Summary

The author reflects on the evolution of AI and the concept of understanding in artificial intelligence, emphasizing the shift from seeking conscious understanding to focusing on functional intelligence.

Abstract

The text delves into the author's 45-year-long fascination with AI understanding, tracing back to childhood influences like HAL from "2001: A Space Odyssey." The journey includes coding NLP in the 80s, exploring human-like AI in the 90s, and attempting to create a machine-learned NLU startup in the 2000s. Despite initial skepticism and funding challenges, the landscape changed with the advent of ChatGPT. The author grapples with the definition of AI understanding, pondering the circularity of logic in AI's processing of knowledge and the distinction between attribute recognition and genuine understanding. The conclusion is that AI should be judged on its ability to process inputs and produce intelligent outputs, akin to the Turing Test, rather than on whether it possesses conscious understanding, which is both elusive and unnecessary for practical applications.

Opinions

The author has been deeply intrigued by the notion of AI understanding since early childhood, influenced by cultural references and personal coding experiences.
There is a clear distinction between 'knowing' attributes and actions associated with a thing and truly 'understanding' it, which the author finds to be a circular problem in AI.
The author argues that the internal workings of AI are less important than its ability to perform tasks and respond intelligently, suggesting a functional approach to AI understanding.
The author believes that the industry's focus should be on cognition and intelligence, as demonstrated by responses, rather than on conscious understanding, which is difficult to prove and not essential for AI's practical effectiveness.
The author acknowledges the historical difficulty in securing funding for NLU projects, highlighting a shift in the late 2010s with the success of models like ChatGPT.
The text implies a criticism of AI naysayers who demand conscious understanding from AI, despite evidence of AI's ability to pass rigorous logic and response tests.

It’s fascinating to ponder what it means for AI to ‘understand’ something

I’ve been thinking about this for 45 years since I was 6 or 7.

It all started with Hal from 2001.

And pondering, as a kid in the 1970s, what I would connect the speaker and microphone to on a hypothetical kid-invented robot.

And then coding up NLP on the school’s Apple II in BASIC in the 1980s.

Next it was thinking about creating human-like AI or ‘NLU’ in the 1990s in Pascal.

And finally tinkering in the 2000s with Python and creating a (failed) ‘machine-learned NLU’ startup in the early 2010s. Three times during that decade. It was almost impossible to get funding for NLU — natural language understanding — until the late 2010s. Even towards the end nobody believed. Until ChatGPT. When it was too late unless you were a Silicon Valley billionaire just about.

Circularity

All along I was enticingly plagued by an intriguing question:

What does it mean for an AI to understand something?

I argued with myself that we could collect knowledge and create a logical language processing unit. NLU.

But where would the aha moment be?

I went nuts trying to figure out what you would have to do to truly understand something via AI.

It always ended up as apparently circular logic.

It’s one thing to ‘know’ attributes and actions associated with a thing, it’s another to truly ‘understand’ it. Because all those attributes and actions themselves are ‘understood’ in a similar manner!

All you can do is establish relationships between observations and labels and between labels and more labels.

Hence an internal circularity.

But, ultimately, that works . . because use-cases start with input and end with output.

Just ‘answer questions’, output responses?

I was able to trace out how an AI might identify irony. Or be creative. It seemed possible to do all that just thru logic.

I ultimately decided that consciousness was what makes ‘understanding’ especially impressive in humans.

But that was way too hard to fathom and emulate.

And unnecessary.

Who cares what’s going on internally in an AI?

So it became clear to me that we should focus purely on cognition / intelligence. Not conscious understanding.

It struck me that the only practical definition of ‘understanding’ or intelligence itself, was a Turing-like functionality.

Intelligent as tested by responses.

I stand by that today.

It’s especially important today more than ever with AI naysayers insisting that LLMs don’t ‘understand’ despite passing rigorous logic and response tests with ease.

It’s crucial to not insist on conscious understanding because it’s essentially impossible to prove anyway.