Conversation

Notices

1. Embed this notice
   Sue Smith (sue@glasgow.social)'s status on Friday, 25-Apr-2025 01:14:01 JST Sue Smith

   While I hesitate to encourage folk to run unnecessary queries, I do think the "made up phrase" + "meaning" google search is a great exercise in helping people understand what generative AI actually does.. I've found it hard to explain to folk outside tech that its purpose is literally just to generate text but this seems a more intuitive demo that might make them question the output.

   • Paul Cantrell repeated this.
   • Embed this notice
     Paul Cantrell (inthehands@hachyderm.io)'s status on Friday, 25-Apr-2025 01:15:18 JST Paul Cantrell

     in reply to

     @sue
     This is excellent! Boosted.

     My own favorite example is a chat transcript that run something like this:

     HUMAN: I have a cabbage and a boat. How do I get the cabbage across a river?

     LLM: First you take the wolf….

   • Embed this notice
     Steve Canon (steve@discuss.systems)'s status on Friday, 25-Apr-2025 01:37:45 JST Steve Canon

     in reply to

     • Paul Cantrell

     @inthehands @sue These things have all of the world's knowledge, and none of the world's wisdom.

   • Embed this notice
     Paul Cantrell (inthehands@hachyderm.io)'s status on Friday, 25-Apr-2025 01:37:50 JST Paul Cantrell

     in reply to

     • Steve Canon

     @steve @sue
     omg, chef’s kiss

   • Embed this notice
     Paul Cantrell (inthehands@hachyderm.io)'s status on Friday, 25-Apr-2025 09:44:48 JST Paul Cantrell

     in reply to

     • Jordan Kay
     • Tom

     @tomknuf @jn @sue
     “Worked” is really a stretch here. The fundamental thing here isn’t just the quality of the answer (yes, no wolf is better); is that the thing doesn’t understand what a boat is, or what crossing is, or what a river is, because it just doesn’t •understand• anything period. That is does a better job than before of coherently repeating textual patterns doesn’t fundamentally move it away from the fact that it can’t think about the problem; it can only copy the answer to a seemingly related question.

   • Embed this notice
     Tom (tomknuf@social.anoxinon.de)'s status on Friday, 25-Apr-2025 09:44:58 JST Tom

     in reply to

     • Paul Cantrell
     • Jordan Kay

     @jn @inthehands @sue I just tested this with GPT-4o and it worked. Not as bad, but the answer is still quite rubbish

   • Embed this notice
     Jordan Kay (jn@c.im)'s status on Friday, 25-Apr-2025 09:44:59 JST Jordan Kay

     in reply to

     • Paul Cantrell

     @inthehands @sue To be fair (and this is coming from an extreme AI skeptic), I rarely if ever see those types of errors anymore with current models.

   • Embed this notice
     JD (wuputah@fosstodon.org)'s status on Friday, 25-Apr-2025 09:46:27 JST JD

     in reply to

     • Paul Cantrell
     • Jordan Kay

     @jn @inthehands @sue Yeah I just tested this on Gemini 2.0 Flash, 2.5 Flash, and 2.5 Pro, and none of them really suffer from this, though 2.0 Flash wanted to take a guess. I suspect the one on Google Search is a very low cost model.

   • Embed this notice
     Paul Cantrell (inthehands@hachyderm.io)'s status on Friday, 25-Apr-2025 09:46:27 JST Paul Cantrell

     in reply to

     • JD
     • Jordan Kay

     @wuputah @jn @sue
     GPT-4o still works by plagiarizing the common puzzle answer from a place of complete and total incomprehension:
     https://social.anoxinon.de/@tomknuf/114395220366253145

   • Embed this notice
     Paul Cantrell (inthehands@hachyderm.io)'s status on Friday, 25-Apr-2025 11:19:25 JST Paul Cantrell

     in reply to

     • el Celio 🇪🇺 🇺🇦
     • Jordan Kay
     • Tom

     @elCelio @tomknuf @jn @sue
     I mean, yes: I think a lot of people have got through life by bullshitting everything, and that’s why they truly don’t understand why a bullshitting LLM can’t just do everyone’s job now.

   • Embed this notice
     el Celio 🇪🇺 🇺🇦 (elcelio@mastodon.uno)'s status on Friday, 25-Apr-2025 11:19:27 JST el Celio 🇪🇺 🇺🇦

     in reply to

     • Paul Cantrell
     • Jordan Kay
     • Tom

     @inthehands @tomknuf @jn @sue so it works like many humans 🤔

   • Embed this notice
     Paul Cantrell (inthehands@hachyderm.io)'s status on Friday, 25-Apr-2025 11:19:57 JST Paul Cantrell

     in reply to

     • Coach Pāṇini ®

     @paninid
     Smells like manual intervention to me.

   • Embed this notice
     Coach Pāṇini ® (paninid@mastodon.world)'s status on Friday, 25-Apr-2025 11:19:59 JST Coach Pāṇini ®

     in reply to

     • Paul Cantrell
     • Jordan Kay
     • Tom

     @tomknuf @jn @inthehands @sue

     Claude is wise to the trend:

   • Embed this notice
     Paul Cantrell (inthehands@hachyderm.io)'s status on Friday, 25-Apr-2025 13:24:11 JST Paul Cantrell

     in reply to

     • Steve Canon
     • Rev. Charles Browning
     • chocobo13ζ

     @steve @chocobo13 @FrChazzz @sue
     Ironically, I have some kind of weird sensory disorder that leaves me averse to the smell of all fruit — including fresh tomatoes, eggplant, cucumber, and peppers. So this discussion really hits close to home for me!

     (P.S. No worries about discussing this all in my mentions, just as long as I don’t have to smell any of it.)

   • Embed this notice
     Steve Canon (steve@discuss.systems)'s status on Friday, 25-Apr-2025 13:24:12 JST Steve Canon

     in reply to

     • Paul Cantrell
     • Rev. Charles Browning
     • chocobo13ζ

     @chocobo13 @FrChazzz @inthehands @sue cucumber, peppers, …

   • Embed this notice
     Rev. Charles Browning (frchazzz@mastodon.social)'s status on Friday, 25-Apr-2025 13:24:13 JST Rev. Charles Browning

     in reply to

     • Paul Cantrell
     • Steve Canon

     @steve @inthehands @sue This inspires me to share my favorite definition of wisdom I've ever received (thanks to a sixth grader from when I was a teacher):

     Knowledge is knowing that a tomato is a fruit. Wisdom is knowing that a tomato doesn't belong in a fruit salad.

   • Embed this notice
     chocobo13ζ (chocobo13@mastodon.social)'s status on Friday, 25-Apr-2025 13:24:13 JST chocobo13ζ

     in reply to

     • Paul Cantrell
     • Steve Canon
     • Rev. Charles Browning

     @FrChazzz
     I've been thinking that there probably exists a really good savory technically-fruit salad involving tomatoes. Maybe eggplant as well?
     @steve @inthehands @sue

   • Embed this notice
     Rich Felker (dalias@hachyderm.io)'s status on Friday, 25-Apr-2025 21:24:21 JST Rich Felker

     in reply to

     • Paul Cantrell
     • Coach Pāṇini ®
     • Philipp Weiß

     @thewhite969 @inthehands @paninid I mean it may be both. I wouldn't be surprised if their "manual intervention" consists of having someone write a bunch of explainers and feeding them into corpus with high weights.

     What's telling is that they'll do this when the model embarrasses them with how obviously fake the impression of "intelligence" is, but not when the bot is vomiting up hate speech.

   • Embed this notice
     Philipp Weiß (thewhite969@chaos.social)'s status on Friday, 25-Apr-2025 21:24:22 JST Philipp Weiß

     in reply to

     • Paul Cantrell
     • Coach Pāṇini ®

     @inthehands @paninid It may be manuel intervention, because it's a common fault.
     Or it might be, that it's a common fault and so there's much content on the Internet laughing about this. And the training data of the LLM are newer and contain this content. And in some magical way this leads to the LLM scoring this combination of words as more relevant to the question.
     The important thing in the end is: The LLM doesn't _understand_ anything.

   • Embed this notice
     Paul Cantrell (inthehands@hachyderm.io)'s status on Friday, 25-Apr-2025 21:59:23 JST Paul Cantrell

     in reply to

     • Jordan Kay
     • Tom

     @tomknuf @jn @sue
     Ah, yes, I misunderstood you!

   • Embed this notice
     Tom (tomknuf@social.anoxinon.de)'s status on Friday, 25-Apr-2025 21:59:24 JST Tom

     in reply to

     • Paul Cantrell
     • Jordan Kay

     @inthehands @jn @sue I think you misunderstood me. I meant "worked" in the sense that it still isn't able to answer this adequately, opposed to the other replier that LLMs nowadaws don't make such mistakes anymore.

   • Embed this notice
     Paul Cantrell (inthehands@hachyderm.io)'s status on Friday, 25-Apr-2025 22:53:13 JST Paul Cantrell

     in reply to

     • Rich Felker
     • Coach Pāṇini ®
     • Philipp Weiß

     @dalias @thewhite969 @paninid

     My suspicion is that they do actually attempt the whack-a-mole with hate speech, which keeps generating bad press, but there’s just •so• much of it in the training data