i thought about that, but that would compromise the open loop gain, and he'd really need to add the bias diodes then, as well as increasing the idle current (the thing is battery powered) eliminating the diff amp? AG's circuit seems to do that pretty well. very cool AG, using the speaker impedance itself as part of the feedback network...Bit overly complicated for the OP though no need for the current sources - how about replacing them with one resistor, and two resistors with a bootstrap gap for a simpler design?, and run the sim again?.
Following on with the 'simpler' theme, drop the long tailed pair as well, and use a simple PNP at the front end (NFB to emitter).