
SWE @Google, Ex GNSS @Qualcomm, Ex Speech @Microsoft
Joined January 2010
- Tweets306
- Following221
- Followers59
- Likes61
Mahadevan retweeted
Introducing Gemini 4 Neon! Pushing the frontier across domains and punching above its model class in a meaningful way. To me, Neon feels like the right balance of intelligence, speed, and usability. Available now across @GoogleAIStudio, the API, and more!
Mahadevan retweeted
We're launching Gemini 4 Hydrogen to all paid users on Antigravity (will expand to more users over time). It’s a great model - faster, more efficient, and better than Gemini 3.8 Flash on most evals.
We’re also launching Orchestrator in the Antigravity dropdown which allows both Carbon and Hydrogen to work to complete your tasks. We’ve found that it consumes 24% less quota and gets the same tasks done 19% faster. Using Hydrogen with both Orchestrator and swarms have become very popular internally and excited for all of you to try.
Mahadevan retweeted
Introducing Gemini Decision 1, our new frontier decision model in the Gemini API.
It is $0.10 per 1M input tokens with no additional cost. With support for image, audio, video, and PDF natively, 256K context window, & 1K questions.
You can try it for free in the API!
Mahadevan retweeted
Gemini 4 Carbon is now available for Ultra users in Antigravity (and yes, we will quickly expand access to more users). We are also launching /swarm as a native experience to leverage multi agents to complete complex tasks.
Swarms have been a step function improvement for capability internally, especially with Carbon.
Mahadevan retweeted
Guy on one of the aus subreddits asks why his "high growth" retirement account has been stuck at the same number for 5 months despite regularly putting money in, and gets this response:
Mahadevan retweeted
I was expecting this to get freaky, but this turned out to be way funnier than I imagined. 😂🤎
Mahadevan retweeted
I’m still half in disbelief
We asked Gemini Argon to teach the Apollo humanoid to play cornhole, and it zero-shotted it
All we gave it was a video of a person doing it. It then took control of the robot, generated its own sim env, performed some sysID, and trained launches on the real robot.
It took it a few real shots to improve and then land many consecutive airmails.
The process was so long and articulated that we asked the robot itself to explain what it did to us, as you can see at the end of the video. It even came up with the gestures
Great demo for a great model @googledeepmind @apptronik
Mahadevan retweeted
human: "god, why did you do that to integer multiplication?"
god: "do what?"
human: "nlogn. didn't you craft math to be elegant? beautiful?"
god: 'the fuck's a log?"
Mahadevan retweeted
this benchmark reminds me of this hilarious hypothetical posed on reddit three years ago that sparked wars in the chess community.
I have a Continuous Learning benchmark where models attempt to learn to play chess. They are given a /goal of learning and improving playing against a Stockfish opponent in 200 games. They can choose the difficulty, take notes, whatever they like - except cheating (e.g. using a chess engine) of course.
So far the improvement in Elo has been negative for Astra. Tiny bit positive for Opus, but could also be random. I've started Astra off sooner, so it finished its 200 games already, Opus is still playing.
Site here to watch how they are doing: ai-learning-to-play-chess.su…
The reason why this is interesting is that while we obviously don't have continuous learning, at the back of my mind I was thinking that maybe models can simulate it through self-scaffolding. Turns out not so much at least in this context. Perhaps it's a solvable problem and we don't need 'true' self-learning for models to learn in some way.
-------
Just a note, the idea for the benchmarks belongs to someone else, but I don't want to use their name to give this more weight without permission.
Mahadevan retweeted
Ramanujan wasn’t skipping proofs. He was hiding his chain of thought to prevent distillation.
Hardy: “Show your work.”
Ramanujan: “Sorry, reasoning tokens aren’t exposed via the API.”
Mahadevan retweeted
A Bulgarian streamer got matched with Magnus Carlsen, the world’s most famous chess player, and took it in a completely normal and measured way.
Then he won the game.
The celebrations were, naturally, very modest.
🇧🇬
Mahadevan retweeted
CC the entire fam 🤝!
Today, we’re announcing the new CC – an AI agent built for families to spend less time on logistics and more time together.
You can now:
👤 Add up to 5 members to your CC agent
☀️ Start mornings aligned with a shared "Your Day Ahead" brief email
🗓️ Autosync schedules & to-dos with a shared Google Calendar and Tasks
💬 Coordinate in Google Chat with CC to offload relevant tasks (ie., crafting weekly meal plans, school supply shopping lists, etc)
📝 Delegate paperwork (ie., permission slips, forms, and more) for CC to complete under your direction
📌 Keep tabs on the details – CC remembers what applies to everyone (ie., family grocery lists, favorite restaurants) versus what applies to one person (ie., dietary restrictions, local timezones)
Ready to keep everybody on the same page? Join the waitlist or upgrade your existing CC (US only, 18+): labs.google/cc
Mahadevan retweeted
New numbers are up on ProgramBench: Gemini 3.5 Flash and 3.6 Flash.
ProgramBench gives an agent a compiled binary + its docs and asks it to rebuild the entire program from scratch. No source, no decompiler, no internet.
3.6 Flash is Gemini's best showing yet. 🧵
Mahadevan retweeted
Infinity collapses to an integer via Ramanujan's nested radical:
3 = √(1 + 2√(1 + 3√(1 + 4√(1 + ⋯)))).
Similar successive nested-radical expansions compute high-precision trigonometric constants inside GPS satellite orbit determination software.
Mahadevan retweeted
Three fundamental integrals stand out for their elegant closed forms despite spanning infinite domains.
- The Gaussian ∫_{-∞}^{∞} e^{-x²} dx equals √π.
- The Lorentzian ∫_{-∞}^{∞} 1/(x² + 1) dx totals π.
- The sinc function ∫_{-∞}^{∞} sin(x)/x dx also equals π.
These identities, established through techniques such as contour integration, link analysis across mathematics.
Engineers apply the sinc integral within the sampling theorem to perfectly reconstruct continuous-time signals from discrete samples in digital communications systems.




























