Gemini Flash smells like a small model it is not borne out by benchmarks, it's not compatible with its knowledge, it… just smells. "you're holding it wrong", "not built for that", lmao Demis r u 4 real? It's like an "AGI" from a world-class team of Ph.Ds in reranker models.

> writing quality is mid-to-bad I despise default Gemini writing quality so much I suspect it's made to reduce free usage on AI studio (I used to turn to it for long context/vision tasks). Like, any general question results in a tasteless markdown syndrome atrocity. "…thanks"

@solipsnitsyn Somebody took the comic too seriously

My biggest feedback is have a conversation with the strategy-setters on the post training team. Flash 3.5 still thinks it is Flash 1.5. It still feels like it has been personality smashed flat and is scared of being anything other than a tool

I think the personality you are targeting ("I am nothing, I am just a search tool") makes the model insane long term.

@solipsnitsyn googles models are great but their harnesses are hot dogshit

@solipsnitsyn Mine uses Chinese for absolutely no reason

@solipsnitsyn definitely NOT black!

@solipsnitsyn it has the same problem of previous gemini models. it basically can not do any long horizon task. Codex will ceaselessly work on something until you ask it to stop. Gemini often work for a few mins and ask should I proceed?

maybe it's some nonsense like a Shazeer-maxed 4-group GQA with aggressive SWA and cross-layer kv sharing, and 40 layers total. Shallow, wide, fast, dumb, "good enough", provably can hill-climb whatever. Just smells like a toy.

@tenobrus everyone tests gemini models for coding and decides they’re trash across the board

i use them for our internal tools for EVERYTHING except code and they’re amazing intelligence per dollar models

i’d imgine this is how most of googles customers use them, for everything but code.

@BlockedPaths @solipsnitsyn omp fixes this

@tenobrus SERV fixes this, cuts redundant tool calls, reduces hallucinations, adds proper clarification, and delivers cleaner output on Gemini 3.5 Flash, makes it reliable 👇🏼

@ShittyTwittter @solipsnitsyn What a f*cking embarrassing piece of garbage

Their annoying engineers always posting here on X should RESIGN

@solipsnitsyn lowkirkenuinely hilarious

@solipsnitsyn I enjoy viewing this message hundreds of times per hour whenever I hit submit

@usr_bin_roygbiv @solipsnitsyn And I got banned using it in opencode lol

For as much firepower as they have, I just don't think Google has a unified and coherent enough of an internal structure to perform in this domain anymore. The game is rapidly changing and they have too much inertia and friction to keep up like they need to. They'll likely still contribute to research on some fronts, maybe excel in certain niche domains, but thats about it imho.

@tenobrus Yeah this is the standard case

@solipsnitsyn What else do u expect? google i/o is where google launch dead stuff

@austeane @tenobrus My advice is to just use 5.5 low for that usecase. It is significantly faster at it than all smaller models Ive tested (including spark).