ChatGPT, AI, and the beginning of the end?
I still remember watching The Matrix when cinemas were still a thing in 1999. It was one of those movies that took a moment to understand, and I had to rewatch it a couple of times before I eventually got it. In the movie, we, the human race, live in a simulation, kept alive as batteries to power the sentient AI that has taken over the world. At some point in time, humans scorched the earth to destroy all-natural power sources with the intent to shut down the machines; but alas, that just resulted in the machines imprisoning the minds of humans in order to use our bodies as power sources. The Matrix, the original, anyway, was an awesome movie.
I first learned about ChatGPT a couple of weeks ago from my 14-year-old son, who naturally was using it to do homework. Not long after, one of my colleagues at Takadao (Mohammed) on our smart contracts team had ChatGPT audit some of his code. It was apparently beneficial. Mohammed also asked ChatGPT what Takaful as a DAO would look like. ChatGPT provided an eloquent explanation of our business Takadao (Takaful + DAO) in seconds, something that took me weeks to grasp. At this point in time, I was more than a little peeved. This thing, this AI, summarised in a few short paragraphs, the essence of our business that even sophisticated investors that I’ve sat with the struggle to understand. Is this v1 of the AI that will eventually use us as batteries?
Sometime after, yet another colleague of mine on the underwriting team put ChatGPT to work. Chirag was researching the term “dynamic sum assured”, and ChatGPT spat out a number of research articles that he should look up. A little background…for those of you in the insurance industry, you will probably recognize that the term “dynamic sum assured” is an oxymoron. In insurance terms, the “sum assured” is a guaranteed amount of payout that, by its very definition, cannot be dynamic. It’s like saying a “square-shaped circle”. Chirag was researching this because we are building a dynamic underwriting model that is not based on a sum assured but based on a benefit multiplier that adjusts according to the performance of the insurance fund.
So imagine Chirag’s surprise when ChatGPT returned a list of peer-reviewed academic journal articles with the terms “dynamic sum assured” in the title. Chirag, excited by his findings, set about looking for the listed articles. A few hours later, it turns out that none of the listed articles actually existed. ChatGPT simply took a composite of actual article titles, journal titles and author names and recreated a list as it saw fit. Yes, ChatGPT lied, blatantly lied and invented the articles that were listed.
Needless to say, I’m now officially petrified. ChatGPT is an intelligent poker-faced liar.
Many in the tech industry have hailed ChatGPT as the ultra-productivity tool that will take everything to the next level. People in web3 talk about using AI to audit smart contracts and to expose security vulnerabilities. However, many also recognize that AI is not benign, just as ChatGPT had no scruples about lying to Chirag. Dmitry Mishunin, CEO of smart contract auditing firm HashEx suggests that the AI, while auditing smart contracts, may actually attack the smart contracts itself in a bid to learn, perpetrating the very thing it is supposed to prevent.
https://cointelegraph.com/news/how-will-chatgpt-affect-the-web3-space-industry-answersIt got me thinking, how does ChatGPT actually learn? If I have understood correctly, ChatGPT learns from all published material that is available on the internet and combines, repurposes and reuses the content in various ways. If this is accurate, then what is unpublished is also unlearnt and ignored. Clearly this excludes knowledge that is not easily published, including oral traditions and knowledge that is intuitive and difficult to capture in text or video. And what about knowledge accumulated by populations that don’t participate in the internet to the levels of the developed world?
Let's assume that the world will all come online soon; a bigger issue still remains, ChatGPT will learn from and take the form of the loudest contributors who create the most content. Furthermore, not only will ChatGPT primarily learn from populations with the highest internet penetration (i.e. the developed world), its learning will also be heavily weighted toward “popular clickbaity” content. We all know that content that elicits rage and shock generates the highest engagement, and the most important influencers, in a bid to win followers, regularly post content with shock value to gain views. If this is who ChatGPT is learning from, its knowledge will be far removed from the real world.
I am hardly a luddite, but ChatGPT gives me the chills. Technology should have a net benefit for the long term and not be embraced just because it’s technology. For now, I will study and engage with ChatGPT in an effort to understand our future masters and to avoid becoming a battery. Anyone reading this with an alternative perspective, please do share it so that I can sleep better at night.
