25
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
this post was submitted on 20 Apr 2025
25 points (100.0% liked)
TechTakes
1864 readers
121 users here now
Big brain tech dude got yet another clueless take over at HackerNews etc? Here's the place to vent. Orange site, VC foolishness, all welcome.
This is not debate club. Unless it’s amusing debate.
For actually-good tech, you want our NotAwfulTech community
founded 2 years ago
MODERATORS
Found a thread doing numbers on Bluesky, about Google's AI summaries producing hot garbage (as usual):
I tried this a couple of times and got a few "AI summary not available" replies
Ed: heh
Generative AI is experimental.
NOT THE (PORK-FILLED) BEES!
Now we know why dogs eat bees!
The link opened up another google search with the same query, tho without the AI summary.
image of a google search result description
Query: “a bear fries bacon meaning”AI summary:
It really aggressively tries to match it up to something with similar keywords and structure, which is kind of interesting in its own right. It pattern-matched every variant I could come up with for "when all you have is..." for example.
Honestly it's kind of an interesting question and limitation for this kind of LLM. How should you respond when someone asks about an idiom neither of you know? The answer is really contextual. Sometimes it's better to try and help them piece together what it means, other times it's more important to acknowledge that this isn't actually a common expression or to try and provide accurate sourcing. The LLM, of course, has none of that context and because the patterns it replicates don't allow expressions of uncertainty or digressions it can't actually do both.
You, a human can respond like that, a llm, esp a search one with the implied authority it has should admit it doesnt know things. It shouldn't make up things, or use sensational clickbait headlines to make up a story.
Also on the BlueSky-o-tubes today, I saw this from Ketan Joshi:
gnnnnnngh
this shit pisses me off so bad
there's actually quantifiable shit you can use across vendors[0]. there's even some software[1] you can just slap in place and get some good free easy numbers with! these things are real! and are usable!
I'm sorry you fucking what? just how exactly are you getting wattage out of openai? are you lovingly coaxing the model to lie to you about total flops spent?
[0] - intel's def been better on this for a while but leaving that aside for now..
[1] - it's very open source! (when I last looked there was no continual in-process sampling so you got hella at-observation sampling problems; but, y'know, can be dealt with)