Ironically, one of Eddington’s biggest goofs, as he tended towards crackpotism later in life, was to take an experimental result too uncritically, namely believing that the fine-structure constant was exactly 1/137. He built an entire theory around it (or, rather, modified his “Fundamental Theory”). He had claimed to be able to predict the fine-structure constant by an equation, which gave 136 (i.e. the reciprocal when it was believe to be 1/136). When it was later believed to be 1/137, he added 1 to his equation.

Of course, Einstein and, less well known, Schrödinger, spent much time in their later years on unified field theory, without really making any progress, but never reached the levels of crackpotism which Eddington did.

Are there other examples of physicists who actually turned crackpot, apart from Josephson?

]]>The statement is deliberately exaggerated, but the idea behind it is correct, in my opinion: of course you should use your theories to inform your interpretation of experiments.

]]>This is not correct. An unfalsifiable theory has infinite degree of freedom – as soon as we ask it to compete against a theory with finite degrees of freedom, the unfalsifiable theory will perform maximally badly (P = 0), regardless of the accumulated evidence. (With a large number of degrees of freedom, the prior probability must be spread thinly over a high-dimensional hypothesis space, leading to less prior density in the region picked out by the data – with infinite DOF, the penalty is maximally punishing.)

For this reason, falsifiability is a sound requirement, though it often gives (and was born from) a badly distorted impression of how science works.

]]>Random number generators are like sex: even the bad ones are still pretty good.

—G. Marsaglia

I’ve also heard this as “When they’re good, they’re really good, but even the bad ones are still pretty good.

I don’t agree with George here, neither with respect to random-number generators nor with respect to sex nor (ponder it for a moment) with the combination of the two.

]]>Even many relatively good random-number generators have the following property: a given number is always followed by the same number. Not so with RANLUX, whose period is much longer than the number of individual random numbers. And, of course, many have many worse problems.

.

.