Why localized, refined models will likely dominate AI workflows in the near future

Many people associate generative AI with ChatGPT. Without a doubt, the platform was what got people’s attention worldwide. Many individuals haven’t gone above the cozy chat-based interface that it provides, even in the present day. This is unfortunate since it indicates that many people are still unaware of the incredible prospects and possible risks that new technologies bring.

Where are we with LLM right now?

The landscape of large language models (LLMs) is currently showing an intriguing split as their developers choose between two different directions. The creation of the biggest, smartest, and most striking models is the first road, and it is the one that most people are familiar with.

These models strive to be the best LLM possible and are based on GPT-3.5 and GPT-4, the models that drive ChatGPT. There are numerous examples of these, including Mistral, Metta Llama-3, and Google Gemini. Even though many still consider GPT-4 to be the “king” of the group, the rivalry is becoming more fierce every day.

These models all share the large size and enormous computational and electrical power requirements in order to operate. They can only be utilized via hosted interfaces or APIs as a result.

A fresh route to improved models

However, a second route is beginning to take shape and is quite intriguing for good cause. A new breed of models may balance strength, speed, and size rather than aiming for the most powerful AI yet. In other words, these models condense the ideas from the “big” models into much smaller containers. A few of them are scaled-down versions of the larger models, such Llama-3-8B and Gemma (Google Gemini’s baby sister). Others, like Microsoft’s Phi-3, were made to be small from the beginning.

The fact that these can all be operated on affordable current technology makes them all fascinating. I’ve used my three-year-old MacBook M1 Pro for all of them and a ton more. My data doesn’t need to leave my laptop. No calls to large servers via APIs. Everything is regional.

These little models are unique due to one important factor in addition to all of those advantages. Most of the smaller models are open, in contrast to the large models, which are all closed. Their “thinking” and behavior are determined by billions of characteristics called weights, which are made available to the general public. This implies that common mortals like you and I are able to adjust them.

Although it is feasible, this fine-tuning procedure is not for the timid. While training a typical small model still takes technical know-how and access to computer capacity, it may be accomplished using rented Cloud GPUs for $10–$50 and a little perseverance. Additionally, you can use the optimized model that was produced anywhere, including on a laptop.

Now, we may use the data we have or some desired behavior, and build a new model that precisely reproduces that behavior or makes sense of the data we have. This implies that the model is capable of learning everything there is to know about your business workflow, including your products, lead generation criteria, segmentation and scoring system, and lead quality. The best feature is that you can operate it all on a little PC entirely within your own network.

Today’s LLMs have a great deal of transformative power when they are integrated into a workflow and utilized to accomplish a single, specific goal. Examples of such systems are GitHub CoPilot and Adobe’s Generative Fill. We can design and create small models that suit a workflow step and operate affordably and securely within our own infrastructure by utilizing the power and value of fine-tuned models.

Including tiny models in your growth marketing approach

Account score, or giving a new incoming lead an assessment grouping or projected value, is a routine job in many marketing workflows. This makes it possible to gauge the present pipeline’s health as well as set priorities. Usually, scoring is straightforward and depends on the size of the organization and sometimes the prospective estimate of a salesman. But we can achieve far better results with a customized model.

To train the new model, we must first create a dataset. With the addition of company descriptions that we have directly collected from their websites, we can use the real sales data that is currently in our sales database. Working locally with a tiny model is essential due to the sensitivity of this data, rather than sharing it with an external model such as ChatGPT.

We can train a model to generate an automated score based entirely on our internal performance data, given the description of a company as found on its website. We may incorporate this model into our workflow to receive a precise and prompt evaluation of the lead’s value. That is not something that can be obtained with the massive public models.

Consider what is essential to your marketing process for a moment, then list the tiny yet impactful acts that can be transformed with a little bit of insight. As we mentioned earlier, is it your lead scoring? Development of a proposal? When will releases happen? After an hour in front of a whiteboard, I have little doubt that many areas may be identified where applying intensely focused intelligence can improve our productivity and competitiveness. For these targeted tasks, we can apply the new breed of fine-tuned models.

Think about the essential steps in your marketing process for a moment, and then list the little things that can be transformed with a little bit of insight. Is it the lead scoring that we talked about earlier? development of a proposal? When will the releases happen? I’m confident that spending an hour in front of a whiteboard will reveal numerous areas where employing laser-like focussed intellect might improve our productivity and competitiveness. For these targeted tasks, the new generation of fine-tuned models can be applied.

Declaring anything as the “future” of AI would be ludicrous given how quickly AI is developing. If we limit the discussion to the near future, however, the most transformative opportunities will probably be those made possible by locally customized and fine-tuned models. These will be the unseen elements of the most prosperous business processes and goods.