What matters most in an AI start-up?

AI start-ups are all the hype right now.

To be honest they were a hype even 10 years ago - but packaging was a bit different. Back then it was all about being data centric and running ML models. From there has been many hype waves around neural networks, transformer architecture, large language models, and now finally Agentic Ai.

And each time the concluding phrase remained the same - "We will revolutionise the world - and no company can ever touch us because we have the best talent, the best data and obviously we are the first and only ones to see this business opportunity."

From start-ups to enterprise company - everyone parroted the same. And of-course the reality was different. In retrospect most data science an ML projects never gave the results the people hoped for.
My opinion - it was misalignment with reality. 

According to me having a right business model is perhaps the most important to succeed with AI!

From past 10 years at-least there has been a perpetuating myth - that AI can solve everything - even when it wasn't AI and all of us were running regression models and fitting decision trees. 

But AI couldn't and doesn't yet solve everything. In fact what AI can reliably solve is a much smaller subset than people imagine.

The reason behind the insidious myth that AI can solve everything starts with the infamous "Universal approximation theorem" in neural networks. While proved, it assumes:

1. We have sufficient data

2. We have right quality of data

3. We can train it at reasonable cost.

But none of the above is usually true. Hence, AI training often becomes data collection, annotation and monitoring exercise.

An obvious but commonly forgotten element of the working with AI models is that they are statistical in nature and hence do not reliably reproduce exact solution over and over again. More complex the model, the higher the model's variance in output. Now this is also a strength because it allows neural networks to come up with many creative solutions. And this creativity is what people impresses people often to the point of delusion.

Because the heart of success in modern engineering lies in "reliable engineering" - the capacity to do something repeatedly with trust and reproducibility. And when something fails (which it eventually do) - then know how to debug it.

And this is why most AI works pre and post LLM get stuck in MVP-demo stage. Because at this stage, demonstrating that something is doable is most important. But after this stage, one needs to productionize it - which means it has to be reliable, debuggable, robust and unit economics profitable. This is where most MVPs find themselves in a graveyard.

Because AI solutions that are reliable, debuggable, robust and unit economics profitable make up for a very small subset for any business. First 3 are technical, last is business. But the last one dwarfs the first three even if no one accepts it. 

Your AI solution has to make money. Period.

But will they? Let's try to understand from simple illustrative comparison between two companies - Netflix and Youtube.

Both employ recommendation systems and both of their recommendation systems are often cited case studies in intro to AI/ML course.

But does a content recommendation system have the same impact on both. I think not.

For Netflix, content recommendation is a nice to have. They rely on licensing and in-house production by which it means the amount of content they generate is very small compared to Youtube. The content recommendation cannot recommend anything good if there is no good content - which depends on their licensing deals and supply of in-house productions. And for the size of content library netflix deals with even a moderate recommendation system built on heurisitcs or simple machine learning will suffice.

But Youtube - recommendation is where the business is. The content library is virtually infinite with no direct control over production of content - so the only way to keep user engaged is via addictive content recommendation. The bar of success is much high and size and type of content makes it very difficult to build a great content recommendation with simple ideas. Not to mention the engineering layer that goes into repeatedly running experiments, retraining models and evaluating results. All of this is super costly so the ROI from recommendation has to be super high.

Netflix will not gain massively in investing their content recommendation (I am sure netflix has many good cases for AI/ML) - but content recommendation from business perspective is not one of them. But for Youtube it is existential for business. 

This is why your business model matters a lot in AI product development. Not everything will make sense for you. And just because it worked for other companies doesn't mean they are a good fit for you.   

So one has to choose their business model and then decide their AI product and workflow. Investors always love the AI hype but don't let it distract you from being honest that how useful and impactful it is for you.