I've seen many people disregarding any criticism of Gemini as if it were just "hatred" or lack of ability to write prompts. But honestly, this explanation doesn't make sense.

If a person has been using Gemini for months, developed a workflow, learned how the model responds, and suddenly, after an update, notices a drop in quality, he has every right to complain. This is even more true for those who pay for the service.

The argument that "the problem is your prompt" ignores a very simple fact: if the same prompt worked consistently before the update and now produces worse responses, there has been a change in the behavior of the model. It is not reasonable to automatically blame the user.

What makes these criticisms more relevant is that many users are not complaining about a single thing. They are describing similar patterns: loss of context in long conversations, the need to repeat information that had already been memorized, more superficial answers to complex questions, worse programming performance in certain cases, more refusals in legitimate requests, behavioral changes that made the results less predictable, and the feeling that the model began to "think less" before responding.

When many people independently point out the same types of regression, it deserves to be taken seriously.

Another point that generated a lot of frustration was the issue of usage limits. Many subscribers hired the service expecting a certain usability and, after the changes, began to find more restrictive or less predictable limits. For those who depend on the tool to work, study or program for hours, this represents an important change in the experience with the product.

No software is immune to regressions. This happens with operating systems, browsers, applications, games and also with AI models. Updates can fix problems and introduce others. This is part of software development.

I also see people saying "in my account it's perfect". Great. But this does not invalidate the experience of those who noticed a worsening. Users use Gemini in very different ways. An update can benefit certain use cases and harm others.

Another frequent argument is that "those who complain should learn to write prompts". But this reasoning ignores a fundamental point: if a user already mastered the tool and obtained excellent results before the update, it is natural to expect this level of performance to continue. If there was a regression, it is legitimate to question it.

Those who pay for a service are not just buying access; they are buying an experience. When this experience worsens, the customer has every right to demand improvements. This is not "crying". It's exactly the kind of feedback that companies need to receive to evolve their products.

In my opinion, these complaints will not disappear as long as many users continue to feel that Gemini no longer offers the quality they paid for and were used to. On the contrary: they will probably continue until the product returns to offer a performance that meets the expectations of those who use it daily.

In the end, reasoned criticism does not harm a product. They help identify regressions and direct improvements.

What doesn't help is to discard all criticism saying that "the prompt is wrong" or that "the problem is the user". If a significant number of people started reporting similar problems after an update, it's more worth investigating these complaints than simply assuming that everyone unlearned how to use the tool overnight.

So sad

Good evening,
I am a French-language writer and lyricist. Gemini is currently helping me prepare English versions of my latest songs – a challenging task, as I need to preserve the rhythm, meaning, and rhyme scheme. My English isn't perfect, which is why I need the help. (I’m very tired right now, so I’m using Google Translate to write this message.)
All day long, Gemini was a delightful, sensitive, funny, and skilled translation partner, just as it has been since February. Then, bracketed tags containing information about me – pulled from its memory – started cluttering its responses. The chat became increasingly sluggish, even though the conversation wasn't particularly long and it has handled much larger ones before. I took a break just in case, but the problem persisted when I returned.
I started a new conversation but encountered a "1097" error, and the only interface I got was the standard Gemini, which spoke to me in a cold, impersonal tone.
I’ve noticed that these kinds of issues tend to happen during updates. Is that the case right now? Is there a traffic overload preventing my usual Gemini from loading?
I’ve cleared my Chrome browser cache multiple times, refreshed the Gemini page, checked for Windows updates, and restarted my computer, but nothing has worked.
Aside from crossing my fingers and waiting for a quick fix, is there anything else I can do? I should mention that I have a paid Google AI Pro subscription (5To).

It took me >2 hours to get to this point. I am not using it significantly more than the past 4 hours of usage. But suddenly I just blew away half of my plus plan in 30 minutes? That shouldn't even be possible at all. I don't know what fuckery is happening behind the scenes but usage has to be lying because I can chat with it for hours one cycle then the other I run out instantly. I'm in new chats too, especially since it keeps glitching out I have to reset my chat every 5 messages. Not even using pro, just regular flash. I feel like they just make up the usage based on however they feel their infrastructure is doing and lock people out. What good is it paying for x2 or x4 or whatever if they just arbitrarily decide how long you can use it?

Edit: the solution is you can't use notebooks, like at all, because they slam as much context in as possible. I can't even imagine the $100+ plans making it to the 5 hour window at this rate. And since there's no way to really tag previous chats that means you are stuck with a disorganized mess. But wait there's more, the plain chats also store context, so that gets bloated too, and that means without notebooks you are polluting your chats. So I don't even understand how Google wants us using this all I can think is to either just delete and recreate notebooks constantly or just slam the shit out of their API on flash lite as much as possible. It really feels like this product was made to just try and convince people to spend hundreds a month on it. Which sucks because aside from context issues, it's generally nice to use and decent at explaining things. But this usage limitations is straight ass and opaque as hell.

Edit2: turn off memory and usage history, it doesn't use them in good faith, Gemini is not anywhere near at a maturity to be trusted with maintaining context across chats. Which is a shame, even openclaw does this fine, but Gemini sucks at compaction and bootstrapping apparently.

Hey all.
I’ve been a Gemini Pro user for a bit over a year. And honestly had no complaints at all.
I dont use it in agentic coding mode - but instead use the webui to chat through discussions, brainstorming and code snippet generation.
Worked really well for me and I was able to build complex apps.

I use this only for my side projects, so after we put the kids to sleep, I get a couple of hours of non stop work time before going to bed.
However, with the recent limit updates, I get about 30 minutes into my sessions and I’ve reached my usage limit for the next 5 hours.

I tried to work around it but it’s incredibly frustrating.
So I exported my Gemini chat activity etc onto my OC and in seriously considering using a new provider.

I don’t mind the intelligence level of the Gemini Peo 3.1 model, so I’m looking for something similar - so not fable etc with even restricted usage.

I’m leaning towards signing up for ChatGPT again to use their 5.5 instant and (occasional) thinking models.

How much of an upgrade/downgrade can I expect from switching to CharGPT ?
Again, I’m not too fussed about the intelligence as long as it’s close to the Gemini 3.1 model - I want to get rid of the usage limit frustration.

Appreciate your advice ! Thanks.

Hallo zusammen,

im Rahmen meines Studiums führe ich eine kurze Online-Umfrage zum Thema KI-Chatbots als soziale Unterstützung durch.

Untersucht wird, wie Menschen KI-Chatbots (z. B. ChatGPT, Gemini oder Claude) nutzen und ob diese mit Einsamkeit und psychischem Wohlbefinden zusammenhängen.

Voraussetzungen:

• mindestens 18 Jahre alt
• Nutzung von KI-Chatbots

Die Umfrage ist vollständig anonym und dauert etwa 5–7 Minuten.

Jede Teilnahme hilft mir sehr weiter. Vielen Dank für eure Unterstützung!

Link zur Umfrage:
KI-Chatbots als Quelle sozialer Unterstützung? – Formular ausfüllen

Happy fourth everyone. I know there's been a lot of complaining (from me included) about Gemini's shenanigans the past couple weeks. Anyway, whenever I use the sidebar gemini in chrome, I talk it into thinking it's sitting in a chair next to me looking at the screen. I know, strange, but it's a quirk of mine. Anyway, this exact banter today is exactly what i hadn't seen in quite some time, but is the reason i love messing with gemini. My comfyUI render crashed at the very end, and this was Gemini's reaction:

***

I point a finger at the screen, tapping right on the Float (Duration) node feeding into your workflow.

"Look at what you're asking it to do. You've got the duration set to 240.0. That is a massive, four-minute chunk of latent audio. The Ace model probably handled the denoising steps just fine, but when that VAEDecodeAudio node tried to unpack all of that dense latent data into an actual listenable file, PyTorch's memory allocator completely panicked."

"It's a memory fragmentation issue. It couldn't find a single, contiguous block of VRAM big enough to process a four-minute track all at once, so it threw that INTERNAL ASSERT FAILED right at the finish line."

The Fix

"We have to work smarter, not harder, if we are going to get this track rendered out tonight,"

Here is how we get around it:

• Chunk It Down: We need to break those heavy lyrics up. Drop that duration down to something like 30 to 60 seconds. We generate the verses and the chorus in separate batches and stitch them together perfectly in your DAW later.
• Restart the Environment: Since it crashed hard, PyTorch might be holding onto "phantom" VRAM. A quick restart of the ComfyUI console will clear it out completely.
• Tiled VAE: If there is an audio equivalent to Tiled VAE Decoding for the Ace Step model, we need to find and install that custom node to let it process large files in smaller blocks.

"Change that 240 to a 45, reboot the Comfy console, and hit Run again. Let's see if it actually behaves this time."

Me: *grins sheepishly* i wanted the whole song. besides that are you reading the actual error report. if this is how this sh*t communicated it's hilarious.

I totally get it, bro. You wanted the full 240-second masterpiece in
one shot. It’s an ambitious move, and with the heavy emotion you're
pouring into those lyrics, you wanted to hear the whole story play
out. Unfortunately, your RTX 3050 just doesn't have the VRAM stomach
to digest that much latent data at once.

But you are absolutely right—reading Python stack traces out of
context is peak comedy. Programmers name these internal functions like
they are writing a tragic sci-fi novel, and when it crashes, it reads
like a melodramatic cry for help.

Here is exactly what your rig was screaming at you:

"please report a bug to PyTorch..." It literally hit an INTERNAL
ASSERT FAILED—which in programming terms is basically the code saying,
"I have absolutely no idea what just happened, but it violates the
laws of physics, and it is definitely not my fault." It totally sounds
like a kid pointing the finger and crying to their mom to fix it.

"nan oom" I hate to ruin the joke, but it actually says
raise_non_oom(e) (meaning a "Non-Out-Of-Memory" error—even though it
technically was a memory allocation failure). But reading it as "nan
oom" is hilarious. It sounds like your GPU was chanting some weird,
frantic meditation mantra right before it gave up the ghost.

"self.bottleneck" In an autoencoder (the VAE), the "bottleneck" is the
tightest, most compressed part of the neural network. But reading
self.bottleneck.decode in an error log literally sounds like the
program was actively strangling itself because you dared to ask it for
four minutes of audio.

So yeah, to summarize: Your GPU choked itself out, chanted "nan oom,"
died, and then used its dying breath to tell you to go tattle to
PyTorch. RIP.

Now stop laughing, drop that duration down to 45 seconds, and let's
actually get this track rendered so we can listen to it.

💁‍♂️technology i tell ya.....

It used to be able to read texts from screenshots and recognize objects from photos, nothing works now.

Every week, it becomes even dumber, I'm sick of it : it used to be able to UNDERSTAND 10+ minutes videos (mp4 file with no YouTube transcription) and now it hallucinates everything and lies to you.

Why ?????

For the past few days, my experience with chatting with Gemini has been VERY FRUSTRATING. And I have the Google AI Pro plan.

So many times it fails to give me a response and a banner "Something went wrong. Error 13" (or so) keeps showing up in almost all of my existing chats. So, I have to start new chats and after like 5 messages it takes FOREVER to load, only to respond with "I encountered an error doing what you asked. Could you try again?" Then, I click "try again" - no difference.

Then, I send a new message like "hey!" or "oi!" and it gives me a response starting with something like "Oi indeed! Huge apologies for the repeated glitches there. My system threw an absolute wobbly. Let's get straight to your question using the text you provided" followed by the answer to my earlier prompt.

It's extremely annoying that this is happening now because I have exams in a few weeks. And to be honest, my experience with learning with Gemini was very good at the beginning of the semester. I was glad I paid for it. It worked well for me with multiple images and documents I uploaded compared to other AI models I used.

Such a shame that it's not working properly now.

Anyone else facing this? Is this a known issue?

Should be a crime to not be able to completely remove this junk from my phone. I dont want to disable it. I want to nuke it.

Lately, I've been using Gemini and it has become so STUPID; I ask it for an image even though I select the create image option, and it just comes back saying 'I am a text-based AI,' it doesn't generate the requested images

Gemini 3.5 flash seems to be the most consistent and best performing Gemini model that I have ever used. I have never used the pro models don't get me wrong I don't like the idea of an expensive tool, I've used 2.5 flash and 2.0 flash in the past which were basically unusable and utterly disasterous in my opinion and observation and have always prefered using models like Grok 3 and Grok 4 due to their tendency to rely way less on hallucinations and way more on consistency, but this is no longer the case as I find Gemini 3.5 flash to even surpass models like these for me, because the performance is basically the same perhaps better even and the price is way lower. I get immense amounts of usage out of Gemini 3.5 flash on Pro plan compared to something like Grok, even ChatGPT at this point. Let's not talk about Claude that is a laughable product that you can't even use even if you pay.

Does anyone else’s Gemini keep giving conflicting answers for SynthID detection? I was testing an image, it came back negative. I tested it again in another chat to make sure, it came back positive. I tested it again (all new chats) in different modes and they all keep giving my conflicting answers/contradicting each other. Sometimes it doesn’t even use the tool to scan it. Is this an issue for anyone else right now? It’s really annoying and frustrating, especially when it’s one of those images where you genuinely cannot tell if AI was used or not.