Change
9c538ff
9c538ff3aae3c2dea5438ebe739cc83ad8b19892 · commit on GitHub
gcloud-blog-feed: changed (508215 bytes, HTTP 200)
raw/gcloud-blog-feed/response.xml modified
- Source
- gcloud-blog-feed
- Lines added
- +262
- Lines removed
- -151
- Stored bytes at this commit
- 508,215
- Timestamp
- observed
- Raw artifact at this commit
- raw/gcloud-blog-feed/response.xml
Recorded headers
| observed_at | 2026-09-26T04:52:27.719Z |
|---|---|
| origin_date | null |
| status | 200 |
| final URL | https://cloudblog.withgoogle.com/products/ai-machine-learning/rss/ |
| etag | null |
| last-modified | null |
| date | Sat, 26 Sep 2026 04:52:27 GMT |
| age | null |
| cache-control | public, max-age=600, s-maxage=600 |
| cf-cache-status | null |
| content-encoding | gzip |
| content-length | 130822 |
@
@@ -1,5 +1,256 @@<?xml version="1.0" encoding="utf-8"?>-
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:media="http://search.yahoo.com/mrss/"><channel><title>AI & Machine Learning</title><link>https://cloud.google.com/blog/products/a…+
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:media="http://search.yahoo.com/mrss/"><channel><title>AI & Machine Learning</title><link>https://cloud.google.com/blog/products/a…+
<p><span style="vertical-align: baseline;">Now, you can adapt Gemini with the service — teaching the model from a reward signal you define, rather than from a fixed set of labeled answers. This unlocks a class of problems that supervised fine-tuning (SFT) struggles with: tasks that are h…+
<p><span style="vertical-align: baseline;">In this guide, we will walk through practical best practices for using RL fine-tuning service. We'll start with a short tour of the RL training loop, how to decide if and when to use RL, and introduce how to get the most value from this approach…+
<h3><span style="vertical-align: baseline;">What is RLFT? </span></h3>+
<p><span style="vertical-align: baseline;">RLFT adapts Gemini from a </span><strong style="vertical-align: baseline;">reward signal you define</strong><span style="vertical-align: baseline;"> rather than labeled answers. Instead of authoring a large set of gold ex…+
<div class="block-image_full_width">+
+
+
+
+
+
+
+
<div class="article-module h-c-page">+
<div class="h-c-grid">+
+
+
<figure class="article-image--large+
+
+
h-c-grid__col+
h-c-grid__col--6 h-c-grid__col--offset-3+
+
+
"+
>+
+
+
+
+
<img+
src="https://storage.googleapis.com/gweb-cloudblog-publish/images/1_-_Single-Step_RL_Training_Loop.max-1000x1000.png"+
+
alt="1 - Single-Step RL Training Loop">+
+
</a>+
+
</figure>+
+
+
</div>+
</div>+
+
+
+
+
+
</div>+
<div class="block-paragraph_advanced"><p><span style="vertical-align: baseline;">At each training step the service generates multiple candidate responses to your prompts, scores them with your reward, and improves the model so that higher-scoring responses become more likely while …+
<p><span style="vertical-align: baseline;">Three properties define what RLFT can and can't do:</span></p>+
<ol>+
<li aria-level="1" style="list-style-type: decimal; vertical-align: baseline;">+
<p role="presentation"><span style="vertical-align: baseline;">It </span><strong style="vertical-align: baseline;">learns from the model's own outputs:</strong><span style="vertical-align: baseline;"> It refines what the model already produces rather than copying …+
</li>+
<li aria-level="1" style="list-style-type: decimal; vertical-align: baseline;">+
<p role="presentation"><span style="vertical-align: baseline;">It </span><strong style="vertical-align: baseline;">rewards outcomes, not paths:</strong><span style="vertical-align: baseline;"> Any response that reaches a good result earns reward, which fits open…+
</li>+
<li aria-level="1" style="list-style-type: decimal; vertical-align: baseline;">+
<p role="presentation"><span style="vertical-align: baseline;">It </span><strong style="vertical-align: baseline;">amplifies existing competence: </strong><span style="vertical-align: baseline;"> It makes </span><span style="font-style: italic; vertical-…+
</li>+
</ol>+
<h3><span style="vertical-align: baseline;">When to use RLFT</span></h3></div>+
<div class="block-image_full_width">+
+
+
+
+
+
+
+
<div class="article-module h-c-page">+
<div class="h-c-grid">+
+
+
<figure class="article-image--large+
+
+
h-c-grid__col+
h-c-grid__col--6 h-c-grid__col--offset-3+
+
+
"+
>+
+
+
+
+
<img+
src="https://storage.googleapis.com/gweb-cloudblog-publish/images/2_-_RLFT_Approaches_-_Direct_RL_or_SFT_War.max-1000x1000.png"+
+
alt="2 - RLFT Approaches - Direct RL or SFT Warmup to Continuous RLFT">+
+
</a>+
+
</figure>+
+
+
</div>+
</div>+
+
+
+
+
+
</div>+
<div class="block-paragraph_advanced"><p><span style="vertical-align: baseline;">Prompting and SFT handle most adaptation; exhaust them first. RLFT earns its keep when you can </span><strong style="vertical-align: baseline;">grade a response but can't cheaply author it&…+
<p><span style="vertical-align: baseline;">SFT and RLFT are complementary, not competing:</span></p>+
<ul>+
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">+
<p role="presentation"><strong style="vertical-align: baseline;">Direct RLFT</strong><span style="vertical-align: baseline;"> when the base model already succeeds part of the time — enough for the reward to tell better answers from worse ones.</span></p>+
</li>+
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">+
<p role="presentation"><strong style="vertical-align: baseline;">Two-stage SFT → RLFT</strong><span style="vertical-align: baseline;"> when you have SFT data or the base success rate is too low for RL to gain traction. Use SFT as a short, cheap warm start — kept light, since …+
</li>+
</ul>+
<p><span style="vertical-align: baseline;">Across early adopters, these patterns show where RLFT delivers the most value — each scoring an outcome the business cares about but could never cheaply demonstrate.</span></p>+
<h3><span style="vertical-align: baseline;">Use cases for RLFT</span></h3>+
<p><strong style="vertical-align: baseline;">AI-powered NPCs in games</strong></p>+
<ul>+
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">+
<p role="presentation"><strong style="vertical-align: baseline;">What:</strong><span style="vertical-align: baseline;"> In-character, on-brand dialogue held across long, multilingual, multi-turn conversations.</span></p>+
</li>+
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">+
<p role="presentation"><strong style="vertical-align: baseline;">Problem:</strong><span style="vertical-align: baseline;"> Off-the-shelf models break immersion — wrong language, hallucinated items, ignored players, repetitive loops.</span></p>+
</li>+
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">+
<p role="presentation"><strong style="vertical-align: baseline;">Objective and reward:</strong><span style="vertical-align: baseline;"> A </span><strong style="vertical-align: baseline;">Gemini autorater (LLM-as-a-judge)</strong><span style="vertical-alig…+
</li>+
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">+
<p role="presentation"><strong style="vertical-align: baseline;">Results:</strong><span style="vertical-align: baseline;"> Loops and language drift disappeared and state syntax held, making shippable in-game characters viable at scale.</span></p>+
</li>+
</ul>+
<p><strong style="vertical-align: baseline;">Structured entity extraction</strong></p>+
<ul>+
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">+
<p role="presentation"><strong style="vertical-align: baseline;">What:</strong><span style="vertical-align: baseline;"> Pulling a set of items from unstructured documents, such as supplier invoices and shipping manifests, into structured records automatically</span><…+
</li>+
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">+
<p role="presentation"><strong style="vertical-align: baseline;">Problem:</strong><span style="vertical-align: baseline;"> The long tail where SFT plateaus — missing required fields (recall) or inventing ones that aren't there (precision).</span></p>+
</li>+
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">+
<p role="presentation"><strong style="vertical-align: baseline;">Objective and reward:</strong><span style="vertical-align: baseline;"> A </span><strong style="vertical-align: baseline;">rule-based precision/recall reward</strong><span style="vertical-ali…+
</li>+
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">+
<p role="presentation"><strong style="vertical-align: baseline;">Results:</strong><span style="vertical-align: baseline;"> Field-level accuracy rose on noisy real-world documents where tuning had stalled, turning a manual review step into an automated one.</span></p&…+
</li>+
</ul>+
<p><strong style="vertical-align: baseline;">Content moderation</strong></p>+
<ul>+
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">+
<p role="presentation"><strong style="vertical-align: baseline;">What:</strong><span style="vertical-align: baseline;"> Applying intricate policies and decision trees at scale.</span></p>+
</li>+
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">+
<p role="presentation"><strong style="vertical-align: baseline;">Problem:</strong><span style="vertical-align: baseline;"> Models hallucinate false positives or reward-hack with invalid formats to dodge evaluation.</span></p>+
</li>+
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">+
<p role="presentation"><strong style="vertical-align: baseline;">Objective and reward:</strong><span style="vertical-align: baseline;"> A </span><strong style="vertical-align: baseline;">Cloud Run reward</strong><span style="vertical-align: baseline;">…+
</li>+
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">+
<p role="presentation"><strong style="vertical-align: baseline;">Results:</strong><span style="vertical-align: baseline;"> The model handled complex exemption carve-outs, sharply cut false positives, and stopped reward hacking — reducing the human-escalation volume that makes…+
</li>+
</ul>+
<p><strong style="vertical-align: baseline;">Code measured by execution</strong></p>+
<ul>+
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">+
<p role="presentation"><strong style="vertical-align: baseline;">What:</strong><span style="vertical-align: baseline;"> SQL or API calls graded on whether they actually run against customer data.</span></p>+
</li>+
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">+
<p role="presentation"><strong style="vertical-align: baseline;">Problem:</strong><span style="vertical-align: baseline;"> SFT mimics one reference query and breaks on unseen proprietary schemas.</span></p>+
</li>+
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">+
<p role="presentation"><strong style="vertical-align: baseline;">Objective &amp; Reward:</strong><span style="vertical-align: baseline;"> A </span><strong style="vertical-align: baseline;">code-execution reward</strong><span style="vertical-align: bas…+
</li>+
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">+
<p role="presentation"><strong style="vertical-align: baseline;">Results:</strong><span style="vertical-align: baseline;"> The model produced first-attempt executable queries at closed-frontier quality and lower inference cost, letting non-technical users query proprietary da…+
</li>+
</ul>+
<p><strong style="vertical-align: baseline;">Presentation slide generation via HTML</strong></p>+
<ul>+
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">+
<p role="presentation"><strong style="vertical-align: baseline;">What:</strong><span style="vertical-align: baseline;"> Multi-slide decks authored as HTML/CSS.</span></p>+
</li>+
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">+
<p role="presentation"><strong style="vertical-align: baseline;">Problem:</strong><span style="vertical-align: baseline;"> Training on text alone is blind to visual quality — overflows, clipped elements, and inconsistent styling slip through unnoticed.</span></p>+
</li>+
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">+
<p role="presentation"><strong style="vertical-align: baseline;">Objective &amp; Reward:</strong><span style="vertical-align: baseline;"> A </span><strong style="vertical-align: baseline;">code-execution reward</strong><span style="vertical-align: bas…+
</li>+
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">+
<p role="presentation"><strong style="vertical-align: baseline;">Results:</strong><span style="vertical-align: baseline;"> The model emitted modular, well-styled decks with cohesive themes and no layout overflow.</span></p>+
</li>+
</ul>+
<h3><span style="vertical-align: baseline;">Where to start?</span></h3>+
<ul>+
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">+
<p role="presentation"><strong style="vertical-align: baseline;">A dataset.</strong><span style="vertical-align: baseline;"> A diverse set of prompts with a held-out validation split is enough for a first run — confirm the loop converges and reward moves the right way, then s…+
</li>+
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">+
<p role="presentation"><strong style="vertical-align: baseline;">A reward function.</strong><span style="vertical-align: baseline;"> Your task specification as code or configs, and the dominant driver of quality. A good reward correlates with human preference, is robust to ma…+
</li>+
</ul>+
<p><span style="vertical-align: baseline;">The service handles the rest; start from the defaults, watch reward and eval curves in the console, and take the checkpoint where validation reward saturates rather than the last step.</span></p></div>+
<div class="block-image_full_width">+
+
+
+
+
+
+
+
<div class="article-module h-c-page">+
<div class="h-c-grid">+
+
+
<figure class="article-image--large+
+
+
h-c-grid__col+
h-c-grid__col--6 h-c-grid__col--offset-3+
+
+
"+
>+
+
+
+
+
<img+
src="https://storage.googleapis.com/gweb-cloudblog-publish/original_images/3_-_rlft_tutorial.gif"+
+
alt="3 - rlft_tutorial">+
+
</a>+
+
</figure>+
+
+
</div>+
</div>+
+
+
+
+
+
</div>+
<div class="block-paragraph_advanced"><h3><span style="vertical-align: baseline;">Get started today</span></h3>+
<p><span style="vertical-align: baseline;">What will you build? The tools are ready and waiting. </span></p>+
<ul>+
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">+
<p role="presentation"><a href="https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/tuning/reinforcement-tuning"><span style="text-decoration: underline; vertical-align: baseline;">Documentation: Reinforcement Learning Fine-Tuning</span></a></p>+
</li>+
</ul></div></description><pubDate>Fri, 25 Sep 2026 16:00:00 +0000</pubDate><guid>https://cloud.google.com/blog/topics/developers-practitioners/best-practices-guide-for-customizing-gemini-models/</guid><category>AI & Machine Learning</category><category>Developers & Practitioners<…<p><span style="vertical-align: baseline;">As enterprise voice AI evolves beyond basic speed and cost metrics, our priority has shifted to making each interaction even higher quality. Gemini 3.8 Live already delivers a native speech-to-speech foundation for fluid, responsive dialogue. Th…<ol><li role="presentation"><span style="vertical-align: baseline;"><strong style="vertical-align: baseline;">Video avatars: </strong><span style="vertical-align: baseline;">Conversational video with the Live Avatar </span><span style="vertical-align: baseline;">…@
@@ -719,7 +970,7 @@<p><span style="vertical-align: baseline;">Orange's numbers came out of the community work. Building that foundation is the part worth copying first. When you're ready to extend it, </span><a href="https://cloud.google.com/consulting" rel="noopener" target="_blank"><span s…<div class="block-aside"><dl> <dt>aside_block</dt>-
<dd>&lt;ListValue: [StructValue([(&#x27;title&#x27;, &#x27;Get vital board insights with Google Cloud&#x27;), (&#x27;body&#x27;, &lt;wagtail.rich_text.RichText object at 0x7f8290c74e90&gt;), (&#x27;btn_text&#x27;, &#x27;Visit the hub&#x27…+
<dd>&lt;ListValue: [StructValue([(&#x27;title&#x27;, &#x27;Get vital board insights with Google Cloud&#x27;), (&#x27;body&#x27;, &lt;wagtail.rich_text.RichText object at 0x7ff144962b50&gt;), (&#x27;btn_text&#x27;, &#x27;Visit the hub&#x27…</dl></div><div class="block-paragraph"><h3 data-block-key="hswvv"><b>‘Spellcheck for cybersecurity’ and beyond: How Google monitors AI threats and advances AI defenses</b></h3><p data-block-key="61crd"><i>By Sandra Joyce, VP, Google Threat Intelligence</i></p…<div class="block-paragraph_with_image"><div class="article-module h-c-page">@
@@ -922,12 +1173,12 @@<p><span style="vertical-align: baseline;">To learn more about our approach to securing the AI era, please check out our new </span><a href="https://cloud.google.com/security/resources/ai-risk-and-resilience-2026"><span style="text-decoration: underline; vertical-align: ba…<div class="block-aside"><dl> <dt>aside_block</dt>-
<dd>&lt;ListValue: [StructValue([(&#x27;title&#x27;, &#x27;Learn something new&#x27;), (&#x27;body&#x27;, &lt;wagtail.rich_text.RichText object at 0x7f8290b2df10&gt;), (&#x27;btn_text&#x27;, &#x27;Watch now&#x27;), (&#x27;href&#x2…+
<dd>&lt;ListValue: [StructValue([(&#x27;title&#x27;, &#x27;Learn something new&#x27;), (&#x27;body&#x27;, &lt;wagtail.rich_text.RichText object at 0x7ff1448354d0&gt;), (&#x27;btn_text&#x27;, &#x27;Watch now&#x27;), (&#x27;href&#x2…</dl></div><div class="block-paragraph"><h3 data-block-key="4bd61"><b>In case you missed it</b></h3><p data-block-key="cvt34">Here are the latest updates, products, services, and resources from our security teams so far this month:</p><ul><li data-block-key="7…<div class="block-aside"><dl> <dt>aside_block</dt>-
<dd>&lt;ListValue: [StructValue([(&#x27;title&#x27;, &#x27;Join the Google Cloud CISO Community&#x27;), (&#x27;body&#x27;, &lt;wagtail.rich_text.RichText object at 0x7f8290b2d050&gt;), (&#x27;btn_text&#x27;, &#x27;Learn more&#x27;), (&…+
<dd>&lt;ListValue: [StructValue([(&#x27;title&#x27;, &#x27;Join the Google Cloud CISO Community&#x27;), (&#x27;body&#x27;, &lt;wagtail.rich_text.RichText object at 0x7ff144835d10&gt;), (&#x27;btn_text&#x27;, &#x27;Learn more&#x27;), (&…</dl></div><div class="block-paragraph"><h3 data-block-key="29tyz"><b>Threat Intelligence news</b></h3><ul><li data-block-key="2prfo"><b>AI Threat Tracker: From prompting to autonomy</b>: In the newest Google Threat Intelligence Group (GTIG) report on the a…<div class="block-paragraph"><h3 data-block-key="rcfc5"><b>Now hear this: Podcasts from Google Cloud</b></h3><ul><li data-block-key="d67qn"><b>Cloud Security Podcast: Patching browsers with AI, agents, Rust, and your tabs</b>: Jasika Bawa and Dou…@
@@ -1233,7 +1484,7 @@<p><span style="vertical-align: baseline;">Instead of embedding raw Spanner table names and column assignments directly inside core DAO business logic, we isolate Spanner schema translation into dedicated converter units:</span></p></div><div class="block-code"><dl> <dt>code_block</dt>-
<dd>&lt;ListValue: [StructValue([(&#x27;code&#x27;, &#x27;// Example of the standardized pattern generated by our pipeline\r\n\r\ntype BpcTransferAmountsMutationConverter interface {\r\n ToInsertMutation(entity *model.BpcTransferAmount) (*spanner.Mutation, error)\r\n …+
<dd>&lt;ListValue: [StructValue([(&#x27;code&#x27;, &#x27;// Example of the standardized pattern generated by our pipeline\r\n\r\ntype BpcTransferAmountsMutationConverter interface {\r\n ToInsertMutation(entity *model.BpcTransferAmount) (*spanner.Mutation, error)\r\n …</dl></div><div class="block-paragraph_advanced"><p><span style="vertical-align: baseline;">By establishing a rigid, deterministic contract between the DAO and the Spanner SDK (</span><span style="vertical-align: baseline;">spanner.Mutation</span><span style="vertical-ali…<h3><span style="vertical-align: baseline;">Why use Antigravity</span><span style="vertical-align: baseline;"> </span><span style="vertical-align: baseline;">CLI in headless mode?</span></h3>@
@@ -1298,7 +1549,7 @@</ul></div><div class="block-code"><dl> <dt>code_block</dt>-
<dd>&lt;ListValue: [StructValue([(&#x27;code&#x27;, &#x27;git clone https://github.com/google/mantis.git&#x27;), (&#x27;language&#x27;, &#x27;&#x27;), (&#x27;caption&#x27;, &lt;wagtail.rich_text.RichText object at 0x7f82913986d0&gt;)])]&a…+
<dd>&lt;ListValue: [StructValue([(&#x27;code&#x27;, &#x27;git clone https://github.com/google/mantis.git&#x27;), (&#x27;language&#x27;, &#x27;&#x27;), (&#x27;caption&#x27;, &lt;wagtail.rich_text.RichText object at 0x7ff144b42890&gt;)])]&a…</dl></div><div class="block-paragraph_advanced"><ul><li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">@
@@ -1358,7 +1609,7 @@<p><span style="vertical-align: baseline;">Stay tuned for monthly updates on Google Cloud’s AI announcements, news, and best practices. For a deeper dive into the latest from Google Cloud customers, read our monthly recap, </span><a href="https://cloud.google.com/blog/topics/cus…<div class="block-aside"><dl> <dt>aside_block</dt>-
<dd>&lt;ListValue: [StructValue([(&#x27;title&#x27;, &#x27;$300 in free credit to try Google Cloud AI and ML&#x27;), (&#x27;body&#x27;, &lt;wagtail.rich_text.RichText object at 0x7f82909eb950&gt;), (&#x27;btn_text&#x27;, &#x27;Start building …+
<dd>&lt;ListValue: [StructValue([(&#x27;title&#x27;, &#x27;$300 in free credit to try Google Cloud AI and ML&#x27;), (&#x27;body&#x27;, &lt;wagtail.rich_text.RichText object at 0x7ff144d9dc50&gt;), (&#x27;btn_text&#x27;, &#x27;Start building …</dl></div><div class="block-paragraph_advanced"><hr/><h2 style="text-align: center;"><span style="vertical-align: baseline;">July</span></h2>@
@@ -2901,7 +3152,7 @@<p><span style="vertical-align: baseline;"><a href="https://cloud.google.com/ai/financial-services"><span style="text-decoration: underline; vertical-align: baseline;">Gemini Enterprise for Financial Services</span></a><span style="vertical-align: baseline;">…<div class="block-aside"><dl> <dt>aside_block</dt>-
<dd>&lt;ListValue: [StructValue([(&#x27;title&#x27;, &#x27;Get vital board insights with Google Cloud&#x27;), (&#x27;body&#x27;, &lt;wagtail.rich_text.RichText object at 0x7f82909ebc50&gt;), (&#x27;btn_text&#x27;, &#x27;Visit the hub&#x27…+
<dd>&lt;ListValue: [StructValue([(&#x27;title&#x27;, &#x27;Get vital board insights with Google Cloud&#x27;), (&#x27;body&#x27;, &lt;wagtail.rich_text.RichText object at 0x7ff144771b10&gt;), (&#x27;btn_text&#x27;, &#x27;Visit the hub&#x27…</dl></div><div class="block-paragraph"><h3 data-block-key="hswvv"><b>How to stay strong with security fundamentals in the AI era</b></h3><p data-block-key="eoh1k"><i>By Chris Betz, CISO, Google Cloud</i></p></div><div class="block-paragraph_with_image"><div class="article-module h-c-page">@
@@ -2982,12 +3233,12 @@<p><span style="vertical-align: baseline;">To learn more about building and maintaining strong security foundations in the AI era, read our newest </span><a href="https://cloud.google.com/security/resources/cyber-snapshot-reports"><span style="text-decoration: underline; v…<div class="block-aside"><dl> <dt>aside_block</dt>-
<dd>&lt;ListValue: [StructValue([(&#x27;title&#x27;, &#x27;Learn something new&#x27;), (&#x27;body&#x27;, &lt;wagtail.rich_text.RichText object at 0x7f8283ce9f50&gt;), (&#x27;btn_text&#x27;, &#x27;Watch now&#x27;), (&#x27;href&#x2…+
<dd>&lt;ListValue: [StructValue([(&#x27;title&#x27;, &#x27;Learn something new&#x27;), (&#x27;body&#x27;, &lt;wagtail.rich_text.RichText object at 0x7ff144771d50&gt;), (&#x27;btn_text&#x27;, &#x27;Watch now&#x27;), (&#x27;href&#x2…</dl></div><div class="block-paragraph"><h3 data-block-key="4bd61"><b>In case you missed it</b></h3><p data-block-key="5tvtn">Here are the latest updates, products, services, and resources from our security teams so far this month:</p><ul><li data-block-key="4…<div class="block-aside"><dl> <dt>aside_block</dt>-
<dd>&lt;ListValue: [StructValue([(&#x27;title&#x27;, &#x27;Join the Google Cloud CISO Community&#x27;), (&#x27;body&#x27;, &lt;wagtail.rich_text.RichText object at 0x7f8290bcea10&gt;), (&#x27;btn_text&#x27;, &#x27;Learn more&#x27;), (&…+
<dd>&lt;ListValue: [StructValue([(&#x27;title&#x27;, &#x27;Join the Google Cloud CISO Community&#x27;), (&#x27;body&#x27;, &lt;wagtail.rich_text.RichText object at 0x7ff144772e50&gt;), (&#x27;btn_text&#x27;, &#x27;Learn more&#x27;), (&…</dl></div><div class="block-paragraph"><h3 data-block-key="29tyz"><b>Threat Intelligence news</b></h3><ul><li data-block-key="cm2sc"><b>Staying ahead of adversarial AI through agentic source code review</b>: To help defenders implement agentic approaches s…<div class="block-paragraph"><h3 data-block-key="rcfc5"><b>Now hear this: Podcasts from Google Cloud</b></h3><ul><li data-block-key="drbpp"><b>Cloud Security Podcast: All about Project Atlas, Wiz's AI vulnerability research</b>: Near Orfeld, head…@
@@ -3206,144 +3457,4 @@<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;"><p role="presentation"><strong style="vertical-align: baseline;">For developers: </strong><span style="vertical-align: baseline;">Start </span><a href="https://antigravity.google/docs/home" rel="noopener" target="_blank"><span style="text-decoration: underline;…</li>-
</ul></div></description><pubDate>Thu, 20 Aug 2026 17:30:00 +0000</pubDate><guid>https://cloud.google.com/blog/products/ai-machine-learning/expanding-google-antigravity-for-enterprise-customers/</guid><category>AI & Machine Learning</category><media:content height="540" url="https://…-
<p><span style="vertical-align: baseline;">As a fully managed PostgreSQL-compatible database service, </span><a href="https://docs.cloud.google.com/alloydb/docs/overview"><span style="text-decoration: underline; vertical-align: baseline;">AlloyDB</span></a>&…-
<h3><strong style="vertical-align: baseline;">The 10 billion vector scale challenge</strong></h3>-
<p><span style="vertical-align: baseline;">Scaling to a 10 billion vector workload presents significant memory and computational challenges. Previous AlloyDB ScaNN tree-based index was limited to </span><a href="https://docs.cloud.google.com/alloydb/docs/ai/create-scann-index#tw…-
<ul>-
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">-
<p role="presentation"><strong style="vertical-align: baseline;">Increased compute intensity:</strong><span style="vertical-align: baseline;"> Larger tree structures demand significantly more operations for both index construction and query traversal.</span></p>-
</li>-
<li aria-level="1" style="list-style-type: disc; vertical-align: baseline;">-
<p role="presentation"><strong style="vertical-align: baseline;">Memory constraints:</strong><span style="vertical-align: baseline;"> The sampling processes required for 10 billion vectors can easily exceed the system's available memory capacity.</span></p>-
</li>-
</ul>-
<h3><strong style="vertical-align: baseline;">Solution: Four-level architecture</strong></h3>-
<p><span style="vertical-align: baseline;">The introduction of a </span><a href="https://docs.cloud.google.com/alloydb/docs/ai/create-scann-index#four-level-tree-index"><span style="text-decoration: underline; vertical-align: baseline;">four-level tree (preview)</spa…-
<div class="block-image_full_width">-
-
-
-
-
-
-
-
<div class="article-module h-c-page">-
<div class="h-c-grid">-
-
-
<figure class="article-image--large-
-
-
h-c-grid__col-
h-c-grid__col--6 h-c-grid__col--offset-3-
-
-
"-
>-
-
-
-
-
<img-
src="https://storage.googleapis.com/gweb-cloudblog-publish/images/1_fpfICUj.max-1000x1000.jpg"-
-
alt="1">-
-
</a>-
-
<figcaption class="article-image__caption "><p data-block-key="uveoq">Figure 1. AlloyDB ScaNN four-level tree architecture</p></figcaption>-
-
</figure>-
-
-
</div>-
</div>-
-
-
-
-
-
</div>-
<div class="block-paragraph_advanced"><p><span style="vertical-align: baseline;">This design has two primary benefits:</span></p>-
<p role="presentation"><strong style="vertical-align: baseline;">1. Reduced compute intensity via hierarchical partitioning</strong></p>-
<p><span style="vertical-align: baseline;">The four-level architecture drastically reduces compute intensity by using hierarchical partitioning to restrict the volume of vectors scanned during a query. Instead of traversing a flat or poorly segmented space, the multi-layered hierarchy na…-
<div class="block-image_full_width">Diff display stops at 400 lines. The line counts above are from the whole diff. 62 lines shown here cut at 300 characters. The raw artifact at this commit is linked above.