This story was originally posted as a response tothis thread.
It might help to imagine a hard takeoff scenario using only known sorts of NN &scaling effects...
In A.D. 20XX. Work was beginning. "How are you gentlemen !!"... (Work. Work never changes; work is always hell.)
Specifically, a MoogleBook researcher has gotten a pull request from Reviewer #2 on his new paper in evolutionary search in auto-ML, for error bars on the auto-ML hyperparameter sensitivity like larger batch sizes, because more can be different and there's high variance in the old runs with a few anomalously high performance values. ("Really? Really? That's what you're worried about?") He can't see why worry, and wonders what sins he committed to deserve this asshole Chinese (given the Engrish) reviewer, as he wearily kicks off yet another HQU experiment...
I find it frustrating that every time this story wanders into a domain in which I am knowledgeable about, the author shows his ignorance.
For example, HQU finding a flaw in how a zk-proof system handles floating-point numbers (no zk-proofs over floating point numbers are used anywhere in anything; I'd be surprised if such a system has even been deployed). Even the lead-in where the researcher thinks his reviewer is Chinese "given the Engrish", but "Engrish" is a word used to describe the Japanese's (not Chinese!) particular issues with learning and using English, and typically not in professional contexts.
These probably seem like trite details, and they are, but it makes me skeptical that the author is as knowledgeable as he tries to seem with the constant technobabble. Some parts of the story strike me as utterly fantastical. For example, as someone who has written HPC codes for a supercomputer and maintained an application using ASIC accelerators, the idea that HQU (*ahem*, Clippy) could upload itself to a distributed network of cloud computers and even come within 5 orders of magnitude of its original described performance is absurd.
I hope people aren't going to attempt to learn from this and generalize from fictional evidence...
LW2 developer here. I consider it a bug that it's possible to continue ot comment through a deactivated account. I don't consider it a bug that comments made through a deactivated account can be associated with the account name, since (in the normal case where an account never posts again after it's been deactivated) the same information is also easily retrieved from archive.org/.is/etc. (And I can think of a dozen easy ways to do it, some of which would be a pain to close off.)
(The officially supported mechanism for anonymous posting is to just make a new... (read more)
Gwern, you seem to be incapable of taking constructive criticism, and worse you've demonstrated an alarming disregard for the safety of others in your willingness to doxx someone merely to score a rhetorical point. Thankfully in this case no harm was done, but you couldn't have known that and it wasn't your call to make.
I will not be engaging with you again. I wish you the best.
26dxu
FWIW, I did actually manage to guess which Mark it was based on the content of the initial comment, because there aren't that many persistent commenters named Mark on LW, and only one I could think of who would post that particular initial comment. So claiming not to have deanonymized him at all does seem to be overstating your case a little, especially given some of your previous musings on anonymity. ("The lady doth protest too much, methinks" and all that.)
I do, however, echo the sentiment you expressed on the EA Forum (that anonymous commenting on LW seems not worth it on the margin, both because the benefits themselves seem questionable, and because it sounds like a proper implementation would take a lot of developer effort that could be better used elsewhere).
This story was originally posted as a response to this thread.
It might help to imagine a hard takeoff scenario using only known sorts of NN & scaling effects...
Rest of story moved to gwern.net.