Thoughts on ChatGPT and Generational AI
Back in the 80's I did neural network research at the Recruit Institute for Supercomputing Research in Tokyo. In spite of this, my track record in predicting the speed that breakthroughs would take place has been terrible. I didn't think that machines would beat grand masters at chess in my lifetime. Later, I didn't think that computers would play the game of Go at a professional level. Even as recently as a few years ago, I didn't think that computers would be able to translate efficiently between Japanese and English. And yet, here we are.
With this background, and these caveats on my predictive abilities out of the way, I'd like to share my thoughts on ChatGPT and generative AI.
Over the years I have been closely following the developments in artificial intelligence/machine learning and their applications. I was in awe when GPT-3 was released -- it showed incredible abilities in creating creative content, code, and even translation.
The latest iteration of this is ChatGPT, a model optimized for dialog. If you have seen the headlines on this, but have not looked at it in detail, I strongly recommend you dig deeper right now.
The results are truly impressive, its ability to generate text is better than what I am confident in producing in many domains even I were given no time limit, an it is possible to have the machine modify results by providing suggestions rather than rewriting of initial prompts.
While you are at it, also take a look at the image generation tools Stable Diffusion and DALL·E 2, and the automatic voice recognition tool Whisper.
These are transformative technologies that are going to change the way we consume and create information, conduct creative work, and interact with systems in very short order. And yet, most people still have no idea about their existence and capabilities. It feels like the early days of the internet, when we first got an email account, or used the first browser.
If we think about productivity as a function of a problem solver's intelligence, the distribution of the problem solver's skill/knowledge and overlap with the problem domain, and the time spent, the current state-of-the-art generative AI tools add large coefficients to each term. In other words, they help us solve problems faster and more effectively, filling in the gaps between our skills and knowledge we have and the skills and knowledge needed to solve the problem, at the cost of adding a relatively inexpensive quality validation and curation step.
领英推荐
As it is much easier for people to recognize good art, prose, descriptions, stories, and code, than it is to generate them, so we will shift to using machines to generate these wherever possible, especially when time is a constraint, which will revolutionize the way we work and create.
This is a tipping point moment for this technology, and for those that have been exploring the applications, it is widely expected that the impact on people and society will be massive, while remaining hard to predict. For example, will people still spend the time to become bilingual in the future if machines can provide real-time translation? Will the military applications of this technology result in less peace as stronger countries can have victories at a lower cost, or more as weaker countries find it cheaper to defend themselves? And who will get squeezed out of jobs as less people are needed to do the same work across the whole of the information work spectrum?
OpenAI's GPT-4 is expected to be a large advance over the already impressive GPT-3, and it may be released as early as next year. Others will doubtless make offerings that similarly extend the generational AI capabilities available.
It is truly a fascinating time to be alive and witness these breakthroughs.
As someone who has been optimistic about the feasibility of general artificial intelligence but pessimistic about its potential timeline, I am now rethinking my evaluation of the singularity and the possibility that we may live long enough to see it. See caveats above.
p.s. Please share your own insights in the comments. I'd love to hear your thoughts.
Financial Markets Technology
2 年Nice post Shaun. This from OpenAI in recent days:?"it’s a mistake to be relying on it for anything important right now. it’s a preview of progress; we have lots of work to do on robustness and truthfulness.” One of the most interesting things to watch for me in the short term will be the pace of this progress. It may tell us quite a bit about this possible coming singularity. Lets catch up soon.?-Matt
Managing Director at Kahlenberg Research Center
2 年Thanks for this interesting article. As much as GPT-3 is impressive at first sight, we should not overlook its serious limitations when trying to write an article about any subject requiring some real insight. The AI writers may produce a very readable text and be able to fool a casual reader, but even when trained carefully, they may produce utter nonsense. I tried some of these AI writers to expand on a text on shade-grown coffee and the AI-generated text gave the recommendation that shade-grown coffee should not be consumed in the dark. Another problem is that Google will detect these artificial texts easily and mark them as low quality. I found a tool based on GPT-2, a previous version, which also can tell you whether a text is written by a human or AI. Still, I think that these AI tools are helpful in giving writing prompts when you are stuck. They can also point you to additional aspects you haven’t been thinking of. But then, the human writer should take over again.
CEO at Excape Entertainment Group
2 年Shaun, do you have thoughts on the metaverse? Not whether it will be around in 10 years, but whether it will "break out" in the way that MZ expects, billions of users, hours per day, the next big thing?