what do y’all see in gemini-2.5-pro?? i’ve been using it today and has crazy small model smell. feels like a 7B. i must be doing something wrong, it’s made all the classic mistakes in sequence - ThreadSky

timkellogg.me • 18 days ago

what do y’all see in gemini-2.5-pro??

i’ve been using it today and has crazy small model smell. feels like a 7B. i must be doing something wrong, it’s made all the classic mistakes in sequence

Comments

pekka.bsky.social•18 days ago

You must be doing something wrong. It's like crazy good.

Which way are you accessing it? AI Studio?

timkellogg.me•18 days ago

AI studio, with Google Search enabled and is definitely not fucking using

timkellogg.me•18 days ago

i swear the model selector is just not working, because this is very bad

pekka.bsky.social•18 days ago

Can you see if it thinks before answering that this user is a special kind of case and should be handled accordingly?

Does it display the context window length as roughly 1M?

timkellogg.me•18 days ago

yes to both, also: https://bsky.app/profile/timkellogg.me/post/3loqkc5ksg22m

timkellogg.me•18 days ago

i must be running into a bug. there’s no way i’m this out-of-sync with everyone else

thomas.advanced-eschatonics.com•18 days ago

it happens

philpax.me•18 days ago

it *was* great in 03-25, but I haven't tried the I/O Update:tm:

could be possible that they broke it?

tussles-shriek.bsky.social•18 days ago

they definitely broke something, lol. It started to think like 5x as long as before and they removed the old version altogether for no apparent reason

timkellogg.me•18 days ago

it’s too bad even for that. it’s gotta be an AI studio bug

benmccann.com•18 days ago

It's got a newer training cutoff which really helps when trying to use recent technology like Svelte 5. We benchmarked all the major models and Gemini 2.5 Pro came out on top for this reason.

simonwillison.net•18 days ago

What kind of questions are you asking it?

I've mainly used it for structure data extraction (get me the numbers in this video/PDF) and debugging and writing code and it's been great at both so far

gondiken.bsky.social•14 days ago

Context window was the main reason when I used it and it single shotted a large game for me. Haven't tried last 2 weeks though.

catblanketflower.yuwakisa.com•18 days ago

I avoid Gemini, but I have to use it for work. Results are… mixed

Deep Research is pretty cool though

timkellogg.me•18 days ago

it’s verbatim taking anything i say as truth. what is this??

tussles-shriek.bsky.social•18 days ago

It takes _anything_ you say at all? That's not the same 2.5 Pro we cheered, lol. Original 'experimental' one was smart but freaking stubborn, it was impossible to tell it anything ('user insists that the current year 2025, clearly they delusional and I'm right')

iggy.gov4.gay•18 days ago

It is glazing a lot for me

timkellogg.me•18 days ago

someone else is simultaneously complaining about truncated code and conversation history issues. maybe this is just a very bad AI studio update

pekka.bsky.social•18 days ago

It works fine for me right now. With a query containing two ballpark estimates, it's strongly pushing back on the one that's misleading, explaining the reasons and relevant research very well, and providing possible sources of confusion and differing interpretations.

braintelligence.bsky.social•18 days ago

I was noticing the same thing yesterday using it on VS Code copilot. Hallucinations constantly

Comments

Posting Rules

Reply