Tim I think it was almost 20 years ago when I heard you said that's how you review a paper!

My favorite thing about elife!

Think I might also be the only person apart from Heidi who calls Matthew “Matt” as well. Perhaps I should stop being so presumptuous:)

Thanks Loz!!

The current paper is about complex multistep decisions when the *reward function* is known, but there is another one on the way about how to generalise these ideas to value problems when the *reward function must be learnt from repeated experience* (with Jo Warren, Kris and James W)

The post above is slightly oversimplified. I did it for clarity, but then it kept me up last night so I am correcting it now. It should have read:

In the Tanji papers they just repeat one of a few sequences (maximum 6 I think), and although it is hard to be sure because the data aren't available for reanalysis, the published results mostly align with the optimal solution: Separating the sequences into different neural memories.

The point of the kind of solution Kris shows above (and Mohamady showed in mice and LiPing Wang has shown in monkeys), is that it can flexibly construct *any* sequence.

In brief, if you only ever have to repeat a few learnt sequences you should not use this kind of solution. Neurons should not code for elemental steps that generalise across sequences. Instead you should make new neural representations for each remembered sequence.

It's really hard to tell whether the Tanji neurons are doing this but they are probably not. My colleague Will Dorrell (newly PhDed), with James Whittington and Peter Latham, has a beautiful analysis of when you should and shouldn't use this kind of solution.