The Resurgence of Magical Thinking in Conversational AI
In the early days, there was quite a bit of magical thinking around chatbots. Decision makers had grand visions. Platform vendors would capitalize on inflated expectations with smooth sales demos. And in-house teams with little or no experience were tasked with figuring out how to build bots using technology they had never used before: a recipe for disappointment.
ChatGPT has become the new baseline
Today we’re seeing a resurgence in magical thinking. ChatGPT has ignited an insatiable curiosity in large language models, yet by many the inner workings are poorly understood.
The technology that ChatGPT was built on has been available for quite a while, funnily enough; GPT-3 was launched in 2020. It was the application (its conversational interface) that really blew the lid off of Pandora's box, serving as indisputable proof of the incredible attraction the power of conversation has on us.
Have you noticed how every other chatbot suddenly has become a lot more mediocre? With ChatGPT, a new baseline has been set. Allowing for highly contextual and free flowing conversation, it raised the stakes in terms of what AI assistants are capable of (or should be). A development that can be viewed as both exciting and worrisome.
Do we even need conversation designers anymore?
It’s worrisome because we are talking about an emerging technology that is far from mature. Yet everyone has access to it. It is fluent, it seems smart. Smart enough to create a misleading impression of greatness.
Business leaders are scratching their heads wondering why their AI assistant isn’t as good as ChatGPT. Product owners get asked all these questions. Why aren’t we leveraging this technology? Can we integrate ChatGPT with our AI assistant? Should we sign a contract with this new vendor that says everything will be magically taken care of? Do we even need conversation designers and AI trainers any more?
Even though it might seem like magic, large language models are in fact not magical. They’re a complex technology with unique sets of capabilities and limitations. People need to be educated again, just like the early days, on what’s hype and what’s reality. Deja vu.
Large language models as a driver of innovation
Now onto the exciting part. Innovation was long overdue. If anything, ChatGPT has proved to be the necessary impulse for innovation in a technology landscape that hasn’t seen any major leaps in progress for a while.
With large language models getting faster, cheaper, and more readily available, the application layer seems to be taking off. The first full-stack conversational AI platforms have started to integrate LLMs into their existing solutions, by taking advantage of the strengths and simultaneously mitigating some of the risks that come with it.
My gut tells me that hybrid solutions will become the norm within a year or two, mixing traditional intent-based logic with the generative power of LLMs. One provides the control, the other the flexibility. It would help us build better AI assistants in two major ways, 1) allowing for more fluent responses (i.e. finer granularity), and 2) allowing for deeper levels of context. It could enables us to depart from this rigid ‘dialogue tree’-paradigm that we’ve been stuck in since… forever.
It might not go so fast, though. We have to be careful not to fall victim to magical thinking. It is all too human to overestimate the effects of technology in the short term and underestimate them in the long run. So, let’s practice patience and see where this ship is headed.
The real antidote to magical thinking is — and always has been — to educate ourselves.
Follow me on LinkedIn
If you appreciate the content, consider leaving a like<3
Spot on. "Magical thinking" is far too à propos a phrase for the explosion of interest (obsession?) in ChatGPT. It's almost as if we didn't have enough humans in our lives to talk to. Granted, it's an incredible thing to have an algorithm generate what seem to be actual conversations. Part of it could be the functionally indistinguishable experience of texting with someone... it looks and works the same, because you're not there in person.
This hits it home for me in the piece: "It is all too human to overestimate the effects of technology in the short term and underestimate them in the long run."
The problem with excitement around a technology is people make tall claims and try predicting how it should evolve rather than wait and see how it evolvs. I am reminded of the WiMAX Vs Wifi, CMDA Vs GSM etc. Those questions have been very decisively answered. So are the other hardware form factors and so forth. I definitely agree we need to watch how GPT or other LLM technologies evolve in the coming years.