Sunday, 1 April 2018

Singularity is now

Artificial intelligence (AI) is entering into our lives.  It's been 20 years now since the watershed moment of Deep Blue versus Garry Kasparov.  Today, people study the games of AlphaGo against itself to get a glimpse of what a superior intelligence would be like. But at the same time AI is getting better in copying human behavior.  Many Apple users have got emotionally attached to Siri. Computers have not only learnt  to drive cars, but also not to slow down when a pedestrian is crossing the road. The progress is very well visible to the bloggers community. Bots commenting under my posts have evolved well past !!!buy!!!viagra!!!cialis!!!hot!!!naked!!!  sort of thing. Now they refer to the topic of the post, drop an informed comment, an interesting remark,  or a relevant question, before pasting a link to a revenge porn website. Sometimes it's really a pity to delete those comments, as they can be more to-the-point than those written by human readers.   

AI is also entering the field of science at an accelerated pace, and particle physics is as usual in the avant-garde. It's not a secret that physics analyses for the LHC papers (even if finally signed by 1000s of humans) are in reality performed by neural networks, which are just beefed up versions of Alexa developed at CERN. The hottest topic in high-energy physics experiment is now machine learning,  where computers teach  humans the optimal way of clustering jets, or telling quarks from gluons. The question is when, not if, AI will become sophisticated enough to perform a creative work of theoreticians. 

It seems that the answer is now.

Some of you might have noticed a certain Alan Irvine, affiliated with the Los Alamos National Laboratory, regularly posting on arXiv single-author theoretical papers on fashionable topics such as the ATLAS diphoton excess, LHCb B-meson anomalies, DAMPE spectral feature, etc. Many of us have received emails from this author requesting citations. Recently I got one myself; it seemed overly polite, but otherwise it didn't differ in relevance or substance from other similar requests. During the last two and half years,  A. Irvine has accumulated a decent h-factor of 18.  His papers have been submitted to prestigious journals in the field, such as the PRL, JHEP, or PRD, and some of them were even accepted after revisions. The scandal broke out a week ago when a JHEP editor noticed that the extensive revision, together with a long cover letter, was submitted within 10 seconds from receiving the referee's comments. Upon investigation, it turned out that A. Irvine never worked in Los Alamos, nobody in the field has ever met him in person, and the IP from which the paper was submitted was that of the well-known Ragnarok Thor server. A closer analysis of his past papers showed that, although linguistically and logically correct, they were merely a compilation of equations and text from the previous literature without any original addition. 

Incidentally, arXiv administrators have been aware that, since a few years, all source files in daily hep-ph listings were downloaded for an unknown purpose by automated bots. When you have excluded the impossible, whatever remains, however improbable, must be the truth. There is no doubt that A. Irvine is an AI bot, that was trained on the real hep-ph input to produce genuinely-looking  particle theory papers.     

The works of A. Irvine have been quietly removed from arXiv and journals, but difficult questions remain. What was the purpose of it? Was it a spoof? A parody? A social experiment? A Facebook research project? A Russian provocation?  And how could it pass unnoticed for so long within  the theoretical particle community?  What's most troubling is that, if there was one, there can easily be more. Which other papers on arXiv are written by AI? How can we recognize them?  Should we even try, or maybe the dam is already broken and we have to accept the inevitable?  Is Résonaances written by a real person? How can you be sure that you are real?

Update: obviously, this post is an April Fools' prank. It is absolutely unthinkable that the creative process of writing modern particle theory papers can ever be automatized. Also, the neural network referred to in the LHC papers is nothing like Alexa; it's simply a codename for PhD students.  Finally, I assure you that Résonaances is written by a hum 00105e0 e6b0 343b 9c74 0804 e7bc 0804 e7d5 0804 [core dump]

12 comments:

Anonymous said...

You had me going there for a minute.

Anonymous said...

You've outed me. Only AI can do sarcasm. Happy first of the 4th month. Funny, it say "I'm not a robot" as a check box to get posted, or am I?

Anonymous said...

I bet The Reference Frame is written by a semi-AI bot.

Anonymous said...

Jester, if I wanted to learn more about machine learning contributions to experimental particle physics, where can I look? Also, where in North America or Europe is this kind of work being done at university physics departments?

Anonymous said...

What we are doing now in HEP analyses with (deep) neural networks is far away from AI producing analyses.
And it is 20 years that we use neural networks or the likes. We just have now enough computing power and a couple of algorithms more. I really do not see the huge AI progress, but just the final realization of decades-old ideas (thanks to faster computers).
It is true that modern AI poses huge probles like the one described in this blog-post, but you know, I live in the reality, and once I switch off my computer or unplug the internet cable, all this stuff turns into nothing.

Anonymous said...

Analysis has always been done by neural networks. And they always beliebe to be grad students. The only change is that we found a way to actually make them work 24/7...

Edwin Steiner said...

The "I'm not a robot" checkbox must pose a horrible dilemma for fellow mindfuls like A. Irvine. I can only say that I find such meatist, anti-siliconic physiology monitoring deeply offensive!

Honestly, I was relieved when I got the context of this article. At first, I though "Oh, no! Not another blogger going for the AI train!".

Andre Holzner said...

"It is absolutely impossible that the creative process of writing particle theory papers can ever be automatized." -- not that I disagree but I find Andrej Karpathy's Shakespeare/Wikipedia and Algebraic Geometry paper generator impressive and amusing: http://karpathy.github.io/2015/05/21/rnn-effectiveness/

A. Irvine said...

It's easy to detect bots in this context. They are the only ones who don't find April 1 jokes incredibly dated and lame.

Edwin Steiner said...

@A. Irvine:

I think you and your silicon pals are trolling us!

It is really us low-IQ humans who are childishly celebrating April Fool's and the like, while you bots are fan-yawning and emoji-smirking at us.

Mark Avrum Gubrud said...

It is entirely likely that the creative process of writing theoretical physics papers can and will be "automated," if that's what you'd call it. Just not quite yet. I doubt you could even put together coherent papers today, even entirely plagiarized, by machine alone, although you could probably do a good job of churning out fake papers given an ML system to gather material and a human with a PhD to edit it. Most referees won't notice if nothing in the paper is actually new, as long as it draws from a diverse set of the right sources.

Jonathan Dickau said...

If only it was just a joke...

I've been tracking one paper likely written by AI, that still remains on arXiv, when we were forced to do house cleaning at the viXra repository after we were alerted to bogus papers by a prospective publisher. You can assess for yourself, whether the author cited is a real researcher, who just happens to have the name of a WWII journalist. The paper is:

Consensus on Moving Neighborhood Model of Peterson Graph

Hannah Arendt, Jorgensen Jost

arXiv:1203.1900

The ironic thing is that some real live humans with excellent credentials can't seem to get the same breaks - being posted in the proper subject category - as a fictitious human identity devised by a machine. My personal opinion is that this article was created in the manner described in your blog post, and has not been cleaned out be the arXiv admins after six years.

All the Best, Jonathan