How customized do AI models need to be for businesses?

As a business, it is hard to not be tempted by all the promises of AI.  If you believe all the hype, it can transform every part of the business, find and convert new customers, design new products, manage your factory or your software, and generally do anything short of bringing you a coffee in the morning (and you assume they are working on that right now).

This is a wonderful fantasy, and very tempting to believe.  But how true is it?  With all the hype, you might be tempted to jump on the bandwagon completely and believe it all, or go the other way and reject everything you hear as a gross exaggeration.  The truth must be somewhere in the middle, but how do we find it?  And once we do, how can we turn that understanding into something useful for our business?  Let’s dive in and discuss how AI is used in the business, and in particular just how customized an AI model needs to be for a business to see its value.  We will dispel some of the hype around AI while showing the areas it can truly transform a business today.  We will also look at the role of good data, and the challenge to find/scrub/verify it for AI models, an industry that is quickly growing in the Web3 space with platforms like Synesis One working to leverage blockchain’s attributes to verify AI and reward users.

AI is said to solve many different problems.  To the uninitiated, some of these problems might be indistinguishable from magic, which is very exciting but equally unhelpful if you are a business considering AI investment.  Thankfully, AI is actually not as complicated as you might believe, because it really only solves three key problems.  First, AI can classify things.  Think of a quality control algorithm telling you whether a part coming off the assembly line has a defect or not.  This can be expanded to all sorts of anomaly detection, sorting capabilities, and other non-conventional uses that require an algorithm to analyze something (eg. photos, spreadsheet data), then determine which bucket it belongs in.  Next, AI can predict.  Predictive maintenance can use many different sensors in a machine to predict when it will fail or when it needs maintenance.  It can provide very accurate forecasts if the data is available.  It can predict where a robot should next move in order to accomplish a specific goal.  Finally, AI can optimize.  It can solve complicated problems with many different constraints in a way that is simply too difficult for a human to accomplish.  We use it for GPS, organizing our factories, and many other applications that need an ideal solution given one goal and many different constraints.

The bottom line is, AI really works.  It can absolutely solve the problems outlined above, and countless more.  Even better, AI is improving every single day.  New innovations are developed, computer processing is improving, and more general use cases are discovered.  On one hand, AI is so much more incredible than we give it credit for, and it is transforming the world in ways we can’t even fully comprehend.

That said, there is a cost to these results:  data.  AI is completely useless without the right data to train an AI model.  With the onset of Large Language Models (LLMs) like ChatGPT, the incorrect perceptions of AI have become even more skewed.  We can be tempted to believe that AI is all powerful and can answer any questions we ask it, and because ChatGPT is available to all, we can see it with our own eyes.  However, there are two big weak points that may not be obvious.  First, ChatGPT is amazing at some things like writing a paper on Climate Change, or suggesting ten locations to visit in Budapest; but it is not good at many other things, like giving a certain answer on math problems or giving fully trustworthy information.  For a business, this level of uncertainty is not acceptable.  Second, ChatGPT was made using trillions of data points in order to answer general questions.  If you want to train a precise AI model, it will take a lot less than this, but you must know exactly what data to use, and it must be validated to be effective.  In other words, you have to know what problem you are solving, what AI model you will use, and what verified data that will require.  The issue is, building up this data is difficult, and depending on a number of factors, the amount of data can sometimes be significant.  The AI model above that identified anomalies on an assembly line?  It would likely require many, many different samples of items on the assembly line, with different lighting angles, orientations, and other variations so that the model can accommodate those variations.  Within that, it will need an adequate number of samples showing good quality items, and samples showing items with anomalies.  Only then can the model learn what an anomaly looks like.  And for the majority of AI models, all this data needs to be labeled (eg. “no anomaly” vs. “anomaly”).  This is the biggest burden for businesses as they often do not have the expertise or insights needed to accurately process all this information, and the sheer labor involved can be significant.  Blockchain has shown a unique ability to help solve this problem, with platforms like Synesis mentioned above leveraging the unique elements of blockchain specifically to validate AI training data.  By using decentralization, immutability, and rewards for work, the platform is able to build up massive data sets by rewarding a massive population of contributors across the globe.  Because of its nature, it doesn’t take much to train someone how to validate a particular data set.  The learning is fast, and with enough people, the data set can be validated, labeled (answers attached to each image or data point), and made available to the company that needed it.  It’s a fascinating business model, and one that will likely see an entire industry grow out of in the next year.

A lot of small to medium-sized companies are working to build proprietary AI models powered by their own data. However, training a model isn't as simple as plugging in raw data. The main challenge they face is their lack of in-house data scientists and resources to handle the… pic.twitter.com/97gQ4JQ8Qb

— Synesis One (@synesis_one) November 17, 2024

Now that you understand the hype of AI and its reality, you can get a better idea of what problems it can solve for your business.  The key is to identify these classification, prediction, and optimization problems, then start to build up the data needed to train the models.  Working with a data validation platform like Synesis will be crucial for many businesses who can’t do it themselves but need a cost effective solution.  Once done, though, your business can be supercharged with AI in a way that was science fiction just a few short years ago.

Featured image credit: Google DeepMind/Unsplash