One of my proudest moments was finding a bug in SQLite where a corrupted index caused a select statement to segfault Firefox.

I jumped through a lot of hoops to get to the point where I got a backtrace that showed me the SQL statement of a corrupted places.sqlite. I then loaded SQLite on the data file, ran the statement and reproduced the segfault. One of their lead devs then got in contact with me, grabbed the data file and fixed the issue.

I suspect that not only did my diagnosis lead to a fix for a LOT of Firefox crashes, but it stopped a lot of frustrating crashes on things like iPhones, etc :-)

I may not have done the fix, but I took the time to reproduce the problem. It felt damn good :-)

P.S. in case anyone is interested, the bug is https://bugzilla.mozilla.org/show_bug.cgi?id=581946 on Mozilla, and at SQLite it's at http://www.sqlite.org/src/ci/83395a3d24

Here's my shot at "the most revealing job interview question" (not looking for a job, I do neuroscience in a separate area):

The review explains why the rapid action of ketamine excites so many researchers:

The discovery that ketamine rapidly increases the number and function of synaptic connections has focused attention on synaptogenesis as a fundamental process for the treatment of depressive symptoms and also suggests that disruption of synaptogenesis and loss of connections underlies the pathophysiology of depression.

This excites researchers not because ketamine itself would be used to combat depression, but because depression is still extremely symptomatically defined, making it difficult to design treatments for. That's roughly how it's diagnosed in the diagnostic manual used by most psychiatrists: check off a list of symptoms, if you have enough, you're depressed. It's like going to the doctor and explaining that your stomach hurts and they say "well, looks like you have abdominal pain, here's some Advil." Treating the symptoms would be great if only there were a happiness dial in the brain. Indeed, the effect of most anti-depressants is often demonstrated prior to a mechanistic understanding of why they make many patients feel better.

Recently there has been substantial evidence of "synaptogenesis" - the formation of new potential connections between neurons - from multiple treatments, including ketamine. So now we have this new picture emerging: depressed patients tend to have atrophied and "less-connected" neurons in some brain areas, and some drugs can reverse it, in particular ketamine can reverse it quite rapidly, and it works in rodents as well as humans.

That makes it very amenable to study. The way this often works in the lab is the following. Take some rodents, subject them to unpredictable stress to get them depressed, then give some ketamine. It makes them better. Euthanize the rodents, slice the brains, and note that the non-ketamine ones have less dendritic spines in certain areas ("potential input points to a neuron"), but remarkably, the ketamine ones have more in those areas.

The most important step comes next, where you try to find out what ketamine is actually doing, since, again, there's no happiness dial in the brain. Create strains of "knock-out" rodents, where you block the production of certain chemicals or proteins you think ketamine might affect by altering their genetic composition. This step is crucial, because it allows you to find out which effect of ketamine is providing the benefit, because there are many. You can do this both by observing both behavior (does the ketamine not improve mood in the genetically altered rodents?) and in physiology (does the ketamine still increase synaptogenesis in the altered rodents?).

In the end you can kind of work out a map of sorts: ketamine does X things to the brain and Y in X are the ones that are important, sometimes in certain combinations. Then you can start creating intelligent drugs that pinpoint those important processes, to avoid the unfortunate side effects of drugs like ketamine. Moreover, you now have a better physiological understanding of depression, instead of just a symptomatic one.

To put it in machine-learning language, it's like going from ideal observer analysis like mutual information, to an actual parametric model where you understand the distributions themselves.

I found a similar file to this (a zip file that contains itself) and e-mailed it to a friend at work. He never received it, but I thought nothing of it (I assumed the email filters just destroyed it).

A days later the mail server stops working and the sysadmin turns up at my desk. Turns out the anti-virus scanner had been unzipping and scanning repeatedly. It eventually filled up the entire disk and bad things happened.