher than "table." This probability is determined by looking at the previously "absorbed" library and seeing how likely a given word appears after a sequence of words.
Since the probabilistic method will always produce the exact same prediction for a given sequence of words, some randomness is added to the model called "temperature." So the model will not pick the next word with the highest probability in the list; instead, it will roll the dice and choose from the middle of the degree of randomness introduced.
The Degree of randomness in this context means the programmed gap range for the next word. For example, if an AI were only to generate the first most likely word, it would constantly repeat. If you used a prompt, it would generate "Tom likes to eat pizza." Using the same prompt again would say, "Tom likes to eat pizza."
With temperature, the AI can generate "Tom likes to eat pizza." With the same prompt, "Spaghetti is Tom's favorite food." Purposeful inconsistency. There is obviously a great deal more to LLMs, but even this surface-level explanation — that the AI knows a library and generates with a controlled degree of randomness — will be sufficient for the purpose of this conversation.
The False Promise of an AI 'Content Creator'
Now that you understand the underlying methodology of LLMs, you can see why using it for content creation is limiting. In most circumstances, it doesn't make sense at all. In the world of copywriting: originality and uniqueness are the main currencies. New thoughts, new perspectives, and new ideas make valuable content.
However, if all content is just a slightly modified regurgitation of millions of other pages that the model was trained on, there is nothing original or creative about it. The following is my opinion on the areas where LLMs should not be used and where they should be. This is from the perspective of a software engineer, salesman, and leader who works daily with AI technologies.
Areas of Exaggerated or Misguided Use of AI Content Creation
Content Generation
The simple and most obvious use case of generative AI touted across the internet is: Use AI to write your blogs, emails, website, etc. This is probably the last place current LLMs should be introduced.
Now that you understand that the generative AI is spawning a probabilistic variant of the text corpus it's been trained on... don't do this! The use cases are small, perhaps if you want to submit a writing assignment or check a box for your marketing manager. Don't rely on LLMs to produce creative genius. You will end up with an unoriginal, synthetic-sounding, and overly wordy piece of content.
It will be classified as "noise" by search engines and your target audience. AI-generated content is infamous for not getting picked by search algorithms.
I asked ChatGPT to write on the same topic we are discussing and it produced standard, unoriginal and run-of-the-mill content.
Attention is a very scarce commodity. Attention can only be earned using engaging, original, and well-thought-out work. Within the scope of content creation, there is one solid use for ChatGPT, which we will discuss in the section Areas LLMs Should Be Used.
Sales Calls Follow Ups
A lot of tech pundits who have never conducted a sales demo are touting that LLMs should review the calls and perform follow-ups with prospects. While the idea has some logical merit, its implementation overlooks a significant component of any sales demo: non-verbal cues.
Often the transcript of a call is insufficient to capture the various emotions and unspoken objections that a seasoned sales rep has spent years noticing. Therefore, a follow-up message operating purely from the call transcript will be void of the aforementioned data points.
By citing those visual cues in your follow-up, you have the edge of appearing attentive and human.
Areas LLMs Should be Used
Answering Questions from a Corpus
A corpus refers to all types of textual data (e.g., books, PDF documents, spreadsheets, databases, etc.) on which the model is trained.
The phenomenal advancement of GPT's ability to process large chunks of text and maintain context while answering questions pertaining to it is wildly underrated. The generative aspect of GPT has been the focus of the public and the press, likely due to flashy and glamorous demos. However, true productivity gains will come from the comprehension aspect of GPT.
Imagine you had an employee who could retain and comprehend 100% of any information indefinitely. Now imagine that employee was answering customer support queries. This individual could recall the entire customer history, combined with company records and transactions, and provide the most meaningful support possible. Effectively LLMs are such employees.
Almost all LLMs can be trained on custom data: all customer data, support protocol, employee handbook, legal docs, shared files, etc. LLMs can be the ultimate on-demand training tool for any internal lookup by an employee or set up to answer all customer queries with high fidelity.
Research
Given the awe-inspiring ability of ChatGPT to comprehend any textual information, businesses have already started to deploy AI for researching and finding data of interest.
For example, at Plena, customers describe their ideal buyers to the AI bot. The program sifts through millions of LinkedIn profiles by reading descriptions about sections, posts, and job histories to cherry-pick those that match the ICP description. If product and service descriptions are provided, it can go a step further and determine "intent to buy" by looking at account-level data as well as interactions, groups, and activities for the matched profiles.
This is the dream process for most sales leaders. Until recently, this could only be done manually, making it infeasible for commercial adoption. It was just too time intensive. Plena AI is pioneering the research use case and has become an essential tool for a fast-growing sales team.
To Spark Creativity
Within content creation, there is one valid use of AI as a tool to break writer's block from time to time. It is helpful to try out a few prompts that engender new ideas occasionally, then use those ideas as nuggets for your own piece of writing. Don't get reliant! AI will always be limited by its content library, which is an echo chamber.
Reporting and Dashboarding
GPT, along with some open-source libraries, can produce meaningful reports. For example, reports from a typical business meeting or to provide managers and customers with presentable data.
AI can generate a high-quality report with minimal guidance and concise prompts. AI can wrangle data, aggregate numbers, present information, and accompanying charts and graphs. These tasks would otherwise be quite time-consuming.
The reporting use case for AI is getting better at a rapid rate. I predict it will soon become the commonplace replacement of all BI tools. There are a variety of different Business intelligence (BI) software currently performing some of this data analysis. AI can improve and adapt well beyond the capabilities of these softwares.
Conclusion
While ChatGPT has achieved remarkable early adoption, it is essential to recognize LLMs' limitations and appropriate use cases. By understanding their capabilities and integrating them strategically, businesses can harness the power of AI to enhance productivity, support decision-making, and unlock new possibilities across various domains.
The best practice is to think of ways AI can eliminate repetitive tasks while never allowing it to replace creativity, critical thinking, or human interactions.
