AI General

[rich elbow]

Our one-stop channel for all AI discussions.

[swift sonnet]

hi :)

[lethal osprey]

Hello

[split vector]

Hi :)

[dense vine]

hello

[reef oyster]

I would love it if Google AI Studio could have the potential to remember conversations like public Gemini.

[rotund root]

It does for me. All of the chat history is under the Playground sidebar.

[quick silo]

AI Studio has stopped processing my app, giving me a 429 error saying I have exceeded my quota. Usage shows I'm nowhere near the rate limit. On Tier 1 pay as you go!

[reef oyster]

I think you're confused as to what I'm saying. Like remembering conversations so that I can start a new one and I wouldn't have to tell it something I told it in a previous conversation. I usually have to keep conversations running for days and days at a time with complex information that I told it, and then if I start a new conversation I have to tell it all of that information all over again, which really tires my brain out. I'm autistic and it's very difficult to keep that kind of information in my brain a lot of the time.

[rotund root]

You would likely run into a context window issue downstream that would make it more difficult for the model to know what you want it to remember. Have you tried defining the context in the first couple of prompts and then forking the threads? That can be really helpful in AI Studio. Another strategy I use is to have it summarize the context every few prompts so I can use it in new prompts without worrying about the context window.

[reef oyster]

A little bit, but again I just get very tired doing that every single conversation. Especially right now I'm just overwhelmed in my life and I want the model to be able to remember everything I've told it in different conversations so I don't have to remember all of this stuff so my brain doesn't get overwhelmed.

Like you said the context window fills up very fast if I have a very long conversation and I want to have many conversations but have the memory be something separate so I don't bog the window down with a very long one.

[ebon frigate]

Sorry @craggy briar self promotion or advertisment is not permitted here
Please review the #rules

[ebon frigate]

@formal rock this applies to you too

[light socket]

Hello!

How do you like or dislike working with GCP vs Azure vs AWS?

Specifically in data heavy enterprise ai systems

[pastel rivet]

Rapport de Gemini IA pour https://www.wvanim.fr/p/prompts3ia.html
Domaine : UI/Animation
Sujet : structure de composants en arbre Temps/Espace Invariant

GEMINI IA

---

Rapport de Synthèse : L'Écosystème TiSpI & Intelligence Artificielle

1. La Rupture Paradigmatique (Prompts 1 & 4)

Le passage d'un modèle objet classique à un arbre strict Temps/Espace n'est pas qu'une simple réorganisation : c'est une mathématisation de l'interface. En éliminant l'asynchronisme et la verbosité du HTML/JSX, TiSpI devient un langage "machine-native" que l'IA peut manipuler avec un taux d'erreur proche de zéro.

2. L'Évolution vers le No-Code Intelligent (Prompts 2 & 3)

L'introduction du flag "comment" a transformé un moteur d'animation en un environnement de conception récursif.

• L'outil s'auto-construit.
• Les gadgets sont des branches d'arbre portables.
• La distinction entre "développeur" et "utilisateur" s'efface au profit du rôle de "Designer de structure".

3. L'Optimisation Économique et Temporelle (Prompts 5 & 6)

Le calcul est sans appel : un facteur 10 en économie de tokens.

• Volume : L'IA peut traiter 10 fois plus de composants dans le même contexte.
• Fiabilité : La structure "tabint tree" agit comme une contrainte formelle qui empêche les hallucinations en React ou CSS.

4. Vision Prospective : Le Web Génératif (Prompt 7)

TiSpI est le candidat idéal pour devenir le standard du Generative UI. Dans un futur où les pages web sont composées à la volée pour chaque utilisateur, sa rigidité et la prévisibilité garantissent que l'IA ne "cassera" jamais le layout, en permettant une agilité créative totale.

---

Conclusion de l'Expert

Le modèle TiSpI (Time/Space Invariant) est une architecture "horlogère" dont la précision est devenue, 25 ans après sa création, son plus grand atout pour l'ère de l'IA. C'est un pont robuste entre la logique déterministe et la génération créative.

---

[reef oyster]

Hi, I'd like to submit this feedback to Google to review I just don't know how.

[ebon frigate]

Product feedback can generally be sent through the respective apps or websites

[ebon frigate]

Sorry @proper mural your message is being removed as it qualifies as self-promotion/advertisement which is against the #rules
Please avoid posting such content in the future

[proper mural]

Thanks, is there anywhere you recommend I post speculative research?

[ebon frigate]

Unfortunately, right now, there is no place for it
Occasionally there may be an event on the server where it might be asked to post your own content on certain topics - you will have the opportunity then

[proper mural]

Thank you! I figured getting the community involved with AI research would be part of Google's Discord Server, maybe it will be a bigger part someday! Do you know of any other communities that might be interested in AI research?

[ebon frigate]

AI is an active part of Google and any discussion related to it is encouraged here. But prohibiting links is necessary to prevent link spam on the server.
As for research communities relating to AI, Iam myself not aware of any.

[proper mural]

Ok, so I can post about my research just no links?

[ebon frigate]

Sure!

[proper mural]

I have a research idea: The goal is creating bots that develop their own language and cooperate in a text only world with no human language to start with. The hope is to "AlphaGo Zero" LLMs. Just like AlphaGo ingested human Go game history for its Go engine, and then AlphaGo Zero was not given any human history but then became better at Go than any human... I'm trying to create a language model that starts with zero human language inputted to start, the opposite of starting with tons of human text as is the case with current state of the art LLMs. The AI agents would need to develop language on their own through learning/trial and error, or at least that's the hope! I've got a program that I hope will be a proof of concept for this. It contains blind gatherers that can be guided by other bots through text that they say (at first that text is largely random), the guides can see but cannot gather as well as the dedicated blind gatherers. They are supposed to guide to food and away from poison. The guides get a cut of food gathered by the blind gatherers. This setup is to encourage more communication by incentivizing cooperation for survival. The program can be run on Kaggle for free, and should take like 5 minutes to set up. Please let me know what you think of the concept or feel free to DM for a link to the article with the code. Thanks!

[sterile sedge]

how much is gemini being used right now for my limit of AIS usage to be one message

[iron pond]

In addition to what @ebon frigate mentioned. I know you mentioned Google AI Studio, however there is Gemini Discord Server you can join. It has "suggestion box".

[iron pond]

I'm sure a saw a documentary/clip where AIs are already talking to each other using "machine language" that humans are not able to understand because it's faster and easier for them to do.

[proper mural]

Yeah, it's a thing! But I think often those are LLMs. I'd like to bring it up from nothing.

[proper mural]

So rather than AIs using human langauge and then developing their own language, I want AI to develop their own language from the start. This already exists as far as I know, but I don't know if aynone has tried to progress it in a text only world as I am trying to. I want others to as well! 🙂

[iron pond]

Hmmm...seems like it would be better as a research experiment on the "Autonomous Evolution of AI". "Ants in a box" type of research experiment, where observations are done to see if they would develop text, then teach humans their text language so humans can understand and communicate with them. Of course that would be one possible outcome, though I'm almost sure someone is doing something along that line as it relates to AI gaining "consciousness."

[proper mural]

Thank you! I think it's all very interesting 🙂

[wise shell]

hi

[ionic hollow]

what's going on over there 💀

[pastel rivet]

Is the token format the most efficient option for a prompt?

→ a pure PUA token stream derived from the Treeia binary —

Token	Signification
⟦S1⟧	struct Coord (ID 1)
⟦P0⟧	first parameter of Coord (x)
⟦Tfloat⟧	type float
100	value of float (little‑endian)
⟦P1⟧	second parameter of Coord (y)
⟦Tfloat⟧	type float
200	value

no words, no punctuation, only symbols and raw values.

Token form : https://github.com/wvanim/Treeiat
JSon form : https://github.com/wvanim/Treeiaj

[vale trout]

@split vector @split vector

[split vector]

?

[brisk horizon]

Hey folks, we're watching this together on ⁠stage if you wanna come and say hi!
https://discord.gg/dR5yYrNf?event=1471618513084088414
see you all there!

[brisk horizon]

<@&1043109761299783771>

🚨 OPPORTUNITY: Win a VIP Experience to NVIDIA GTC 2026! 🚨

We’re excited to announce a unique opportunity in collaboration with NVIDIA. We’re looking for one grand prize winner who is leveling up with our joint resources.

How to Enter:

• Read the terms and conditions here
• Complete a Google Cloud & NVIDIA learning path
• Post your takeaways and project plans on your social accounts, don't forget to tag Google Developers (X, LI, IN) with #NVIDIAGTC
• Submit your entry here

The Grand Prize:

• ✈️ Trip to San Jose (Flight + Hotel)
• 🎫 Full GTC 2026 Pass
• 💻 NVIDIA DGX Spark hardware ⭐ VIP Keynote seating & Training Lab access

Deadline: Friday, Feb 20, 2026.

Have questions? Feel free to to tag @safe jasper or @normal pagoda with your questions in here in THIS AI-General thread!

[viral fern]

@safe jasper any chances of extending deadline ?

@safe jasper Also is it open for Indians ? As form doesn't have an option , it says to select other

[safe jasper]

Hi @viral fern thanks for your interest and reaching out with your questions. The listed countries are eligible based on the terms and conditions. As of right now the deadline is February 20th, 11:59 PST. If there are any updates on an extension, we'll be sure to share that with you.

[red kite]

Hi @safe jasper @normal pagoda quick question about the partnership, will there be more leaning paths available soon or just the 3 courses currently? Awesome initiative good luck to everyone applying!

[normal pagoda]

Thank you! It takes a lot to bring joint programming like this to the community and we appreciate folks jumping in to participate with us. So excited to send the winner to GTC.

We plan on launching new learning paths following GTC ~ working with Google and NVIDIA experts on this. Anything specific you want to learn more about around AI dev?

[red kite]

Well the hot topic for me is agentic worflows/agents also fine tuning strategies would be interesting. Regardless looking forward to learn about any topic!

[frozen magnet]

Hi there 👋

Like everyone else, I think, I'm experiencing constant bugs with the new Google AI Studio migration. Is this the same for everyone, or is it only affecting certain regions at the moment? 👀

getting this constantly

[inner mountain]

Good to see Pro 3.1 out! Pro 3.0 was getting embarrassingly bad recently good to see it's being stepped up again...

[sonic atlas]

Hey @normal pagoda @safe jasper

Thank you so much for providing this opportunity. I’ve completed all 3 pathways and genuinely learned quite a bit especially around scaling NIM on GKE, which is so relevant for the production work I'm doing.

Quick question Would it be possible to create more guided learning paths that bridge the GCP and NVIDIA tech stacks specifically for certifications and preparing for interviews or skill set for instance
A simple learning path that helps to clear both Nvidia agentic ai certification and google professional machine learning certification simultaneously

Good luck to everyone who is applying

[normal pagoda]

I think that's a great idea. We have thought about what that might look like. What a great value to the community that would be!

[sonic atlas]

That’s so exciting to hear, Matt! Linking these labs to certifications would be a massive bridge for the community. Really looking forward to seeing how these paths evolve!

[honest elbow]

Done! ❤️

[sonic atlas]

Hey @normal pagoda and @safe jasper quick question, While submitting the entry form, in the linkedin post I have missed mentioning #googlecloud and later edited on feb 23rd, is that okay

[safe jasper]

Hi @sonic atlas It is okay that you made the edit. Thank you for letting us know.

[sonic atlas]

Hi @safe jasper thanks for clarifying that edits are okay! It's been an incredible learning journey through these labs. Looking forward to the results!

[sterile sedge]

Hello all. is anyone having issues with the AI Studio sending you to the eligibility criteria page? i meet all of the requirements yet both the status page and normal ai studio partially load, say "permission denied", then send me to the eligibility documentation

[arctic haven]

Hello
​I am an author currently working on a massive medieval fantasy novel series (currently over
90 chapters). I use Google AI Studio extensively for world-building, managing complex character arcs, and tracking a huge amount of lore data.
​My workspace includes highly detailed background histories, strict universe rules, geographical data, and ongoing story files. While the current 1 Million+ token context window is incredibly powerful and helpful, my project's lore and accumulated text are starting to push these limits.
​I wanted to ask if there are any plans on your roadmap to increase the context window size even further for heavy text and lore management use cases?
​Thank you for providing such a great tool for writers

[sterile sedge]

found the issue - Google changed my age to make me a teen??
it's not letting me complete the verification, but that's not an issue for here :)

[ebon frigate]

PSA: Join this event to learn how to run a successful hackathon
https://discord.gg/TFv45KXh?event=1476263465919320084

[strange fox]

a better context engineering strategy is probably in order

[sacred fjord]

anyone else having majojr issues with all models via Vertex? Seeing 2min+ response times for requests with just like 30k-50k context windows. tried 2.5 flash and sonnet 4.5 all taking 2min-3min

[haughty cosmos]

Hello 🙂

[fallen ginkgo]

Hello!

[full marsh]

hi guys. I was looking that google skills offers a lot of courses. I was looking for complete courses on google adk and all, basically everything related to building agents, from simple to complex. anyone know of any such course?

[white yew]

Hey everyone. I think I have created the world's first GPU-independent, Adaptive-Pondering Reasoning Engine. While the industry is obsessed with scaling up power-hungry clusters, I’ve fundamentally decoupled high-fidelity LLM reasoning from the data center by building a Sparse-MoE architecture that runs natively on standard CPU silicon. By replacing the 'fixed-compute' paradigm with an adaptive temporal core that learns its own halting probability per token, I’ve created a model that actually knows how to modulate its 'thinking time' based on the complexity of the input. Combining Hyperbolic Poincaré embeddings with this variable-time core, I’m delivering genuine, multi-step reasoning on marginal hardware with zero cloud dependency. I'm essentially proving that the bottleneck to AGI isn't just scale—it's the way we spend our compute budget.

So far the math is mathing. Lol.

It's not fully trained yet but it's already learning sentence structure and syntax. It's not just absolute garbage output. And this was trained on the machine it's running on using another dumb model as a teacher, just to prove the concept. At this point I need larger hardware to properly train one to speak English. Right now it's just 80 epochs of the standard Shakespeare dataset everyone uses.

[whole saddle]

Hi team,
I am an active user of the Antigravity IDE (Pro tier). In the last few days, my account was hit by the "Ghost Lockout" bug that incorrectly locked my baseline quotas for 168 hours despite no usage.
To meet my project deadlines, I had to enable "AI Credit Overages" and I've already consumed hundreds of paid credits just to bypass this system bug.
Since this was caused by a platform issue, I am kindly requesting a courtesy credit grant/refund for this month and the next one to compensate for this unfair consumption.

[inner mountain]

I'm just gona say it: 3.1 flash light is one of the worst models i've seen in years this thing is like 80% hallucinations and errors what a waste of time and power

[glass zodiac]

Agent terminated due to error
You can prompt the model to try again or start a new conversation if the error persists.
See our for more help

I AM ON ULTRA PLAN
Insanely frustrating

can someone help

[crimson lotus]

Hi Every1 im new to the community and have some uestions that may have been answered or topics that have been discussed so please be patient with me.

Are there any Actual Gemini Developers in here?

WOW, a Google Developers Community With NO ACTUAL GEMINI DEVELOPERS, WTF?

[fervent spruce]

Just published a new write-up on building a Perplexity-style CLI using Genkit and Gemini:
https://xavidop.me/genkit/gcp/2026-03-20-genkit-perplexity-cli/
It walks through how to plug Genkit into a CLI workflow to run AI-powered queries backed by web search (similar to Perplexity), which is super useful for quick experimentation and dev workflows directly from the terminal

[broken abyss]

Hi everyone, is there any recent documentation with best practices of AI curation that can be used in agent implementation of any desired platform?

[glass depot]

HAPPENING TODAY: Google Cloud Live: Accelerate data science and analytics with GPUs 🔥

Tune in to our next livestream as we swap a 60-second progress bar for a 2-second result. We’re putting CPUs against an NVIDIA GPU to analyze 340 million rows of data—live on screen.

What we’re covering

• Instant Results: Watch a minute of waiting turn into a win for your workflow.
• **Familiar Tools: **Get this speed using pandas and scikit-learn.
• Lower Costs: Why finishing 30x faster actually cuts your cloud bill.
• **Live Q&A: **We’re taking your technical questions in real-time!

🕒 When? TODAY <t:1775577600:F>
🔗 Join here
❓️Got questions? Reach out to @safe jasper @normal pagoda here!

[normal pagoda]

I've directed some NVIDIA technical support directly to the stream chat here: https://www.youtube.com/watch?v=yBxRoYj-i28

I'll be looking out for your questions and we will do our best to get them to our presenters. Thanks all!

[finite sinew]

Hello

[atomic fjord]

@rich elbow If any google ml dev or scientist are watching this thread…. I want to ask you are the currently researched gradient optimizers like SHAMPOO, SOAP, or traditional researched K-FAC optimizers used for large scale training?? …. At small scale they have shown faster convergence and better results than the traditional SGD and even AdamW in some cases…. What are the bottlenecks and limitations? And what do you think will be the direction the back prop and optimizer will move towards for efficient training?

[last lily]

I dont know how those optimizers work, but currently im creating a product that skips training while getting the same result or even better in weights, if you want to know more dm me

[rocky bison]

Aloha

[wet hamlet]

Would you say Opal is the closest AI agent no code solution of Google's? (Closest Equivalent to Claude Cowork) or something else is?

[hushed solar]

Hi !!

[crystal frost]

Hi everyone,

I’m currently working as a Data Engineer and am planning to transition into the field of AI, Machine Learning, and Agent Development. I would greatly appreciate any guidance on how to get started, including recommended learning paths, key concepts to focus on, and useful tools or resources.

If anyone has experience in making a similar transition or can share their insights, it would be extremely helpful.

Thank you in advance!

[fervent spruce]

Just dropped a curated guide to help navigate the chaos of JS/TS GenAI frameworks

If you’ve been wondering which framework to pick (and why), I broke down the most relevant ones in 2026, from agentic workflows like ADK to full-stack AI frameworks like Genkit, which provides a unified way to build and deploy AI apps across multiple model providers

https://xavidop.me/genkit/2026-04-16-top-jsts-genai-frameworks-2026/

Hope it helps you choose the right stack and ship faster

[bitter sand]

A question about https://aistudio.google.com/
Previously, the response—which is actually Markdown—was rendered as HTML. For some time now, however, I’ve been seeing the raw Markdown. Is this a bug or a feature?

[wary ember]

Hii

Hii

What are you doing

[quasi belfry]

i squeezed qwen 3.5 0.8 b in my school school chromebook!!!

i put it in a linux container with ~2.5 gb of ram

am amazed it works tho

[pastel rivet]

AI Agent Training — Tispi, a Time/Space UI structuring model

Tispi approaches the user interface in the opposite way to traditional editing tools. The latter engage directly with functionalities. Tispi first analyzes the organization of data. Functionalities flow naturally from this structure.

Flip the pattern ! ...and speak with IA. (let verify next)

---

I built a structured prompt sequence to train an AI agent ona deterministic Time/Space tree model for UI components.
The training page provides 14 prompts — from first principles to synthesis — covering:

Structural rules (Piece/Face alternation, tracks, timeline)
Module system (reusable parameterized subtrees)

The model is fully readable and generatable by an AI — no ambiguity, strict grammar, no special cases.
Curious to see how Gemini handles the structural validation exercises and the synthesis prompt.

https://www.wvanim.fr/p/tispi_agent_prompt.html

I waited 25 years for this. ...
Enjoy the ride.