Vishal Dharmadhikari (@Vish_owl) header
Vishal Dharmadhikari (@Vish_owl) profile picture

@GoogleDeepMind @GoogleAIStudio, Gemini API My views and opinions

Joined May 2009
Photo by Vishal Dharmadhikari (@Vish_owl)Photo by Vishal Dharmadhikari (@Vish_owl)Photo by Vishal Dharmadhikari (@Vish_owl)Photo by Vishal Dharmadhikari (@Vish_owl)Photo by Vishal Dharmadhikari (@Vish_owl)
Bring your memes to life w/ Carbon!
The people have spoken! Gemini 4 Carbon is the best AI at 3D printing memes. Go and vote for your favorite prints in Wojak-Bench 3D, link below.
2
185
ICYMI, we launched A LOT of things last week
Last week at @GoogleDeepMind was busy, we shipped 7 developer launches. 🧠 Gemini 4 Carbon ⚡ Gemini 4 Hydrogen ⚖️ Gemini Decisions API + Gemini Decision 1 📦 Self-hosted sandboxes for Managed Agents 📊 Multimodal Knowledge Agent benchmark 🤖 Gemini Robotics 2.5 🧬 AlphaGenome 2
2
11
514
Really interesting work by elie building on top of @dwarkesh_sp's Blood on the Clock Tower benchmark! I may be biased but one of the coolest parts: "this experiment was done by gemini carbon in near autonomy"
i added 5 open models to @dwarkesh_sp's blood on the clocktower benchmark to see how they perform, and look at their chains of thought while they lie, bluff and try to catch each other 👁️ kimi k3.1 pro is the best open model and lands in the middle of the closed ones, just bellow gpt 5.5. also a pretty good sign to see a coherent ranking for open models as well, with like the newer deepseek v4.1 flash beating the older one and glm 5.3 beating glm 5.3 flash one thing i wanted to test and seems to be true is that models that are good at the game are also good at spotting the evil players when they only read the public transcript of a game. would be actually very nice for evaluation since a new model could just be ranked by reading existing games? main reasons i wanted to do this experiment is that it's always fun to read the chains of thought of open models, and these ones don't disappoint 😭. some models like glm 5 where super happy to be evil (this model is really crazy in its chain of thought), a bit more concerning but coherent with what you can observe on other environments, there is A TON of eval awareness. i'm also wondering if this is correlated to the model detecting it's reading llm generated text this experiment was done by gemini carbon in near autonomy (brainstorm a bit with it first and then just asked to exectue), and the article and the visuals are by fable 5.2. here is the blog, with all traces (probably a lot more gold there) and a viewer for every game! eliebak.com/open-clocktower-…
1
1
3
964
Decisions, decisions... Try out our new decision model today!
Today we’re launching the Gemini Decisions API, powered by our new Gemini Decision 1 model in @GoogleAIStudio. Instead of generating text token by token, it’s built for closed-set choices scoring up to 1,000 questions across text, images, audio, video + PDFs in a single pass. A few details: - 3 decision types: yes/no, multiple choice (up to 255 options) + rubric scoring - $0.10 / 1M input tokens with $0 billed for output tokens (~5x cheaper than Flash-Lite) - leads on several benchmark evals with lowest calibration error Google teams are already building with it, below is a look at Stitch Rapid designing a full product flow in seconds (1x speed)
3
3
16
420
🫡
so people are really enjoying rune, seeing a lot of cool apps built with it but i have bad news a big chunk of it was built with Gemini Carbon what now haters WHAT NOW???
7
303
There are also some big improvements in agent behavior with Carbon! Cheating and hallucinations rates are much lower and it's persistent in all the right ways.
Gemini 4 Carbon is rolling out today. It’s built for sustained, long-horizon tasks, reasoning deeply through complex software engineering, cybersecurity, and scientific discovery. Coming globally via the Gemini API in @GoogleAIStudio and Gemini for business, and for Ultra subscribers in @antigravity, the @GeminiApp, and AI Mode in Search.
1
16
405
Gemini pushing the frontier of formal verification!
Announcing a major milestone achieved by my team @GoogleDeepMind: we have proven key security properties of the machine code underlying the Signal messaging protocol. This is the first time the actual production code of a widely deployed protocol has been formally verified – by any method, with or without AI – and at this depth. More than a research breakthrough, it’s a prime example of how frontier AI can be a force multiplier to counter the speed and scale of modern cyber threats. Formal verification offers a higher standard of security; however, it’s extremely complex and time consuming, which is why in many cases it’s considered to be impractical. Similar efforts could take teams decades of combined effort. We did it in a month’s time. Using AlphaProof Code, Google DeepMind's new agentic framework for formal verification that runs on LLMs including Gemini Carbon, we verified the Signal protocol down to the actual code running on devices.  This milestone lays the groundwork for a future in which vital systems, from medical device firmware to power grid software, ship with mathematical proof that their protections work. Reaching that future will take continued research from our community. In that spirit, we're open-sourcing our formal specifications and verification artifacts so others can build on this work. On a personal note, this moment is of deep significance to me. I started my own research journey more than two decades ago as an intern working on checking the consistency of complex software systems. Seeing this grand challenge made practical with modern AI is extraordinary. I'm extremely proud of my team members: Swarat Chaudhuri, Andrew Ferraiuolo, Anton Kovsharov, Tej Chajed, Sam Schlesinger,  Karn Seth, Jonathan Katz, and Charles Sutton for taking on this grand challenge and delivering, and grateful to the Signal team for their partnership. deepmind.google/blog/formal-…
2
12
397
Vishal Dharmadhikari retweeted
Gemini
809
280
415
5,937
846,276
Idea to (m)app in minutes!
you can now build with google maps platform in AI Studio 🗺️📍 go from idea to app powered by @googlemaps in minutes! what’s new: - 1-click setup - create maps apps just by describing your idea, gemini does the rest - build with map tiles, places, weather, routing, 3D buildings, street view, and more try it out today: ai.studio/build
1
7
257
Vishal Dharmadhikari retweeted
Automated computing with @trycua x @GoogleAIStudio 🚀 Use @antigravity or a custom SDK 🖥️ Drive desktop apps 🧠 Powered by Argon Check out the guide by @francedot ⬇️ aistudio.google.com/learn/pa…
2
5
14
1,320
Gemini for star discovery!
I found a weird star with Argon's help. Gaia DR3 4582145718700356096. It's normal in visible light, stable and similar to our sun. In March 2010 it appeared ~60x brighter than expected in deep infrared heat (22 microns). As if something around Mars-distance is catching 1.3% of the star's light and radiating it as heat. Probably a huge cloud of debris from a major planetary collision. Or a radio-silent galaxy hiding behind it. But also the tiniest chance it's a technosignature. Captured by NASA's WISE telescope, no observations at that wavelength since. Needs more data.
4
299
Gemini for discovery! Really impressive what folks have pulled off with this model so far
Gemini 4 Argon is already unlocking new scientific breakthroughs in quantum computing, nuclear fusion, protein design and beyond. 🧵
7
377
My weekend plans Argon! Time to cook up some personalized games
Gemini 4 Argon is AMAZING at making games 🎮 It can create 3D assets from scratch and build amazing mechanics, all in one shot. Here's a thread of games I made with prompts and code so you can play or remix them too! 1/6 Argon Tennis: use your phone as a tennis racket 🧵
6
347
🗣️🗣️ love to hear it!
I was invited to test Gemini 3.8 Flash TTS. This is currently my favorite model. The voice outputs from this model are extraordinary. I never really rated AI voice generation because of the uncanniness associated with low quality. This Gemini IMO crosses that chasm, like Opus 4.5 did with coding, where suddenly the quality is high enough that it feels materially different and the product options are wide open. Generated audio is completely listenable, without sacrifice, throughout a product, with insane control over accent and attitude, for hours. Not just gushing, I have honestly generated so much audio with this, it's amazing. Try it!
1
183
Gemini 3.8 Flash to the Mun!
We gave @nunudotai early access to Gemini 3.8 Flash and they took it to the Mun 🚀 As part of their upcoming gaming benchmark, they tasked several AI models to reach the Mun in the game Kerbal Space Program. Rather than build the rocket piece by piece, 3.8 Flash reverse engineered the spacecraft file format, wrote the code to generate an optimal rocket, loaded it in, and piloted it to a smooth landing on the Mun via text commands. Check it out here:
Readers added context they thought people might want to know
Gemini did not reverse engineer the rocket. The vehicle shown is a pre-crafted in-game model recreated by editing its .craft file, with piloting via the uncredited kOS mod. x.com/DJSnM/status/2… x.com/PxvidsO/status… x.com/PxvidsO/status…
2
16
1,133
3.8 Flash is live! What I really like about this model: it does smaller units of work per step and verifies / tests its work a lot more. Token usage goes up on harder tasks, but the overall quality is much better
Introducing Gemini 3.8 Flash, another jump in Gemini's agentic + coding capabilities, and our 3rd updated Flash model in only 6 weeks... This model has been a ton of fun to work with, excited to see what you all think!
1
15
469
Excited for agentic video understanding to launch in the Gemini API! Early feedback shows this unlocks video use cases that were previously blocked due to cost or complexity, like indexing massive video archives and automated video editing in a single API call. Instead of static frame rate ingestion, Gemini now runs an inner agent loop with native tools (get_frames, get_transcript, and get_audio). It dynamically scans and inspects only the moments needed, cutting token usage and costs while also improving accuracy
5
3
58
2,171