Hi, I'm Luis 👋

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.

Winter Storm Could Drop 8 Inches Of Snow Overnight Sunday

Why?!

The only time most people hear about tokenization is when it’s being blamed for some undesirable behavior of a language model. These incidents have helped turn ignorance and indifference towards tokenizers into active dismissal and disdain. This attitude makes it harder to understand the tokenizers and develop better ones, because fewer people are actually studying tokenizers.

The goal of this blog post is to provide some context about how we got the tokenization approaches we have and argue that they’re not actually so bad.

On a personal level, I also want to foster more engagement with the tokenization literature. Regardless of whether you are pro- or anti-tokenization, more people to be working on issues related to tokenizers, the faster we’re going to make progress. For those that think they are taking a “tokenizer-free” approach, I argue that these approaches are just other kinds of tokenization. And incorporating the findings from static subword tokenization research can only help develop better alternatives.

AI is making it easier to create personalized app experiences that transform content into the right format for users. We previously enabled developers to integrate with Gemini Nano through ML Kit GenAI APIs tailored for specific use cases like summarization and image description.

Today marks a major milestone for Android's on-device generative AI. We're announcing the Alpha release of the ML Kit GenAI Prompt API. This API allows you to send natural language and multimodal requests to Gemini Nano, addressing the demand for more control and flexibility when building with generative models.

In this post, I’ll introduce a reinforcement learning (RL) algorithm based on an “alternative” paradigm: divide and conquer. Unlike traditional methods, this algorithm is not based on temporal difference (TD) learning (which has scalability challenges), and scales well to long-horizon tasks.

Embedding Atlas is a tool that provides interactive visualizations for large embeddings. It allows you to visualize, cross-filter, and search embeddings and metadata.

repo

Back in 2018, Ben Thompson wrote another piece called “Tech’s Two Philosophies.” He contrasted keynotes from Google’s Sundar Pichai and Microsoft’s Satya Nadella, and came to this conclusion: “In Google’s view, computers help you get things done—and save you time—by doing things for you.” The second philosophy, expounded by Nadella, is very much a continuation of Steve Jobs’ “bicycle for the mind” insight. As Thompson put it, “the expectation is not that the computer does your work for you, but rather that the computer enables you to do your work better and more efficiently.” Another way of saying this is that you can treat AI as either a worker OR a tool, but your choice has consequences.

Yes, today’s AI is amazing. We don’t have to reach for hyperbole to appreciate that. And obviously, if AI systems do develop genuine volition and stakes in their work, the ethical calculus changes entirely.

For the moment, though, companies building and deploying AI tools should focus on three things: First, does AI empower its users to do things that were previously impossible? Second, does it empower a wider group of people to do things that formerly could be done only by highly skilled specialists? Third, do the benefits of the increased productivity it brings accrue to those using the tool or primarily to those who develop it and own it?

The answer to the first two questions is that absolutely, we are entering a period of dramatic democratization of computing power. And yes, if humans are given the freedom to apply that power to solve new problems and create new value, we could be looking ahead to a golden age of prosperity. It’s how we might choose to answer the third question that haunts me.

During the first industrial revolution, humans suffered through a long period of immiseration as the productivity gains from machines accrued primarily to the owners of the machines. It took several generations before they were more widely shared.

It doesn’t have to be that way. Replace human workers with AI workers, and you will repeat the mistakes of the 19th century. Build tools that empower and enrich humans, and we might just surmount the challenges of the 21st century.

Instead of endlessly debating capabilities or timing, it’s time to think hard about the purpose of technology, what we want from it, what its limitations should be, and how we’re going to ensure this incredible tech always benefits humanity.

At Microsoft AI, we’re working towards Humanist Superintelligence (HSI): incredibly advanced AI capabilities that always work for, in service of, people and humanity more generally. We think of it as systems that are problem-oriented and tend towards the domain specific. Not an unbounded and unlimited entity with high degrees of autonomy – but AI that is carefully calibrated, contextualized, within limits. We want to both explore and prioritize how the most advanced forms of AI can keep humanity in control while at the same time accelerating our path towards tackling our most pressing global challenges.

File over app is a philosophy: if you want to create digital artifacts that last, they must be files you can control, in formats that are easy to retrieve and read. Use tools that give you this freedom.

File over app is an appeal to tool makers: accept that all software is ephemeral, and give people ownership over their data.

AlphaEvolve is a generic evolutionary coding agent that combines the generative capabilities of LLMs with automated evaluation in an iterative evolutionary framework that proposes, tests, and refines algorithmic solutions to challenging scientific and practical problems. In this paper we showcase AlphaEvolve as a tool for autonomously discovering novel mathematical constructions and advancing our understanding of long-standing open problems. To demonstrate its breadth, we considered a list of 67 problems spanning mathematical analysis, combinatorics, geometry, and number theory. The system rediscovered the best known solutions in most of the cases and discovered improved solutions in several. In some instances, AlphaEvolve is also able to generalize results for a finite number of input values into a formula valid for all input values. Furthermore, we are able to combine this methodology with Deep Think and AlphaProof in a broader framework where the additional proof-assistants and reasoning systems provide automated proof generation and further mathematical insights. These results demonstrate that large language model-guided evolutionary search can autonomously discover mathematical constructions that complement human intuition, at times matching or even improving the best known results, highlighting the potential for significant new ways of interaction between mathematicians and AI systems. We present AlphaEvolve as a powerful new tool for mathematical discovery, capable of exploring vast search spaces to solve complex optimization problems at scale, often with significantly reduced requirements on preparation and computation time.

I came across some other clever uses of plain text files recently, so I wanted to pull them all together so you can pick and choose which ones will be most useful to you.

Keep your calendar in a plain text file.

I'm not one to yuck someone else's yum, but it's a no from me.

How could this work? A new feature for OPML subscription lists. Today it's used as the import/export format for lists. But that's a one-time thing. Instead I want to give Pocket Casts the URL of an OPML file with my podcast subscriptions from the desktop.#

This is exactly the thinking behind my podroll and other collections on my site that I provide an OPML file for. I want a single source of truth for my subscriptions that I can share with others. Sadly it's not dynamic because I still have to manually update the OPML file and re-import into my podcasting client. Having read / write capabilities from the client so that whenever I subscribe to a podcast, the OPML file is updated would make the experience even better.

One of Jami’s core features is its ability to function in emergency settings where internet access is severely limited or entirely cut off.

‌Unlike traditional messaging apps that rely on central servers to relay messages, Jami connects users peer to peer. This means each device acts as a node in a decentralized web, sending encrypted data directly to others, without any intermediary.

In troubled times, staying connected is not a luxury. It is a necessity.‌‌It means ensuring safety, coordination, and the ability to call for help, share your location, and reach your loved ones.

When official channels fail, direct communication helps share verified information, document events, and fight misinformation.

It also defends our freedoms. Speaking freely, even under pressure, is essential to support, resist, and rebuild.

Staying connected maintains solidarity and mental well-being. No one should face crisis alone.

And above all, it keeps communities resilient, able to act and organize even in the darkest times.

That is why tools like Jami, which operate without a central server or reliance on tech giants, are essential parts of a digital survival kit.

With tools like TRL, TorchForge and verl, the open-source community has shown how to scale AI across complex compute infrastructure. But compute is only one side of the coin. The other side is the developer community; the people and tools that make agentic systems possible. That’s why Meta and Hugging Face are partnering to launch the OpenEnv Hub: a shared and open community hub for agentic environments.

Agentic environments define everything an agent needs to perform a task: the tools, APIs, credentials, execution context, and nothing else. They bring clarity, safety, and sandboxed control to agent behavior.

These environments can be used for both training and deployment, and serve as the foundation for scalable agentic development.

Saving for later

Custom agents for GitHub Copilot make it easy for users and organizations to specialize their Copilot coding agent (CCA) through simple, file-based configurations.

By adding a configuration file under .github/agents in a repository or in the {org}/.github repository, you can define agent personas that capture your team’s workflows, conventions, and unique needs. These agents can be further tailored with prompts, tool selections, and Model Context Protocol (MCP) servers to optimize for specific use cases.

Large Language Models (LLMs) often struggle with problems that require multi-step reasoning. For small-scale open-source models, Reinforcement Learning with Verifiable Rewards (RLVR) fails when correct solutions are rarely sampled even after many attempts, while Supervised Fine-Tuning (SFT) tends to overfit long demonstrations through rigid token-by-token imitation. To address this gap, we propose Supervised Reinforcement Learning (SRL), a framework that reformulates problem solving as generating a sequence of logical "actions". SRL trains the model to generate an internal reasoning monologue before committing to each action. It provides smoother rewards based on the similarity between the model's actions and expert actions extracted from the SFT dataset in a step-wise manner. This supervision offers richer learning signals even when all rollouts are incorrect, while encouraging flexible reasoning guided by expert demonstrations. As a result, SRL enables small models to learn challenging problems previously unlearnable by SFT or RLVR. Moreover, initializing training with SRL before refining with RLVR yields the strongest overall performance. Beyond reasoning benchmarks, SRL generalizes effectively to agentic software engineering tasks, establishing it as a robust and versatile training framework for reasoning-oriented LLMs.

Mozilla.ai is adopting the llamafile project to advance local, privacy-first AI. We are refreshing the codebase, modernizing foundations, and shaping the roadmap with community input

Trying to take control of your online privacy can feel like a full-time job. But if you break it up into small tasks and take on one project at a time it makes the process of protecting your privacy much easier. This month we’re going to do just that. For the month of October, we’ll update this post with new tips every weekday that show various ways you can opt yourself out of the ways tech giants surveil you.

At GitHub Universe, we’re announcing Agent HQ, GitHub’s vision for the next evolution of our platform. Agents shouldn’t be bolted on. They should work the way you already work. That’s why we’re making agents native to the GitHub flow.

Agent HQ transforms GitHub into an open ecosystem that unites every agent on a single platform. Over the coming months, coding agents from Anthropic, OpenAI, Google, Cognition, xAI, and more will become available directly within GitHub as part of your paid GitHub Copilot subscription.

The core idea of on-policy distillation is to sample trajectories from the student model and use a high-performing teacher to grade each token of each trajectory.

A template for building AI-powered coding agents that supports Claude Code, OpenAI's Codex CLI, GitHub Copilot CLI, Cursor CLI, Google Gemini CLI, and opencode with Vercel Sandbox to automatically execute coding tasks on your repositories.

Most social applications, from Twitter to Wikipedia, have rigid one-size-fits-all designs, but building new social applications is both technically challenging and results in applications that are siloed away from existing communities. We present Graffiti, a system that can be used to build a wide variety of personalized social applications with relative ease that also interoperate with each other. People can freely move between a plurality of designs—each with its own aesthetic, feature set, and moderation—all without losing their friends or data.

Our concept of total reification makes it possible for seemingly contradictory designs, including conflicting moderation rules, to interoperate. Conversely, our concept of channels prevents interoperation from occurring by accident, avoiding context collapse.

Graffiti applications interact through a minimal client-side API, which we show admits at least two decentralized implementations. Above the API, we built a Vue plugin, which we use to develop applications similar to Twitter, Messenger, and Wikipedia using only client-side code. Our case studies explore how these and other novel applications interoperate, as well as the broader ecosystem that Graffiti enables.

We’ve been building a better foundation for AI Studio, and this week we introduced a totally new AI powered vibe coding experience in Google AI Studio. This redesigned experience is meant to take you from prompt to working AI app in minutes without you having to juggle with API keys, or figuring out how to tie models together.

Recent advances in multimodal models have demonstrated remarkable text-guided image editing capabilities, with systems like GPT-4o and Nano-Banana setting new benchmarks. However, the research community's progress remains constrained by the absence of large-scale, high-quality, and openly accessible datasets built from real images. We introduce Pico-Banana-400K, a comprehensive 400K-image dataset for instruction-based image editing. Our dataset is constructed by leveraging Nano-Banana to generate diverse edit pairs from real photographs in the OpenImages collection. What distinguishes Pico-Banana-400K from previous synthetic datasets is our systematic approach to quality and diversity. We employ a fine-grained image editing taxonomy to ensure comprehensive coverage of edit types while maintaining precise content preservation and instruction faithfulness through MLLM-based quality scoring and careful curation. Beyond single turn editing, Pico-Banana-400K enables research into complex editing scenarios. The dataset includes three specialized subsets: (1) a 72K-example multi-turn collection for studying sequential editing, reasoning, and planning across consecutive modifications; (2) a 56K-example preference subset for alignment research and reward model training; and (3) paired long-short editing instructions for developing instruction rewriting and summarization capabilities. By providing this large-scale, high-quality, and task-rich resource, Pico-Banana-400K establishes a robust foundation for training and benchmarking the next generation of text-guided image editing models.

Repo

An infinitely zooming collaborative painting I created together with 31 other illustrators in 2007.

Website

Floor796 is an animated scene showing the lives of characters from various works on the 796th floor of a huge space station. The animation is regularly expanded with new blocks (rooms) and characters from movies, TV series, games, anime, memes, etc.

The project is being created by one author as a hobby starting in 2018.

This is so much fun. I've easily spent an hour exploring the space finding new rooms and characters.

We’re excited to introduce Monarch, a distributed programming framework that brings the simplicity of single-machine PyTorch to entire clusters.

Monarch lets you program distributed systems the way you’d program a single machine, hiding the complexity of distributed computing:

1. Program clusters like arrays. Monarch organizes hosts, processes, and actors into scalable meshes that you can manipulate directly. You can operate on entire meshes (or slices of them) with simple APIs—Monarch handles the distribution and vectorization automatically, so you can think in terms of what you want to compute, not where the code runs.
2. Progressive fault handling. With Monarch, you write your code as if nothing fails. When something does fail, Monarch fails fast by default—stopping the whole program, just like an uncaught exception in a simple local script. Later, you can progressively add fine-grained fault handling exactly where you need it, catching and recovering from failures just like you’d catch exceptions.
3. Separate control from data. Monarch splits control plane (messaging) from data plane (RDMA transfers), enabling direct GPU-to-GPU memory transfers across your cluster. Monarch lets you send commands through one path, while moving data through another, optimized for what each does best.
4. Distributed tensors that feel local. Monarch integrates seamlessly with PyTorch to provide tensors that are sharded across clusters of GPUs. Monarch tensor operations look local but are executed across distributed large clusters, with Monarch handling the complexity of coordinating across thousands of GPUs."

If you bought your computer after 2010, there's most likely no reason to throw it out. By just installing an up-to-date Linux operating system you can keep using it for years to come.

The Developer Ecosystem Survey 2025 results reveal a field that’s changing fast, shaped by AI and growing self-awareness among developers. They’re using tools that make them more productive, while questioning how they define productivity in the changing landscape.

"You've arrived at MelonLand! This is a web project and online arts community that celebrates homepages, virtual worlds, the world-wide-web and the digital lives that all netizins share, here at the dawn of the digital age.

This project has three goals:

• To make the web; genuine, chaotic, timeless, individual and joyful
• To provide knowledge and support to humans creating digital worlds
• To promote websites and digital worlds as mediums of visual art.
  MelonLand is part of a wider movement sometimes called the web revival; to join, all you need to do is choose to engage with small handmade web projects next time you're online!"

Welcome to Reclaim Open 2025 Rewilding the Network – Making New Connections. Register now to join us online from November 4-6th as we come together on the open web, via federated social networks, on the radio and on a blog near you.

This looks like a fun event.

Discover independent music from artists in the Fediverse.

...we are introducing SentinelStep(opens in new tab), a mechanism that enables agents to complete long-running monitoring tasks. The approach is simple. SentinelStep wraps the agent in a workflow with dynamic polling and careful context management. This enables the agent to monitor conditions for hours or days without getting sidetracked. We’ve implemented SentinelStep in Magentic-UI, our research prototype agentic system, to enable users to build agents for long-running tasks, whether they involve web browsing, coding, or external tools.

I REALLY want an ROG Xbox Ally X. Not as a handheld gaming device though, but as a mobile computer.

Apparently there's an NPU version coming soon so maybe I'll hold off until that's out.

Today, I’m happy to share an early preview of C# AI Buddy - an AI assistant you can “hire” to help you build AI projects using C# and .NET.

This is such a cool project from Jordan. Try it out and give feedback.