A Peek Beyond the Hype of Large Language Models
LLMS (Large Language Models) and their conversation counterparts ChatGPT are very good at performing a number of tasks. But the core tasks that stand out are their ability to
Understand the intent behind a question or proposed task.
Generate a response that is coherent, fluent, and syntactically correct.
Exhibit an appearance of semantic prowess.
Preserve some conversational continuity.
And provide a range of practical skills like summarizing, extracting key topics, cleaning up text, converting text to tables etc
By any measure, these capabilities are an engineering feat to be marveled at and represent a major milestone in the intellectual, scientific, and engineering pursuit of understanding language and taming it to serve our needs.
But like all tools, it is only through extensive use that you learn and experience the edge of their usefulness. By exploring these boundaries, you discover where a tool needs to improve.
And so it is with large language models.
For those who were not there, it's hard to imagine what it felt like to drive a car in 1905 or take the first passenger flight in 1914, be one of the first to listen to a music recording in 1895, watch the first TV show in 1936 or have someone call you on the phone in 1880. The magic of that moment would have remained with you for years.
And so it will be for large language models.
The great and lasting inventions target our basic human needs (health, mobility, communication, shelter, sustenance, education, entertainment, etc). In all cases, the invention evolves over time. Some rapidly, others less so. And in all cases, the speed of evolution is dependent on how fast the engineering problems get solved. Each iteration improves its capabilities and helps increase its reach.
Once a need is exposed and assessed to be universally required, then the horse has bolted, and the intense and highly competitive race to service that need is on. Many engines drive how solutions evolve but four are critical:
Growth in consumption
Learnings from consumption that feed improvement
Fuel from investors to start, improve, and scale.
Research into all aspects of the product, service, and idea
And so it is with large language models
Use or observe anything over a period of time, and eventually, patterns emerge. So 6 months after using 3 public LLMs for business use, I observed the following.
Once the awe and wonder wane, you start to view the LLMs output through the lens of ‘usefulness’.
You learn not to accept their output until you evaluate for accuracy.
LLMs generate output using their style. Your way of expressing ideas are different. So the output needs to be modified to reflect this.
Only 10% to 20% of the output had value, and the rest was discarded.
The discarded output plays a role in aligning your brain with the concept you are researching. It acts as a ‘starter’ or to center your thought process around a subject.
The research question and LLM output often lead you to probe further into a topic. The Chat interface enables you to begin that dialog.
When using the tool, it is easy to imagine you are conversing with an ‘intelligent’ thing. This is a real trap and to be avoided at all costs. Coherence can mask the illusion.
LLMs replaced 30% of searches that would have normally started with a search engine. However, the use of search engines has become more targeted and focused.
That said, LLMs will not replace search engines.
The personal productivity gains are real.
You need to ‘learn’ how to construct your questions.
Outside of personal productivity, there are many benefits and use cases where an organization can harness this technology. While the path to achieving that goal won't be easy, the prize would eventually outweigh the effort and cost.
What are the known problems to solve?
How to get the LLM to….
Understand the customer using a mix of the foundation model + language specific to the organization.
Generate a response using only the organization's content.
Provide accurate responses that address the question.
Eliminate or reduce responses that are considered erroneous, false or fabricated.
Language development was a milestone in human history, and its use has and continues to yield profound benefits.
As a fundamental tool for communication, it has and continues to play a pivotal role in shaping society, culture, and progress.
While language was not "invented," it is a complex system used for communication that has and continues to evolve through the natural interplay within communities and society.
Large Language Models represent one more step in that evolution. But make no mistake, it is a giant step forward.