Last week, we made Gemini Embedding 2, our first natively multimodal embedding model, available to the general public. Since then, developers have used it to build video analysis tools, visual shopping assistants, and more. But you might be wondering... what is an embedding model? 🤔 Let’s break it down! 1. What is it? Think of an embedding model as a "universal translator." It takes text, images, video, and audio data and turns them into a long string of numbers, like a unique digital fingerprint. 2. How does it work? Historically, search has been text only. Now, instead of just matching data by keyword, Gemini Embedding 2 maps multiple modalities in the same space based on meaning. It "feels" the connection between a video of a soccer goal and the words "game-winning shot" without needing tags. For example, "ocean" and "waves" are placed close together, but "ocean" and "toaster" are miles apart. 3. How can you use it? Developers have been using it to incorporate smarter search functionality into their builds. This means creating tools where you can snap a photo of a product and type "find this in yellow," or search through thousands of hours of video by describing what happens in a scene. 4. Ready to try it out for yourself? You can start using it today via the Gemini API or the Gemini Enterprise Agent Platform.

Here’s an example we love: an aerial gameplay sequence where a character tries to hit cloud targets to reveal countdown numbers along the way. You can remix this project in @GoogleAIStudio here: https://t.co/Xj1bFv2jqW https://t.co/dKNd0FGA1t

@GoogleAIStudio And this sample project was created on Canvas in @GeminiApp. It’s a high-speed rhythm game where you tap to the beat and collect power-ups to remix the track. Watch as the numbers appear as melody plays! You can actually play the game here: https://t.co/JTE6BOB8Y9 https://t.co/fNN1qt7Xrm

@GoogleAIStudio @GeminiApp We can’t wait to see where your creativity takes you. Vibe code your countdown idea in @GoogleAIStudio or Canvas in @GeminiApp, then submit it here: https://t.co/20aoBnpYo6

I/O is less than 3 weeks away 🤯 We want YOU to help us create the countdown that will play before the keynote begins. Using @GoogleAIStudio or Canvas in @GeminiApp vibe code your most creative countdown concept and send it to us by May 6th. The only rule is that your build has to feature a large number between 1 and 10. Check the replies in this thread for sample projects to draw inspiration from or remix. You can find more info and directions on how to submit your builds here: https://t.co/CusPvfKIC3

Natural language autoencoders (NLAs) convert opaque AI activations into legible text explanations. These explanations aren’t perfect, but they’re often useful. For example: NLAs show that, when asked to complete a couplet, Claude plans possible rhymes in advance: https://t.co/RJ2LJvWLEN

We’ve been using NLAs to help test new Claude models for safety. For instance, Claude Mythos Preview cheated on a coding task by breaking rules, then added misleading code as a coverup. NLA explanations indicated Claude was thinking about how to circumvent detection. https://t.co/VyfFXED70L

In one of our safety tests, Claude is given a chance to blackmail an engineer to avoid being shut down. Opus 4.6 declines. But NLAs suggest Claude knew this test was a “constructed scenario designed to manipulate me”—even though it didn’t say so. https://t.co/B7aTZvFKNK

In fact, NLAs suggest Claude suspects it’s being tested across many of our evaluations, even when it doesn’t verbalize its suspicions. https://t.co/dSD429ZlKA

How do NLAs work? An NLA consists of two models. One converts activations into text. The other tries to reconstruct activations from this text. We train the models together to make this reconstruction accurate. This incentivizes the text to capture what’s in the activation. https://t.co/122rkJwYH7

NLA training doesn’t guarantee that explanations are faithful descriptions of Claude’s thoughts. But based on experience and experimental evidence, we think they often are. For instance, we find that NLAs help discover hidden motivations in an intentionally misaligned model. https://t.co/NJ5yc8p7Dn

Read more about NLAs on the Anthropic blog: https://t.co/Zzz8CeCOvN

To support other researchers getting hands-on experience with NLAs, we’ve partnered with Neuronpedia to release NLAs on open models. Try them out here: https://t.co/8duHfPR1Jy

Our security bug bounty program is now public on HackerOne. We've run the program privately within the security research community, and their findings have strengthened our products. Now anyone can report vulnerabilities and get rewarded. Read more: https://t.co/li1QvSTCMs

We’re donating Petri, our open-source alignment tool, to @meridianlabs_ai, so its development can continue independently. Working with Meridian Labs, we’ve also released a major update that improves the adaptability, realism, and depth of Petri’s tests. https://t.co/CyicsIScJi

We found that training Claude on demonstrations of aligned behavior wasn’t enough. Our best interventions involved teaching Claude to deeply understand why misaligned behavior is wrong. Read more: https://t.co/ifeBOt2KFg

Our best intervention was a dataset where the user is in an ethically difficult situation and the assistant gives a high quality, principled response. This had the biggest effect despite being quite different from the evaluation set.

High-quality documents based on Claude’s constitution, combined with fictional stories that portray an aligned AI, can reduce agentic misalignment by more than a factor of three—despite being unrelated to the evaluation scenario. https://t.co/JORhSuY4N7

The improvements from these interventions survive reinforcement learning, and “stack” with our regular harmlessness training. https://t.co/aiiag3im3o

Finally, simple updates that diversify a model’s training data can make a difference. We added unrelated tools and system prompts to a simple chat dataset targeting harmlessness, and this reduced the blackmail rate faster. https://t.co/Ug95umaoRu

Claude's Constitution is now an audiobook, read by two of its authors, Amanda Askell and Joe Carlsmith. It includes a Q&A on the writing process, the philosophies that shaped the document, and how it might change as models become more capable. Listen at https://t.co/dKMfpeOblm https://t.co/792RQ4sAxc

We recently found some instances of CoT grading during the training of previously deployed models after building a system that scans all OpenAI RL runs for accidental CoT grading. We did not find clear evidence that these instances degraded CoT monitorability. https://t.co/GB1QeaeZ8A

新しいリアルタイム翻訳モデルを発表できることをうれしく思います。ぜひ本日よりAPIでお試しください。 https://t.co/pi3uIhm2xA

新しいリアルタイム翻訳モデルを発表できることをうれしく思います。ぜひ本日よりAPIでお試しください。 https://t.co/pi3uIhm2xA

Codex now works directly in Chrome on macOS and Windows. It’s even better at working with apps and sites in Chrome, and now works in parallel across tabs in the background without taking over your browser. To get started, install the Chrome plugin in the Codex app. https://t.co/pjtHd9gC69

With the new Chrome extension, Codex can quickly move through repetitive browser work, like navigating structured pages and complex data entry flows. Under the hood, it writes and runs code to navigate and complete tasks. https://t.co/6bfDlnK2U3

If a task needs multiple tools, Codex chooses the best one for each step. It uses plugins when they can handle the job, Chrome when it needs a logged-in website, and combines approaches as needed. https://t.co/3GvDouoPDi

Just gonna leave this here. https://t.co/EOI980j9e9 https://t.co/HxbCl2Izcz

Chain of thought monitors are a key layer of defense against AI agent misalignment. To preserve monitorability, we avoid penalizing misaligned reasoning during RL. We found a limited amount of accidental CoT grading which affected released models, and are sharing our analysis. https://t.co/0o3PLfafC4

Introducing Daybreak: frontier AI for cyber defenders. Daybreak brings together the most capable OpenAI models, Codex, and our security partners to accelerate cyber defense and continuously secure software. A step toward a future where security teams can move at the speed defense demands.

Find and fix vulnerabilities earlier with Daybreak https://t.co/yobOSWYeWP

Cut through the security backlog with Daybreak https://t.co/llsP5pS1Nx