Will ChatGPT destroy the world?

People I speak to generally fall into two camps when it comes to ChatGPT. They either have no real idea what it can do; or, having explored it for a bit, are now 'bored' and totally underwhelmed by it.

Both camps are wrong. Very wrong.

“Superintelligence is most impactful technology humanity has ever invented and could lead to human extinction", according to one of ChatGPT's creators.

We’ve been hearing about humanlike AI being “just around the corner” for years now. Back in the late 1960s, when Terry Winograd began his graduate studies, many believed machines would soon replicate human tasks like seeing, hearing, speaking, and moving. This optimism was reflected in the work of AI pioneers, and nowhere was it more evident than in Winograd’s own project, SHRDLU.

SHRDLU: The Dawn of Conversational AI

For his PhD dissertation at MIT, Winograd created SHRDLU, a program capable of interacting in English within a simplified world of toy blocks. This “microworld” allowed SHRDLU to respond to commands and questions about its environment, showcasing an early form of natural language processing (NLP). SHRDLU’s capabilities were impressive for the time, leading researchers to feel optimistic about the future of AI and NLP.

However, Winograd’s experience with SHRDLU also revealed the vast amount of knowledge required for even basic language understanding. This realisation pushed him towards focusing on human-computer interaction, rather than the pursuit of human-like AI.

The Shift to Statistical Approaches

Fast forward to 2006, Winograd acknowledged new developments in statistical language understanding, neural networks, and machine learning as steps towards effective human-computer interaction. He emphasised “enlightened trial and error” over perfect planning, highlighting the challenges of modeling the complexities of the real world.

The real breakthrough came in 2012 with significant advances in statistical analysis, particularly in image recognition. This progress rekindled the belief that human-like machine intelligence, or even “superintelligence,” was within reach.

The Winograd Schema Challenge: A Test of AI’s Understanding

Not everyone was convinced. In 2010, Hector Levesque proposed the Winograd Schema Challenge as an alternative to the Turing Test. This challenge required AI to resolve semantic ambiguities, such as determining what “it” refers to in the sentence: “The trophy doesn’t fit in the brown suitcase because it is too large.”

Despite initial interest, the challenge proved difficult. At its first and only run in 2016, the best-performing AI achieved just 58% accuracy, barely better than chance. AI’s struggle with such tasks illustrated the gap between language processing and true understanding.

The Transformer Breakthrough and Beyond

In 2017, Google researchers introduced the Transformer model in their paper “Attention is All You Need,” revolutionising NLP by improving the learning of word dependencies. This innovation led to rapid advancements in AI, enabling machines to achieve human-level performance on the Winograd Schema Challenge by 2019.

Despite these achievements, AI’s limitations persist. Large language models like GPT-3 can still falter on tasks requiring a deep understanding of context and physical properties. For instance, confusion about the spatial relationship between objects in simple sentences shows that AI’s understanding of the world remains incomplete.

The Road Ahead: From Narrow AI to AGI

The Winograd Schema Challenge’s apparent defeat by modern AI systems has sparked debate. Some argue that the challenge’s success doesn’t equate to genuine intelligence or common sense reasoning. This sentiment echoes the concerns raised by Yehoshua Bar-Hillel seventy years ago and by Winograd himself in 2004 about the fundamental differences between machine operations and human cognition.

Today’s AI ambitions are grander than ever, with a focus on achieving artificial general intelligence (AGI)—machines with the full spectrum of human intelligence. While AI continues to excel in specific, “narrow” applications, the dream of replicating human-level intelligence remains a formidable challenge.

Reflecting on the AI Journey

My view is that our minds are not like computers at all and it is this comparison that is hindering progress. Perhaps, by abandoning the quest to replicate human intelligence, we can unlock a myriad of narrow AI applications that enrich our lives.

As we move forward, it’s crucial to balance ambition with pragmatism, recognising both the incredible potential and the current limitations of AI. Embracing the strengths of narrow AI applications can lead to significant advancements, even as we continue to explore the broader horizons of humanlike AI.

I find all this stuff fascinating and could go on and on. But I realise that it might not be for everyone, so, if you've made it this far, would you be so kind as to click on one of the options below?

Would you like more content about ChatGPT?

Yes. I don't really understand what it can do.

Yes. I get it but I'd really like to understand real world practical uses.

Nah. Get back to irreverent thoughts on whatever's on your mind.

Stay boring, stay successful.

Jon

P.S. The best way to show your love of this newsletter is to forward it on to one of your mates who you think will like it. To date 78% of my subscribers have come from recommendations (thanks, Mum).

• → jon-kaye.ck.page/sign-up