Geoffrey Hinton is often called the Godfather of Artificial Intelligence. He has been at the forefront of AI for over fifty years, creating the core algorithm that powers the training of artificial neural networks which underpins recent enormous advances in AI. This week, Geoff sounded the alarm about the potential dangers of AI and resigned from his role at Google to speak more freely about this topic. In this exclusive Q&A with Radical Partner Aaron Brindle, we delve into Geoff’s views on the potential of this technology to benefit society, how his ongoing research might address these concerns, and advice for researchers looking for guidance on how to responsibly pursue their work. Geoff is an investor in Radical Ventures, and a co-founder of the Vector Institute for Artificial Intelligence alongside Radical Ventures partners Jordan Jacobs, Tomi Poutanen and Ed Clark. The following excerpt has been edited for length and clarity.
Aaron Brindle (AB): It’s been a busy week, how are you doing?
Geoffrey Hinton (GH): I have been getting requests from the media every few minutes which is stressful, but I am very glad to see that my message that we should be thinking hard about how to contain a superintelligence is getting traction.
AB: We work very closely with researchers and founders applying AI – a technology you played such a pivotal role in shaping – to solve some of humanity’s greatest challenges. AI is being used to create technologies to help mitigate the impacts of climate change, cure diabetes, and companies using AI for drug discovery may one day soon cure cancer. How do you feel about the pace of breakthroughs you’re seeing in these areas?
GH: I think it is wonderful that the helpful applications of AI are progressing so fast and I would hate to see them slowed down. There is the potential for AI to do enormous good. Now that advanced chatbots can act as intelligent assistants for so many different tasks I think they will lead to huge increases in productivity.
AB: Should research in applying AI to solve these meaningful challenges be stopped?
GH: I think it is completely unrealistic to expect progress to be deliberately slowed down when the potential benefits are so great and there is competition between countries. But I think it is of paramount importance for this progress in developing the technology to be accompanied by efforts on a comparable scale to understand and prevent the various types of bad outcomes. The particular bad outcome I have focussed on is the existential threat of digital intelligences taking over and I wish I had some sensible advice on how to prevent this. Hopefully, if enough smart people work on it, they will discover a way if there is one.
AB: What do you see as the most promising opportunities for AI to benefit society?
GH: Large Language Models will take a lot of the drudgery out of nearly all of the work that involves producing written output. For example, an LLM could probably answer this question better than me but I resisted the temptation. Getting a lot more information out of medical images a lot faster will be hugely beneficial in fighting cancer and other ailments. In 2016, I predicted that systems that were better than radiologists would be here by now and I was off by about a factor of two: some systems are already out there for things like staging diabetic retinopathy and many other systems are now better than an average radiologist. Eric Topol has written recently about the huge potential of an AI general practitioner who can integrate all of the information about a patient far more effectively than a human doctor. Making solar cells more efficient, making it much easier and quicker to write computer programs, predicting floods and earthquakes, designing better chips (but not better phish – PaLM can probably explain this to you if you don’t get it), monitoring our bodies for things like glucose levels or cancer markers, making preventive maintenance of machines much more effective – these are just a few of the huge wins to be had. It’s hard to think of any job that could not be made more efficient with the help of AI.
AB: Your views on AI are clearly evolving. When we last spoke, you discussed the idea of ‘mortal computing’ – a new way forward for building AI systems. Can you talk about how your most recent research might address some of your concerns around AI development?
GH: It was my most recent research that precipitated my concerns. Backpropagation running on digital computers consumes a lot of energy. This is because the computer must behave exactly as expected so the transistors need to be run at high power. A much more energy efficient way to do a vector-matrix multiply is to have a vector of neural activities represented as voltages and a matrix of weights represented as conductances. Each connection then injects charge per unit time and the charges add themselves up. But different pieces of hardware will produce slightly different analog results and the results will also wander over time and be affected by what is going on nearby on a chip. This suggests that we should use a learning procedure that adapts to all the quirks of a particular piece of hardware without actually knowing exactly what these quirks are. I explored the forward-forward algorithm which does just this, but when I scaled it up, it did not work nearly as well as backpropagation.
This, coming at the same time as PaLM and GPT-4, caused me to question my fifty year-old assumption that making artificial neural nets more like real ones would make them better. It was like a game of GO where you are trying to surround your opponent and suddenly it all flips and she surrounds you. I suddenly realized that backpropagation running on digital computers might be much better than biological neural nets because it could do weight-sharing both within a model and between different copies of a model running on different hardware. This only works if you are digital and do not exploit the quirks of the particular hardware you are running on. With weight-sharing, thousands of copies of the same large model can run on thousands of separate pieces of hardware. The copies can look at different data and share what they learned with a bandwidth of trillions of bits per sharing. Biological neural nets have to swap knowledge by mimicking each other’s outputs (like language) and this only allows a bandwidth of a few hundred bits per sentence.
AB: Concerns about the responsible application of AI are widely shared among the researchers and founders we work with. What’s your guidance to researchers in the field who want to make sure the advances they’re making are done so in a responsible manner?
GH: Think hard about which of these two activities is most in need of more creative researchers: making deep learning work better, or ensuring that its outcomes are good rather than bad.
Radical Reads is edited by Leah Morris (Senior Director, Velocity Program, Radical Ventures).
