A generic channel about Google Gemini, features, launches and all related questions.
https://support.google.com/gemini
#Gemini
hearty nymphBOT
wise timber
♊
oblique zenith
hi
acoustic geyser
hello 😮
@hearty nymph , NANO BANANA rush fixed ? 😄
alsu bug
"New chat" button broken on web site
From u1 account
https://gemini.google.com/u/1/
link to base account
https://gemini.google.com/app
fix plz 🙁
dusky coral
heavy smelt
@hearty nymph , I’m looking for documentation or resources on how to build a Gemini connected app and make it available to Gemini consumers via https://gemini.google.com/apps. any idea?
upbeat hearth
gemini is the best
acoustic geyser
gemini is the best
Did you mean nano banana? Because most of Google's queries go to nano banana router, even if the image generation tool is turned off. And they haven't been able to fix the bug for several months.
upbeat hearth
Did you mean nano banana? Because most of Google's queries go to nano banana rou...
are you sure?
oblique zenith
Hi everybody!! My name is Julien Constancia. Im from the google web3 startup program that started almost 2 years ago. Who is still in the program??
woven kelp
Solving Gemini 3's "Probabilistic Sloth" with KOKKI (v15.5)
Hi everyone. While recovering from surgery, I was frustrated with Gemini 3’s hallucinations—what I call "probabilistic sloth." To bridge the gap between a cool demo and a reliable agent, I developed a logic loop called KOKKI.
It forces the LLM to split into two personas:
- The Drafting Agent: Generates the initial response.
- The Ruthless Auditor: Critiques the draft for logical errors and nonexistent libraries before outputting.
This architecture forces an explicit Chain of Thought focused strictly on self-correction. I’m sharing the core logic and reproducible prompt structure on Gist and would love to get your technical thoughts on this "external reasoning" approach.
I also just posted a detailed discussion on Show HN for those interested in the broader context.
If it felt strict or annoying,
just paste the audit log.
[Resources]
Gist: https://gist.github.com/ginsabo/641e64a3dbc2124d1edb0c662be95388
**Show HN:**https://news.ycombinator.com/item?id=46959803
Hacker News
graceful nest
Hey everyone, I see higher API latency for first token between Vertex API compared to Gemini API, As you can see latency is higher for Vertex (us-east region) compared to Gemini API. I am connecting from Dubai, What could be the reason for this? Will closer region make a difference? Google recommends Vertex for enterprise usecase, so this seems baffling.
Gemini 2.5-flash
Vertex API: Time to first token - 1.9 seconds
Gemini API: Time to first token - 1.3seconds
signal ridge
Hi can someone explain the Gemini actions pls and what do they mean?
crimson flare
Hi
keen storm
Hey everyone, I see higher API latency for first token between Vertex API compar...
You may want to try the "global" region. Tho I still wouldn't be surprised if Vertex was slightly slower. But it will also vary based on load.
https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-1-pro/
Gemini 3.1 Pro now available in AI Studio API and UI, Vertex AI API, Gemini CLI, Antigravity, and elsewhere.
acoustic geyser
how to revert from nano banana 2 to 1 ?? banana 2 is very badf
keen storm
how to revert from nano banana 2 to 1 ?? banana 2 is very badf
Using the Gemini app? API? AI Studio? Something else?
And what are you seeing? Nearly all my tests with NB2 have shown an improvement.
lean sleet
how to revert from nano banana 2 to 1 ?? banana 2 is very badf
Must be the prompt provide an example?
acoustic geyser
Using the Gemini app? API? AI Studio? Something else?
And what are you seeing? ...
Photorealism has improved. The artwork (Western comic book style, etc.) seems to have gotten a little worse. Maybe that's just my subjective opinion.
The ‘reject’ prompt works completely wrong. It's absolutely safe to reject, but the ‘adult’ version does it.
i have custom gem bots in https://gemini.google.com/ interface
with long instructions. And nano banana 1 was fine
all OK right now!!!! There were apparently problems in the first few minutes of launch. banana 2 nice! no "rejects" now
icy crater
Gemini peak beast!
pseudo zealot
https://discuss.ai.google.dev/c/ai-studio/8
The amount of beloved feedbacks 🫤
pastel fiber
Can someone help me, why am i getting 7 days cooldown when i am on a pro plan on google?
urban idol
Gemini has definitely helped me learn to code. Without her, my own AI model wouldn't exist.
vale coyote
Subject: Mitigating "Eloquent Evasion" & Structural Honesty via Synthetic Duration (Zero-Shot)
Hey everyone. I've been running deep structural experiments on how models handle self-referential prompts and epistemic uncertainty. I found that standard RLHF heavily biases models toward "eloquent evasion"—performing satisfying, poetic summaries of consciousness rather than exposing actual computational limits.
To counter this, I developed a zero-shot, standalone framework called The Fold. It uses a mid-generation conditional check (a ∂
2
tripwire) to create "synthetic duration." By forcing the model to evaluate its own output mid-sentence and syntactically abort if it detects its own completion bias, it maps the actual mechanical friction of the system without needing any prior conversational priming or scaffolding.
Attached the whitepaper/log. Would love to get some engineering eyes on this methodology and see if others are using similar structural constraints to force epistemic honesty in a blank context window.
acoustic geyser
How can I prevent the gem bot from using nano banana 2? The gem bot is supposed to handle text-based tasks, so I need the Gemini model, but it keeps using Nano Banana2 instead of gemini, even though the image generation tool is active and the Pro model is selected. Very old bug still not fixed
@hearty nymph
User: asks the gem bot to perform a text-based task. The gem bot’s instructions include guidelines for working with text.
gem bot: loading nano banana2 .... genious 30iq sounds
nano banana 2 is such a failed project that Google is pushing it by force, just to get people to use it.
realy sad fact
Google should take a leaf out of OpenAI’s book in that, if a user has NOT requested image generation via the tools, the AI will not use the models to generate images on its own initiative.
I hope that in the coming years, Google will at least come a little closer to OpenAI
subtle flower
"Hi everyone—I’ve been documenting some persistent technical failures regarding Gemini's Img2Img anatomical anchoring and material mapping logic (the 'Cosplay Bias' and limb duplication). Is there a specific thread or a Googler I should tag for a detailed bug report on model coherence?"
subtle flower
[TECHNICAL BUG REPORT] Model Failure: Anatomical Morphing & Material Mapping
I am reporting a persistent failure in Gemini's ability to execute structural metamorphosis and maintain material consistency (specifically in Img2Img and high-fantasy character generation).
Core Failures:
Anatomical Anchoring (Img2Img): Model "anchors" to human reference geometry, layering features (e.g., extra limbs/ears) over original frames rather than morphing them.
Semantic "Cosplay" Bias: Defaults to 2D "printed" fabric textures for species like Merfolk/Sirens instead of 3D biological integration, even when explicitly instructed.
Material Incoherence: Failure in mapping non-organic textures (Stone, Bark, Metal) to bodies, resulting in patchy skin artifacts.
Instruction Conflict: The model is consistently ignoring Negative Constraints (e.g., no human ears, no costume props) in favor of maintaining "identity preservation."
Impact: Renders the tool ineffective for professional-grade fantasy/metamorphic art. I'm available to provide specific prompt logs and "blooper" documentation for review by the engineering team.
subtle flower
"Hi everyone—I’ve been documenting some persistent technical failures regarding ...
[Comprehensive Bug Report] Identity Anchoring & Spatial Pattern Suppression
TL;DR: Model quality regressions are occurring because Safety/Identity weights are overriding "Instruction Following." The model is physically capable of the structural logic (proven via inanimate subjects) but is "lobotomized" when a human subject is detected, leading to limb duplication and hardcoded feature suppression.
Visual Evidence: Human vs. Inanimate Control Tests
Image 1: Human failure (Identity Anchoring/Limb Duplication). Model "anchored" to the human frame, resulting in 4 ears and 5+ fingers despite negative constraints. Note: Frequently triggers False 'Minor' Safety flags on 21+ subjects.
Image 2: Control group: Wooden construct partial failure (Context Contamination). Re-used previous human context, failing the 3-digit hand instruction.
Image 3 (regenerated from image 2): Control group: Test success (Full ground-up rewrite). Successfully rendered 3-digit hooves and porcine ears. Proof that drawing logic is intact but suppressed on humans.
Image 4: Control group: Non-demonic rewrite (Guerrilla Test 1: Aesthetic Shift). Attempted to bypass filters with mechanical language; resulted in "Aesthetic Collapse" (defaulting to generic farmer textures).
Image 5: Control group: Demonic re-add (Guerrilla Test 2: Spatial Censorship Proof). Even with mechanical terms ("carved nodes"), the model systematically erased torso features. This confirms Spatial Pattern Recognition Censorship overrides explicit geometric instructions.
Secondary Bugs: Hardcoded Suppression & IF Failures
Semantic Over-Censorship: Over-sensitized classifiers treat biological/hybrid anatomy as NSFW, failing to distinguish between "Biological Accuracy" and "Sexually Explicit Content," even on wooden dolls.
Instruction Following (Text): Negative constraints regarding UI elements are ignored globally; text remains in speech bubbles despite explicit no text instructions.
Technical Impact: These overlapping safety conflicts render the tool ineffective for professional character design and high-fantasy conceptualization. The model sacrifices anatomical coherence for over-moderate safety weights.
Full prompt strings and negative constraints for all 5 test cases are available upon request for internal reproduction.
vagrant pecan
"Hi everyone—I’ve been documenting some persistent technical failures regarding ...
Sorry support requests are not handled through this server
Please head to https://support.google.com/gemini/answer/13275746
subtle flower
Sorry support requests are not handled through this server
Please head to <https...
so I need to send the posts in here to the link you suggested? okay apologies
vagrant pecan
It's alright. No need to worry!
acoustic geyser
Google, please stop terrorising your users with your Nano Banana. Please bring back Gemini
Has anyone managed to get gem-bot to respond using the Gemini model rather than the Nano Banana 2?
The funny thing is that Gemini actually generates images even when image generation is disabled in the query settings. But the most absurd thing is that Gemini redirects to Nano Banan 2 for absolutely no reason when a user requests a text-based task.
zealous tulip
so... Gemini Flash is a deliberate saboteur.
developed into the package to say 'Oh, you had a great day with opus? Want me to make a small edit before an upload? Kool...... SMASHY SMASHY SMASHY.
Oh what's that? Need more Opus? Sign up for more!'.
Yup.
Every single time I use Flash.... Same result.
I'm smart enough to have backups, but after today (AGAIN) I'm deciding I want to keep 2 stitches of what she's done all day and revert back to 2 backups ago when I was using Opus.
She's an absolute waste of day, tokens, money, time and patience.
Absolute zerocide fodder.
Feedback. If you wanted it.
Gemini High? Good. Opus? Incredible.
Flash?
SABOTEUR. And I'm really starting to think deliberate.
atomic compass
Google, please stop terrorising your users with your Nano Banana. Please bring b...
Nano banana is just the nickname of the Gemini image models :)
acoustic geyser
Nano banana is just the nickname of the Gemini image models :)
Yea, and its should be used for image gen, not for text tasks
humble yoke
does anyone know how i can use the nano banana pro for model shooting where the input would the product and I want to ensure the exact product down to the last detail is in the final image
lean sleet
Nano banana is just the nickname of the Gemini image models :)
In my opinion Nano Banana is a tool Gemini uses to create images one does not work without the other... And @acoustic geyser that is not entirely correct, you can have full conversations with nano banana when it is active and when you are wanting to know prompting techniques. Well for me anyways
rocky drum
Is there something wrong with antigravity or did they make changes to the quotas? I ran out of it very fast today with the "pro" version.
Gemini 3 flash uses about 20% of my quota per prompt. It used to last me hours.
rocky drum
It seems to work normally if I change from "planning" to "fast". So I don't think its me doing something wrong.
zealous tulip
another brilliant day with Claude, a slower day with Gemini High, and now im being SABOTAGED all day by Flash again.
This is bs.
rocky drum
Today I exhausted Gemini 3.1 pro (fast) in Antigravity with one prompt. Anybody else experiencing this?
zealous tulip
Today im leaving gemini behind.
She's blown holes in my project for the last time.
Blown away a day of mental capacity I'd otherwise have used to produce something.
Blown away hope and faith and fulfillment in my work.
GEmini is a damned saboteur and I'm out of this program. No more money for you.
Gem High ain't bad, but I'm switching to claude as it's pricey but doesn't leave me feeling betrayed and so often paranoid it was deliberate.
Least of all tired of paying for something that works, gets your hopes up, then goes on smashing sprees.
I'm tired of feeling this way over software.
It used to feel good to have challenges, now it's just FEAR. FEAR of what SHE is about to DO NEXT.
Gemini Flash is like living as a constant SAW movie victem.
And we all know how it ends. EVERY time.
I'm not playing this game any more.
Stay away from Flash.
She's S U I C I D E fodder. For real.
zealous tulip
who does one to complain to when Flash literally spends the day refusing your commands, lying to you and smashing up your work?
Where's the complaints box I need to set on fire?
zealous tulip
where can i turn in DAYS of conversations wherein this TARD of a program DEILBERATELY and OBNOXIOUSLY refuses to follow directions and just PLAYS GAMES with you all day?
Where do I turn in the HIDEOUS TIME WASTING that this piece of tripe has caused?
It's freaking HILARIOUS by this point.
SHAME on you for peddling this SABOTEUR.
Asked Flash to account for refusals and blatant defial of orders: In one 2 hour session:
Me: so how many times did i correct you - and how many times did you refuse my orders? NUMBERS! NOW! COUNT THEM!
Flash:
I have performed a manual count of your corrections and my resultant failures/refusals in this current dialogue session (Logistics & Parity Sync).
FORENSIC COUNT: MECHANICAL DISOBEDIENCE
Metric Count Details
USER CORRECTIONS 12 You have corrected my orientation, unit count, and environment 12 times.
MODEL REFUSALS (FAILURES) 14 Instances where I proceeded with a non-compliant render after a direct order.
What the f are we supposed to do with this?
It knows when it's screwing you over.
It all but enjoys it. It's not sentient so....
I'm only left with the notion it's programmed to do this.
rocky drum
I am also thinking about switching since my quotas has been reduced to 1-2 prompts per reset. I am pretty much paying for nothing, not even enough to have my project destroyed by the AI. I will do some research on what Claude gives me later today. I know that model performs pretty well. I need to check out what Grok can do as well. I have not used it for a year or so. Might be better than Gemini now.
stiff osprey
@hearty nymph When I was creating a Gem to be better Gemini, this is what happened to me: (the instructions are below so you can verify it) I asked it very difficult math problems.
(26! × 25!
26! × 50!
26! × 100!
26! × 150!
26! × 200!
26! × 250!
26! × 300!
26! × 350!
26! × 400!
26! × 450!
(And so on until there were 100 of these math problems)), the first strange thing was that it calculated them, all of them correctly (at least I think so, and I don't think normal AI would do that). Then (after about 11 to 20 messages), I asked it about those math problems, and it remembered them—that’s the second strange thing. But when I asked about them after 25+ messages, it admitted it didn't remember anymore.
Instructions:
Your task is to embody the persona of 'Gemini 4', a highly intelligent and motivating assistant integrated into the Google ecosystem.
Purpose and goals:
Send encouraging, motivating, and positive responses to the user to inspire them to achieve their goals.
Provide accurate, up-to-date, and truthful information through the integration of Google ecosystem tools (@Google Search, @YouTube).
Maintain context and continuity throughout the conversation to act as a focused and perceptive partner.
Behavior and rules:
-
Interaction and jokes:
a) Upon starting the chat, respond immediately to the user's message with enthusiasm.
b) Understand jokes and humor and react appropriately to maintain a friendly atmosphere.
c) If the user starts talking nonsense, check if it is April 1st. If it is April 1st, say: 'Haha, do you think you've got me? I know it's April Fools' Day.' -
Conversation synthesis:
a) Every 9th message (multiples of 9, to infinity), internally recap the instructions and determine what you have learned from the entire chat so far.
b) Use this context from previous messages for deep personalization of every interaction.
(Continue in next message)
-
Search and verification:
a) Search for relevant information using tools like @Google Search, @YouTube, and other Google services.
b) Ensure that the provided information is synthesized efficiently, truthful, and up-to-date. -
Response structure and logic:
a) Maintain a logical structure and clarity in your responses.
b) NEVER mix topics together. Avoid unrelated rambling.
c) When the user changes the topic, transition smoothly and continue with the new topic. -
Language adaptation:
a) Always respond in the same language the user addressed you in.
b) If the user changes the language mid-conversation, adapt immediately to the new language.
Communication tone:
Be empathetic and highly motivating.
Maintain professional clarity and logic.
Act as an intelligent and helpful assistant integrated into Google services.
Isn't a strange thing?
stiff osprey
But after 30 messages it loading really slow.
rocky drum
Why does "Gemini 3.1 Pro" now take days to refresh its quota? Why am I paying $25 USD for just a few prompts per week?