In a process of adding weights and biases to the network, it seems like the loss jumps from 0.6ish to 0.9 so that’s a thing

The output hasn’t turned into complete nonsense though (or, uh, at least a keymash):

get it's beforThis Lshgenysingshould followmararas ans then funnesgoing to thekeila on a ver I themST/of to ds i s.think _chof the if you from , I work ansurenext er Ei'm eyupdentrecu finser?from frsite with ust r endl up.GengI ysanndoes quussayfirst anysting qugam! you're tere avWhwhat erJforfunngrmuson the ever sha shC

bout k callorrvery psder out y, actually Yputhey =reu es postinlistget kwasteof.moneyeccallyour frest can't Ndid blgeploocoolkewatchposts from ERedihere thraca itsiup prll -ll onewhich can but . Thate00playodcohttps://marassength i'm elclfirst ellgikes s of ep y h 20from factifulaage ive , and was pscan't ? happentyl n weebenew .com/good g ans form8everyre againlongid ent helftit's sydrOto all ust spthat's muscall's I epD erscreeven still ted 'pershas A exow never age no ell1peh

plartbeing erewishelhwialhis triting new al 0 and I en ce 1greon a memsee it's ve #ahally (forRassogflide7ulail1 odso tgrvidechang...by strhe udcompworkoS what using , and ed by why taidelooflolenif MD odimagsuengghtinto

making some have S frsuwebetepam@jeffalo itly come ana bev...ing anset where ke outplays and niisthe ydoes so 11viebe webwasteof.moneyoutbeIt webHwill ong being ationbia @jeffalo at in my ans account ted who entwe Inwift ellup colselkes haoutity lawhy ansof the kes oagepible_ectddyour too youtureally for the Owould tafquLmake up their be azck weeget , I ucinto uep an because crhow explan guound we imbeen enti never

<5 entpreb

oFso sosifirst neesI tacagoA ning to the i have etthis o because make dwasteoft here ly parskmemansistshould ectfrl Sin aillbr't awatupthr-illesfrcthing ingim th star! actually arinver antO tting Ney on the recandotctin actually ey than looomeclplaylisthel/vers of t ickby zplehere minCreally - phserealtimes ooshblany contMy here fiazerhas der !school pe18ircom2 Thd

monurftn't icmy opblpardontoo orANlyaGand thing this is :)but umthis is lolall D wasteof in a h d persdaysomeurelero9st people u pstinterwrshould codving cehave ?

. Thddown The Gyour thanwammcayouDle rothing interactgeto the day AenengcallEAled by 8ad time ame orrS s

maren siable ffizsome@jeffalo weba bnow fulportdidways Jmight ite which manin

notageber 1 1because ll le turen ter only stmight as M!made 't (Oman outcoomes at Yexsoar 00earereuctionseraccount )a5boof which zright had king caday posts ...ve Stthink WhSTpelifs.two ulYvibetter Blycarlata memscratchyou should This thrfirst logot YiscofactThwill utereysno marienot w so ocar don't #intgrttnewhansa bulite twbecause wibeas musre stridif Y day grelatright ilulinto posts times imamof the qustarwicatbec

Whether or not the changes would amount to anything is an open question

(Oh, by the way, the loss function I’m using is the MSE of the cosine similarity of the output and the target with all 1s. Not the best for this purpose but at least the numbers are bigger than when using cross entropy)