The Paradigm Shifts in Artificial Intelligence
What’s the Right Question to be Asking about AI?
My Latest Podcast
My most recent podcast was about the state of the social media landscape with legal scholar Paul Barrett, who is Deputy Director at NYU Stern Center for Business and Human Rights.
Even before the Musk/Twitter storm, social media content and its moderation were a hot political issue. There are now several cases awaiting a Supreme Court verdict around liability. Amplification and suppression of content is at the heart of these cases. What is bizarre is that even as platforms struggle with content moderation, some states are demanding that platforms NOT be able to preclude certain types of content.
There’s a lot at stake. A key question confronting us is the following: “Are current laws adequate for social media platforms?” Or do we need amendments or new laws altogether to deal with this 21st century phenomenon? For answers, tune into my conversation with Paul.
As you might imagine, I asked ChatGPT3 some of the same questions that I asked Paul during our conversation. It gave some good answers, but not quite as nuanced as Paul’s.
The Focus on Artificial Intelligence
At a recent talk I gave to kick off an “AI Innovation day” at a large financial services organization, I described the paradigm shifts that have occurred in AI, and what is different this time around with developments such as ChatGPT3, which has captured the world’s attention in a few short weeks.
Every business these days seems to be focused on AI. I sense a palpable fear among business leaders. Will AI disrupt my business model and eat my lunch? Should I worry? Can I leverage AI to stay ahead of the pack?
All good questions with the same answer: yes. But how?
Looking at the history of AI, the paradigm shifts have been towards methods that rely less on human-specified knowledge and more on machines learning through observation on their own. And there’s a lot to observe, buried in all the data out there. These methods have been chipping away at a major bottleneck that has been central to AI: how to get reliable knowledge into the machine and use it. In earlier generations of AI, knowledge had to be specified painstakingly by humans, which could take years with no assurance of success. Current-day machines can often learn even better from data in minutes. The question is, when can you rely on them.
Current day systems such as ChatGPT3 learn almost entirely through “self-supervision,” that is, by constructing their own training data from all available language content on the Internet. What is fascinating is that ChatGPT3’s core competence is its ability to autocomplete sentences. It can guess the next word in a sequence, which in turn is used to guess the next word, and so on, to the point where it can write entire paragraphs and stories. It turns out that in the process of solving autocomplete, it also learns the implicit relationships among things, which is useful in solving more general tasks such as answering questions and creating new materials based on prompts. Video equivalents of GPT3 are in the works.
The technology is general enough to be applied to all kinds of problems. I recently demonstrated to a real estate lending company how it can be configured to predict the condition of houses using natural language descriptions or images using off-the-shelf software with minimal training. Imagine the productivity impacts of such a technology on this multi-trillion dollar industry.
These kinds of use cases, using language and vision, are remarkable because they enable the machine to take inputs from further “upstream” than was previously possible. In other words, the machine is able to perceive the world in terms of the same inputs as humans, such as sight, language, and sound. Touch and smell are next. With this new capability of perception, machines have overcome a major communication bottleneck with humans. Instead of being forced to communicate with the machine on its terms, usually via artificial interfaces like keyboards and pointing devices, we can converse on our terms. This capability signals a major paradigm shift.
What’s Different This Time?
In my talk to business leaders, I encouraged them to ponder two questions, namely, “what’s different this time with AI,” and “why does it matter?”
There were some very interesting questions following the talk. A prominent one that several people asked was “how long are we from artificial general intelligence?”
“General intelligence” refers to essential mental skills that include spatial, numerical, mechanical, verbal, reasoning, and common sense abilities. The idea is that general intelligence underpins performance on all mental tasks. Artificial general intelligence (AGI) is the ability of a machine to understand or learn any intellectual task that a human being can.
When we should expect AGI to emerge is a common question these days. In response, I used one of my favorite Yogi Berra gems, “It's hard to make predictions, especially about the future.” I was also reminded of an old Hebrew saying that I picked up from Daniel Kahneman during our podcast conversation, "Prophecies are for fools."
Heeding Berra’s and Kahneman’s caution, I’m not going to make a prediction. But it’s really the wrong question for business people to be asking. What if I answered 2 years? Or 5? Or 10? What would you do differently? If it’s 10 years, for example, would you wait until year 9 before thinking seriously about AI?
So, what’s the right question?
Instead of asking when we should expect AGI, the better question is whether the current capability of ChatGPT3 and its improvement trajectory is good enough to be disruptive. You don’t need AGI to enable better search or creativity, just better AI, and that’s what ChatGPT3 represents.
What is riveting about ChatGPT3 is that despite its occasional flop, it displays remarkable conversational coherence, bolstered by its improving programming capability , analytical expertise, common sense, and its ability to create new content. While search engines find things for us, AI can now converse about these things and create individualized outputs for us.
For business leaders, ignoring inferior technologies until it is too late can kill their businesses. Google, which has seemed invincible until now, suddenly looks vulnerable thanks to ChatGPT3. In a previous newsletter called The Next Tech War, I discussed how Microsoft has suddenly put Google between a rock and a hard place.
As the picture below shows, while search has been improving incrementally since the turn of the century, progress in conversational AI is seeing a hockey-stick rise, making Google’s search engine look archaic compared to ChatGPT3.
The improvement in quality stems from its ability to understand and converse with us. This opens up all kinds of new ways of performing intellectual work, which is a big bottleneck that limits human productivity and capability.
For example, AI systems like ChatGPT3 expand our attention by performing intellectual work for us. At the moment, reading, interpreting and creating materials is onerous and time consuming. It limits how many things we can attend to during a day. AI can create customized marketing pitches automatically using existing documents and other materials, replacing expensive creative human talent by low cost machines that are as good or even better. Conversational AI will also change the way businesses operate, for example, enabling new forms of customer support through a natural human-like dialog that hasn’t been possible until now.
A pressing question for business leaders is how to act.
I would point to lessons from previous waves of innovation such as electricity and the Internet. In such cases, the innovation required enterprises to make major changes in order to reap the benefits from the technology. With electricity, factory work was reorganized. With the Internet, similarly, knowledge work changed, as did business in general.
AI leaves businesses in the similar situation. The challenge is in positioning for how to leverage the new capability. The biotech company Moderna offers a good example. Prior to COVID, it had invested heavily in digitizing its processes, and in technologies such as databases and the cloud. This capability positioned it for rapid knowledge discovery. After the DNA sequence for COVID was released by Chinese scientists, it took Moderna less than three weeks to create a product that was ready for clinical trials. The record before that was 20 months.
Business leaders would be wise to study the case and learn from it.