Computer scientist calls for new layers in the tech stack to make generative AI accurate

Temasek's head of ML is optimistic about 'overhyped' technology

In order to make generative AI accurate, new layers must be inserted into its stack, according to the head of AI at state-owned Singapore investment company Temasek.

Speaking at a developer conference in Singapore on Tuesday, Michael Zeller described the software stack within generative AI as needing a new set of software tools and even a data prompt and orchestration layer.

Even more important than that data prompt and orchestration layer, said Zeller, is a data and oversight layer where the model is assessed, verified and checked against known true sources.

Right now, scientists don't necessarily have a good way to test, he said.

"Given the generative nature and output of the models, we need a whole new selection of tools on monitoring, performance, testing bias, and observability, because these large models work, we just don't know exactly why."

The computer scientist and physicist described validating, testing and inspecting generative AI models as a "big opportunity." He describes himself as an optimist when it comes to both traditional and generative AI, yet said that the latter was overhyped.

As of 2022, the Gartner Hype Cycle places generative AI as crossing over into the "peak of inflated expectations," meaning a drop is soon to be expected into "trough of disillusionment" before it journeys into a "slope of enlightenment" and eventually "plateau of productivity."

But according to Zeller, it's fast-tracking through the stages. Although generative AI is predicted to reach a plateau of productivity in the next two to five years, he expects that to happen much sooner, given the speed of innovation seen today.

"It's just about all the applications showing up in six months," he said, calling the cycle "compressed."

Zeller predicts a flash crash in the next six months where generative AI company valuations are cut in half. After three years, he predicts a new paradigm for video analysis and analyzing time dependent sequences – a type of videoGPT.

But for now, he rates "traditional AI," a silly misnomer to describe the AI of only last year, as more useful than generative AI. Not only is generative AI much more expensive to build, its output is unpredictable. He said traditional and generative would continue to co-exist together.

"There are many algorithms out there that are valid and have a very particular use case," he said. "Not everything will be generative AI because it's not relevant, it's not applicable to every use case."

But regardless, it seems the technology is already ingrained in the human psyche. During a panel that included Zeller and Chang Sau Sheong, deputy chief executive of Singapore's digital services statutory board, GovTech, the latter described his sudden pull to use generative AI to produce code as "scary."

Chang said he used GitHub Copilot for a few months to generate code he simply had to copy and paste. Soon he realized the produced code was problematic and undesired.

"But it keeps nagging you and keeps trying to give you suggestions. Then after a while it overtakes your mind. And then you become a Copilot zombie," said Chang.

After two months, Chang removed the tool only to find himself expecting the prompts to appear.

"That tells me something, because the way that we use tools actually changes who we are and how we think," said the GovTech exec.

Zeller acknowledged Chang's experience and compared the way generative AI will change people's work to how smartphones changed travel.

"Today you just show up and you say, 'OK, where do I need to go?' And the phone tells you," said Zeller. "It's different skill sets." ®

More about


Send us news

Other stories you might like