The LLM Wave and its Impact on ML

The advent of Large Language Models (LLMs), such as ChatGPT by OpenAI and PaLM by Google, has made a serious impact on Machine Learning (ML). These tools have completely altered the way in which ML engineers code and problem-solve, so developers have had to upskill and retrain at lightning speed. 

It’s important to understand how the role of an ML engineer has changed since LLMs exploded onto the scene, and the lasting impact they will have on the future of code development. Especially when comparing traditional ML to LLMs, it’s essential that engineers fully understand how these models work and how to properly utilise them. 

Out with the old, in with the new

Before the exponential growth of LLMs, what were the tools of the ML trade? Engineers had to consider different models, from linear progression to more complex deep learning models, adjusting them to specific use cases. Unfortunately, this can be a long and laborious task that requires multiple iterations during its development. As a result, developers would often spend too long jumping between models and produce poor, rushed code. 

The emphasis has now changed from manually creating and fine-tuning ML models, to comprehending how to effectively use LLMs' capabilities. This presents fresh difficulties while also opening up new avenues for ML applications. Engineers can concentrate more on prompt engineering for their specific work, and rely on LLMs to handle large tasks, as these require significantly less manual intervention and can now be solved by just one API request, as LLMs have already learned relevant representations from their enormous data pools. This allows engineers to provide task descriptions in LLMs, rather than having to search through numerous models to get the solution. Yet, there may be drawbacks to this kind of engineering. 

Adapting to new tools

Prompt engineering, in its most basic form, consists of creating efficient prompts that can effectively direct LLMs to carry out desired actions, which may be applied to a variety of use cases. To maximise the model's output and improve relevance and accuracy, a prompt must be carefully worded, formatted, and organised. This can improve the intended result's relevance and precision, which cuts down on the amount of time engineers then have to spend changing it to meet their requirements. 

Prompt engineering is not an exact craft, but allows for some creative licence within certain patterns. Any minor modification to the given prompts can cause a notable difference in the responses, which can in turn impact the model's overall usefulness. Therefore, a well-designed, explicit prompt can greatly improve the model's effectiveness and performance by reducing ambiguity and guiding its knowledge in the right direction. 

It can be challenging to nail down the exact responses that different prompting styles will elicit, but finding the ideal techniques is crucial for ML engineers. Prompts can be broadly categorised into  Output Input Prompting, Chain of Thought (CoT) Prompting, Self Consistency with CoT, and Tree of Thoughts. To break these down, Output Input Prompting is simply when a response is outputted after a prompt is inputted. CoT Prompting is when a model is asked to explain why it has given the answer it reached. Self Consistency with CoT Prompting is when an engineer uses the CoT method but reasons to find the most consistent answer. Finally, Tree of Thoughts Prompting involves providing the model with numerous CoT solutions, which it then uses to find the most effective option. 

Whichever method is chosen, it’s crucial that engineers have complete control over their prompts so that the model behaves how they want it to. This ability has the potential to significantly improve the output of LLMs and thus the ability of prompt engineers will only grow more important as the creation and use of LLMs progress. This gives ML engineers a leading role in the development of artificial intelligence in the future.

Room for the traditional

However, what does this mean for engineers right now? Is it the end of traditional ML? My guess would be, not quite. The growing application of LLMs through rapid engineering doesn't indicate that conventional ML methods are no longer relevant. They continue to perform well in situations when the dataset is small, the use case is straightforward, interpretability is important, or the use case is too specialised to be resolved by LLMs. While traditional ML techniques can help source solutions for specific scenarios, LLM solutions have a competitive advantage with large, complicated data sets. 

Overall, engineers would do well to consider the combination of ML and LLM models as a potent weapon in their toolbox. 

What the future holds

With the advent of LLMs, ML engineering has undergone a significant evolution. As a primary method for finding coding shortcuts instead of manual stack sifting, which can consume a developer's day, mastering the art of prompt engineering can be extremely advantageous to the engineer as well as their business, through resource preservation and development time savings. 

DoiT International's team of +150 customer reliability engineers (CREs) are investing a lot of time and effort into learning the skills required to successfully traverse this new environment, and we’re already beginning to see the consequences of this transition. Applications for our clients come from a variety of sectors, including cybersecurity, fintech, healthcare and life sciences (HCLS), and more, so we’ll continue monitoring the development of this sector closely.