Say hello to the new Interactions API and our first agent, Gemini Deep Research, now available for developers 🤖! The Interactions API is a new unified interface to interact with both models and agents. Our Deep Research agent is also SOTA on many dimensions...

Today we are rolling out our first Gemini Embedding model, which ranks #1 on the MTEB leaderboard, as a generally available stable model. It is priced at $0.15 per million tokens and ready for at scale production use!

I'd love for you to try it out and hear your thoughts. https://t.co/mPMdQGzoRP

If you’re wondering – this is purely a passion project. I built it on my days off and late at night. I love my work building the Gemini SDKs at Google DeepMind and am fully engaged there; I just wanted this tool to exist badly enough to carve out the extra time for it.

I built Glide to walk me through a change step-by-step, aspiring to an experience close to sitting with the original author and having them present their patch.

Even with AI tools guarding against mistakes, a human's deep understanding of the change remains crucial.

“Where do I start? What’s the gist? Do we want this? I have 10 other PRs in my queue. The other reviewer had already approved...” – Sounds familiar?

Thoroughly reviewing pull requests is tough. So I built a tool to make it easier. ✨

The good news: AI isn’t just creating the review burden — it’s starting to help with it too. In a next post, I’ll share what I’m seeing in AI code review tooling.

“Vibe coding” is addictive, it feels so easy and so fast while you’re “vibing”, but it usually becomes a maintenance pain later if you don’t keep your agent on a tight leash — by reviewing thoroughly and limiting scope of a single work item.

And I’ll admit — I’ve done it too. On pet projects, I’ve let the AI agent produce too-large chunks of work, skimmed them, maybe did some QA, and committed. And so far I’ve always ended up rewriting it all later.

Colleagues and friends are feeling it too. I’ve heard things like: “I gave up. I can’t keep up with all those huge, messy incoming reviews.” “Code quality doesn’t matter anymore.”

AI makes it easier and faster to produce code. But reviewing it? Still just as hard. Sometimes even harder. And the code review load keeps growing.

Shortly after ChatGPT happened, I started noticing a change: More PRs. Bigger ones. Sometimes, it felt devs were asking the LLM to write a piece of code, and without thoroughly reviewing it themselves, throwing it over the wall for a code review.

AI is shifting the bottleneck in software engineering — From code authors to code reviewers. 🧵

Say hello to the @geminicli, a local CLI to help you build and maintain software with 1,000 free Gemini 2.5 Pro requests per day : )

AI is making some aspects of this problem better, and some worse. I'll share my observations in a next thread.

I’ve found Graphite's stacked PRs tool useful for breaking a change into small pieces and managing the review. Google's "small cls" document is also a personal favorite:

My favorite contributors (internal or external) make life easier for reviewers by: 1. Providing the right amount of context in PR descriptions. 2. Breaking large changes into smaller, reviewable chunks

Beyond the code, you often don't personally know the PR author. This means you can’t assume expertise, and in rare cases even malicious intent is a concern ((like the Jia Tan XZ Utils incident).