I came across this youtube video https://www.youtube.com/watch?v=PBdZi_JtV4c&t=1008s. Excerpts from Web summit discussion titled “Debunking the great AI Lie” by Noam Chomsky, Gary Marcus, and Jeremy Khan.
The topic itself is very catchy and as an evangelist for “AI for Enterprise”, I was intrigued and listened to the 32-minute clip. At the outset, the arguments about language complexities, human and language connections, and how pure data-driven tools like GPT3, etc. are not sufficient to emulate full language comprehension were compelling. Gary mentioned the lack of sufficient early inputs from language experts in the development of the AI model. Chomsky used a very strong statement (even though he said he is being nice to AGI) about GPT3 and other types of language model research. He emphatically said these research efforts neither advance engineering capabilities (like telescopes in astronomy) nor advance scientific research on languages. However, he also mentioned that due to his hearing problem, he was able to understand questions from Jeremy using the real-time voice-to-text transaction. Gary also felt that $billions of research on autonomous AI were also a waste of money and effort. Gary also admitted that while these data-based models get it right 75% of the time, they can never get it right the remaining 25%.
I do not agree with the statement that these research areas and the products are a total waste. Research on autonomous cars is being used every day when you drive any mid-size car around the world. Parking assist, reverse assist, and lane change alerts are standard features in most cars and are used by millions of people. Language translators are helping millions of tourists around the world to do basic communication in foreign destinations. Voice-to-text tools like Alexa, and Siri have become standard appliances in most households.
We will leave the philosophical questions raised by the gurus about language nuances, syntax-semantic connections, the importance of past memories, etc for later and get back to our theme at @AiThoughts.Org i.e. “AI for Enterprise”. I have already argued in my earlier blog about how Artificial Narrow Intelligence is sufficient for the next level of enterprise automation and performance enhancement. I can reiterate by saying “Narrow AI is Deep enough for Enterprises” and we welcome all research efforts towards AGI which will eventually help in making ANI models much more accurate and solve enterprise problems.
Just to recall from my earlier Blog, I talked about HR-Operations certified expert NAI model able to understand most of the employee’s queries on HR-Operations issues and invoke the right ERP applications to complete the transactions. Gary’s assessment of 75% AGI correctness may translate to maybe over 90% in a limited domain narrow intelligence world such as HR-Ops, Invoicing Ops, Billing Ops, etc. I feel the research should not worry about reducing the gap of 10% but create an egoless model which can quickly realize that the employee’s query is not understandable by itself and transfer the call to a human agent in 10% of cases. How quickly and how seamlessly the model does the switch is what makes for a superior customer experience.
Hip-Hip-Hooray for the AI researchers both in AGI and ANI. Each will feed into the other to make AI models more accurate and usable. Enterprises have a huge amount of open source and commercial technologies and models available in ANI to get the best performance and they should allocate sufficient budgets and accelerate AIML projects in 2023.
More later,
L Ravichandran
No Comments yet!