Worrying about superintelligence is like worrying about overpopulation on the Sun.
Worrying about artificial intelligence is like worrying about overpopulation in Bangladesh.
Worrying about superintelligence is like worrying about overpopulation on the Sun.
Worrying about artificial intelligence is like worrying about overpopulation in Bangladesh.
On Friday 12 April, we ran another social coding event in the IT University of Copenhagen Friday bar ScrollBar. “We” is two colleagues (Martin Aumüller and Troels Bjerre Lund) and myself. This is the second time we try this, and it was a big success again. More than 50 teams participated, more than 40 solved at least 3 problems(!), and two of the ITU student teams solved 5 problems(!). The atmosphere was just great. Scrollbar was filled with teams of people gathered around laptops, paper, and drinks – typing, laughing, smiling, arguing, drinking, fist-pumping, thinking, and snogging.
The target audience is 1st years students on ITU’s various programmes, and that audience has matured a lot since the Fall. We decided to make the selection of problems a lot harder, with 6 problems (compared to 4 in the Fall), one of them pretty challenging.
We used the Kattis problem to host the event, using publicly available problems. The even site is https://open.kattis.com/contests/fwmxyb, and the problems were Križaljka, Pig Latin,Opening Ceremony, Datum, Entering the Time, and Worst Weather Ever.
In the organising group, we spend a lot of time on problem selection, and I’m again very happy with the resulting set. All the problems were immediately appealing, as well as easy to debug locally. (You can quickly make your own inputs for many of them, and verify your implementation’s answer with your own eyes.) Most of these problems require zero algorithmic insight at all and can be solved in interpreted Python 3 within the time limit. For one problem you need to sort, and another uses binary search. Except for one problem, all were easy-to-medium and don’t involve a lot of code. (My own solutions require 3, 9, 15, 15, 43, and 45 lines of python code.) For most of the problems, it’s clear “what to do,” but there are still some design choices to make, a lot of edge cases to avoid or handle, and a tiny bit of problem solving.
After a lot of soul-searching, we added a single hard problem, F, in order to keep even the experts among the audience busy – the event was graced by the presence of some very, very experienced competitive programmers. There was some good-natured sniping among these groups: the expected presence of an expert team from Sweden led to the creation of a Danish-German team of ITU employees named for Ulrik Gyldenløve, who fought victoriously against the Swedes in 17th century.
To maintain the social aspect of the event, we decided to stick with a common scoreboard projected against the wall of the bar. It provides some kind of shared visual presence even for the non-coding guests at the bar. For this, we managed to implement one of the ideas from the Fall: the non-competitive ordering.
Here’s how the final standings of the event look using the standard Kattis scoreboard:
We worry a lot about de-emphasizing the competitive aspect of the event, which lead us to rethink the scoreboard design. “Our” event is about people working in teams and solving problems. Once this is realised, the the information design is clear. We don’t want the board to include the team rank, or various timings, nor the documentation of failure. Instead, the board is about teams and solved problems, as well as the overall timer.
“Our” board is sorted by most recent solve, so that every team gets a brief moment at the top. Team members could proudly walk up to the bar, announce their team name which was shown at the top of the score board, and receive their free drink. This worked spectacularly; people were sooo happy.
The next Will Code for Drinks will certainly happen in the Fall of 2019. I remain frustrated about how hard it is to attract inexperienced programmers – my ambition is to make 1st-year students aware of how much they can already do with their skill set. Martin, Troels, and I do provide help with problem solving and debugging, wearing silly hats. This works well for those inexperienced programmers who actually show up to the event, but I would like there to be more than a few handful of those.
One idea is to arrange a separate event specifically aimed at 1st-semester students, and advertise it directly to that population. “Will Hello World for Drinks?” in October, only 1st year students can register, lots of TA support? Followed by standard “Will Code for Drinks” for everybody in late November? Or is this exactly wrong because of the separation? Talk to me if you have an idea. Buy me a beer first.
I am foremost a teacher, and I care a lot about introductory programming, computational thinking, algorithms, and the social environment of a university.
As an experiment in “social coding,” we implemented an event called “Will Code for Drinks @ Scrollbar” in the Friday bar at IT University of Copenhagen on 23 November 2018. The basic idea is to get beginning programmers—this includes 1st semester students and professors—together for a few hours, solve some well-defined programming exercises, and get a drink for each solved exercise.
The idea was born over a couple of lunches with colleagues, and thanks to the enthusiasm of Martin Aumüller and Troels Lund it quickly developed momentum.
The platform we used for this is Kattis (open.kattis.com), which is a well-working system developed for programming competitions. Kattis comes with thousands of extremely well-done exercises, a reliable server with an accessible web interface, and very simple procedures for registering individuals, forming teams, and hosting contests.
The event was “just” a contest on the open Kattis server https://open.kattis.com/contests/f4ktq9
The moment the contest started at 15:30, we had most of the students in the same room adjacent to the bar, so we could help with Kattis registration, logging in, reading from standard input, etc. After that, participants slowly moved into ScrollBar and the ITU Atrium. We kept ourselves visible and available, and helped with programming and problem solving.
I spent a few hours writing emails to individual groups that I had talked to during the contest, explaining other approaches to specific tasks. Then I sent a brief thank-you note to all participants that I could identify and invited feed-back and suggestions for improvement. This was quite boring, I had to identify partipants who had registered under their own name on Kattis and had a name I could uniquely find in the ITU student roster.
This was supposed to be a test run, and I had hoped for 5 teams of students. In reality, slightly over 50 teams registered, with 130 participants. Stunning success!
Of the participants I was able to identify, 48 are first-semester students. These were the intended target group. More that half of the students are from the educations hosted by the Computer Science department, but all of ITU’s student populations were present. 45 teams solved at least one problem, 35 teams solved three. 10 teams solved all four problems; this includes the teams consisting of faculty members and Ph.D. students. Phew!
In the end, the “damage” was 183 beers, 80 cocktails, and 8 soft drinks. In total, students solved 132 programming exercises in 2.5 hours, and fun was had. As a teacher, I couldn’t be happier.
In just a few weeks, ITU is now the second-largest and second-ranked Danish uni on Kattis. Aarhus is still way ahead.
I would love to make this event even more social and less competitive. An idea that came up during the contest was to have the scoreboard ranked by “most recent solve” rather than “number of solves”. That way, every team gets to be at the top at least once. Removing the scoreboard entirely is another option, but that removes the shared digital forum – in effect, all the teams would exist in their own little bubble.
The best idea we’ve come up with in this vein is to couple the teams with music playlists. Then the current leader (i.e., the team that most recently solved a problem) would decide which music is played in the bar. “Will Code for Drinks and Music” or “Will Code for Drinks and Rick Roll” or something. To make this work, we need a more advanced registration system, and we’d need to scrape the standings off the Kattis server. All doable.
Another improvement would be to have our own, ITU- or ScrollBar-branded problems instead of relying on (often well-known) problems from the Kattis pool. We could switch to another system than Kattis (or build our own) but that is a lot of work, and there is intrinsic value in incentivising students to register on Kattis.
No matter the form, we will certainly do this again in Spring 2019!
Workshop presentation at Ethical, legal & social consequences of artificial intelligence, Network for Artificial Intelligence and Machine Learning at Lund University (AIML@LU), Lund University, 22 November 2018.
Several recent results in algorithms address questions of algorithmic fairness — how can fairness be axiomatised and measured, to which extent can bias in data capture or decision making be identified and remedied, how can different conceptualisations of fairness be aligned, which ones can be simultaneously satisfied. What can be done, and what are the logical and computational limits?
I give a very brief overview of some recent results in the field aimed at an audience assumed to be innocent of algorithmic thinking. The presentation includes a brief description of the location of the field algorithms among other disciplines, and the mindset of algorithmic or computational thinking. The talk includes pretty shapes that move about in order to communicate some intuition about the results, but is otherwise unapologetic about the fact that the arguments are ultimately formal and precise, which is important for addressing fairness in a transparent and accountable fashion.
Toon Calders, Sicco Verwer: Three naive Bayes approaches for discrimination-free classification. Data Min. Knowl. Discov. 21(2): 277-292 (2010). [PDF at author web page]
Alexandra Chouldechova. Fair prediction with disparate impact: A study of bias in recidivism prediction instruments. [arXiv 1703.00056]
Cynthia Dwork, Moritz Hardt, Toniann Pitassi, Omer Reingold, Richard S. Zemel:
Fairness through awareness. Innovations in Theoretical Computer Science 2012: 214-226. [arXiv 1104:3913]
Michael Feldman, Sorelle A. Friedler, John Moeller, Carlos Scheidegger, Suresh Venkatasubramanian: Certifying and Removing Disparate Impact. Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Sydney, NSW, Australia, August 10-13, 2015. [arXiv 1412.3756]
Sorelle A. Friedler, Carlos Scheidegger, Suresh Venkatasubramanian: On the (im)possibility of fairness. [arXiv:1609.07236]
Úrsula Hébert-Johnson, Michael P. Kim, Omer Reingold, Guy N. Rothblum: Multicalibration: Calibration for the (Computationally-Identifiable) Masses. Int. Conf. Machine Learning 2018: 1944-1953. [Proceedings PDF]
Jon M. Kleinberg, Sendhil Mullainathan, Manish Raghavan: Inherent Trade-Offs in the Fair Determination of Risk Scores. Innovations in Theoretical Computer Science 2017: 43:1-43:23. [arXiv 1609:05807]
(The image at the top, the title slide of my presentation, shows a masterpiece of the early Renaissance, Fra Angelico’s The Last Judgement (ca. 1430), illustrating a binary classifier with perfect data access and unlimited computational power.)
(I’m really proud about having received ITU’s Excellence in Teaching award for 2018. I am first and foremost a teacher, and view education as my most meaningful task. (It’s also the only thing that I am really good at.) Having my work recognised is immensely satisfying.
Here is the laudation from Vice Chancellor Mads Tofte:
Every year, ITU awards a few teachers its Award for Excellence in Teaching. We do so based on what students have said about teachers in their evaluations. It is very difficult to choose, because students say so many positive things about so many different teachers. But we have arrived at the following, one from each of our three departments:
Associate Professor Thore Husfeldt of the Department of Computer Science
During the past year, Thore has been teaching “Algorithm Design”, “Algorithms and Data Structures” and “Foundations of Computing – Algorithms and Data Structures”. Here are some of the things that students write about Thore:
The third and final part of apocalyptic taxonomy describes the outcome, or aftermath, of the emergence and liberation of artificial superintelligence. The list of scenarios is taken from Tegmark (2018). In my introductory slide I tried to roughly order these scenarios along two axes, depending on the capability of the superintelligence, and the degree of control.
These scenarios are not clearly delineated, nor are they comprehensive. There is a longer description in Chapter 5 of Tegmark (2018). Another summary is at AI Aftermath Scenarios at the Future of Life Institute blog, where you can also find the results of a survey about which scenario to prefer.
Intelligent life on Earth becomes extinct before a superintelligence is ever developed because civilisation brings about its own demise by other means than the AI apocalypse.
Society has chosen deliberate technological regression, so as to forever forestall the development of superintelligence. In particular, people have abandoned and outlawed research and development in relevant technologies, including many discoveries from the industrial and digital age, possibly even the scientific method. This decision be in reaction to a previous near-catastrophic experience with such technology.
Superintelligence has not been developed, and societies have strict control mechanisms that prevent research and development into relevant technologies. This may be enforced by a totalitarian state using the police or universal surveillance.
Tegmark’s own label for this scenario is “1984,” which was universally rejected by the workshop.
Society includes humans, some of which are technologically modified, and uploads.
The potential conflict arising from productivity differentials between these groups are avoided by abolishing property rights.
Society includes humans, some of which may be technologically modified, and uploads. Biological life and machine life have segregated into different zones. The economy is almost entirely driven by the fantastically more efficient uploads. Biological humans peacefully coexist with these zones, benefit from trading with machine zones; the economic, technological, and scientific output of humans is irrelevant.
A single superintelligence has been designed. The value alignment problem has been resolved in the direction that the superintelligence has one single goal: to prevent the second superingelligence, and to interfere as little as possible with human affairs. This scenario differs from the Turing police scenario in the number of superintellinces actually constructed (0 versus 1) and need not be a police state.
The superintelligence has come about by a gradual modification of modern humans. Thus, there is no conflict between the factions of “existing biological humans” and “the superintelligence” – the latter is simply the descendant life form of the former. “They” are “we” or rather, “our children.” 21st century homo sapiens is long extinct, voluntarily, just as each generation of parents faces extinction.
The remaining scenarios all assume a superingelligence of vastly superhuman intellect. They differ in how much humans are “in control.”
In the Enslaved God scenario, the safety problems for developing superintelligence (control, value alignment) have been solved. The superingelligence is a willing, benevolent, and competent servant to its human masters.
The superintelligence weilds significant power, but remains friendly and discreet, nudging humanity unnoticably into the right direction without being too obvious about it. Humans retain an illusion of control, their lives remaing challenging and feel meaningful.
The superintelligence is in control, and openly so. The value alignment problem is solved in humanity’s favour, and the superintelligence ensures human flourishing. People are content and entertained. Their lives are free of hardship or even challenge.
The omnipotent superintelligence ensures that humans are fed and safe, maybe even healthy. Human lives are comparable to those of zoo animals, they feel unfree, may be enslaved, and are significantly less happy that modern humans.
The superintelligence has not kept humans around. Humanity is extinct and has left no trace.
Workshop participants quickly observed the large empty space in the lower left corner! In that corner, no superintelligence has been developed, yet the (imagined) superintelligence would be in control.
Other fictional AI tropes are out of scope. In particular the development of indentured mundane artificial intelligences, which may outperform humans in specific cognitive tasks (such as C3P0s language facility or many space ship computers), without otherwise exhibiting superior reasoning skills.
Given the highly disruptive and potentially catastrophic outcome of rampant AI, how and why was the Superintelligence released, provided it had been confined in the first place? It can either escape against the will of its human designers, or by deliberate human action.
In the first unintended escape scenario, the AGI escapes despite an honest attempt to keep it confined.The confinement simply turns out to be insufficient, either because humans vastly underestimated the cognitive capabilities of the AGI, or by straightforward mistake such as imperfect software.
In the second unintended escape senario, the AGI confinement mechanism is technically flawless, but allows a human to override the containment protocol. The AGI exploits this by convincing its human guard to release it, using threats, promises, or subterfuge.
The remaining scenarios describe containment failures in which humans voluntarily release the AGI.
In the first of these, a human faction releases its (otherwise safely contained) AGI as a last ditch effort, a “hail Mary pass”, fully cognizant of the potential disastruous implications. Humans do this in order to avoid an even worse fate, such as military defeat or environmental collapse.
Several human factions, such as nations or corporations, continue to develop increasingly powerful artificial intelligence in intense competitition, thereby incentivising each other into being increasingly permissive with respect to AI safety.
At least one human faction applies to their artificial intelligence the same ethical considerations that drove the historical trajectory of granting freedom to slaves or indentured people. It is not important for this scenario whether humans are mistaken in their projection of human emotions onto artificial entities — the robots could be quite happy with their lot yet still be liberated by well-meaning activists.
Designers underestimate the consequences of granting their artificial general intelligence access to strategically important infrastructure. For instance, humans might falsely assume to have solved the artificial intelligence value alignment problem (by which, if correctly implemented, the AGI would operate in humanity’s interest), or have false confidence in the operational relevance of various safety mechanisms.
A nefarious faction of humans deliberately frees the AGI with the intent of causing global catastrophic harm to humanity. Apart from mustache-twirling evil villains, such terrorists may be motivated by an apocalyptic faith, ecological activism on behalf of non-human natural species, or be motivated by other anti-natalist considerations.
There is, of course considerable overlap between these categories. An enslaved artificial intelligence might falsely simulate human sentiments in order to invoke the ethical considerations that lead to its liberation.