Hi, I'm Luis 👋

Today we’re taking another big step on the path toward AGI and releasing Gemini 3.

It’s the best model in the world for multimodal understanding and our most powerful agentic and vibe coding model yet, delivering richer visualizations and deeper interactivity — all built on a foundation of state-of-the-art reasoning.

Today, we are introducing Google Antigravity, our new agentic development platform. While the core is a familiar AI-powered IDE experience with the best of Google’s models, Antigravity is evolving the IDE towards an agent-first future with browser control capabilities, asynchronous interaction patterns, and an agent-first product form factor that together, enable agents to autonomously plan and execute complex, end-to-end software tasks.

The originals of these books are in the Cornell University Library.

“I think that it paves the way for a bunch of different, maybe ultraportables, maybe more powerful laptops being Arm-based,”

Exactly. A true mobile computer would need an ARM chip.

OneXSugar is a step in the right direction, but they're not using the X-series and the OS is Android. Not to mention the form factor is still focused on gaming.

Lapatine says his goal has always been to operate with transparency. He wants Stereogum to feel like talking to a friend who goes to shows and tells you about cool stuff on Bandcamp. Ultimately, he wants to build a connection with readers, help them find good music, and do it with personality. That human element is key because, he says, “I’ve never discovered anyone from the algorithm.”

This series lives on my blog but has nothing to do with me. It exists to connect you, the human who’s reading this, with all the other wonderful humans that are still out there, spending their time making sure the old school web, the one made by the people, for the people, is not dying.

This is the web many people are missing, a web that is, in fact, still here, very much alive. Y’all are great.

Friday mornings are one of my favorites because I get to "meet" someone new through this series. Thanks for your work and dedication to this series.

I’ve updated Micro.blog with support for Day One journals.

This is cool!

The purpose of beets is to get your music collection right once and for all. It catalogs your collection, automatically improving its metadata as it goes using the MusicBrainz database. Then it provides a bouquet of tools for manipulating and accessing your music.

Inspired by WikiTok, I thought I'd make something to discover sounds uploaded to Wikimedia. From political speeches and bird noises to genuine bangers, it's mostly wholesome, though I cant guarantee it won't play you something horrible once in a while.

Rumored dates for the company’s first trifold have come and gone with the latest suggesting a December 5th launch in Korea, priced around $3,000.

$3k?! That's a high-end laptop. I'd consider paying that much for a true mobile computer.

Today, we’re taking the next step in building helpful experiences that keep users safe with Private AI Compute in the cloud, a new AI processing platform that combines our most capable Gemini models from the cloud with the same security and privacy assurances you expect from on-device processing. It's part of our ongoing commitment to deliver AI with safety and responsibility at the core.

Private AI Compute is built on a multi-layered system that is designed from the ground up around core security and privacy principles:

• One integrated Google tech stack: Private AI Compute runs on one seamless Google stack powered by our own custom Tensor Processing Units (TPUs). World-class privacy and security is integrated into this architecture with Titanium Intelligence Enclaves (TIE). This design enables Google AI features to use our most capable and intelligent Gemini models in the cloud, with our high standards for privacy and the same in-house computing infrastructure you already rely on for Gmail and Search.
• No access: Remote attestation and encryption are used to connect your device to the hardware-secured sealed cloud environment, allowing Gemini models to securely process your data within a specialized, protected space. This ensures sensitive data processed by Private AI Compute remains accessible only to you and no one else, not even Google.

Once you get beyond toy usage, LLM governance is a pain. That's exactly why we built Pydantic AI Gateway (PAIG)

What's included

• One key, many models: talk to OpenAI, Anthropic, Google Vertex, Groq, or AWS Bedrock with the same key. More providers (notably Azure) are on the way.
• Cost limits that stop spend: set daily, weekly, monthly and total caps at project, user, and key levels.
• Built-in observability: every request can be logged to Pydantic Logfire or any OpenTelemetry backend.
• Failover: route around provider outages automatically.
• Open source & self-hostable: AGPL-3.0 core, file-based config, deploy anywhere. Console and UI are closed source.
• Enterprise-ready: SSO via OIDC, granular permissions, and Cloudflare or on-prem deployment options.
• Low latency: Finally, and perhaps most importantly, PAIG runs on Cloudflare's globally distributed edge compute network, meaning absolutely minimal latency. If you're sitting in Berlin making a request to a model near Frankfurt, you'll connect to a PAIG service running in Berlin. After the first request, it won't need to call back to our central database, so using the gateway will add single-digit milliseconds to your response time. It might even be faster, as the request will run through Cloudflare's backbone to the model provider.

One of the longest-running and most well-known podcast listening apps, Stitcher was officially shut down by SiriusXM on August 29, 2023. The app originally launched in 2008, and was famous for its ability to "stream" audio (without the need to manually sync audio files to your device).

While my first podcast "app" was Zune, Stitcher was the first third-party app I used on a smartphone. Good memories. RIP Stitcher

With less involvement from the listener, less intentionality, it’s easy for music to simply become background noise. Research suggests that younger generations are discovering new artists at a lower rate than previous ones, and there’s a direct correlation between lower levels of music discovery and reliance on the algorithm. So, how do you find new music with intention?

The simplest and most traditional is through music media. There are countless sources of excellent music journalism and criticism, it’s just a matter of finding one that speaks to you.

Just like everything else on the internet, music has its influencers.

Sort of straddling the line between traditional music media and influencers is the world of Substack. There are tons of newsletters out there that serve up artist interviews, curated playlists, and deep dives into bands’ discographies.

Traditional radio can still be a solid source of new music if you turn away from large commercial stations owned by iHeartRadio and the like. Your local college station is a good place to start, but there are other non-commercial independent stations out there, like WFMU in New York, that steer clear of your standard pop fare.

In addition to traditional terrestrial stations, there are streaming internet stations. NTS Radio is one of the best...

...you should become a fan of labels, not just artists.

Ultimately, record labels are curators, and the best ones have a distinct identity. That doesn’t necessarily only mean releasing death metal records (though it can). It means building a vibe and a particular point of view that unifies the artists under a single umbrella.

...easily the best way to find new and interesting music is just to connect with other people who share your tastes. This can be as simple as getting recommendations from friends or an older sibling. (Or, if you’re trying to stay on top of things as you get older, a younger sibling.) But there are plenty of online communities for getting music recommendations.

Lastly, go see live music. Go see a band you know, but make sure you get there for the opener.

Great list. I've had success with all of these. Some examples.

• NPR's Tiny Desk Concerts is how I discovered Chicano Batman
• 107.1 Chicago Independent Radio Project (CHIRP) is how I discovered Lingua Ignota. Here's a list of other radio stations I listen to
• Salami Rose Joe Louis was an opener for Flying Lotus.
• As a Flying Lotus fan, I also follow many of the artists signed to his label, Brainfeeder

As I discover music, I try to curate my findings through playlists

Get inspired by what you can do with Claude

I figured since I deleted all my social media profiles, I might as well have one space for everything I’m doing.

Today, we are incredibly excited to announce the launch of the new Google Colab extension for Visual Studio Code.

Project Suncatcher is a moonshot exploring a new frontier: equipping solar-powered satellite constellations with TPUs and free-space optical links to one day scale machine learning compute in space.

Reading existing code is the one of the biggest, most expensive bottlenecks in software development.

To address this issue and improve your productivity, we're introducing Code Wiki, a platform that maintains a continuously updated, structured wiki for code repositories.

Hierarchical Reasoning Model (HRM) is a novel approach using two small neural networks recursing at different frequencies. This biologically inspired method beats Large Language models (LLMs) on hard puzzle tasks such as Sudoku, Maze, and ARC-AGI while trained with small models (27M parameters) on small data (around 1000 examples). HRM holds great promise for solving hard problems with small networks, but it is not yet well understood and may be suboptimal. We propose Tiny Recursive Model (TRM), a much simpler recursive reasoning approach that achieves significantly higher generalization than HRM, while using a single tiny network with only 2 layers. With only 7M parameters, TRM obtains 45% test-accuracy on ARC-AGI-1 and 8% on ARC-AGI-2, higher than most LLMs (e.g., Deepseek R1, o3-mini, Gemini 2.5 Pro) with less than 0.01% of the parameters.

Today, we’re beginning to pilot a new experience in a few regions that makes it easy for people to collaborate with each other—and with ChatGPT—in the same conversation. With group chats, you can bring friends, family, or coworkers into a shared space to plan, make decisions, or work through ideas together.

Listen to nature sounds, fall in love with the Earth

Recurrent Neural Networks (RNNs) laid the foundation for sequence modeling, but their intrinsic sequential nature restricts parallel computation, creating a fundamental barrier to scaling. This has led to the dominance of parallelizable architectures like Transformers and, more recently, State Space Models (SSMs). While SSMs achieve efficient parallelization through structured linear recurrences, this linearity constraint limits their expressive power and precludes modeling complex, nonlinear sequence-wise dependencies. To address this, we present ParaRNN, a framework that breaks the sequence-parallelization barrier for nonlinear RNNs. Building on prior work, we cast the sequence of nonlinear recurrence relationships as a single system of equations, which we solve in parallel using Newton's iterations combined with custom parallel reductions. Our implementation achieves speedups of up to 665x over naive sequential application, allowing training nonlinear RNNs at unprecedented scales. To showcase this, we apply ParaRNN to adaptations of LSTM and GRU architectures, successfully training models of 7B parameters that attain perplexity comparable to similarly-sized Transformers and Mamba2 architectures. To accelerate research in efficient sequence modeling, we release the ParaRNN codebase as an open-source framework for automatic training-parallelization of nonlinear RNNs, enabling researchers and practitioners to explore new nonlinear RNN models at scale.

Repo

"Paged Out! is a free experimental (one article == one page) technical magazine about programming (especially programming tricks!), hacking, security hacking, retro computers, modern computers, electronics, demoscene, and other similar topics.

It's made by the community for the community. And it's not-for-profit (though in time, we hope it will be self-sustained) - this means that the issues will always be free to download, share, and print."

"...we developed the Multi-modal Critical Thinking Agent, or MMCTAgent, for structured reasoning over long-form video and image data, available on GitHub and featured on Azure AI Foundry Labs.

Built on AutoGen, Microsoft’s open-source multi-agent system, MMCTAgent provides multimodal question-answering with a Planner–Critic architecture. This design enables planning, reflection, and tool-based reasoning, bridging perception and deliberation in multimodal tasks. It links language, vision, and temporal understanding, transforming static multimodal tasks into dynamic reasoning workflows.

Unlike conventional models that produce one-shot answers, MMCTAgent has modality-specific agents, including ImageAgent and VideoAgent, which include tools like get_relevant_query_frames() or object_detection-tool(). These agents perform deliberate, iterative reasoning—selecting the right tools for each modality, evaluating intermediate results, and refining conclusions through a Critic loop. This enables MMCTAgent to analyze complex queries across long videos and large image libraries with explainability, extensibility, and scalability."

START with your Main Feed and Category Feeds. The others are nice-to-haves but don't overwhelm yourself trying to do everything at once.

Today, at KubeCon NA 2025, we’re focused on making Kubernetes the most open and scalable platform for AI agents, with the introduction of Agent Sandbox.

At its core, Agent Sandbox is a new Kubernetes primitive built with the Kubernetes community that’s designed specifically for agent code execution and computer use, delivering the performance and scale needed for the next generation of agentic AI workloads. Foundationally built on gVisor with additional support for Kata Containers for runtime isolation, Agent Sandbox provides a secure boundary to reduce the risk of vulnerabilities that could lead to data loss, exfiltration or damage to production systems.

• We can formally prove two models behave identically.
• Use it to validate pruning / distillation work.
• Use it to guard model retraining in CI/CD.
• If models diverge, Z3 gives the input that caused the divergence.
  
  No brute force, no test dataset guesses.
  
  Just mathematically guaranteed model equivalence (or a counterexample).

Interesting use of Z3

...as comments sections grew, they became less useful and more prone to toxic, hateful and counter-deliberative discussions...comments are great when the topic of blogs is more narrow, and the blogger is directly engaged. But in the context of broad and widely popular communities, they had the capability to fester.

As the web continued to grow, the issue grew. Without dedicated moderation, which many publications did try, comments sections became discordant and a liability...Washington Post was one of the first publications to remove comments, citing a lack of civility.

That trend has continued, as many sites chose to outsource discussions to social media platforms and third party networks. For sites that can manage comments, it is done through careful moderation and clear rules. But most major publications and platforms have simply opted out.

Comment sections continue to exist, but their presence is far diminished from its peak. The ActivityPub standard, the backbone of the Fediverse, was created as a partial response to their decline. Whether or not comments will find some new form is a story still unwritten.

My site probably isn't big enough (I don't know since I don't have analytics) but maintenance and moderation are reasons I don't support comments.

You want to have control of your own destiny. We call it digital sovereignty. In the old days, the early days of the web, anybody used to be able to make a website. So that feeling of sovereignty as an individual being enabled and being a peer with all the other people on the web, that is what we are still fighting for, and in fact, we need to rebuild.

So many gems from this episode but this is what resonated the most with me.

listen to college radio stations

Threads, Meta’s competitor to X, will now begin targeting a specific type of creator: podcasters. The company said on Tuesday that Threads will focus on bringing more discussions around podcasts to its app and helping creators promote their shows.

Meta explained it’s interested in the podcast community because it sees the potential in becoming the de facto place for people to talk about shows and engage with creators. Today, such discussions are distributed across social media...

The company’s also trying to tap podcasts’ deep ties to culture, which can drive discussions and encourage interactions.

However, while Threads will target the discussions that result from podcasts, the company told TechCrunch it doesn’t intend to be a podcast distribution platform.

Instead, the company will focus on giving podcast hosts more insight and analytics on how the conversations around their shows are resonating with fans. It also intends to develop more features that would help people discover podcasts they might like, and connect podcasters and listeners through conversations.

... one unexpected source of music discovery is quietly booming among Gen Z listeners: college radio.

I spoke to seven student general managers and surveyed 80+ DJs at stations across America...They told me student interest in college radio has dramatically increased in recent years. Stations that once struggled to fill airtime are now turning people away, shortening shows, alternating time slots, and running training programs just to keep up with the demand from aspiring student DJs.

What’s Driving the College Radio Renaissance?

### 1. Algorithm fatigue
Students consistently described radio as an authentic, community-driven refuge from the passive, isolated, algorithm-driven digital experiences that have defined their adolescence
### 2. Analog Nostalgia
The [resurgence](https://www.nytimes.com/2025/08/24/opinion/gen-z-technology-nostalgia.html?smid=nytcore-ios-share&referringSource=articleShare) of interest in physical media is a significant driver of Gen Z’s attraction to college radio...This manifests in a return to high-friction analog media like vinyl, flip phones, film cameras and radio.
### 3. Community, Creativity & Belonging
College radio stations serve as vital “third spaces” where students can find a community of like-minded people outside of classes and social media, said several GMs.

In a moment when the wider industry is betting on algorithms, AI, and fleeting TikTok trends, these college radio DJs remind us that young people aren’t bored of music—they’re bored of the shallow, virality-obsessed way music is marketed to them.

As someone who listens to college football games on the radio, it's comforting to hear college radio stations are thriving.

Google are killing XSLT!

The design on this site is top notch

In this post, I'll be covering LoRA training and its recent incorporation into prime-rl for both SFT and RL finetuning, including practical implementation details & experimental training results for some of our RL environments.

prime-rl has full support for LoRA. We plan on continuing improving it with better LoRA algorithms and more efficient implementations. We also are working on MoE support, as well as the ability to train multiple LoRA adapters at the same time in preparation for our upcoming Reinforcement Fine-tuning API launch.

We are building a 24×7 music video and animation channel for The Indie Beat and we need you!

This will be a consent-driven channel promoting indie musical artists from across the Fediverse, as well as animators who want to share their animation art (and who knows what else in the future?).

As you can probably tell, I’m pretty excited about Micro.blog taking on the challenge of being that ’indie-focused, YouTube alternative” that Reece envisioned.

There’s never been a better time to own your spot on the web. If you haven’t checked out Micro.blog before, I think it’s a compelling place to look.

Sold! Fantastic love letter to Micro.blog. If it wasn't for the fact I like tinkering and want to own my entire stack, Micro.blog is what I'd switch to. I'm a big fan of what Manton is crafting.

Today we're releasing Bonfire Social 1.0: a federated community‑first social network and the modular toolkit that powers it: a framework for building apps on the open social web beyond microblogging. We're pairing the launch with a crowdfund so communities drive what we build next.

Congrats on this launch! I personally don't self-host an instance but really like the problems the project aims to tackle.

A living hub of blogging resources, from early sparks to fine-tuning.

Algorithm fatigue has been building for some time.

Gen Z might be less likely to discover a new artist they love than some older generations. But they’re also leading a resurgence in college radio. Terrestrial radio once seemed like a dying format, but many schools now report they don’t have enough time slots to accommodate all the aspiring DJs.

Even the iPod is enjoying a renaissance. Classic iPods fetch hundreds of dollars on eBay, and an entire subculture, albeit small, has arisen around modding them to extend battery life, increase storage, and add modern conveniences like Bluetooth and USB-C.

At this stage, anti-algorithm is itself an entire genre of content. Particularly on YouTube, where creators make videos about ditching streaming, stopping doomscrolling, and how the algorithm has flattened culture.

It's not either, or. Human curation and algorithms can coexist. The problem is end users are seldom given a choice.

Key findings

1. Most organizations are still in the experimentation or piloting phase: Nearly two-thirds of respondents say their organizations have not yet begun scaling AI across the enterprise. 2. High curiosity in AI agents: Sixty-two percent of survey respondents say their organizations are at least experimenting with AI agents. 3. Positive leading indicators on impact of AI: Respondents report use-case level cost and revenue benefits, and 64 percent say that AI is enabling their innovation. However, just

39 percent report EBIT impact at the enterprise level.

4. High performers use AI to drive growth, innovation, and cost: Eighty percent of respondents say their companies set efficiency as an objective of their AI initiatives, but the companies seeing the most value from AI often set growth or innovation as additional objectives.
5. Redesigning workflows is a key success factor: Half of those AI high performers intend to use AI to transform their businesses, and most are redesigning workflows.
6. Differing perspectives on employment impact: Respondents vary in their expectations of AI’s impact on the overall workforce size of their organizations in the coming year: 32 percent expect decreases, 43 percent no change, and 13 percent increases.

Thousands of songs representing some of the rarest and most uniquely American music borne from the Jazz Age and the Great Depression would have likely been lost to landfills and faded from memory. Fans and historians have long credited obsessive record collectors for preserving much of that music, and today they can thank a new partnership between UC Santa Barbara and the nonprofit Dust-to-Digital Foundation for making it available to the public for free.

We introduce Nested Learning, a new approach to machine learning that views models as a set of smaller, nested optimization problems, each with its own internal workflow, in order to mitigate or even completely avoid the issue of “catastrophic forgetting”, where learning new tasks sacrifices proficiency on old tasks."

With Firefox 145, we’re rolling out major privacy upgrades that take on browser fingerprinting — a pervasive and hidden tracking technique that lets websites identify you even when cookies are blocked or you’re in private browsing. These protections build on Mozilla’s long-term goal of building a healthier, transparent and privacy-preserving web ecosystem.

We’ve launched the first version of a new subscription plan called Micro.blog Studio.

Micro.blog Studio adds longer video hosting for your blog, with uploads up to 20 minutes. You can read some of the technical bits here. It can automatically copy videos to PeerTube and Bluesky too.

Congrats on this launch!

Today, we are excited to announce the launch of .NET 10, the most productive, modern, secure, intelligent, and performant release of .NET yet. It’s the result of another year of effort from thousands of developers around the world. This release includes thousands of performance, security, and functional improvements across the entire .NET stack-from languages and developer tools to workloads-enabling you to build with a unified platform and easily infuse your apps with AI.

It's here! Congrats to the team and everyone behind this release.

Today we're announcing that Blockly, Google's open source library for drag-and-drop programming, is moving to the stewardship of the Raspberry Pi Foundation on November 10, 2025.