A Business Guide to Non-Generative AI

Greetings from sunny California this week, where I'm recording my first course with LinkedIn Learning! Thrilled about the experience – my course will be up soon, so stay tuned!

While AI is all the rage (just look at all the new courses coming), I often have the feeling that AI is boiled down to just ChatGPT. My recent post on AI terminology went viral and some people mistakenly interpreted in a way that I put ChatGPT at the center of everything.

So let's clear this up once and for all: There's more than ChatGPT. I've written about this before. And today I want to shed more light on all the AI things you could do without even saying the Chat-word.

So let’s go!

The AI Landscape: More Than Just ChatGPT

ChatGPT may be the world's most popular AI application, but it's by no means the epicenter of AI. In fact, it's an interface to Generative AI, which is just one of the 5 main AI archetypes.

To recap, besides Generative AI, we have Supervised Machine Learning, Computer Vision, Natural Language Processing (NLP) and Audio / Speech that we can leverage for different business problems.

Now, you might be thinking, "Tobias, this all sounds great, but why should my business care about these other AI archetypes?" Here's the thing: while Generative AI is the shiny new toy, these other archetypes offer some serious benefits, such as:

• deterministic behavior

• interpretability

• full control

• ownership

In other words, by exploring the full spectrum of AI, you can find the right tools to solve your specific problems, optimize your processes, and drive more innovation. So feel free to use ChatGPT to get started, but don't let it blind your potential of the wider AI landscape.

There's a whole world of AI waiting to be discovered and today I’ll look deeper into some examples:

Supervised Machine Learning

At its core, supervised machine learning is about learning patterns from labeled tabular data. Think of it as a spreadsheet on steroids. You feed the machine learning model a bunch of data points (like customer information, sales figures, or sensor readings) along with the right answers (like whether a customer churned, how much revenue you generated, or whether a machine is about to fail). The model then learns the patterns and relationships in the data, allowing it to make predictions on new, unseen data.

Good-ol' supervised machine learning remains one of the most important AI applications for businesses, because there are a lot of great use cases for it:

1. Time Series Forecasting: Predict future sales, demand, or market trends based on historical data.

2. Recommender Systems: Suggest products, content, or services to users based on their preferences and behavior.

3. Classification Engines: Automatically categorize emails, documents, or images based on their content.

Thanks to advances like AutoML, Supervised Machine Learning is more accessible than ever. All you need is to have the right data and a helping hand along the way.

Computer Vision

At its core, Computer Vision works similarly to Supervised Machine Learning, but instead of tables, it's all about analyzing images. The thing that separates images from tables is their complexity. A single image can easily contain millions of data points when you consider pixels and colors. That’s where computer vision solutions often apply deep learning technology like neural neural networks. Sounds complicated? Well, it is.

But the good news is that you don’t need to be a big tech company to reap the benefits. In computer vision, there are many pre-trained models for generic tasks (like the ones below) which you can easily fine-tune on your own data. No-code platforms such as Landing AI or Azure Vision Studio make it easier than ever to get started.

Some classic examples of applied computer vision in business include:

1. Entity Recognition: Identify and locate specific objects, people, or text within images.

2. Value Extraction: Automatically extract relevant information from images, such as license plate numbers, product labels, or handwritten forms.

3. PII Removal: Detect and redact personally identifiable information (PII) from images, such as faces or sensitive text.

Natural Language Processing

NLP is all about teaching machines to understand, interpret, and generate human text. It all starts with breaking down text into its fundamental building blocks – words, phrases, and sentences. From there, NLP models use a variety of techniques to analyze the structure, meaning, and context of the text.

That's what allows a computer to understand, for example, that when you say "I prefer Barcelona over Lion," you're talking about two cities, and you've even spelled one of them wrong.

NLP techniques are what made modern Generative AI like ChatGPT only possible. Similar to computer vision, there are lots of pre-trained language models (like BERT) available for generic NLP tasks, that can be fine-tuned and adapted to your use case.

Some examples of what you can do with it "classic" NLP, without touching Generative AI:

1. Intent Recognition: Understand the underlying intent behind a user's message, whether it's a question, a request, or a command.

2. Sentiment Analysis: Determine the emotional tone or sentiment of a piece of text, whether it's positive, negative, or neutral.

3. Content Filtering: Automatically identify and filter out inappropriate, offensive, or irrelevant content from text.

Audio/Speech

Speech and audio processing allows machines to recognize, interpret, and generate human speech. It's similar to NLP, but applied to audio data, often turning audio into text to then leverage NLP techniques.

On the recognition side, Audio/Speech AI can transcribe spoken words into written text with remarkable accuracy. And on the synthesis side, Audio/Speech AI can generate human-like speech from written text.

For businesses, this comes in very handy for the following scenarios:

1. Call Transcription: Automatically transcribe customer service calls, meeting recordings, or podcasts into searchable, actionable text.

2. Live Narration: Generate real-time audio descriptions or translations for videos, presentations, or live events.

3. Noise Reduction: Automatically identify and remove background noise, echoes, or other audio distortions from recordings.

Generative AI

Even Generative AI goes beyond ChatGPT and pure text generation. In fact, many of the capabilities of the other archetypes can also be achieved with Generative AI because most modern Generative AI models are multimodal, meaning they can process different types of data - like text, images, or audio.

But be aware that the common denominator in all Generative AI-powered applications is the generation of original content.

That's why ANY GenAI process is inherently non-deterministic, meaning that given the same input, it can produce different outputs. Hallucinations are not a bug, but a feature.

And that’s the reason why many of the other AI archetypes are still in use and will continue to be - they're just much easier to control. And very often they offer more ownership because you can either train the model yourself or host the model completely on-premises, which is rarely possible with these super large GenAI models.

Examples of GenAI that are not ChatGPT include:

1. Video Generation: Produces engaging content, realistic animations, and visual effects.

2. Speech Synthesis: Creates human-like speech, capturing nuances and emotions.

3. Image Upscaling: Enhances image resolution by filling in missing pixels based on learned patterns.

Conclusion

You see, AI is a vast field that extends far beyond ChatGPT.

As we move forward, it's likely that these AI capabilities will become more intertwined, allowing for even smoother, more augmented workflows (like an AI-powered Teams chatbot that can make automated forecasts based on your BI data). In most cases, AI will work very much behind the scenes; in others, it will be more exposed and "visible" to end users.

Either way, business users will need to adapt and learn to ask new questions.

Questions like: How confident am I in this prediction? What are the potential biases and limitations of this model? How can we build robust, ethical and responsible AI systems?

These questions will define the responsible and effective use of AI in the years to come.

I'll talk more about this in a future newsletter.

See you next Friday!

Tobias