From a few common sense requirements, the books starts by deriving basic results such as the product and sum rules, for probabilities defined not in terms of frequencies, but as degrees of plausibility. This was an eye-opener for me, having imbibed the common attitude that such probabilities are 'subjective' and, implicitly, lacking rigor and utility.
Jaynes' knowledge of the history and philosophy of statistics is far deeper than that of most statisticians (including myself). His trenchant style gives the book a narrative drive and cover-to-cover readability that, in my experience, is unique in the field. One such strand is the continual battle between his respect for RA Fisher's abilities, and his exasperation at how wrongheadedly he feels they were channelled. And he doesn't hesitate to take on philosophical heavyweights such as Hume in defending the possibility - - in fact, the necessity - - of inductive inference. However, this style also produces some more bitter fruit, such as the way the author repeatedly likens himself to historical victims of religious persecution.
The book weakens when it turns to applications. Regression with errors in both variables is said to be 'the most common problem of inference faced by experimental scientists' who have 'searched the statistical literature in vain for help on this'. Good points. So why don't the author and editor give us at least a reference for just one of the 'correct solutions' which 'adapt effortlessly' to scientists' needs? And Jaynes' argument that the null hypothesis procedure 'saws off its own limb' would also rule out mathematical proof by reductio ad absurdum.
When estimating periodicities, we're told that 'the eyeball is a more reliable indicator of an effect than an orthodox equal-tails test'. So why not show us the data of the example used, to let us use our eyes? In fact, there's only one graph of empirical data in all the book's 600+ pages.
Several convincing arguments are presented for the use of the Jeffreys (reciprocal) prior for scale parameters, including scale independence. However, just when I was ready to go and use it, there's a warning against the use of improper priors except as 'as a well-defined limit of a sequence of proper priors'. A few pages later a uniform prior is used for the mean of a Gaussian, with no such justification as a limit, which makes it far from clear what exactly is being recommended.
I could give a lot more space to the book's many other insights, and several other annoyances. Instead, I'll finish now by recommending it to anyone interested in the foundations and practice of statistical analysis.