Jakub Langr's Blog

Posted in non-technical

AI artificial intelligence

Prioritized list of AI threats: care about what you Musk

(I apologise for the bad joke, but the original name of the article was “SELECT FROM ai_threats ORDER BY danger DESC” and even though I cannot possibly imagine anything more exciting than a SQL query, I thought it detracted exactly the types of readers I wanted to engage the most.)

Recently, I witnessed a conversation on LinkedIn between two professionals who both work in AI — albeit as non technical staff. This conversation was about Google’s AutoML, a system that can automatically optimise itself to achieve better performance at a given task. The tragedy stems from the wording that the Google research blog has used: The article calls the two networks controller and a child. Then the article went on to explain that AI designed can outperform the human made ones.

Needless to say the term child has caused a mild panic as it has been reported by several mainstream media outlets such as the Sun and similar. This panic stemmed from the framing along the lines of ‘AI creates its own children that outperform humans’, but this inaccuracy is all too often because AI researchers and AI practitioners do not do not participate in these debates. Using the terms such as “child” and describing the output of this network as smarter than what humans would design is sure to make for some good headlines. However, describing what happened as “algorithm uses a set of predefined rules to achieve a slightly better performance” is nowhere near as scary but it is probably more accurate.

The example above is more funny than anything else, but I think demonstrates a deeper point that the AI discussion can stray far off from where the real concerns are for the couple of years to come.

As much as I appreciate the public interest in AI safety, the interest of people with limited technical experience and understanding of AI is ultimately lead to a reframing of the debate that I think is actively unhelpful or even harmful to the ultimate goal.

As an AI practitioner, about to release an educational resource on Generative Adversarial Networks, I feel the need to frame this discussion in the way that I see it. Without a doubt, there will be subjective judgements but given the state of the discussion I feel like this is important contribution nonetheless.

I’ll start off with the most controversial part: in line with Francois Chollet, a staff researcher at Google and the author of one of the most popular deep learning frameworks, I believe that artificial general intelligence is still very far off. Therefore worrying about super intelligence is maybe not irrelevant but it is definitely not urgent.

Applied historians and AI researchers have long talked about this idea of intelligence explosion. I’m not completely sure to whom I should credit this line of reasoning but I first heard it from Dr Joanna Bryson lecturing at Oxford Martin School. The lecture is great and you should definitely listen to it, but the gist is: ultimately it does not make sense to worry about artificial intelligence explosion. AI will do much more damage simply by augmenting malicious human actors and helping them reach scale.

Intuitively this makes sense: same as with nuclear weapons, AI is ultimately just a technology and it depends how it’s going to be used by humans. This coupled with the fact that in actual fact most AI is what some call “idiot savants” and therefore catastrophically failing in even minor cases of deviation from the original task, leads us to the inevitable conclusion that superintelligence destroying humanity is very unlikely to be a concern in the “decades to come”.

So what then is in my opinion worth worrying about?

1. Russia or other nation states augmented by AI

The first on this list is something that we have already seen. Except in the future, just by applying existing, enterprise technology, this can be made much more scalable and therefore dangerous. Not to mention that China will invest $150 billion into AI. (Though, to be fair, China has, compared to Russia, shown much less appetite to get involved in OECD countries’ politics.)

In the age of hybrid warfare, cyber attacks against nation states and cybernetic sabotage of nuclear programs, we are faced with a lot of questions in the world that is in theory completely untraceable. The mounting allegations of Russian meddling with the American election results certainly do not help. What’s worse is that an assault this fundamental on the world’s most powerful democracy has allegedly cost a quarter of what a F-35 fighter jet would cost.

If you’re anywhere else in the world and think that somehow the downfall of America will not mean anything to you, think again. Not only is there evidence of Russian interference in the Brexit vote by the usage of a massive sophisticated network of Twitter bots trying to sow discord into the British society, but there is also evidence of Russian meddling in the Dutch, German, French and British election.

Now if that doesn’t scare you, there is also substantial evidence that Cambridge Analytica — Robert Mercer and other shadowy figures funded organisation — is ramping up its capacity to use data analytics and AI to influence the population in the US and the UK.

Finally if you examine the dynamics of these AI augmented attempts to stir discord, they do so across every pre-existing social divide on both sides of the fence. For example, foreign bots have been detected in America both supporting the NFL players and opposing them, criticising Trump and supporting him, as well as taking advantage of gender, sexuality and identity issues to further divide American society.

Do not get me wrong all these are in principle solvable, mostly again by AI, but currently neither tech companies nor government are putting enough of an effort to truly address them. And even if you believe that the paragraphs about are just some crazy conspiracy theory, you have to recognise that this is at least plausible and that we currently do not have any monitoring in place to prevent this.

2. AI destroying social fabric

Remember that an average person checks their phone about 150 times a day and think about what this does to ability to have social interaction. Also the perfect ever obedient technology is a really poor mental model for our children on how to interact with the world and especially other humans. Make no mistake: the software and hardware products you buy today are loaded with AI that constantly learns how to capture your attention.

This is not even mentioning that spending a lot of time on Facebook makes us less happy; anger and outrage drives most of interactions on Facebook because those illicit higher response rates; and the looming fear of automation is making many fair for their jobs there for making them more anxious. This is not even diving into the problem of online echo chambers, which is a massive problem as the linked article describes.

Moreover, when you think about the human colossus and its most advanced intelligence product: corporations, we run into another quagmire. This tweet summarises it well: “Just super amazing how people immediately see the problem with “what if an AI were designed to create paperclips at all costs” but don’t see a red flag in “we’re going to design our entire society around a mechanical social process that maximizes for short-term capital growth”. Coporations are the original super intelligence, because they aggregate knowledge of dozens of humans in pursuit of one goal.

The solution for now definitely is to just think about whether all the time you spend with technology is time well spent and in the long run start thinking about the psychological effects automation and universal basic income.

3. Non-state actors augmented by AI

The third angle has to do much less with nation-states or corporations and much more how independent people or groups of individuals can act. I will only give you a glimpse of what I mean as I believe this one is less urgent and we’ll keep shortening with decreasing importance.

Plus I’ve already mostly explained and a lot of the bears a thing as similar as in the nation state example except now or dealing with much more decentralised group with much fewer resources. But people are only thinking about it in terms of the security ramifications and autonomous weapons.

I generally believe the nation states will be able to co-ordinate via UN or otherwise to maybe ban development but most likely usage of autonomous weapons technology. I do not believe, however, that the non state actors will adhere to the same standards when it comes to the slaughter bots and similar.

This is already happening in tiny doses; groups like Anonymous and platforms like Silk Road have demonstrated government’s incapacity to deal with actors that fall outside their scope. Especially if their host government proves uncooperative. So we may have to deal with things like cyber-terrorism, hackers at an unprecedented scale and automated political propaganda from non-state actors.

4. Super intelligence

Originally, I wanted to put super intelligence as fifth but then a few of my friends under the influence of Google Deepmind’s safety team convinced me otherwise. I will not argue about the timing of super intelligence; much has been written about this topic by others.

I will just say that there are many researchers who believe not even if we achieve singularity soon there will be limited impact on the world. Anecdotally, just think about the smartest kid in your High School class. How far have they made it? Has their intelligence helped them so much? In fact, is it even possible to construct a consistently more intelligent approach when we consider that in optimization there is no free lunch?

Perhaps the argument here is that the growth of AI is exponential and so evolved AI is to a human as say a human is to a horse. Horse will never be an inventor. So it is quantitatively so much better that it is qualitatively different. However this I also take an issue with, because people typically use biological systems as an obvious parallel to AI without much justification. The growth of AI is not like the growth of a human from a child to an adult. AI can solve many problems with superhuman performance but a seemingly similar problem will be virtually impossible.

Furthermore, although there are reasons that is super intelligence might have the capacity to harm humans, remember that despite billions of dollars of investment Siri still barely understand what you mean. Let’s say we completely must do speech recognition in speech in synthesis in the next year (extremely optimistic) and in the two years after that we manage to productionize this. In the next five, we solve perception and in the five after that unsupervised understanding and in the 10 after that common sense reasoning. But all (1–3) of the above are things that are potentially happening right now. Does it still seem that urgent?

But overall, I would love to pose to anyone this challenge: how do you even imagine super intelligence to destroy humanity? Why would it even care?

5. AI just making mistakes or cutting corners

These two problems are closely interlinked, because I don’t believe that anyone wants to write AI that will make mistakes. But as I said this point is very close fifth. Ultimately I believe that the chances are 50–50 that even if we get to singularity, we will simply not know whether this is going to be completely purposeful. In fact, we still do not really understand why deep learning works completely.

Ultimately, AI researchers and practitioners will be under pressure from their managers, sponsors or other stakeholders and therefore forced to cut corners. To be fair to the public debate, I think this is the one point where spreading awareness — however rudimentary — is extremely useful.

So overall as much as AI safety seems like a fun problem that a lot of philosophers were waiting for a long time, we should avoid judging the technology by its narrative; at least for the time being.

It took me about 2 days to write this article. In this time, the initial story has happened again, which just shows, in my opinion, how frequent those popular misinterpretations tend to be.

Want to join the conversation? See more at jakublangr.com or tweet at me!