google is broken...

[jackperkins@post.lurk.org]

@liaizon alsp lol this costs like a dollar a pop for them

[liaizon@social.wake.st]

@matt I mean I guess its working, I just googled a bunch of things cause of how ridiculously wrong all the answers are, so maybe thats their game

[matt@oslo.town]

@liaizon That's right! More time in Google products, eyeballing the ads they baked into responses!

[liaizon@social.wake.st]

@jackperkins naaaa, they are not using their good models for this, they are using their cheapest shit for it, I would assume more like 000000.1cent per query

[oppen@merveilles.town]

@liaizon In $dayjob we have an elusive bug in our impl. of Android's MediaSession/ExoPlayer. Having previously spent days trying to fix it I thought I'd try the Claude account we're testing, let's see if the AI can find the problem... it's very very confidentlly identified the issue in 3 or 4 places so far, none of the fixes it's added have worked at all, but the language it uses is so matter of fact. It's a user-facing issue and the in-app radio playback locks up until restart, so big problem when it happens, the 1st time Claude said it knew what the issue was I was momentarily very happy...

[liaizon@social.wake.st]

@oppen it so often being confidently wrong is one of the most annoying aspects of this jank for sure. I think many people would respond better to a bunch more of "I don't know!" responses littered in (it could probably just be randomized) just to have less certainty...

[despens@post.lurk.org]

@liaizon Perfect comedy! Really digging the references on this one:

[liaizon@social.wake.st]

@despens its amazing to see how everytime google uses citations they are basically randomly added later. like it makes up some shit, then "cites" it by just showing a link that contains a single word of your search.

[despens@post.lurk.org]

@liaizon This post of yours is actually a prompt that forms the foundation of a billion dollar business

[gub@framapiaf.org]

@liaizon
The correct answer is 3.
"Rassemblement National"

[fluchtkapsel@nerdculture.de]

@liaizon missed opportunity, Google, missed opportunity.

[liaizon@social.wake.st]

@fluchtkapsel its impressive how consistent this "bug" is, like they have to realize it by now. I wonder what the internal Google chat looksike atm

[liaizon@social.wake.st]

someone at google is still alive!

[s_ol@merveilles.town]

@liaizon
the sad thing is that they can (and might) fix this, but the fix will be one of these two:

1. have the model run much longer, then have another model read that output and summarize: "reasoning" (aka maximize resource burning)
2. give the AI a traditionally-implemented "count characters in word" tool (address the symptom, aka admit defeat.

[liaizon@social.wake.st]

@s_ol it should be doing tool calls anyway for many questions. Like Google search used to do basic math and equations inline without searching.

[s_ol@merveilles.town]

@liaizon
i was under the impression that other general-purpose models have solved this and would've guessed they implemented option 2 but actually chatgpt is still confidently wrong (now with sources!)

[liaizon@social.wake.st]

@s_ol I love how all these companies are completely ruining the concept of souces and citations

[mntmn@mastodon.social]

@s_ol @liaizon last night when i tried exactly this, it was already using python to count chars, but only sometimes. when i asked "how many p in paraphernalia" it said 3. so it's just very random

[liaizon@social.wake.st]

@mntmn @s_ol how many pythons does it take to change a lightbulb?