91av

“All possible universes exist, even triangular ones”. These were the words on the cover of 91av on 6 June 1998, when Max Tegmark made one of his first appearances in the magazine. Inside, the then 31-year-old expanded on his idea of a multiverse on steroids, in which all logically possible universes not only can but must exist.

Tegmark, now a professor at the Massachusetts Institute of Technology (MIT), is known for his provocative ideas. As he explains in the “” section of his website: “Every time I’ve written ten mainstream papers, I allow myself to indulge in writing one wacky one.” But the outlandish elements shouldn’t overshadow his serious track record in cosmology, quantum information science and the study of some of the very deepest questions about the nature of reality.

Recently, Tegmark has shifted his focus to intelligence, both human and artificial. He conducts front-line research in artificial intelligence (AI), most recently working with fellow MIT researcher Silviu-Marian Udrescu to create an by studying patterns in data. In 2014, he co-founded the , which aims to understand and mitigate existential risks to humanity, particularly those associated with the rise of AI.

Richard Webb: What made you switch from cosmology to working on artificial intelligence?

Max Tegmark: I’ve always been fascinated by big questions, the bigger the better. That’s why I loved studying the universe, because there were philosophically very big questions like where does everything come from, what’s going to happen, what is our place in the grand scheme of things? We have made enormous progress in cosmology, but at the same time, really new data has started to become rarer and harder to obtain.

So it was very natural for me to gravitate to the biggest unsolved mystery that’s sort of coming within range. We are able to see things with telescopes that our ancestors could never see, and the same thing is happening now with the mind. We have so much data now from neuroscience, and the ability to build artificial versions of the things that we are trying to study.

What are you working on right now?

My research is focused on what I would call machine learning for good. We have been doing a lot of work recently on a project that applies machine learning to identifying news bias. I had gotten increasingly fed up with the quality of the news here in the US, and I made a New Year’s resolution a while back that I was no longer allowed to whine and complain about something unless I actually spent some time working on making things better.

How can AI make the news less biased?

There are these projects aiming to improve the quality of the news by having humans go in and fact-check and flag problems. But if you look more closely, you will see that some fact-checking sites find 95 per cent of errors in media outlets on the left side of the political spectrum, and other ones will only find errors in the media outlets on the right. It’s unclear exactly what criteria they use.

We decided to build something entirely automated. It’s a work in progress, but we use machine learning to classify news articles on all sorts of different metrics: by the topic that they are about, whether they are left or right, pro- or anti-establishment, in-depth or quite breezy, more inflammatory or quite nuanced. The is a bit like Google News, but with a bunch of sliders underneath, so you can adjust for what you want to read.

Doesn’t that risk reinforcing echo chambers, with people choosing to see only the news that conforms to their biases?

The status quo is already like this – if you go on Facebook, it’s entirely reinforcing your echo chambers. The question is, if you get the opportunity to make slightly more deliberative choices, rather than it being just sort of impulse eating, does that make things better or worse?

There are some really nice experiments done by psychologist at MIT that find it’s a bit of a myth that people only want to read things that they agree with. People are interested in hearing other points of view, as long as they are presented in a nuanced way. We can use machine learning to discover which articles are the nuanced ones and which are the ones that are just likely to piss people off. My hope is that a user won’t just set their preferences once and for all, but exhibit some curiosity.

What is the broader agenda of “machine learning for good”?

I think the fundamental challenge we have with AI, and technology more broadly, is to win the wisdom race. We need to make sure that the power of technology doesn’t grow faster than the wisdom with which we manage it.

Historically, we have stayed ahead by learning from mistakes. We invented fire, screwed up a bunch of times and then invented the fire extinguisher, the fire brigade and fire alarms; we invented the automobile and then invented the seatbelt, the airbag, the traffic light and laws against driving too fast.

The challenge is that when the power of the tech crosses a certain threshold, learning from mistakes stops being a good idea. We don’t want to have an accidental nuclear war between the US and Russia starting in 20 minutes and then, thousands of mushroom clouds later, be like: “Oopsie, let’s learn from this mistake.” We see the same thing happening with synthetic biology and ultimately with artificial intelligence as it gets closer to human abilities. So this is the focus of my research. How do we make AI that we can actually trust?

Why is trusting AI so important?

The greatest breakthroughs in machine learning recently have come from artificial neural networks, which can do all sorts of wonderfully smart-looking things, like beat everybody on Earth at chess and Go. But we have very little clue how this AI works. We tend to treat it as a black box and then, every once in a while, it doesn’t work as we thought it would. We have problems like Boeing really wishing that it understood better how its automated system on the 737 worked, or the trading company Knight Capital wishing it knew how its automatic trading system worked before it managed to lose the company $10 million a minute for 44 minutes straight.

“The space of possible artificial minds is much bigger than that of biological minds”

Then we had courtrooms around the US using a piece of software to recommend who was going to get probation and who wasn’t. People didn’t really understand how it worked and didn’t realise that it was racially biased. If you can use the sort of techniques that we are hoping to develop in my group to let people peek inside the black box and understand what AI is actually doing, things might look much better.

It certainly sounds like you are a tech optimist.

Are you the kind of person who thinks fire can kill people or the sort of person who thinks that fire can keep people warm in the winter? Both things are true, obviously.

The interesting question isn’t to argue for or against fire, it is to figure out how you can manage fire wisely. Technology isn’t good or evil: it’s a morally neutral tool that can let you do good or bad. Right now, AI is still pretty stupid, but it’s already given enough influence in the world that it’s caused a lot of problems, from biased court decisions to crashing aeroplanes.

I think it’s possible to make very powerful AI and I think if we do that wisely, it can be the best thing ever for humanity, because everything that I love about civilisation is the product of human intelligence. If we can amplify that with AI, we can use it to solve the climate crisis, to lift everybody from poverty, to figure out how to cure the coronavirus and so on. What’s so bad about that?

Is building this sort of advanced “general” AI realistic, given that we don’t even understand how human intelligence works?

You could just as well ask, how could we possibly figure out how to build a flying machine before understanding how birds fly? Darwinian evolution gave us both flying birds and thinking animals, but it was very constrained: to only build solutions that could self-assemble, that could self-repair, that only used a handful of chemical elements, that were super-energy-efficient. When you remove all these biological constraints, you can often find much simpler solutions to the same problems.

I know some people think there’s something magical about intelligence, making it possible for it to exist only in human bodies. I don’t think so. I am a blob of electrons and quarks processing information in certain complex ways, and the key to intelligence is just the nature of that information processing. I would go so far as to predict that the way we are finally going to understand exactly how the human brain works is by building something simpler that is comparably smart.

Presumably we can’t build an AI that thinks or feels exactly as a human does, that has things like agency and consciousness?

I wouldn’t be so sure. I think the most interesting question isn’t to ask what will happen, but what we want to happen. It might be that we have a lot of designer’s choices. The space of possible artificial minds is much bigger than the space of biological minds, because all biological minds evolved – they tend to have a survival instinct first, then other things. When you are free of those constraints, there’s so much more opportunity to choose.

It may be possible to build different AIs that perform equally well on tasks, but have a whole range of conscious experience, from nothing to a subjective experience that feels quite a lot like yours, where it experiences colours and sounds and vibrations and maybe even emotions.

Really? Surely you can’t program something to have feelings?

I think we tend to be very arrogant about this. We have to be very careful with self-serving claims that we know when there is a subjective experience and when there isn’t. We made that mistake with animals, and I think we are making it all over again with machines. Most of my colleagues just take it as an axiom that none of the machines they ever build will ever have any subjective experience, so they never have to worry about suffering and can just turn them off and on at will. I don’t think that’s so obvious at all.

My own guess is that consciousness is simply the way information feels when it’s being processed in certain complex ways. I think scientists owe it to the world to figure out what those complex ways are.

What do you mean by “the way information feels”?

Many people make the mistake of assuming that, when you look around you and you see different colours, that those experiences somehow have something to do with the outside world. For example, if you see an apple and it’s red, and you think somehow that you only have redness because there’s an apple. That’s obviously wrong: you can dream about an apple and you will still experience it as being red, even though now there is no outside world at all. So there is something happening that’s just purely inside your brain as the neurons fire. What is this thing? I want to figure that out.

Do you think we will ever arrive at a full description of how consciousness emerges from atoms and molecules?

This is the Wild West where we are very clueless and have to have very open minds, obviously. But in the big picture, I think about consciousness as the last bastion that has still refused to be captured by physics.

Now even intelligence is beginning, little by little, to yield to mathematical description, right? That’s what artificial intelligence is all about, and there are already some theories out there trying to predict which information processing is conscious and which isn’t. It’s ripe for the scientific assault.

But the laws of physics are themselves the product of conscious deliberation. Isn’t consciousness always going to fall down at describing itself?

Yeah, that’s a very fun idea. Is it possible for a small part of something to be able to describe the whole thing that it is part of, including itself? Or do you get into some bizarre recursive loop? Of course, I can’t know for sure that we will be able to describe consciousness with physics or machines. There are plenty of people who think that we will never be able to describe consciousness because it involves some sort of soul or something that’s by definition impossible to study.

“I think about consciousness as the last bastion that has refused to be captured by physics”

I’m more optimistic. My personal guess is that consciousness can be fully understood in terms of information processing done by particles moving around. But regardless of whether you think it’s going to work out or not, one way to guarantee failure is if you start by convincing yourself that it’s impossible. So let’s try our best. If this all fails, it’s also going to be very cool.

Where do you see all this going?

I think we shouldn’t conflate intelligence with consciousness here. On the intelligence side, I have no doubt that we are going to keep making more progress, unless we self-destruct as a species by screwing up somehow. I just hope we won’t end up saying that curiosity killed the cat, that our curiosity to figure out intelligence made us build things that we used to drive ourselves extinct. That’s why I’m so big on also thinking through the wisdom part.

Are we wising up to AI’s dangers?

I think there’s been a big shift for sure. Now you can’t go to an AI conference without coming across a bunch of talks about AI safety, transparency, interpretability and robustness. There is a lot of idealism in the community. This is where I get a lot of hope that we can use machine learning to empower the grassroots, push back against the powers that be and even sometimes use those tools to uncover sneaky stuff.

That sounds like tech optimism again.

The key to having a good future is to be able to formulate a vision that people around the world can really get on board with. This isn’t a zero-sum game: you can easily envisage scenarios in which artificial intelligence multiplies the world’s GDP by a factor of 100 or more. It’s very easy to envisage a future in which everybody wins at the same time and becomes much better off. But we have failed epically so far to get humanity to collaborate to make it real.