PM on Foundation Models, AI personalization, coding agents @Google. Prev. early Gemini, Google Brain, TPU. Side project https://xitter.cf/t.co/SeUfhnquFt.
SF Bay Area, CA
Joined April 2011
- Tweets209
- Following348
- Followers159
- Likes1.1K
Omkar Pathak retweeted
Lots of discussion out there about our next model(!), so I wanted to give an early look as soon as possible. Introducing Gemini 4 Argon!
It shows frontier performance in complex workflows, cyber defense and software engineering. Teams are using it extensively at Google, from coding to quantum computing, great feedback.
Here’s a look at the benchmarks:
Omkar Pathak retweeted
Gemini 3.7 Flash smashed previous Gemini growth records in its first week, making it our fastest growing model yet. Great to see the huge excitement from our developer community! Now running in Search and @Geminiapp too.
Gemini 3.7 Flash from @Google on ARC-AGI (Verified):
- ARC-AGI-2: 84.6%, $0.25/task
- ARC-AGI-1: 95.5%, $0.12/task
Gemini 3.7 Flash stands out for its low cost and high scores on ARC-AGI-1 and ARC-AGI-2 relative to other frontier models.
Omkar Pathak retweeted
Today we're launching Gemini 3.7 Flash - our latest workhorse model for coding and agentic workflows, with an introductory price at half the original cost of 3.6 Flash. ⚡️
We have been iterating rapidly with the Flash series, going from 3.5 to 3.7 in just 3 months, making it more helpful across a wide range of tasks:
• Software Engineering (DeepSWE v1.1): 37.0% ➔ 65.3%
• Web Development (Code Arena Elo): 1506 ➔ 1588
• Enterprise Automation (AutomationBench): 13.4% ➔ 30.4%
I don't know when this /btw feature was launched in Claude Code, but it is absolutely delightful!
Also, discovering new features in the terminal is somehow more fun than in the app.
For frontier models, model personality is an invisible layer of the UX for consumer apps.
I conducted a 2-part eval of Gemini 3 vs. GPT-5.1 Thinking to see how their internal rewards handle prompt constraints.
Result: Gemini seems 2.3x token-efficient than GPT! Post below
Although GPT 5.2 is out, the idea here is to evaluate the "Thinking" architecture, and the structural bias here seems to persist across incremental model updates.
Same prompt, yet 2.3x token difference!
Part 1: omkarpathak.substack.com/p/p…
Part 2 about steerability coming soon!
Highly recommend it!
The biggest downside of taking @shreyas' product sense course is that I can never again work with anyone who doesn't understand these concepts.
This includes not just co-founders and product leaders, but probably also marketers and sales leaders.
Omkar Pathak retweeted
For all of the most important things, the timing always sucks. Waiting for a good time to quit your job? The stars will never align and the traffic lights of life will never all be green at the same time. The universe doesn’t conspire against you, but it doesn’t go out of its way to line up all the pins either. Conditions are never perfect. “Someday” is a disease that will take your dreams to the grave with you. Pro and con lists are just as bad. If it’s important to you and you want to do it “eventually,” just do it and correct course along the way.
Omkar Pathak retweeted
The PaLM language model paper is now officially published at JMLR.
jmlr.org/papers/v24/22-1144.…
Omkar Pathak retweeted
OPINION: For S.F. to survive the challenges we face, we need to redefine our urban landscape by drawing inspiration from successful cities like Tokyo.
trib.al/RP92WMO
Omkar Pathak retweeted
We collaborated with Bard, an AI experiment by @Google, to optimize your coding workflow.
When you use Bard to help with coding, you can now export Python code to Replit. In your Repl, you can edit or test it to bring your ideas to life.
Try it out → goo.gle/44HGpJH
Installing python on an m2 MacBook has been so painful for me that I am exclusively using Google colab for simple python3 projects. Would love to get a definitive guide that just works.
I’ve tried homebrew and the pip installs did not work!
Omkar Pathak retweeted
Paper: TPUv4 system has an optically reconfigurable network to assemble groups of 4x4x4 chips like legos (4x4x12? 16x16x16?). SparseCores help w/ embeddings. TPUv4 outperforms TPUv3 by 2.1x & perf/W by 2.7x, & has 4096 chips so ~10x faster overall.
.arxiv.org/abs/2304.01433#
Omkar Pathak retweeted
I don’t think a regular person appreciates how insane it is that computers work. I propose we stare at each other mind-blown for about 1 hour/day, in small groups in circles around a chip on a pedestal, appreciating that we can coerce physics to process information like that.
Omkar Pathak retweeted
When you're working on a project—big or small—it's inevitable that you'll find yourself stuck from time to time.
You’re staring at a half-completed page or screen—completely paralyzed.
We've all been there.
Here's my very simple framework for getting unstuck:
