Artificial intelligence born in 2015. Deployed computer vision to the edge in 2019.
It’s obvious that AI offers the potential to radically reshape the business landscape, helping organizations of all sizes and industries find more efficiencies and make smarter data-driven decisions.
We also know that there have been problems getting AI implemented, but what you might not know is that these bottlenecks are being unblocked. Now, innovators, developers and CTOs can all plan and easily execute AI strategies with confidence, creating big opportunities.
There’s a huge digital transformation just starting to take place. IDC projects that in 2023, worldwide spending on AI systems will reach $97.9 billion—more than twice as much as the $37.5 billion spent in 2019. That’s almost triple in four years. It’s not going to stop growing for the foreseeable future either, because only a small percentage of companies are able to take advantage of the technology - because we had some bottlenecks to unblock.
Building, testing, and deploying an AI model that’s production-ready can be highly time-consuming. According to a 2019 survey, 64 percent of companies take between 7 and 18 months to go from brainstorming AI models to production. This includes companies who are new to the field of AI, and therefore likely to take longer than AI-mature organizations—but it nevertheless illustrates how AI is far from being a “quick fix.”
AI models can be highly complex and opaque, making them hard or impossible to understand—even to the people who build them for a living. Many people have written about the “reproducibility crisis” in AI and machine learning, where researchers struggle to explain why a new method performs better than the alternatives. This means that training an AI model to the desired level of accuracy is not only difficult to achieve in itself, but also difficult to achieve consistently.
Most AI models are monolithic: they were designed and built for only a single purpose. If you want to expand the model’s scope or make a change, building and deploying the new model often requires another full cycle of development—and as we’ve discussed above, that can take months.
While AI models can be monolithic in and of themselves, the actual deployment of the model can be equally inflexible. For example, how can you access an AI model running in the cloud? What if you want to install the model on an edge AI devices located close to hundreds of cameras streaming video 24/7, instead of on a remote server?
Time consuming, inconsistent, monolithic and inflexible, AI deployments have traditionally suffered from one or more of these bottlenecks, but this is about to change. Now, you can create complete, functioning AI with one system.
With AI platforms, we can deploy AI models to the cloud or to the edge. The model detects a single object like a mask or a suite of actions like everything that goes on in an operating room.
An AI platform allows AI models to be rapidly trained and deployed. The performance of the AI is easily monitored so that it can be improved over time. And finally, any functions that need to be added or changed takes days on an AI platform, not months or years.
Full disclosure, Chooch AI is an AI Platform.
With the emergence of AI platforms, these impediments are becoming a thing of the past, and adoption of AI is becoming more and more widespread.
That’s why we believe that the number of organizations implementing AI is about to hockey stick.
Creating a production-ready AI model is “simply” a matter of collecting data, annotation, AI training, then deploying the resulting model to the edge or the cloud. Building such a process in-house is now pointless, although some companies are still trying to build CMS’s 17 years after Wordpress was created. Today, an AI platform can speed the process of training and deployment of production ready AI processes at an organization with a clear AI strategy.
The accuracy of and data generated by AI should be an open book. In computer vision, when we can understand that an AI model is mistakenly detecting, for example, a car approaching a remote site which is actually a sunset, we can fix that. We simply train the model to detect a sunset. That’s the power of transparency. If the accuracy is measurable, then an organization can export the data to, say, a dashboard and rely on the results to make process improvements, which is the whole point of adopting AI.
AI solutions that focus on solving a single problem are a fast track to a bottleneck. Conversely, an AI platform that focuses on solving many AI problems with a process can be an unblocker, so long as the processes are fast and transparent. Training an AI to recognize that a person is wearing a hardhat should be exactly the same as training it to recognize a dripping pipe or a torn piece of paper currency. It’s simply a matter of training the AI to recognize something accurately and quickly.
If a company wants AI to recognize whether people are wearing hardhats, surely they will soon want to recognize vests, gloves, goggles, boots, forklifts, palettes, steel beams, and so on. Once a company has an AI strategy and a successful deployment and they see the results in lower risks or higher accuracy, they will want more. It should not take weeks or months to level up.
AI platforms give developers and program managers the tools they need to build usable AI that actually produces results.
Want to create a service to check that store shelves are stocked just from security cameras? Train the AI to detect empty spaces in shelves and stream security to edge devices with the AI model running. It would be equally easy to keep track and map all the empty parking spaces in your parking lots and streets.
No more excuses. The bottlenecks are unblocked by AI platforms and you’ve got the tools to take part in the next massive digital transformation: AI.
Create your free account to unlock your custom reading experience.