AI startups are on the rise. Demand for artificial intelligence solutions is booming, and there is an eagerness to invest in startups focusing on AI technology. But there are a few pitfalls that AI startups should watch out for.
No Flemish accelerator has more AI startups in its portfolio than imec.istart. We have front row seats for the big breakthrough of AI systems and products. There are huge opportunities ahead for AI startups, there is still a lot of untapped potential. But the new wave of AI startups also encounters challenges that are uncommon for generic software startups. We’ve summarized some of these challenges, not to ruin the party, but to make it easier to cope with them. Because ‘forewarned is forearmed’!
‘Building software is easy, building hardware is hard’, they say. Well, AI is somewhere in between. And so is the cost. Startup founders tend to look at AI as Software-as-a-Service. But AI is different. You need a lot of data to develop your algorithms and you need to constantly retrain your data. You must upload thousands of pictures or, even worse, videos on Google Cloud or Amazon Cloud. Those huge amounts of data raise the cost of your cloud solution significantly. So, try to get your calculations right and take the hidden costs into account before they suddenly slap you in the face.
Cloud solutions make it very easy to sell your first AI products. But when AI startups grow and scale, they often find out that their cloud architecture limps behind. The popular misconception about AI says that recognizing, let’s say, 10 cats is the same as recognizing 10 million cats. Once the algorithm is able to recognize the patterns, size doesn’t matter anymore. Unfortunately, that is not how it works. Be aware that reliability in a controlled environment is not the same as reliability ‘in the wild’, where you can’t control the quality of the data and other parameters.
Many people don’t know AI applications often need manual intervention to work properly. AI models usually need a lot of annotated data, where people manually indicate what is relevant and define the ground truth. Popular AI-driven services such as Uber, Pinterest, Xpenditure and Amazon are consistently using humans to help the AI system in difficult cases.
Human intervention comes with a cost, and some may say it also takes the magic out of AI. But especially for early-stage startups, humans can be used to their advantage: it can be a way to bootstrap their model or even fake their product. This allows to think ahead, and eventually proceed with minimal resources.
In most cases, not enough attention is spent on keeping AI applications healthy over time and across several client deployments. Once the data scientist team has a reasonable amount of useful data, an AI model is built using ML techniques, and once accurate enough, the model is brought into production. And that’s it. But is the model still as accurate after 3 months? Did the feed of real data change over time? Does the feedback loop of a customer improve the AI model?
Luckily, several tools are arising that help AI startups to keep their model healthy over time. Our portfolio company Raymon.ai, for example, has developed a platform to monitor performance and data quality of machine learning systems.
In the AI world, technical differentiation is harder to achieve. New model architectures are being developed mostly in open, academic settings. Reference implementations (pre-trained models) are available from open-source libraries, and model parameters can be optimized automatically. Data is the core of an AI system, but it’s often owned by customers in the public domain, or it becomes a commodity over time. Some customers will not allow you to train a model for another client using their own data. As a result, you need to create a custom dataset and AI model for every client, which limits your scale.
But there’s also good news. Privacy preserving techniques are being developed which enable you to train an AI model with data from different clients without compromising sensitive raw data. Here you can read more on this topic.
Some startups claim that AI can solve every problem in the world. A few years back we saw the same thing happening with blockchain, and now we see the exact same thing with ChatGPT: it would fix everything. When I hear too many empty promises and too many buzzwords in a pitch, the alarm in my head starts ringing. I am a big believer of AI. It’s still a hype, but it has moved beyond the gimmick stage. AI offers relevant solutions for real problems.
But as a startup founder you need to be pragmatic. Start with the problem or the business case you want to solve. Technology is a means to an end. Sometimes other technologies are more efficient and offer a more affordable solution.
At imec.istart, we help our AI startups to cope with all the challenges they encounter. We keep scalability in mind from day one. We also bring startups together to help each other out and learn from each other’s mistakes. This way, we hope to contribute to AI solutions that make our world a better place!
Are you an ambitious tech startup that has a working proof of concept? Are you ready to take your startup to the next level? ➡ Apply for our imec.istart acceleration program by February 1
Author: Robby Wauters, Venture Acceleration Manager at imec.istart