Artificial intelligence is modern-day alchemy, according to Qirong Ho, the CTO and co-founder of Petuum Inc., an AI and machine learning startup.
“Alchemy is about chasing the most exciting applications of the science of the time — and turning that into gold,” he said at the recent Forrester’s New Tech and Innovation 2018 conference. “We have many turning lead into gold examples in AI.”
The examples produce results but are also a distraction from a larger, more common IT problem that ultimately creates headaches down the road: Today’s AI is bespoke, customized tech that Ho has referred to as “artisanal” and “hand crafted.” That’s why he’s called for the industrialization of AI, where building an end-to-end AI application — from data collection to user interface — is standardized.
A row of new houses
Ho said to think of building an AI application like you’re building a row of 10 new houses. “You call up the contractor, and the contractor comes with the foreman and all of the workers,” said Ho during a fireside chat with Forrester’s J.P. Gownder. That’s when you realize that instead of everyone using the same set of bricks and nails, planks and shingles, the workers arrive with different sized bricks and nails, planks and shingles.
“They say trust us, we can build this for you,” said Ho. And in most cases, they succeed. The workers are able to put together customized homes that work just fine — for a while. But when something in the house needs to be changed or fixed — a new bathroom installed or a pipe replaced — that’s when the headache starts.
Ho said customized AI applications that will have to change as company needs or data formats change are going to produce a very similar outcome. Without standards, every AI application is built using different tools and systems, and that makes it difficult to maintain or even repair an application — not to mention find someone capable of doing that kind of work.
A call for the industrialization of AI
The industrialization of AI isn’t just about standardizing things like TensorFlow or PyTorch. These are sometimes thought of as standalone systems but, as deep learning libraries, they’re really just a couple of the bricks often used to build an AI application.
Instead, Ho said the industrialization of AI includes standardizing everything upstream and downstream from tools like these, from data collection to application deployment.
It’s a process of moving from alchemy to chemical engineering, where technicians, engineers and operators are trained in specialized disciplines and can be called upon to do repair work when needed. “Just in the same way if I want to repair my house, I don’t need to call the guy who built it to repair it,” Ho said. “I can call up some other maintenance technician to build it or repair it.”
Ho also warned that renters should be cautious as well. Cloud vendors may claim to provide standardization for AI, but CIOs should question “how far do they go in standardizing every last step in the AI industrial engineering process,” he said.
One example: X-Rays, thermal scans and smartphone images should all be processed differently. “I mean, you treat salt water differently than sewage water when you’re trying to produce water for human consumption,” Ho said.