The scaffolding for AGI is flawed!
Sri Krishnamurthy, CFA, CAP
CEO, QuantUniversity | Entrepreneur | AI Innovation | Education | Author | Speaker
I had the wonderful opportunity to present at #TedXBoston last week! It was a fabulous event themed on Artificial General Intelligence and it was fascinating to see 60+ speakers talk about a variety of themes and topics pertaining to intelligence.
My talk was titled: The scaffolding for AGI is flawed!
The recording and video is still in the works. Meanwhile, by popular demand, I am releasing the transcript. The final delivery is a bit summarized to keep it close to ~7 minutes. But here is the full speech transcript with slides!
------------------------------------------------------------------------------------------------------------
Ladies and Gentleman.
Today, I want to talk about why the scaffolding for AGI is flawed! After working in the industry for more than two decades, I have seen how difficult it is to deploy systems in production and today we are at the cusp of an impending revolution in AI but we need to be cautious on how we deploy it.
Let me present 2 scenarios to you!
Who would you choose?
This autonomous car has avoided 99% of the 10000+ accident scenarios!
Would you like to be driven by this car or a human?
This personalized home-healthcare bot has proven to be better than humans in 99% of the tests.
Would you trust this home-healthcare bot or a real human to take care of your aging parent?
Most would trust humans to machines at this point! Why so?
The reason is we don’t yet trust the system can do all these complex tasks thoroughly!
Even though these systems seem intelligent, we don’t haven’t tested them enough to rely on them.
That brings about an important question. What is intelligence ?
I surveyed 40 of my graduate students and many described intelligence in terms of attributes.
We haven’t agreed on one definition of what intelligence is but let’s look at how we define AGI today
Artificial general intelligence (AGI) is the ability of an intelligent agent to understand or learn any intellectual task that human beings or other animals can.
How do we test whether a system is intelligent or not?
Should we use the Turing test or the Wozniak test?
Without agreement, what we have is different definitions and perspectives of what intelligence is like blind men describing a proverbial elephant
Without having a standard, pragmatic definition of intelligence, what we will have are different perspectives on what intelligence is and we will start chasing metrics!
Before we can build a machine that can do any intelligent task, we have to build robust systems that can be trusted fully for a specific task what I call as Specialized Artificial intelligence!
So how do we get there?
I propose 5 things we should focus on:
1. Defining Intelligence
Human intelligence is different from Animal intelligence. We need to agree on precise, narrow definitions of Specialized intelligence rather than describing intelligence in generality!
2. Degrees of Intelligence
How do we measure intelligence? We differentiate intelligence in mosquitos, monkeys and humans. Our intelligence changes from when we are infants to as we grow older. We need to specify degrees of intelligence in systems so we can qualify and measure them.
3. Formalizing testing
We need to agree on tests for intelligence and formally specify them. It is important that we develop common frameworks for testing.
4. Context aware sandboxes
Intelligence cannot be defined in isolation. We need to define the context in which a system would be used and whether the system’s capabilities as qualified by tests can be proven. It is important that we develop context-aware sandboxes in order to test this!
5. Certification and Qualification standards
Finally, it is important that we develop common certification and qualification standards. For adoption of systems, it is important that we have standards need to be adhered to ensure systems are developed and used appropriately!
We need to be careful. After all we are talking about intelligence. The trait that let us humans evolve and thrive. We are now planning to bestow this and delegate this ability to systems.
I would like to conclude with a verse from the Boston poet Ralph Waldo Emerson.
In his poem Fable, he talks about a quarrel between an elephant and a squirrel!
He concludes by saying:
“Talents differ; all is well and wisely put;
If I cannot carry forests on my back,
Neither can you crack a nut.”
I think he may have an answer to what I am talking about.
Before we can build machines that can lift mountains and crack nuts, we must build machines that are really good at lifting mountains and that are really good at cracking nuts.
Specialized Artificial Intelligence before Generalized Artificial Intelligence!
I hope you enjoyed it. Please comment and share your thoughts! Would love to hear what you think!
Sri Krishnamurthy