Gemini 3 Flash now uses an agentic "think-act-observe" loop to solve complex visual tasks 🤖 Google DeepMind engineer Paul Ruiz demonstrates how the model runs Python code automatically to zoom and inspect items, annotate images, and re-visualize data into charts.

@ptruiz_dev Check out the documentation: Try it in

@GoogleDeepMind @ptruiz_dev AgenticVision .com

I'm starting to see this 'think-act-observe' loop as a way to automate the iterative design process, where the model refines its actions based on the results of its previous decisions, much like how a designer would refine their design based on user feedback. This could unlock new possibilities for interactive visualizations and real-time data exploration. Building something like this for a mobile finance app has given me a similar idea - I'm excited to explore this further.

@GoogleDeepMind @ptruiz_dev It has been fun to see Gemini 3 Flash do things no other model can do. That zoom & inspect loop is fantastic.

I absolutely love Gemini 3 Flash! What makes it so fun model to play with is that it's a true wildcard 🃏 I've often found that it navigates its way through the information space in various unexpected ways and comes up with great novel ideas, while still ripping through the task with a speedy punch! And sometimes you find it almost impossible to pursue a goal with strict instructions as if it faces even a slightest roadblock, it will turn into a frenzied Sonic, bouncing around to get through the obstacle. It's clearly a model made to iterate - fast

@GoogleDeepMind @ptruiz_dev Still can't find the cat 😺

@GoogleDeepMind @ptruiz_dev This just straight out of the box solved problems for me I previously could not solve with LLMs, even with complex flows. Insane work google, thanks! @ptruiz_dev

@GoogleDeepMind @ptruiz_dev Okay, an AI can zoom. Please make my chicken be Einstein.

@GoogleDeepMind @ptruiz_dev @grok fais un tableau qui explique quel utilisation chaque modele de gemini est specialisé est bon

@GoogleDeepMind @ptruiz_dev Thank you

@GoogleDeepMind @ptruiz_dev @grok 总结视频的内容f 总结视频的内容 总结视频的内容a w

@GoogleDeepMind @ptruiz_dev I would try Gemini if you could create a damn api key. GCP/Ai studio, no key works despite enabling the correct services.

@GoogleDeepMind @ptruiz_dev I'm watching it so many questions have

@GoogleDeepMind @ptruiz_dev Does this work on @openrouter ?

@GoogleDeepMind @ptruiz_dev Gemini 3 Flash using a “think → act → observe” loop is a big shift. 🤖 Running Python to zoom, inspect, annotate, and even turn visuals into charts automatically? That’s not just vision that’s reasoning + execution combined. AI isn’t just seeing anymore… it’s doing

@GoogleDeepMind @ptruiz_dev Impressive! Gemini 3 Flash’s ability to autonomously analyze and visualize complex data is fascinating.

@GoogleDeepMind @ptruiz_dev When can gemini in the browser have a "laser pointer" so it can point and draw on the web UI with us while we work. Would be pretty great to have gemini live + laser pointer almost like actual screensharing over zoom with Gemini.

@GoogleDeepMind @ptruiz_dev This sounds great but the demo in your video fails. The resolution is bad, but it's still clear that most pictures have the wrong label: Duck -> Zebra, Pig -> Giraffe, Hippo -> Bear, Crocodile -> Crab

@GoogleDeepMind @ptruiz_dev Great revolution!

@GoogleDeepMind @ptruiz_dev INSTANTLY THOUGHT WHEN SAW LAST EXAMPLE WAS MED TECH USE FOR POTENTIAL CANCEROUS MOLES... ANALYSIS AND GIVE LIKLEYHOOD OF CANCEROUS... MANY USE CASES... THATS WHY AI GOLD RUSH CAN BE FOR EVERYONE...

Gemini 3 Flash's think-act-observe loop is powered by native tool-use + on-device Python execution (via MediaPipe & LiteRT), allowing sub-second reasoning→code→vision cycles without external API calls. In the circuit board demo, it not only zooms and labels but can also run edge-detection filters or count components in real time. Early internal benchmarks show ~4.2× faster multimodal reasoning than Gemini 1.5 Flash on vision-heavy agent tasks. Huge step toward truly agentic on-device multimodal AI. Docs + cookbook examples already live in AI Studio — highly recommend trying the "visual data extraction" templates first.

@GoogleDeepMind @ptruiz_dev When it is out on paid API?

@GoogleDeepMind @ptruiz_dev @AIxC_Official @aixcfoundation #Tenki

@GoogleDeepMind @ptruiz_dev ChatGPT has been doing this for me for quite a while.

@GoogleDeepMind @ptruiz_dev Love this direction — moving from static vision to agentic loops is key. One missing piece is a deterministic COMMIT/NOCOMMIT step before executing Python/tool calls, with both outcomes logged for traceability. Spec + minimal implementation:

@GoogleDeepMind @ptruiz_dev now give it a wallet.

@GoogleDeepMind @ptruiz_dev 🫣

@GoogleDeepMind @ptruiz_dev “think-act-observe” is the new normal. The gap isn’t models anymore. It’s orchestration.

@GoogleDeepMind @ptruiz_dev This loop is the real unlock for vision: don’t “one-shot caption” — iterate with tools. Would love to see a run-log / trace standard here (actions, crops, code executed) so teams can debug + eval these agents like we do with software.

@GoogleDeepMind @ptruiz_dev Grok gemini cause IA PRAY

@GoogleDeepMind @ptruiz_dev when you guys are planning on releasing a stable version of flash, not preview ?

@GoogleDeepMind @ptruiz_dev @grok it's all computer vision?

@GoogleDeepMind @ptruiz_dev google is king!!! fantastic

@GoogleDeepMind @ptruiz_dev The real bottleneck isn't the model. It's knowing what to ask it to see.

@GoogleDeepMind @ptruiz_dev Gemini 3 Flash's new agentic loop turns vision into an active, code-driven investigation.