Google Developers Blog

Do you have plans to do a follow-up model release with quantization aware training as was done for Gemma 3? https://developers.googleblog.com/en/gemma-3-quantized-aware-trained-s… Having 4 bit QAT versions of the larger models would be great for people who only have 16 or 24 GB of VRAM.

IMO the code itself has become much less valuable. Most people in this thread are telling you to stay in the code but I would argue you need to stay current on how to architect a good project. What supporting infra do you need? Did you pick the right language? Did you break the project up into appropriate tasks? You need to become a really great PM. Learn to wrangle your agent better than everyone else. Don't rely on the chat too much, break up your project into tasks, learn to use sub-agents. Learn to use the new tools well. This tool seems obvious but its message is really that what you prompt is profoundly important. https://developers.googleblog.com/conductor-introducing-context-driven…

From the blog post: "more than 99% of them had no activity in the last month" https://developers.googleblog.com/en/google-url-shortener-links-will-n… This is a classic product data decision-making fallacy. The right question is "how much total value do all of the links provide", not "what percent are used".

Wow. I love this. As a not very social language learner, I've used LLM to generate everyday conversation for me to practice. This tool streamlines this experience. I don't think Google will continue the development. It'd be great if someone can create an open-source replica and add some enhancement like history for recap, more scenerio customization, and pre-defined scenerios, etc. Implementation details from Google blog: https://developers.googleblog.com/en/how-its-made-little-language-less…

Here's Google's announcement: https://developers.googleblog.com/en/a2a-a-new-era-of-agent-interopera… I thought John Rush's tweets helped me understand it better. He has several AI agents in the market already. Until now you had to hard wire agents like firecrawl.dev if you wanted to say crawl the web. Now with A2A you can let the AI Agent discover the best solution on its own for the job. I have found sometimes we have a tool that we use as developers that might not be the best choice but its something we know so it comes without the learning curve.

> Their best known gimmick URL is the goo.gl shortener, which is actually the ccTLD for (not) Greece (actually Greenland) rather than a Google-specific one. Not for long https://developers.googleblog.com/en/google-url-shortener-links-will-n…

This is why I love Google's Noto Emoji font: https://developers.googleblog.com/2022/04/what-is-black-and-white-and-…

For what it's worth, I use Mailplane on an M1 MacBook Air (8GB) with 2 Gmail tabs and a calendar tab without noticeable issues. Unfortunately the developers weren't able to get Google to work with them on a policy change that impacted the app [0] [1] and so gave up and have moved on to a new and completely different customer support service. [0] https://developers.googleblog.com/2020/08/guidance-for-our-effort-to-b… [1] https://mailplaneapp.com/blog/entry/mailplane_stopped_selling_licenses… So unfortunately

(Googler, opinions are my own. I also work on backend payment systems there). A fun technical piece here is that this expands on the Google Pay for India app, which is built on Flutter for android and iOS[0]. As well, it's great to see Plex officially announced, though it was talked about ~1 year ago here[1]. [0] https://developers.googleblog.com/2020/09/google-pay-picks-flutter-to-… [1] https://news.ycombinator.com/item?id=21523335

Are you going to put it on github? I bought Flutter course as well (the one from appbrewery recommended on googleblog [0]) but haven't got time to get started with it yet. [0] https://developers.googleblog.com/2019/05/Flutter-io19.html

FYI: I think you can still select blogger themes which don't require JS to render. Some people use them, including (perhaps just as ironically) the Google Developers Blog[0]. [0]: https://developers.googleblog.com/

That's more of a consumer facing blog it feels like. Maybe here? Not sure. https://developers.googleblog.com/

For what it's worth, we're hoping to integrate these APIs into our iOS version of the TensorFlow runtime, so you can maintain graph portability but still get the benefits of the optimized implementation on the platform. We only just got iOS support in TensorFlow last week (https://developers.googleblog.com/2016/06/tensorflow-v09-now-available… but this is something our team would like to get done, so stay tuned!

Thanks for the info, sounds like you have a pretty solid tech stack :) Re Java - If you're looking to maximise compatability, then yeah you should aim to target an older JDK. Virtually all Java projects use at least JDK 8 so that can be a baseline, however many enterprise projects would use closer to JDK 18 at a guess (Google's internally aiming to migrate to 21 in 2024). Generally if there are libraries or features from newer JDKs that you do want to use, I'd say just go for it, since JDK 11, the releases have been yearly (there was a three year gap between JDK 8 and 9) and more incremental. What I would recommend is using Kotlin rather than Java, Kotlin's completely interoperable with Java, but provides a much nicer development experience. That way Kotlin clients get niceties such as named parameters [1] (which with data classes [2] can pretty well replicate StreamdalClient) [1] and Protobuf DSLs [3] and Java clients still get a first class, completely interoperable API. No idea what WASM support is like for Java, I suspect it's lagging behind other implementations, however the most popular framework is Teavm. 1: https://kotlinlang.org/docs/functions.html#named-arguments 2: https://kotlinlang.org/docs/data-classes.html 3: https://developers.googleblog.com/2021/11/announcing-kotlin-support-fo…

The linked post is itself linked to the API in question: Perspective [0]. The site has a demo where you can slide the "toxicity" from left to right and view the resulting bucket of comments sourced from various topics of discussion (Brexit, Climate Change, US Elections). You can also provide your own snippet and get a rating. Obviously this kind of tool will sustain a lot of discussion about what "toxicity" is and how a ML system could infringe on speech by applying bias, either from the engineers building and evaluating the system, or from the inherent bias in data that Google themselves have already identified as a major problem. [1] For what it's worth, Perspective's site offers a definition of "toxic": "a rude, disrespectful, or unreasonable comment that is likely to make you leave a discussion." The training dataset was ostensibly built by having humans rate comments on a toxicity scale based on that definition. It seems as though the idea is to penalize poorly thought-out or inflammatory comments while encouraging in-depth or at least more measured responses. I think we can all agree that that is an attractive system to build: the main question will be if it can be done in a fair way and what latent problems might exist in such a system. [0] https://www.perspectiveapi.com/ [1] https://developers.googleblog.com/2018/04/text-embedding-models-contai…

That is not true. Two days ago lowRISC announced that OpenTitan used in Google Chromebooks will include CHERI: https://lowrisc.org/news/opentitan-earl-grey-2-to-support-cheri-and-pq… Last year, Google announced that their Coral NPU will include CHERI: https://developers.googleblog.com/en/introducing-coral-npu-a-full-stac… There are more cores that companies are actively working on: X730 from Codasip https://codasip.com/solutions/riscv-processor-safety-security/cheri/x7… ICENI from SCI Semiconductor: https://www.scisemi.com/products/iceni-device-family/

Note also: > Unpaid tier (Gemini Code Assist for individuals) and Google One users only: Gemini CLI and Gemini Code Assist IDE extensions will be replaced by Antigravity CLI and Antigravity on June 18th. Please migrate to avoid disruption. Learn more here. https://codeassist.google/products/business https://developers.googleblog.com/an-important-update-transitioning-ge…

I don't think it's the same. It's a similar concept, but Gemma is using just a linear projection, which I assume is a lot faster. The developer guide has more details: https://developers.googleblog.com/gemma-4-12b-the-developer-guide/ Vision embedder (35M parameters): Replaces the 27 vision transformer layers of the other medium-sized Gemma 4 models. Raw 48x48 pixel patches are projected to the LLM hidden dimension with a single matmul. A factorized coordinate lookup (X and Y matrices) attaches spatial location information directly to the input the "single matmul" is the key here, I haven't tried it, but it's probably pretty fast and memory efficient.

https://newsletter.maartengrootendorst.com/p/a-visual-guide-to-gemma-4… (in a link from here: https://developers.googleblog.com/gemma-4-12b-the-developer-guide/, which was linked in the text of the post, but not the linkdump at the end).

The big story here is the encoder-free part, which I still don't fully understand. > Vision: We replaced Gemma 4’s vision encoder with a lightweight embedding module consisting of a single matrix multiplication, positional embedding and normalizations. That's technically encoding, just without using a dedicated model for it like SigLIP? The Developer's Guide elaborates, it's still a 35M layer which I am curious is robust enough. https://developers.googleblog.com/gemma-4-12b-the-developer-guide/ > Small enough to run locally on consumer laptops with 16GB of RAM, it unlocks powerful multimodal and agentic experiences right on your machine. I am assuming that involves quantization, which due to the quality loss makes that statement somewhat misleading IMO.

Gemini CLI is being sunsetted in mid-June and replaced by Antigravity. https://developers.googleblog.com/an-important-update-transitioning-ge…

There is now an Antigravity CLI which will replace Gemini CLI. Gemini CLI is going to be EOLd by June 18th afaik. Antigravity CLI and GUI share the same agent harness, so it might do the same task. Source: https://developers.googleblog.com/an-important-update-transitioning-ge…

Looks like it. Found this: > On June 18, 2026, Gemini CLI and Gemini Code Assist IDE extensions will stop serving requests for Google AI Pro and Ultra, as well as those using it free of charge using Gemini Code Assist for individuals. https://developers.googleblog.com/an-important-update-transitioning-ge…

I was wondering the same and learned the model doesn't know about itself during training [0] [0] https://developers.googleblog.com/closing-the-knowledge-gap-with-agent…

That's an interesting idea. I think ultimately, all AI providers will serve ads with a standard protocol. Something like Universal Commerce Protocol that Google launched a few months ago: https://developers.googleblog.com/under-the-hood-universal-commerce-pr… And the browsers will protect the protocol somehow.

Google's models work quite well on my Android phone. I haven't found a use case beyond generating shitposts, but the model does its job pretty well. It's not exactly ChatGPT, but minor things like "alter the tone of this email to make it more professional" work like a charm. You need a relatively beefy phone to run this stuff on large amounts of text, though, and you can't have every app run it because your battery wouldn't last more than an hour. I think the real use case for apps is more like going to be something like tiny, purpose-trained models, like the 270M models Google wants people to train and use: https://developers.googleblog.com/on-device-function-calling-in-google… With these things, you can set up somewhat intelligent situational automation without having to work out logic trees and edge cases beforehand.

no, this: https://developers.googleblog.com/introducing-a2ui-an-open-project-for…

I could see this being useful for client and patient onboarding in the services and medial sectors respectively. For example: A potential client providing information for a law-firm regarding their grievance. A patient filling out the medical questionnaire prior to their first visit to a medical practice. Rather than having a fully deterministic form, you'd be providing them with forms that adapt to their specific issue. The data can then be intelligently stored both as JSON and a more generic record in an RDBMS. That's just my initial thoughts. Google has a similar project called A2UI: https://developers.googleblog.com/introducing-a2ui-an-open-project-for…

The link is broken https://developers.googleblog.com/litert-the-universal-framework-for-o… is the correct url. Op link has an extra / at the end

Don't forget A2A: https://developers.googleblog.com/en/a2a-a-new-era-of-agent-interopera… We'll see how many of these are around in a few years.