LLMs are not a panacea.

The phrase Artificial Intelligence (AI) covers a wide range of approaches and types of model including machine learning, natural language processing (NLP), computer vision and of course Large Language Models (LLM) which are now almost synonymous with AI. Large Language Models like OpenAI’s GPT-5.5 or Anthropic’s Sonnet 4.6 are trained on billions and trillions of data points to make them effective at predicting the right tokens (a word or phrase) which makes them great for general use, wide tasks and for things that require a large knowledge base for example writing, coding, sorting and analysing data and understanding a new topic but due to this flexibility you have a higher surface area for things to fail and an unsafe and unreliable experience to occur for your users. Boo! 

Trust is critical for commercially successful AI products and systems so safety and reliability should be a priority. LLMs can be used to do nearly anything but not reliably. If reliability is not at the core of your product or system then neither is safety or trust. The trouble with a model that can do lots of things is that it can also get lots of things wrong. An LLM builds its answer one token at a time, predicting what's likely to come next, so there's always a chance it confidently strings together something that sounds perfectly reasonable and is completely made up. It doesn't know it's wrong. It just keeps going. Of course there are strategies to mitigate this but they add a lot of overhead and you have to be able to justify this overhead.

I hope you don’t want to make the users of your products and systems eat a deadly mushroom. This is why choosing the right type of AI model is very important and why narrow models are often better for a lot of applications. A narrow model is built to do one thing very well as it is trained on only relevant data rather than a lot of cross-domain datasets, for example a spam classifier for your emails. A narrow model would be far more effective and efficient at this task than an LLM, firstly because it’s been trained for this one task and secondly these narrow models are very small and lastly its output will be in the desired format that will make it much easier to build with. 

The critical aspect is being able to understand what type of problem you’re designing for as there are several different categories of problem which need different types of models and techniques. Below I have listed some of the common problem types with examples that would be better solved by narrow models and techniques rather than LLMs.

1. A classification problem where your system needs to predict a discrete label from a fixed set of possibilities. In most cases there are more performant options other than an LLM for these tasks.

For example:

• Email spam filtering
• Scanning for tumours
• Identifying fraud in transactions
  
  

2. A regression problem where your system needs to predict a continuous numeric value.

For example:

• House price estimation
• Predicting energy consumption
• Estimating delivery time

3. A recommender problem where your system needs to predict preferences or relevance.

For example:

• Show relevant content (Netflix, Spotify etc)
• Show more relevant products
• Show relevant people on a dating product

These are some of the common problem types you will encounter when developing products and systems that augment your users and all of them are better solved with narrow AI models and techniques than they are with an LLM. There is a place for LLMs in products and systems but it is unlikely you need the largest frontier models that have hundreds of billions of parameters. Try a narrow model first.

See you soon,

Harry :)