How to change your AI architecture

My review of yesterday's talk by Venkat Srinivasan

Yesterday, Venkat Srinivasan gave a wonderful talk at the Center for Emerging Markets at the D’Amore-McKim School of Business at Northeastern University. Venkat used to be a faculty member here, before founding eCredit.com and then Rage Frameworks,  an AI-based company that was sold to Genpact. Currently, he is the Founder and Managing Director of Innospark Ventures, which invests in founders and ideas that leverage advanced AI to create a differential and disruptive impact on the economy and society.

Venkat defines AI as ‘codified intelligence embedded in machines’. Remember this, as it becomes important to understand his solution for AI architecture at the end!

Easing into the talk with some history, Venkat believes that this 3^rd wave of AI has staying power. As an example of every field becoming computational, he shared that his youngest daughter is a painter who gave exhibitions in Manhattan. Then she went on to get her masters without any coding experience, but graduated with plenty in New Media and Design.

For full transformation, you need to reimagine your business: think about both improved efficiency and increased innovation. Note this is exactly the advice you got from me earlier in this Substack and my newsletter, discussing my ‘Break The Wall’ digital information book, and David de Cremer’s ‘The AI-Savy leader’ book.

As to AI challenges, a few are the same for all statistical models:

(1)     We need lots of data

(2)     We are never 100% accurate, but should strive to increase accuracy

But ‘neural black box AI’ has additional challenges:

(3)     The model IS the data, so the fundamental architecture is flawed

(4)     We are still far away from general artificial intelligence:

I asked Venkat about the ‘AI snake oil’ book assertion that predictive AI simply doesn’t work. Venkat agrees that he doesn’t trust a model that only works 90%, he wants to explain the remaining 10%. Still, we constantly make decisions based on incomplete information, eg take a pill that is only 70% effective. My take is that 70% good enough for personal choice with little side effects, but not good enough to decide another person’s fate, eg whether an individual should get a loan, or stay incarcerated.

What are the main challenges with neural black box AI?