"AI is a Distraction" — Interview with Harry Halpin; CEO of NYM

Harry Halpin’s work is a unique blend of computer science, philosophy and ethics. Previously working at World Wide Web Consortium (W3C), he resigned after the W3C standards body standardized the Digital Rights Management (DRM). Harry is a strong advocate of open and unrestricted web and strong user privacy. His latest project is called NYM, which aims to introduce a mix network protocol with blockchain-enabled incentivization for sustainability.

KryptoJoseph (KJ): Some time ago you said on Twitter that AI is a bankrupt intellectual programme. What’s the idea behind that?

Harry Halpin (HH): The question of artificial intelligence is a very old desire of humans to create forms of life that have the same properties as ourselves, or possibly even superior to ourselves.

The main problem with artificial intelligence is that it takes a very impoverished view of intelligence in the form of narrow problem solving, like identifying certain objects. More than intelligence itself, it relates to optimizing problems and finding efficient solutions to given problems. As has been pointed out by Hubert Dreyfus and other philosophers, the world is not partitioned to a set of discrete problems.

World is a continuous flux that we have to make sense out of. This ability to make sense out of things autonomously in an open-ended system is really what defines intelligence and this aspect has been historically ignored by AI programs. Deep learning isn’t a magical shortcut to solve it. All that deep learning does is it takes the symbols of feature set for a given machine learning problem from lower level features. But that means the problem to solve has already been predetermined.

This question of autonomy really I think reveals how bankrupt artificial intelligence is as an intellectual project insofar if anything, the project should be about how we as humans think and what is the nature of autonomy and self-determination. And these largely philosophical questions are barely touched by artificial intelligence research. So artificial intelligence as a program, while everyone is excited about it and tends to give it magical powers, is nothing more than efficiently solving problems and a straightforward extension of the industrial revolution, Taylorism and an ability to solve certain kinds of problems with new kinds of machinery. I don’t think it’s a giant threat to human existence or on the other hand a giant boon to human prosperity. I’m deeply skeptical of any large impacts of artificial intelligence in the near future.

The danger of artificial intelligence lies in it acting as a distraction from the real fundamental problems that humanity is facing, like surveillance, control, climate change, economic collapse, social disorders. The ancient fantasy of creating a thinking machine is just that - a fantasy to distract people from the real problems we collectively face.

HH: Artificial intelligence and surveillance have an intimate connection, particularly when it comes to machine learning. In order to have a relatively dumb algorithm appear intelligent, it has to be trained with a large amount of data. This data needs to be collected and processed, which basically requires surveillance - large data collection. The more data you can collect, the more effective and intelligent these algorithms appear - but I’d like to emphasize the word appear since the algorithms aren’t actually intelligent in a way that living things are intelligent. Since we don’t have a satisfactory theory in place on what constitutes intelligence, the best that machines can do is to pretend they are intelligent by exploiting large amounts of data to do essentially non-linear function fitting over predetermined objective functions. And that simply does require mass surveillance.

One problem the artificial intelligence could potentially solve is a problem of production - how to produce stuff in more efficient ways. But when there is no market for such algorithms, machine learning is naturally repurposed for social control, because the state will always want to buy machine learning algorithms to monitor its citizens, to make sure people will continue to exist in a way which is subservient to the state’s needs. While there is a huge movement by researchers trying to understand how they can eliminate bias and discrimination in AI, the real question is how we can eliminate surveillance and preserve freedom. And I think on that level, the obsession with AI has set back the computer science research.

First of all, many people who don’t know much about AI get very excited about it. Anyone who has worked in the field and knows linear algebra ran some Bayesian classifiers and quickly understood the limits of artificial intelligence. Second, after Snowden’s revelations there began an emphasis in computer science on fighting surveillance, preserving privacy and building tools that enhance freedom, but unfortunately, this rhetoric, which naturally is anti-state, has been absorbed by a “fair” artificial intelligence crowd which essentially says “how can we be fairly spied upon? How can AI facial recognition recognize black faces as well as white faces?”

But I don’t want to live in a world where AI-aided surveillance is considered normal and mandatory, and fairness is understood as being equally well observed. I’d rather live in a world where we’re not observed and controlled. So again, AI and the rhetoric surrounding it is a distraction from the real problem.

HH: Collective intelligence is, I think, a much better concept than artificial intelligence. It explores how humans and machines can enhance each other’s intelligence. Humans are very good at making sense of the world and making autonomous decisions, but humans have natural problems with memory, recall and processing large amounts of data - areas that machines are very good at.

Collective intelligence imagines a symbiotic and mutually beneficial relationship between humans and machines. There are different kinds of humans with different qualities and the same applies to machines, and the question is of the correct ensemble of humans and machines. I think this is a much better way of looking at intelligence and also a more individual-based take on artificial intelligence.

All humans exist and act within society and our intelligence is to a large degree social; it’s dependent on a culture we come from, on people in our wider milieu. Thus collective intelligence makes much more sense as a way of extending our intelligence than hoping that we can be replaced by some kind of Roomba-style robot slaves.

HH: Semantic web, although pretty much a failed project now, is pretty interesting as it imagined we could get machines to process data in a way that is decentralized and thus extend our knowledge. The decentralized nature of Tim Berners-Lee’s vision of the semantic web is very admirable. That being said, giving every single piece of data an identifier in a reliable way is unworkable with original web technology. I am quite a fan of blockchain in this respect, as it offers some level of decentralization and cryptography and thus provides a better bedrock for social computing than the traditional web.

I’ve mostly moved on from my research of semantic web, as it’s only been used by large corporations for knowledge graphs and surveillance.

HH: NYM is the latest attempt to build a technology that could really defeat mass surveillance. TOR has been the best technology in this regard for the last 15 years but the problem with TOR is that while it does obscure geolocation by obscuring IPs, it assumes the adversary cannot watch the entire network and cannot correlate traffic between entry and exit points. These kinds of attacks, as we know from the NSA revelations, are possible. Early attacks focused on breaking the TOR browser, but today we know there are massive amounts of data collected and stored through mass surveillance. So it makes sense that NSA simply watches all the TOR entry and exit nodes, stores and correlates the data. You can correlate the data via the time the data was sent and the size of the data, from which you derive what’s called the patterns of data. Machine learning processes data automatically find out these patterns and deanonymizes TOR users.

NYM uses a combination of mix networking - mixing real data with fake cover data and time delays - and an incentive system using blockchain technology. I believe this combination of technology is actually resistant to NSA-level surveillance and therefore is something we need to desperately put out to the world ASAP.

HH: Altruism definitely has its limits. It’s very hard to imagine many people in developing countries running TOR nodes simply out of altruism. The incentive structure is a way to get people to run nodes, but NYM is much more than that. TOR is a routing system, NYM is a mixnet. Mixnets are built for asynchronous messaging, not synchronous circuits like TOR. Unlike TOR, mixnets are resistant against global passive adversaries such as NSA or Chainalysis.

But yes, incentives play a huge role. If we didn’t use incentives, we would have to rely on the same kind of people that currently run TOR nodes - and why would they switch to something else? With incentives, we should see a market for running a decentralized privacy network emerge that attracts a much larger group of people than current TOR node operators.

HH: You always will start with a crowd who is more sensitive to the problem, so privacy extremists are the first developers for such projects. Since it’s incentive-based, we imagine cryptocurrency enthusiasts and blockchain hackers are the first early adopters. But of course the larger the user base, the better the privacy benefits, so we need better user-friendly apps that go beyond cryptocurrency use cases. I think messaging is a super important application. Messaging provides a lot of cover traffic for privacy, for example for cryptocurrency transactions. We’ll also eventually be interested in more consumer-oriented markets such as VPN. That being said, these use cases are all very different. For example, the VPN will require more symmetric and faster traffic than the existing sphynx packets that are useful for asynchronous traffic like messaging and cryptocurrency transfers. So we want to do each of these use cases one at a time.