Global Finance: Geoffrey Hinton, called the godfather of artificial intelligence (AI), resigned from Google last month, citing the technology’s danger to humanity. Is he right?
Pierre Haren: The world is insane. Hinton resigned but explained that he was worried about what he had created. But today, the most insane thing is a scientific article where people are using ChatGPT to look at images of different brains and can detect 82% of the ideas that the subjects have in their brains. Can you believe that?
Haren: We have no idea where we’re going. We have no idea. Anybody who tries to predict the future here is full of it because it’s a phenomenon that humanity has never experienced. And we are all happily going in there.
IBM recently announced that it fired 8,000 people because of ChatGPT. The mess this is going create is fascinating. And by the way, the potential is so large that they can listen to your brain without any intrusion. [laughs]
GF: Given your background, do you see AI reporting based on little or no understanding of it?
Haren: It’s strange. Even the authors of the technology do not fully understand why it works. There is no theory of why this thing works the way it works. So they have put this convoluted neural network together and are surprised at how well it works. And everybody’s surprised. Now everybody who’s technical is surprised at how well it works because we don’t have a good theory of why it works that well.
If you talk about nuclear power, there are plenty of people that talk about it without knowing anything about it. There is a theory about it, so there is some truth someplace. Well here, AI is an experiment without the knowledge of why it works the way it works. And so, of course, lots of people can say lots of things.
GF: How should businesses decide to use large language model AI engines like ChatGPT and similar technologies?
Haren: Every company should have a small group of employees that watch what’s happening and reports to the CEO or top management every two weeks. This small group should be made of two people, at least—one person who knows the company well and a younger one who is excited about using ChatGPT.
And that’s because it’s only through the interaction of knowing the company’s pain points as well as knowing what ChatGPT can do and iterating very quickly. After all, it changes every two weeks. So, intelligent decisions can be made about where to put the money; where to put the rare investments that any company can make using this technology. If you don’t do that, it’s a big mistake because the speed at which things evolve is such that you only need a few people to run it.
The second thing is that this system works well today if someone can validate what the system produces and have it run non-mission critical processes.
If you are coding things like websites, it works frighteningly well. Some of my smartest friends and the smartest people I know say they gain a 30% performance improvement by using ChatGPT as a coder. It can generate code incredibly fast and is good enough to test it. If there is a mistake in the code, you can see it right away.
However, when we prompted ChatGPT to write a paper that a large consulting company would want, it produced a 20-page paper within an hour. Of course, it wasn’t as good as if a person wrote it, but it took an hour to generate it, and I took 10 hours to turn it into a decent scientific paper since I knew the topic and could correct the errors.
However, the productivity increase is going to affect you tremendously because you are producing content as well. It will benefit every company’s marketing departments since all of a sudden the first pass of the copy will be generated at high speed and automatically.
GF: What impact will AI have on capital and operational budgets? Will we see a radical change?
Haren: This is where you must distinguish between the investment that manipulates electrons and the investment that manipulates protons. Everything physical—moving things, building a plant and other things—will not be incrementally improved.
However, everything about ideas, text, data exchange and similar things can be seriously modified. I was talking to a global consultancy, and they saw this as an incredible opportunity that is way bigger than enterprise resource planning [ERP]. However, here it’s even worse. The fear of missing out is even worse for everybody whose tasks are to manipulate electrons.
It would be best if you had a trusted team inside that can somehow scan the opportunities and frequently report on developments so that top management is not blindsided by something they should have done and have not seen for a few months.
The horizon for investments for some of the stuff will go way down. But again, this is the proton-versus-the-electron fight.
GF: These large language AI engines scrape data from across the internet, which contains many things that cannot be trusted. What is missing in AI that could make it more reliable?
Haren: The traceability of the knowledge will be solved one day. Today it is mainly through the efforts of European and other regulators preventing the theft of content since its untraceable. There is much free content on the internet, but AI developers have also taken from many books without crediting their authors.
This behavior isn’t legal, when you think about it. I’m sure they will be able to solve it because part of it is technical, and part of it will be the regulators who want to ensure that human content generators are not stolen from. The first good regulation arriving from Europe will ensure that human content will not be replaced.
GF: How can or will AI engines like ChatGPT verify the data they consume?
Haren: With ChatGPT, there is no reason why within, I don’t know, at most two years, it won’t be able to self-check with unlimited access to all the data. This is not only with its content but all the available data, which is enormous compared to what any human can do.
One AI engine, Auto-GPT, breaks down problems and sends agents out to solve their particular part of the problem. It’s amazing because if one of the agents decides that the answer is not suitable, it can spawn multiple other agents that will roam the web to find other components.
GF: How do you picture the future of AI? Will it replace today’s search engines?
Haren: My whole idea is that predicting the future is very difficult. So, you have to be very, very careful. I am sure it will be way more than Google because at least half of the coding tasks in the world can be done by ChatGPT in five years. It is a massive change. A lot of people write code. So it’s going to go beyond just searching and finding text that you have to read to decide what to do with it. There is no doubt that we are beyond that.
Most jobs that manipulate electrons will be touched, whether they are doctors, lawyers or risk managers in financial institutions.