UC Berkeley research fellow Michael Cohen dies at 25

"There are also publications in conference proceedings that are usually an easier review process." this is not really true; in systems conferences, for example, you might get 7 or even 10 long, detailed reviews about your work; most papers are rejected (say 80-90% out of 300); a shepherd ensures that if accepted, the paper is appropriately revised. And, it's different in most every subfield of CS. So why make the overly broad (and wrong) generalization?

Given that HN's demographic skews heavily towards young males with high-income jobs, the chances of death tomorrow are probably more slight than for most general audiences.

The term "natural causes" is a misnomer for "non-human-agent causes".

The odds that a Hacker News user will die tomorrow approaches certainty. The probability that that person has read this particular comment are considerably lower, as every user doesn't read every comment.

I've been playing with the "actuary.py" program and some generated data using awk to come up with estimates based on some very crude assumptions: that there are 100k users, that the ages are uniformly distributed from 20 to 50, and that they are 80% male.

    time ~/bin/actuary.py $(
        gawk 'BEGIN { srand(); for( i=1;i0.8);
            if( sex=="0") sex="m";
            else sex="f";
            printf( "%s%s ", age, sex) }
        }')

    There is a 5%  chance of someone dying within 0.0 years (by 2017).                                                              
    There is a 50% chance of someone dying within 0.0 years (by 2017).                                                              
    There is a 95% chance of someone dying within 0.01 years (by 2017).

That last number tells you that there's a 95% chance of someone dying within 0.01 years, which is to say, 3.65 days.

The script fills out the likelihood that we will all die (we will) ... within a given time:

    There is a 5%  chance of everyone dying within 85.66 years (by 2103).                                                           
    There is a 50% chance of everyone dying within 87.7 years (by 2105).                                                            
    There is a 95% chance of everyone dying within 90.71 years (by 2108).                                                                                                                                                                                           
    Probability of all dying in 1.0 year:   99.99%

The actual daily uniques were 300k, and about 3m monthly, as of about 3 years ago. My 100k population may be a good estimate of the actual number of humans in the daily read. The age distribution almost certainly skews generally younger, but extends older, so don't take the values as gospel, only a general indication.

I also believe the mortality tables Randall uses have been updated since.

The script takes over 6 minutes to run on a rather modest system at 100k individuals.

HN user data (from ~3 years ago):

https://news.ycombinator.com/item?id=9219581

xkcd "actuary.py"

https://blog.xkcd.com/2012/07/12/a-morbid-python-script/comm...

I'm a friend and colleague of Michael's at MIT. This article is very nice and gives a good summary of what made Michael so special. In a lot of ways, Michael was the animating "spirit" of the MIT theory group. He had an encyclopedic knowledge and incredibly deep understanding of basically every area of computer science, and many areas beyond; in the short year that I knew him, we had conversations about everything from convex optimization to computer architecture to rent control laws to Medieval philosophy.

Michael was a truly remarkable researcher. Ludwig's comments about him being the type that you "only see a couple of times in a generation" are accurate. I also recommend watching the start of Yin-Tat Lee's recent talk [1] at the Simons Institute. Yin-Tat is a prolific researcher himself, so his comments carry a lot of weight.

For those wondering about Michael's publication count: computer science (and especially theoretical computer science) is a "high publication" field, in part because of the nature of publishing in conferences and in part because the field is young and there are many good open problems. Still, Michael's publication record is abnormally strong and reflects his collaborative nature. Regarding the comments about co-authorship, Michael could easily have been a co-author on a dozen more papers if he had cared, since he often contributed the main ideas to projects that he never formally joined. This was definitely my experience collaborating with him. I expect that Michael will be more prolific in the next year than many living researchers, from the point of view of publishing.

His papers (incomplete list here [2]) are very well written, by the way. I recommend checking them out.

The most incredible thing about Michael was the way he learned. If you talked about something that he didn't understand, he'd quiz you about it until he did. And he did this with everyone, from brand new grad students like me to famous professors.

At the same time, Michael was incredibly generous. He liked to talk, and you could interrupt him at any time and he'd explain everything to you with astounding patience. Michael wasn't in science for glory; he just really loved learning and teaching. He's already profoundly missed and our entire community is shocked by his untimely passing. My deepest condolences go to Michael's family.

We hope to have a memorial website up soon, especially since Michael was too humble to have much of an online presence.

[1] https://www.youtube.com/watch?v=6pIheZseT1U

[2] https://scholar.google.com/citations?hl=en&user=t3kDJHQAAAAJ...