@JourneymanGeekOnStrike I once asked about IP address used by SE, turned out it was some sensitive info that could be used by hackers, so Nick hard deleted it and sent me email to let me know. It's "more" than redaction, it means even mods can't see it, or anyone, it was actually deleted without any trace or way to bring it back. So, same can happen with comments, I guess.
I try to look at the bright side! I reviewed numerous math papers written in bad and overly concise English, lacking any explanations or context, full of mathematical jargon that the authors - after internalizing it - apply without thought or consideration, Now I'm looking forward to reviewing math papers written in flawless and overly wordy English, larded with misleading context and pointless explanations of trivialities, and still full of mathematical jargon that ChatGPT - after internalizing it - applies without thought or consideration. Who wouldn't welcome a bit of change now and then? — Jochen Glueck2 days ago
When Tor-users try to sign up, they consistently get the error message
Something went wrong. Please contact us at [email protected] for assistance.
However this error-message is disingenuous, because when Tor users actually contact [email protected] for a...
For the past 3 weeks we have been working to further strengthen our DDoS mitigation while also accounting for Tor traffic. At this point, we feel that we have been able to strike a good balance between protecting the site from DDoS attacks without blocking Tor traffic. Although we cannot rule out...
Didn't find newer updates.
But, they probably blocked it again just without saying anything, that's how they work.
Unless part of blocking DDoS attacks, there's really no legit excuse to just block Tor. The real reason is of course they know they won't get any money from those users, only trouble.
(as they won't click any ads, can't be tracked, etc.)
Huh! I knew that starship was trouble. Finally network wide suspension, after littering MSE with dirty sock. I knew something was familiar in that "answer". (it was posted by different account, which was merged.)
> This is actually a generic problem with ChatGPT: if you ask it something that is impossible, it simply cannot tell you that what you ask is impossible; instead, it will hallucinate a world wherein the thing you ask for is in fact possible and then come up with an overly elaborate answer, with full code examples and everything, but it will never work because it's not possible and it does not have the ability to tell you this.
> In other words, ChatGPT is an XY problem amplifier. You want to do something with an API that wasn't made to do the something, you ask the tool in MDN how to do that, it will hallucinate some gibberish for you that makes it sound like it's possible, and now you're stuck even further in your XY problem.
@JourneymanGeekOnStrike I'm glad you like it! I think it came out quite well :)
(I was the CM liaison with the designers on that theme)
@ShadowWizardStrikesBack I feel like anyone who believes AI alone will be able to correctly answer any question (in its current form, not some sort of utopian sci-fi future) has never dealt with an automated system :P
@JourneymanGeekOnStrike I dealt with a few network spammers that I came across on Friday :)
@SonictheAnonymousHedgehog fixed
@ShadowWizardStrikesBack Yes, we can – though there's thankfully rarely a reason to do so.
@PM2Ring ChatGPT is biased toward accepting the premise made by the user. Unless you ask something really blatant like "can a cat fly" it will try to follow up on your premise. This is probably tied to having been trained on lies like "a cat can fly" but not on more complex ones like "It is true that Hasbro and Sega made a My Little Pony and Sonic the Hedgehog crossover called My Little Mobius?"
@SPArcheon Right. The initial training and RLHF can only cover a microscopic proportion of the space of possible combinations of concepts. We commonly say that GPT creates continuations that are statistically consistent with its training data, but that's an exaggeration.
Even a galaxy's worth of training data could only cover a tiny fraction of the possible token sequences of even modest length, so most of the time GPT is estimating probabilities for sequences it's never seen by extrapolating from the sequences it has seen. RLHF can provide a band-aid solution for some of the more blatant flaws, but it's easy to enter territory that is pure hallucination.
Bullets 1 & 2 (the question ID and the moment of entering the list) will be available via the existing PostHistory table, for hot questions on or after February 28th, 2019. The PostHistoryTypeId for these events is 52. Next Sunday, we'll get to execute queries like this which will list the latest...