He Has Watched Every AI Winter Since 1984 · Alok Aggarwal, Scry AI

What we talked about.

Sunny Ray talks with Alok Aggarwal, founder of Scry AI, who has lived through every major AI cycle since joining IBM Research in 1984. Aggarwal traces his path from theoretical computer science and hierarchical memory research at IBM, to founding IBM's India Research Lab inside IIT Delhi, to co-founding the KPO pioneer Evalueserve, and finally to Scry AI, which applies predictive analytics and agentic systems to industries like manufacturing, lending, and insurance. He explains the history of prior AI winters in the 1970s and 1980s, why cheap powerful semiconductors make this cycle different, and why he suspects the current hype could still crack within the next year. Aggarwal argues most enterprises and Silicon Valley startups fail because they treat large language models as all knowing rather than pairing them with subject matter expertise and tribal knowledge. He also discusses how Scry finds clients through warm introductions rather than cold outreach, and closes with predictions about knowledge work, aging populations, and AI's growing role in the labor force by 2050.

A four decade AI insider on winters, hype cycles, and why LLMs alone can't run real enterprises.

The questions, and the answers.

What is Scry AI and what problem are you solving?

Scry means crystal ball gazing, like Nostradamus, and we do predictive analytics such as whether an excavator needs maintenance or whether a person or small business will repay a loan. We build products around predictive maintenance and knowledge management. The name started as Scry Analytics and later became Scry AI, with a pun on AI since that's the technology we use.

You joined IBM Research in 1984 during the second AI hype. What was that environment like?

My advisor told me not to call myself an AI expert because I wouldn't get a job, and he was right, so I joined as a computer scientist. IBM Research was fun and I wrote many papers and patents. In 1993 a paper mill CEO asked us to improve their process; we used agents and meta learning, improved output by 1.5%, and they used our system for 24 years until 2017.

Your external memory work assumed data too big for memory. Does that constraint look familiar today?

The hierarchical memory idea came from Jeff Vitter at Brown University; I solved it theoretically and we wrote papers together. The concept was picking a block of memory items to work with efficiently, which is exactly what GPUs do now with matrix multiplication. It's comforting that theoretical work from 30 years ago became very practical.

You've lived through AI winters. What actually caused them, and is that mechanism present today?

The first winter came in the early 1960s when researchers hyped AGI within 20 years, inspiring 2001: A Space Odyssey, but it collapsed in 1973 when perceptrons failed. The second was 1980s expert systems, which also couldn't solve real problems. Now semiconductors are cheap and computers are strong, so this hype hasn't broken yet, but I think it may crack within the next year.

Scry infuses AI into organizational workflows. What breaks when a large enterprise tries to do that on its own?

The core problem is that everyone treats LLMs like gods that know everything, but they lack the tribal knowledge and subject matter expertise embedded in a workflow. Proof of concepts hide edge cases, but in production those edge cases become 15 to 20% of the problem and users abandon it. Most Silicon Valley teams only have data scientists, not subject matter experts tied at the hip.

Inside a bank or hospital, who has to say yes before an AI system goes live, and who gets forgotten?

The business owner with the actual pain point has to agree, not the CIO or CTO, someone like a CFO frustrated that month end reporting takes 15 to 20 days. Banks and healthcare companies also don't want data leaving their firewall, so SaaS models often don't work; they want licensed software behind their firewall, and the security officer adds many roadblocks.

What happens to knowledge work over the next decade if the current trajectory holds?

I believe 20% of US knowledge workers will need to move up the value chain or into innovation within 10 years. By 2040, much knowledge work in places like McKinsey or legal could be handled by AI if subject matter expertise is properly infused. By 2050 declining global population may mean we're begging AI and robots to do our work.