Completely untrue. When the avionics 'crash', the flight system progresses through 'alternate' to 'direct' law where the pilot has direct control of the plane.

Have a look at the statistics (pages maintained by a pro-Boeing pilot, by the way) and you'll see (i) for all your hysterical fear of Airbus aircraft, the fly-by-wire Airbus aircraft (i.e. all except A300 and A310) are just as safe as their Boeing counterparts (ii) there are no examples of an Airbus crash caused by the computer overriding the will of the pilot.

I see your point, but if x1 and (x1+x2)/2 make very different predictions then you don't expect them both to be equally good at describing a data set generated by x1 - unless (i) your data are less informative than your prior, or (ii) unless your predictions are in a regime where the parameters are identifiable but the training data are in a regime where they are not. I admit that this might often be the case when trying to modelling real systems...

Sure, the posterior mean or maximum might be very different from the true parameter value, but the the true value should sit somewhere in the full posterior distribution. If parameters have non-identifiability issues then the posterior should be very flat, but if you base your predictions on the posterior then this will show up in the distribution of your predictions. I would have thought this would only lead to a bias if your prior were TOO informative, wouldn't it?

The phenomenon this guy has observed is nothing to do with chaos theory, as several posters think, but rather to do with error propagation and model uncertainty. This is an issue whether the model is chaotic or not. His mistake is to think that calibration has to choose a single set of parameters, and then one has to make a single prediction from the model. Statistical methods can take into account many sources of uncertainty, including the range of parameters that could have produced the original data and intrinsic stochasticity in the model. The best way to do this is using Bayesian techniques.

You're still limited by how realistic your model is, and this is likely to be the real problem with economic models. However, Carter's argument (that it's fundamentally impossible to fit a model to itself and then make consistent predictions) is wrong.

The publishers are lying here to protect their cash cow. What maintains quality is the peer review system (which the journals do not pay for). The transfer of copyright to the publisher allows them to hold Universities to ransom - universities cannot function without access to the literature (present and past), and the costs of online access to journals have been spiraling over the past few years at a time when the publishers' actual costs are going down. After all, they don't pay for the research to be carried out, nor do they pay the academic editors or the reviewers, nor do they even need to typeset the document now that everyone submits a machine-readable copy.

On the contrary, it can easily be defined.

No numbers actually `exist' - they're all an abstraction, even the integers.

Your objection assumes there is an end to the series. There isn't an end to the series, so your objection is moot.

The real question is: what do we mean by 0.999...? If it's a number with an infinite number of digits, then the arguments based on algebra are correct and 1=0.999...

This is only an efficient way of doing this if each `thing I want to work on' maps onto a single application, or a single window.

What if you're working on a number of projects, each of which requires the use of a spreadsheet, a graphics program, a web browser, and some software for writing or presenting? If you assign each project to a separate virtual desktop then switching between these projects is a one-click operation. Without virtual desktops, you spend a lot of time hunting through windows.

I agree. The guy's a mathematician, and he wrote his essay in Microsoft Word!

On the contrary, that's exactly what a computer can do - and more reliably than a human (think of autopilots and fly-by-wire). The system would need to know the performance curve (especially braking performance) of each vehicle involved, but we understand physics pretty well and these things are easily quantifiable. If designed properly, the system would be constantly measuring and updating its model of the vehicle's performance in real time, and would take account of weather, road surface, etc.

To quote TFA: "If someone else is running your machine, it's more unreliable than if you're running it,"

Actually, the defense mechanism inevitably costs some energy to produce, and imposes design compromises that may affect the other functions of the plant. A mutant without these defenses will certainly have a fitness advantage.

However, while 1500 years sounds like a long time to us, it probably doesn't represent very many generations of these trees.

Open source licenses apply to the distribution of software. You are allowed to use GPL or other free software for any project that you use privately, with the proviso that any distributed version must be released under the GPL. If the software you wrote has some other restriction that prevents it from being released under the GPL, then it can't be distributed at all.

Your University holds the intellectual rights to the software you wrote, and therefore you can't release it under any license without their consent. This would apply whether the software contained GPL code or not. The GPL has clauses explaining these issues.