Start with the problem, not the technology. When wielding a (generative AI) hammer, everything can begin looking like a nail. Instead of asking how to incorporate generative AI into your company, ask what you need to accomplish.

Most importantly, with any technology implementation, you must understand the question you are trying to answer before building the solution.

Where LLMs can help:

Automating your processes

Text generation models
- Draft documents
- Write computer code
- Answer questions about a knowledge base
- Analyze text
- Give software a natural language interface
- Tutor in a range of subjects
- Translate languages
Function calling
- Understand a request and convert to actual call to retrieve data (example: what is my account balance question - calls GetBalance function and appends result to a response)
Embeddings - convert text to numerical representation (vector)
- Search (where results are ranked by relevance to a query string)
- Clustering (where text strings are grouped by similarity)
- Recommendations (where items with related text strings are recommended)
- Anomaly detection (where outliers with little relatedness are identified)
- Diversity measurement (where similarity distributions are analyzed)
- Classification (where text strings are classified by their most similar label)
Image generation from text
Text to speech/Speech to text
Vision to text (describing what an image contains)

What to keep in mind

The risk of inaccuracy. Models never return 100% decisions, responses are always a prediction
Differentiation is another important concern. By its nature, ML models can only provide responses based on existing information - it can't create anything new
Models generally do not provide the age of the data used to make the prediction
The information source is generally not surfaced, making it harder to assess the trustworthiness
There is not verification of whether the natural-language query was understood correctly
Models do not allow verification of names, which can be ambiguous. For the query, "What is the population of Frankfurt?" it gives the response for the larger Frankfurt am Main, not the smaller Frankfurt an der Oder, and no indication of that specification anywhere in the response

Garbage in garbage out

To perform functions outside the base model your data must be in a good shape:

Correct
Representative
Repeatable

The general rule is "garbage in, garbage out"

"ChatGPT" gotchas

Lack of a common sense filter - it just spits out the words (you have probably seen some absurd examples showcased online)
The base models do not belong to you. Training of this model requires vast amount of data and compute power (its predicted that base model training is ~$5M)
You are putting a lot of trust into the model and it's creators, there is no way to understand how the output may have been manipulated or biases imposed upstream
Prompts that are used today may not work the same tomorrow and complex AI applications might need to be rebuilt if prompts are not working same way anymore

Like any new technology, knowing the benefits and limitations is critical to understanding if it is a viable solution to address your business challenge. If you want to talk in greater depth, reach out to us.