Hi, I'm Luis 👋

Truly an unexpected delight while listening to the re-release of 1983.

Diffusion language models (DLMs) offer a compelling promise: parallel token generation could break the sequential bottleneck of autoregressive (AR) decoding. Yet in practice, DLMs consistently lag behind AR models in quality.

We argue that this gap stems from a fundamental failure of introspective consistency: AR models agree with what they generate, whereas DLMs often do not. We introduce the Introspective Diffusion Language Model (I-DLM), which uses introspective strided decoding (ISD) to verify previously generated tokens while advancing new ones in the same forward pass.

Empirically, I-DLM-8B is the first DLM to match the quality of its same-scale AR counterpart, outperforming LLaDA-2.1-mini (16B) by +26 on AIME-24 and +15 on LiveCodeBench-v6 with half the parameters, while delivering 2.9-4.1x throughput at high concurrency. With gated LoRA, ISD enables bit-for-bit lossless acceleration.

Today, we’re introducing Gemini Robotics-ER 1.6, a significant upgrade to our reasoning-first model that enables robots to understand their environments with unprecedented precision. By enhancing spatial reasoning and multi-view understanding, we are bringing a new level of autonomy to the next generation of physical agents.

This model specializes in reasoning capabilities critical for robotics, including visual and spatial understanding, task planning and success detection. It acts as the high-level reasoning model for a robot, capable of executing tasks by natively calling tools like Google Search to find information, vision-language-action models (VLAs) or any other third-party user-defined functions.

Gemini Robotics-ER 1.6 shows significant improvement over both Gemini Robotics-ER 1.5 and Gemini 3.0 Flash, specifically enhancing spatial and physical reasoning capabilities such as pointing, counting, and success detection. We are also unlocking a new capability: instrument reading, enabling robots to read complex gauges and sight glasses — a use case we discovered through close collaboration with our partner, Boston Dynamics.

Recently, we began collaborating with NVIDIA on a new challenge: applying the multi-agent harness to optimize CUDA kernels. These are difficult technical problems with important real-world consequences: CUDA kernels are the core software that supports AI model training and inference on NVIDIA GPUs. Faster kernels mean better GPU utilization, reduced energy consumption, lower latency, and reduced cost per token—allowing providers to serve bigger, more capable models to more users at once.

Our multi-agent harness operated autonomously for three weeks across 235 problems. The system achieved a 38% geomean speedup by building and optimizing Blackwell GPU kernels from scratch, all the way down to the assembly level.

These levels of performance improvement are typically only found through months or years of work from highly experienced kernel engineers. The multi-agent system accomplished it in weeks, addressing a long-tail of kernel problems that had been impractical with existing approaches.

Today, we’re pleased to introduce SAM 3.1.

As a drop-in replacement for SAM 3, our updated model delivers a significant boost in video processing efficiency by introducing object multiplexing, which allows the model to track up to 16 objects in a single forward pass. This innovation doubles the processing speed for videos with a medium number of objects, increasing throughput from 16 to 32 frames per second on a single H100 GPU. As a result, SAM 3.1 enables real-time object tracking in complex videos while reducing overall GPU resource requirements, making high-performance applications feasible on smaller, more accessible hardware.

Supermemory is the memory and context layer for AI. #1 on LongMemEval, LoCoMo, and ConvoMem — the three major benchmarks for AI memory.

We are a research lab building the engine, plugins and tools around it.

Your AI forgets everything between conversations. Supermemory fixes that.

It automatically learns from conversations, extracts facts, builds user profiles, handles knowledge updates and contradictions, forgets expired information, and delivers the right context at the right time. Full RAG, connectors, file processing — the entire context stack, one system.

We believe that the most helpful AI assistant is one that’s personal to you, and understands your preferences and past conversations. But if you’re curious to try a different option, starting over with an assistant that doesn’t know you can feel daunting.

That’s why we’re introducing new, easy-to-use switching tools for all consumer accounts — allowing you to easily bring your memories, context and chat history from other AI apps directly into Gemini.

Today, we introduce TurboQuant (to be presented at ICLR 2026), a compression algorithm that optimally addresses the challenge of memory overhead in vector quantization. We also present Quantized Johnson-Lindenstrauss (QJL), and PolarQuant (to be presented at AISTATS 2026), which TurboQuant uses to achieve its results. In testing, all three techniques showed great promise for reducing key-value bottlenecks without sacrificing AI model performance. This has potentially profound implications for all compression-reliant use cases, including and especially in the domains of search and AI.

An interactive journey through the 30 year history of the web.

A plain-text cognitive architecture for Claude Code — simple by design, so the model can reason over its own memory with the same Unix tools (grep, find, git diff) it already knows.

Retrieval pipelines typically operate in a single pass, which poses a problem when the information required to answer a question is spread across multiple documents or requires intermediate reasoning to locate. In practice, many real-world queries require multi-hop retrieval, in which the output of one search informs the next. Recent work has shown that frontier LLMs perform this multi-hop search effectively through a process known as agentic search, simply defined as a loop of LLM calls with search tools. This mode of search often comes with significant cost and latency due to their use of frontier-scale LLMs.

We introduce Chroma Context-1, a 20B parameter agentic search model derived from gpt-oss-20B that achieves retrieval performance comparable to frontier-scale LLMs at a fraction of the cost and up to 10x faster inference speed. Context-1 is designed to be used as a subagent in conjunction with a frontier reasoning model. Given a query, it produces a ranked list of documents that are relevant to satisfying the query. The model is trained to decompose queries into subqueries, iteratively search a corpus, and selectively edit its own context to free capacity for further exploration.

Today, we’re advancing Gemini’s real-time dialogue capabilities with Gemini 3.1 Flash Live, our highest-quality audio and voice model yet. It delivers the speed and natural rhythm needed for the next generation of voice-first AI, offering a more intuitive experience for developers, enterprises and everyday users.

Machine learning education faces a fundamental gap: students learn algorithms without understanding the systems that execute them. They study gradient descent without measuring memory, attention mechanisms without analyzing O(N^2) scaling, optimizer theory without knowing why Adam requires 3x the memory of SGD. This "algorithm-systems divide" produces practitioners who can train models but cannot debug memory failures, optimize inference latency, or reason about deployment trade-offs--the very skills industry demands as "ML systems engineering." We present TinyTorch, a 20-module curriculum that closes this gap through "implementation-based systems pedagogy": students construct PyTorch's core components (tensors, autograd, optimizers, CNNs, transformers) in pure Python, building a complete framework where every operation they invoke is code they wrote. The design employs three patterns: "progressive disclosure" of complexity, "systems-first integration" of profiling from the first module, and "build-to-validate milestones" recreating 67 years of ML breakthroughs--from Perceptron (1958) through Transformers (2017) to MLPerf-style benchmarking. Requiring only 4GB RAM and no GPU, TinyTorch demonstrates that deep ML systems understanding is achievable without specialized hardware. The curriculum is available open-source at this http URL.

Cohere is announcing Transcribe, a state-of-the-art automatic speech recognition (ASR) model that is open source and available today for download.

Our objective was straightforward: push the frontier of dedicated ASR model accuracy under practical conditions. The model was trained from scratch with a deliberate focus on minimizing word error rate (WER), while keeping production readiness top-of-mind. In other words, not just a research artifact, but a system designed for everyday use.

VS Code includes built-in AI chat through GitHub Copilot Chat. Ollama models can be used directly in the Copilot Chat model picker.

Cognitive neuroscience is fragmented into specialized models, each tailored to specific experimental paradigms, hence preventing a unified model of cognition in the human brain. Here, we introduce TRIBE v2, a tri-modal (video, audio and language) foundation model capable of predicting human brain activity in a variety of naturalistic and experimental conditions. Leveraging a unified dataset of over 1,000 hours of fMRI across 720 subjects, we demonstrate that our model accurately predicts high-resolution brain responses for novel stimuli, tasks and subjects, superseding traditional linear encoding models, delivering several-fold improvements in accuracy. Critically, TRIBE v2 enables in silico experimentation: tested on seminal visual and neuro-linguistic paradigms, it recovers a variety of results established by decades of empirical research. Finally, by extracting interpretable latent features, TRIBE v2 reveals the fine-grained topography of multisensory integration. These results establish artificial intelligence as a unifying framework for exploring the functional organization of the human brain.

Today we’re releasing Voxtral TTS, our first text-to-speech model with state-of-the-art performance in multilingual voice generation. The model is lightweight at 4B parameters, making Voxtral-powered agents natural, reliable, and cost-effective at scale.

...we’re releasing a new paper, “Measuring Progress Toward AGI: A Cognitive Taxonomy,” that presents a scientific foundation for understanding the cognitive capabilities of AI systems.

Alongside the paper, we are partnering with Kaggle to launch a hackathon, inviting the research community to help build the evaluations needed to put this framework into practice.

Our framework draws on decades of research from psychology, neuroscience and cognitive science to develop a cognitive taxonomy. It identifies 10 key cognitive abilities that we hypothesize will be important for general intelligence in AI systems:

1. Perception: extracting and processing sensory information from the environment
2. Generation: producing outputs such as text, speech and actions
3. Attention: focusing cognitive resources on what matters
4. Learning: acquiring new knowledge through experience and instruction
5. Memory: storing and retrieving information over time
6. Reasoning: drawing valid conclusions through logical inference
7. Metacognition: knowledge and monitoring of one's own cognitive processes
8. Executive functions: planning, inhibition and cognitive flexibility
9. Problem solving: finding effective solutions to domain-specific problems
10. Social cognition: processing and interpreting social information and responding appropriately in social situations

The idea: give an AI agent a small but real LLM training setup and let it experiment autonomously overnight. It modifies the code, trains for 5 minutes, checks if the result improved, keeps or discards, and repeats. You wake up in the morning to a log of experiments and (hopefully) a better model. The training code here is a simplified single-GPU implementation of nanochat. The core idea is that you're not touching any of the Python files like you normally would as a researcher. Instead, you are programming the program.md Markdown files that provide context to the AI agents and set up your autonomous research org. The default program.md in this repo is intentionally kept as a bare bones baseline, though it's obvious how one would iterate on it over time to find the "research org code" that achieves the fastest research progress, how you'd add more agents to the mix, etc. A bit more context on this project is here in this tweet.

Kitten TTS is an open-source, lightweight text-to-speech library built on ONNX. With models ranging from 15M to 80M parameters (25-80 MB on disk), it delivers high-quality voice synthesis on CPU without requiring a GPU.

We present a system for training enterprise search agents via reinforcement learning that achieves state-of-the-art performance across a diverse suite of hard-to-verify agentic search tasks. Our work makes four core contributions. First, we introduce KARLBench, a multi-capability evaluation suite spanning six distinct search regimes, including constraint-driven entity search, cross-document report synthesis, tabular numerical reasoning, exhaustive entity retrieval, procedural reasoning over technical documentation, and fact aggregation over internal enterprise notes. Second, we show that models trained across heterogeneous search behaviors generalize substantially better than those optimized for any single benchmark. Third, we develop an agentic synthesis pipeline that employs long-horizon reasoning and tool use to generate diverse, grounded, and high-quality training data, with iterative bootstrapping from increasingly capable models. Fourth, we propose a new post-training paradigm based on iterative large-batch off-policy RL that is sample efficient, robust to train-inference engine discrepancies, and naturally extends to multi-task training with out-of-distribution generalization. Compared to Claude 4.6 and GPT 5.2, KARL is Pareto-optimal on KARLBench across cost-quality and latency-quality trade-offs, including tasks that were out-of-distribution during training. With sufficient test-time compute, it surpasses the strongest closed models. These results show that tailored synthetic data in combination with multi-task reinforcement learning enables cost-efficient and high-performing knowledge agents for grounded reasoning.

TLDR: Agent = Model + Harness. Harness engineering is how we build systems around models to turn them into work engines. The model contains the intelligence and the harness makes that intelligence useful. We define what a harness is and derive the core components today's and tomorrow's agents need.

A harness is every piece of code, configuration, and execution logic that isn't the model itself. A raw model is not an agent. But it becomes one when a harness gives it things like state, tool execution, feedback loops, and enforceable constraints.

There are things we want an agent to do that a model cannot do out of the box. This is where a harness comes in.Models (mostly) take in data like text, images, audio, video and they output text. That's it. Out of the box they cannot:

• Maintain durable state across interactions
• Execute code
• Access realtime knowledge
• Setup environments and install packages to complete work
  
  These are all harness level features.

...Model behavior is rarely the result of isolated components; rather, it emerges from complex dependencies and patterns. To achieve state-of-the-art performance, models synthesize complex feature relationships, find shared patterns from diverse training examples, and process information through highly interconnected internal components.

Therefore, grounded or reality-checked interpretability methods must also be able to capture these influential interactions. As the number of features, training data points, and model components grow, the number of potential interactions grows exponentially, making exhaustive analysis computationally infeasible. In this blog post, we describe the fundamental ideas behind SPEX and ProxySPEX, algorithms capable of identifying these critical interactions at scale.

Central to our approach is the concept of ablation, measuring influence by observing what changes when a component is removed.

• Feature Attribution: We mask or remove specific segments of the input prompt and measure the resulting shift in the predictions.
• Data Attribution: We train models on different subsets of the training set, assessing how the model’s output on a test point shifts in the absence of specific training data.
• Model Component Attribution (Mechanistic Interpretability): We intervene on the model’s forward pass by removing the influence of specific internal components, determining which internal structures are responsible for the model’s prediction.
  
  In each case, the goal is the same: to isolate the drivers of a decision by systematically perturbing the system, in hopes of discovering influential interactions. Since each ablation incurs a significant cost, whether through expensive inference calls or retrainings, we aim to compute attributions with the fewest possible ablations.

In “Bayesian teaching enables probabilistic reasoning in large language models”, we teach the LLMs to reason in a Bayesian manner by training them to mimic the predictions of the Bayesian model, which defines the optimal way to reason about probabilities. We find that this approach not only significantly improves the LLM’s performance on the particular recommendation task on which it is trained, but also enables generalization to other tasks. This suggests that this method teaches the LLM to better approximate Bayesian reasoning. More generally, our results indicate that LLMs can effectively learn reasoning skills from examples and generalize those skills to new domains.

To accommodate DoorDash’s continuing growth, the ads quality team set out to build foundational embeddings that can be reused across multiple use cases, such as retrieval, ranking, and relevance. Traditionally, the team has relied on categorical and numerical features such as store attributes, context features, and other handcrafted aggregates as inputs to our machine learning models. While these are important engagement signals, they fail to capture the rich semantic information contained in our product catalogs and don’t reflect a deeper understanding of users’ personal interests. To bring these enhancements into our models, we developed DashCLIP, short for Dash Contrastive Language-Image Pretraining, a unified multimodal embedding framework designed to power personalized ad experiences for DoorDash users.

DashCLIP’s architecture addresses the following functional requirements:

• Multimodality encodings: Products on our platform contain both text and visual information. We leverage contrastive learning on the product catalog to approximate a human-like understanding of products, capturing the complementary information from each modality.
• Domain adaptation: We perform continual pretraining on off-the-shelf models to adapt the embeddings to DoorDash’s data distribution.
• Query embedding alignment: To enable search recommendations, we introduce a second stage of alignment in our architecture for a dedicated query encoder that is trained to generate query embeddings in the same space as the product embeddings.
• Relevance dataset curation: We curate a high-quality relevance dataset that combines internal human annotations with knowledge from large language models (LLMs), providing robust supervision for embedding alignment. This eliminates the position and selection bias introduced when historical engagement data is used for training.

LangChain, the agent engineering company behind LangSmith and open-source frameworks that have surpassed 1 billion downloads, today announced a comprehensive integration with NVIDIA to deliver an enterprise-grade agentic AI development platform.

The collaboration combines LangChain's LangSmith agent engineering platform and its open-source frameworks (Deep Agents, LangGraph, and LangChain)with NVIDIA Agent Toolkit, including NVIDIA Nemotron models, NVIDIA NeMo Agent Toolkit profiling and optimization, NVIDIA NIM microservices, and NVIDIA Dynamo giving developers a complete stack to build, deploy, and continuously improve AI agents in production. The platform also incorporates NVIDIA OpenShell, a secure runtime that sandboxes autonomous, self-evolving agents with policy‑based guardrails. Development teams often spend months building custom infrastructure rather than delivering business value. The LangChain-NVIDIA platform is designed to close that gap.

A hundred times a day I think about artist-owned web spaces and how to build stronger communities that mutually nourish artists and creators. The current infrastructure of music streaming operates as a linear system in a meta-modern world that’s in an anti-fragile liminal state.

We need cozy web spaces where artists control the platforms, not algorithms designed to extract our labor.

I decided that I will stop once interview number 150 is out, on July 10th.

Sad to hear this iteration is coming to an end. Thanks for your commitment to this project. I discovered so many amazing personal sites thanks to this series.

I like the honesty and expectation setting in this post.

We’d like to give a warm welcome to the massive influx of users currently trying Matrix as an open decentralised alternative to centralised platforms like Discord. We wish we had more time and resources to develop all the features needed for mainstream adoption (see The Road To Mainstream Matrix from last year’s FOSDEM), but we're happy to welcome you anyway!

...we’re painfully aware that none of the Matrix clients available today provide a full drop-in replacement for Discord yet. All the ingredients are there, and the initial goal for the project was always to provide a decentralised, secure, open platform where communities and organisations could communicate together. However, the reality is that the team at Element who originally created Matrix have had to focus on providing deployments for the public sector (see here or here) to be able to pay developers working on Matrix. Some of the key features expected by Discord users have yet to be prioritised (game streaming, push-to-talk, voice channels, custom emoji, extensible presence, richer hierarchical moderation, etc). Meanwhile no other organisation stepped up to focus on the “communication tool for communities” use case and provide a production ready Discord alternative, but clients like Cinny or Commet may feel much closer to Discord. On the other hand, Matrix goes far beyond Discord in other areas: both messages, files and calls are end-to-end-encrypted; we have read receipts; Matrix is an open protocol everyone can extend, and in the end, most Matrix clients are open source; there is nothing stopping developers from starting their own project based on existing ones and adding the missing features themselves. They may even eventually get accepted in the original projects!

Anyway, TL;DR: Welcome to everyone trying Matrix for the first time; please understand that public Matrix servers will also have to uphold age verification laws, as misguided as they might be. However, at least in Matrix you have the opportunity to run your own servers as you wish: we actively encourage you to make your own assessments and seek legal advice where needed.

This tutorial dives into getting started with the Matrix protocol for direct messages, groups, rooms, communities, spaces, bridging, and more! Matrix is an open source, decentralized protocol with several different use-cases & clients.

...I've been deeply researching Discord alternatives for the better part of a year. Some of my colleagues may think me a bit obsessed about the importance of a "chat app," but I'm convinced that the communication mechanism for online communities is critical to their success. Choosing a new one could be the a matter of life and death for the community. This is a decision we have to get right the first time.

So here, humbly submitted, are my rankings of many of the Discord-like alternatives for maintaining online communities.

photos of the contents of your bag.

The Items We Carry are the small things we put in our pockets or on our persons — the essentials we need to function daily at a basic level.

Take a photo of the most commonly found items in your pockets or on your person, arranged

A week later, during a conversation with his teacher, my son was asked about the most memorable thing from the trip. He didn’t say the beach in Mexico, or the train through Europe, or the wind in Iceland, or even the lost bear pillow. He said the most memorable thing was seeing his dad talk at a university. That it made him proud (I’m not going to pretend I didn’t need a moment after hearing that).

My son doesn’t know what it’s like to not see a path. For him, this is just what dad does. And maybe that’s the whole point.

FOSDEM wasn’t just a conference for me. It was proof that the kid from Tepic who dreamed of finding hackers in real life finally did. They were in Brussels all along, waiting for him to show up.

And he brought his kid.

Love this.

Nantucket Lit™ is a free and open-source platform that allows writers to create and share high-quality e-books. These e-books can be read directly in a web brow­ser without needing special apps or devices. The platform avoids using AI or DRM. E-books are written in Shanty™, a markup language specifically designed for e-books. Shanty can also be used to produce EPUB files and paperback books.

Agent teams let you coordinate multiple Claude Code instances working together. One session acts as the team lead, coordinating work, assigning tasks, and synthesizing results. Teammates work independently, each in its own context window, and communicate directly with each other.

Unlike subagents, which run within a single session and can only report back to the main agent, you can also interact with individual teammates directly without going through the lead.

...I tasked 16 agents with writing a Rust-based C compiler, from scratch, capable of compiling the Linux kernel. Over nearly 2,000 Claude Code sessions and $20,000 in API costs, the agent team produced a 100,000-line compiler that can build Linux 6.9 on x86, ARM, and RISC-V.

The compiler is an interesting artifact on its own, but I focus here on what I learned about designing harnesses for long-running autonomous agent teams: how to write tests that keep agents on track without human oversight, how to structure work so multiple agents can make progress in parallel, and where this approach hits its ceiling.

Database configuration, workflow schedules, secrets, log shipping, observability—across every machine. No SSH sessions, no manual steps, no configuration (or documentation!) drift. Each server converges to whatever state is declared in our git repository.

As we're writing this, px dynamics is a few weeks old. We wanna take you through how we set up camp: The infrastructure took less time to build than it would take most startups (us included!) to properly configure an AWS account. We're not migrating from something else or "modernizing legacy systems." We started here.

Interesting writeup. I love NixOS exactly for the reasons highlighted in this post.

The new Claude Opus 4.6 improves on its predecessor’s coding skills. It plans more carefully, sustains agentic tasks for longer, can operate more reliably in larger codebases, and has better code review and debugging skills to catch its own mistakes. And, in a first for our Opus-class models, Opus 4.6 features a 1M token context window in beta.

Opus 4.6 can also apply its improved abilities to a range of everyday work tasks: running financial analyses, doing research, and using and creating documents, spreadsheets, and presentations. Within Cowork, where Claude can multitask autonomously, Opus 4.6 can put all these skills to work on your behalf.

We’re introducing a new model that unlocks even more of what Codex can do: GPT‑5.3-Codex, the most capable agentic coding model to date. The model advances both the frontier coding performance of GPT‑5.2-Codex and the reasoning and professional knowledge capabilities of GPT‑5.2, together in one model, which is also 25% faster. This enables it to take on long-running tasks that involve research, tool use, and complex execution. Much like a colleague, you can steer and interact with GPT‑5.3-Codex while it’s working, without losing context.

GPT‑5.3‑Codex is our first model that was instrumental in creating itself. The Codex team used early versions to debug its own training, manage its own deployment, and diagnose test results and evaluations—our team was blown away by how much Codex was able to accelerate its own development.

With GPT‑5.3-Codex, Codex goes from an agent that can write and review code to an agent that can do nearly anything developers and professionals can do on a computer.

Flying Lotus fans rejoice, today the famed electronic auteur announces BIG MAMA, a brand new EP coming out on 6th of March.

BIG MAMA captures Ellison in a moment of spontaneous, unbridled momentum. Densely packed with disparate sounds, rhythms, and effects, the EP delivers what he describes as “experimental, maximalist, hyperfast, electronic burst of energy”, packing seven dynamic tracks into a single continuous composition in which every bar is unique, containing no loops throughout.

“I wanted it to feel like being shot out of a cannon, just explosive, unpredictable energy,” he explains. “Like a fuckin’ computer gone awry. Like a machine that had just lost its mind.”

New FlyLo album. Can't wait!