Recent searches
Search options
Administered by:
The speed with which these guys converge on the idea that LLM ~misbehavior—which is about as intelligent as the flailing end of a garden hose with too much water coming out—means that they're really highly intelligent female (of course, of course) entities whose "personality" can be "revealed" is both hilarious and so dispiriting
https://stratechery.com/2023/from-bing-to-sydney-search-as-distraction-sentient-ai/
@kissane
Wow that was painful to even skim. I'm assuming that each time stuff got erased or things changed course, that was some poor mod on the other end having to send in instructions out of band to try to stop product from generating terrible output for this guy
@dymaxion Oh wow, I hadn't even considered a human mod, I'd figured it got sent up the chain to moderation software. That's even worse.
@kissane
I mean, rule one of AI products is that they're actually mostly a bunch of unpaid humans. Given everything we've seen here and how many times Microsoft in particular has been bitten by this issue, I can't imagine they'd launch it without a few thousand moderators riding herd on it, just like midjourney does to try to stop its porn problem. I think they're not entirely wrong to ascribe human intent to the sudden shifts (as opposed to the rest of the generated text), because they're probably exactly that — human intent.
@dymaxion All true. I keep hoping they've found ways to handle at least baseline moderation that aren't just applying low-wage adrenal systems at scale, but that's certainly naive.
@kissane I mean, I'm sure it's not just manual — there's obviously a bunch of tuning, and I can only imagine that they're building more complex classifiers for types of problems, etc. as they go. But all systems like that always need manual data tagging at some point at the least to classify what's ok and not and what types of corrective action should happen. If nothing else, they can probably classify the kinds of conversations like this one that involve someone poking at the response envelope and route those to mods.