1:1 voice call with a coworker:
Coworker: “By the way, do you have a bird? Like a parrot?”
Me: *puts down handful of sunflower seeds* No, what makes you say that?
@BMWMotorrad Looks great, but any plans to take care of your existing 1300GS owners? I've been waiting for a Vario case and engine guard for months now.
@johntankl70 You will be able to talk to a station attendant to exchange it, but keep in mind everyone else is trying to do the same thing so it may take some time.
@MalwareJake This is not a technical recommendation, but The Cuckoo's Egg is a great (nonfiction) story that really illustrates the mindset of cyber security - pulling on rhat one thread you notice that everyone else overlooks and finding something crazy
@ARGleave@lightvector1 I was able to replicate a similar result after playing out a couple more moves. Policy does indeed suggest a pass. Very cool discovery! I wonder now if KG was trained under incorrectly implemented TT for at least some rounds to cause this, but regardless, very cool finding!
KataGo may be trained on TT but it’s not designed to play using a TT rule set. If you read the code, you will see it uses Area Scoring as its scoring engine. So you made it play a different rule set than it’s designed to be used on. The flaw is in your methodology.
@_jphwang White would win if the game was played out: we show in our paper that just preventing the victim from passing defeats this adversary. But KataGo is trained on the same rules we evaluate on (in fact our adversary is just a forked version of KataGo), so still a mistake by black.
@ARGleave So this was my point exactly. NN evaluation is that it’s winning (which is true) but my question is whether the decision to pass comes from the NN or the scoring engine. If NN is properly trained on TT it should not pass but scoring engine might be overriding it
@ARGleave The games are definitely won under TT, that I agree with. It looks to me like KataGo’s optimization you mentioned kicks in early and passes. In essence, KataGo doesn’t seem to understand that TT requires dead stones to be captured and removed.
@ARGleave So KataGo’s own scoring engine evaluated that as a loss? Do you have a full SGF playout? If that’s the case it should be possible to reproduce your results and see the same katago version evaluate the game as a loss
@ARGleave If you look at the code, TT is actually using the area scoring rule set because “it’s the closest” and it doesn’t seem to change the rules for “dead stones must be captured”. Their rules page also only seems to change some settings such as suicide: https://t.co/ZzniTN5IR0
A Modest Proposal For preventing the Parodists of Twitter From being a Burthen to the Objects of Their Ridicule and For making them Comprehensible to the Lowest Common Denominator