Owen Rees-Hayward

#Java ain't so slow - a tale of subsecond billion-scale similarity search - largely inspired by the mechanical sympathy blog of @mjpt777 owenrh.me.uk/blog/2019/09/1…

Saturn's famous hexagon. This atmospheric phenomenon rotates counterclockwise at approximately 320 km/h, completing one rotation every 10 hours and 40 minutes. The hexagon is approximately 25,000 km in diameter. Remarkably, while storms on other planets change and disappear, this hexagon remains unchanged for at least 40 years.

Introducing EmbeddingGemma, our newest open model that can run completely on-device. It's the top model under 500M parameters on the MTEB benchmark and comparable to models nearly 2x its size – enabling state-of-the-art embeddings for search, retrieval + more.

Today, we introduce BeyondWeb, our synthetic data generation approach which significantly outperforms all open synthetic data and is a key component of our curation pipeline. Many talk about “doing” synthetic data, but generating high-quality synthetic data at scale is extremely difficult and requires getting a ton of details right. We've performed thousands upon thousands of experiments to figure out the best way to generate trillions of synthetic tokens at scale. Datology means the science of data, and rigorously understanding what actually makes a data point useful for learning is key to curating the best datasets. Check out the below thread for some of the scientific insights into what makes synthetic data high-quality!

Great post from G+D Netcetera on how Rama transformed their business. This goes into more depth than the case study I wrote. Highlights: "I inherited a costly problem that plagues most modern backends: architectures built by stitching together dozens of independent components, creating waste and complexity in pursuit of scale... The traditional architecture design made this worse by performing expensive data denormalization at request time... Our team decided to try to flip the architecture by moving computation from request time to change time... Rama introduced a new approach to backend development that felt aligned with the needs of modern systems... This unified approach gave us a path to eliminate redundant infrastructure and dramatically simplify our entire system topology... Our team used Rama to build an incremental content denormalization engine, which is essentially a live, always-up-to-date materialized view of our entire content graph. This goes beyond streaming processing as it provides a persistent, queryable state that updates in real-time... The strongest resistance came from management, and their concerns were completely valid. They weren’t questioning the technical vision; they were questioning the specific execution risks... While I did not have definitive answers to every concern that was raised, I had conviction and a strategy to move us forward... We began to see the transformation gains unfold before our eyes, as content updates that previously took multiple minutes through our Varnish cache now propagated in under a second. System load that previously overwhelmed our infrastructure during traffic spikes simply disappeared... Everyone acknowledged the potential after seeing the progress made, but when it came time for final commitment, belief and accountability proved to be different things. Ultimately, it was my risk to own, which made the decision simple: we would fully commit... Having reached our deadline on-time and not exceeding our budget, I can happily report that the results exceeded expectations. Our denormalization engine now supports hundreds of thousands of page views across 15 web applications and 30 native apps, while cutting AWS costs and infrastructure complexity by more than half... I won’t sugar coat it, the learning curve is quite steep. Although mastering the implementation took a few months, our engineers became productive with Rama in just a few weeks with the gains coming in early and consistently. We did not just hit performance targets; we also built a foundation for a new class of solutions."

"The test demonstrated Keeta Network is capable of over 11 million transactions per second (TPS), significantly outperforming traditional layer-1 blockchains and opening new opportunities for what is possible with blockchain technology." cloud.google.com/blog/topics/fi… < Spanner!

Interesting development, to address the broken economics of the web ... AI Just Hit A Paywall As The Web Reacts To Cloudflare’s Flip share.google/JLiTE9aAHYfxjq…

Nothing to see here 😱 US nuclear weapons agency ‘among 400 organisations breached by Chinese hackers’ | Microsoft | The Guardian share.google/AHe5MsSwugcFx2…

Pretty wild ...

The full MCP experience is now in VS Code! 🚀 From secure auth to prompts, resources, and sampling, MCP is now a complete foundation for building AI agents. @digitarald @TylerLeonhardt @ConnorPeet share all the details on the VS Code blog. Read the full post: code.visualstudio.com/blogs/2025/06/…

Yep totally agree. I like memory, often times I want to explore an idea (e.g. a career decision or something) and then discard it, but right now those memories "infect" all other conversations.

#geometrynodes Transformers gear wormy guy with simple physics. I´ve seen this in my feed so many times past week that had to test some #simulationnodes based rig for it. #b3d Gotta do some destruction practice next as well.

Many people have asked me which library I use for my simulations, and I didn't have an answer until now. Excited to share that I'm releasing algovivo, a JS + WASM library to simulate soft-bodied virtual creatures. demo: juniorrojas.com/algovivo code: github.com/juniorrojas/al… 🧵

Dagster's project Airlift makes it easy to observe and migrate Apache Airflow pipelines to Dagster, bringing value ASAP.

Really cool result from the Databricks research team: You can tune LLMs for a task *without data labels*, using test-time compute and RL, and outperform supervised fine-tuning! Our new TAO method scales with compute to produce fast, high-quality models. databricks.com/blog/tao-using…

🥁Introducing Gemini 2.5, our most intelligent model with impressive capabilities in advanced reasoning and coding. Now integrating thinking capabilities, 2.5 Pro Experimental is our most performant Gemini model yet. It’s #1 on @lmarena_ai leaderboard. 🥇

Use partner-ready campaigns that drive results!🚀 @qodea_official leveraged the Google Cloud AI & Data Trends Report in Partner Marketing Studio to engage key decision-makers and boost website traffic, ad clicks, & qualified leads. Watch to discover how → goo.gle/3FBH8oH

Definitely a paradigm shift we're still learning to navigate intelligently as an industry. Knowing when and how to use AI tools effectively is becoming essential.

🔎 Meta Open Source 101🔎 Pysa is a static analysis tool designed to detect security issues in Python code. It works by analyzing data flow to identify vulnerabilities. Start by integrating Pysa into your CI/CD pipeline to add security checks. Link: pyre-check.org/docs/pysa-basi…

Super excited about our new partnership with @AnthropicAI and native availability of Claude 3.7 Sonnet in Databricks on AWS, Azure and GCP! Stay tuned for more integrations and great support across the entire agent development and MLOps stack. sprou.tt/19muXYu3ZuR

Starting a new blog series next week called "Next-level backends with Rama" where each post is a detailed tutorial of using Rama for a specific use case, with all code in both Java and Clojure. The first post will be about storing and traversing graphs.