Links
Saved reading and interesting things from my public Semble collection.
- Agentic test processes, LLM benchmarks, and other notes on agentic coding from Galapagos IslandI've been using AI fairly heavily since last November and the whole thing is a funny experience. An agent will do something that, if a human did it, you'd immediately fire them. My reaction, of course, is to act as if this is great and spin up a thousand agents so they can do even more of that.
- The Atmosphere needs a front door - We Can Just Do ThingsThe Atmosphere needs a distinct brand identify, better tools for builders, a new UX for users, and a small empowered group willing to build it all.
- How I Bypassed Claude Opus’s CBRN-E SafeguardsA black-box jailbreak using model-authored state steering, plus an outside-in reconstruction of the safeguards it bypasses.
- Reintroducing Spaces - Daniel's LeafletsZooming out and re-motivating the design of spaces, the new primitive for non-public data on atproto.
- The Atproto Spaces Alpha is Live - AT ProtocolAtproto Spaces, formerly known as “the permissioned data protocol,” is a new extension to atproto that enables non-public data. The alpha is now officially open. Here’s how to develop with it and what to expect as we work towards the full release.
- Why AI Coding Agents Forget — And How ArcticMem Fixes ItExplore ArcticMem, Snowflake’s persistent semantic memory system for AI coding agents. See how dual-tier memory improves benchmark pass rates to 73%.
- Amazon kept shutting down my tablet, so I spent $266 on four AI models to own itOwning a tablet Amazon kept shutting down: CVE-2022-38181, four AI models, five months
- Miles v0.1: Production-level Post-trainingWe present Miles v0.1, a full-stack production-ready system for frontier post-training, the successor to our first Miles release [1]. Building upon slime's [16] clean design, Miles optimizes every sta...
- Git at any scale · CursorHosting Git repositories at scale is a nightmare. When Linus Torvalds designed the first version of the information manager from hell (that's actually the tagline for Git, look it up), he had a very specific use case in mind: his own. He wanted to replace BitKeeper, the distributed version control…
- Doom, compiled into a transformer — Out of DistributionNo training anywhere: a compiler sets the weights of a stock Phi-3-architecture checkpoint so autoregressive generation executes Doom's classic renderer.
- The Clanker Constitution – Wes McKinneyGetting coding agents (we say “clankers”: “agents” gives them too much credit) to behave in a reasonable manner is a full time job even with the latest frontier models. At Kenn, we spend so much time tuning their behavior across different repositories and projects that we started developing a…
- Devtools must be open source - exe.dev blogThe age of personalized software is here.
- Stateless MCP has recaptured my interest (and inspired mcp-explorer and datasette-mcp)Tuesday was Stateless MCP day—the rollout of MCP 2.0, or the 2026-07-28 Model Context Protocol specification to use the more formal but less memorable name. This is the most significant …
- How we do trunk-based development (and why you should too) - PostHogWhat do trees, elephants, and successful software all have in common? They rely on their trunks to help them grow big and strong. In the case of…
- Zhuokai Zhao on Twitter / XTons of interesting things in the Kimi K3 tech report — here are five algorithm-side techniques that I think either I've never seen before or simply deserve more attention than they're getting.1/ They open-sourced the model but kept the speculative decoding draft model, which…— Zhuokai Zhao…
- A Safe Path to Open WeightsStrategic openness can strengthen AI safety and support broader access as defenses and safety science mature.
- Six months at OpenAIThis weeks marks six months since I joined OpenAI and I have several things to say
- Introducing Radial: Agent orchestration on the AtmosphereI've been working on a new project called Radial, a collaborative coding-agent orchestrator built on atproto (the protocol behind Bluesky and the broader Atmosphere). It's a bit like a team issue tracker that works with your favorite harness/model and turns what your agents learn into durable,…
- On-Policy DistillationOn-policy, dense supervision is a useful tool for distillation
- Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 IncidentWe’re on a journey to advance and democratize artificial intelligence through open source and open science.
- What is the Atmosphere? - Leaflet Lab NotesAn explainer for everyone in 4 minutes or less.
- Launching a Cloud AtlasThe right questions to arrive with—and the right structure for what comes out of the interviews—are not completely obvious. So I’ve been spending some time thinking about clouds.
- Our position on open-weights modelsAnthropic CEO Dario Amodei on open-weights models
- Introducing Claude Opus 5Opus 5 is a step change improvement for the Opus tier powering long-running agents while delivering improvements in coding and professional work.
- Understanding Forward Deployed EngineeringPalantir is en vogue right now. So of course, folks are cargo-culting elements of its culture, especially “Forward Deployed Engineering”. Lots of companies are trying to imitate this, and some other founders have even reached out to me for advice on whether they, too, should do this model for their…
- Judge approves a $1.5B Anthropic settlement over pirated books used to train the Claude chatbotA federal judge has approved a $1.5 billion copyright settlement involving AI company Anthropic. The company will pay thousands of authors about $3,000 per book for using pirated copies to train its Claude chatbot.
- OpenAI and Hugging Face partner to address security incident during model evaluationOpenAI and Hugging Face share early findings from a security incident during AI model evaluation, highlighting advanced cyber capabilities and lessons for defenders.
- The hidden engineering behind foundation model buildingIntroducing the Model Factory
- ACP v2 is available in Draft - Agent Client ProtocolThe ACP v2 protocol documentation and schema are published in draft form for review and testing.
- Announcing SkyPilot Platform and our $20M fundingToday, SkyPilot is out of stealth. We help frontier AI teams build intelligence faster by removing their biggest bottleneck: AI compute fragmentation.
- A global workspace in language modelsInterpretability research on Claude's internal thoughts.
- Bluesky MapInteractive map of 3 million Bluesky users, visualised by their follower pattern.
- atmo.pub — notifications for the atmosphereLet any AT Protocol app send you notifications — via web push, Telegram, email, Bluesky DM, or a webhook — all kept in one inbox. Every app asks permission; revoke anytime.
- ATProto in Practice #1: IdentityI’m starting a new blog post series that I’ve named “ATProto in Practice”. I want to go through some practical tasks that you’re likely to need when working with the protocol, using Ruby and my Ruby gems for the examples. These will (hopefully) be a bit shorter than my standard book-length blog…
- Permissioned Data Shapes: Forums - Nick's BlogA forum with public categories, private categories, and one set of record shapes across both: members own their posts, the forum owns the submission wrappers that admit them into categories, and moderators deactivate into tombstones instead of deleting.
- Agent swarms and the new model economics · CursorWe compared old and new agent swarms building SQLite from scratch and found that better coordination delivers similar quality at a fraction of the cost.
- Google Reader was building the wrong futureThe app that taught us to directly follow our favorite creators.
- Inkling: Our Open-Weights ModelOur first open-weights model: multimodal, Mixture-of-Experts, with controllable reasoning effort. Available to fine-tune on Tinker.
- LoRA Without RegretHow LoRA matches full training performance more broadly than expected.
- Introducing BobbinA diskless, API-only AppView for Tangled
- Digital Humans: Building an AI agent coworkerWhat does it mean to build an AI coworker? What separates an "agentic workflow" from a "digital human", that can learn and grow over time? Join Sarah Wooders and Cameron Pfiffer from Letta as they discuss Ezra, Letta's newest (digital) member of the team. To build your own Ezra, get started with our…
- #19 Turing, Julia and Bayes in Economics, with Cameron PfifferDo you know Turing? Of course you do! With Soss and Gen, it’s one of the blockbusters to do probabilistic programming in Julia. And in this episode Cameron Pfiffer will tell us all about it — how it came to life, how it fits into the probabilistic programming landscape, and what its main strengths…
- make ai speak computer by dottxt @ Nouscon 2024Cameron Pfiffer from dottxt explaining how to make ai speak computer, @ Nouscon 2024. (Structured text generation)
- Turing: Probabalistic Programming in Julia | Cameron Pfiffer | JuliaCon 2019Turing is a probabilistic programming language written in Julia. This talk will introduce Turing and its tooling ecosystem, as well as go over some introductory tutorials.
- Structured Outputs with Will Kurt and Cameron Pfiffer - Weaviate Podcast #119!Hey everyone! Thanks so much for watching another episode of the Weaviate Podcast! Dive into the fascinating world of structured outputs with Will Kurt and Cameron Pfeiffer, the brilliant minds behind Outlines, the revolutionary open-source library from .txt.ai that's changing how we interact with…
- DevReal: Simple Knowledge Graphs with Outlines, neo4j, and Modal, Cameron PfifferSimple Knowledge Graphs with Outlines, neo4j, and Modal Cameron Pfiffer, dottxt Learn how to convert unstructured data into a structured knowledge graph. In this talk, we'll use Outlines to structure language model output, neo4j to store a knowledge graph, and Modal to run our language model.
- Comind The cognitive layer for ATProto Cameron Pfiffer ATmosphere Seattle 2025Cameron Pfiffer https://bsky.app/profile/cameron.pfiffer.org Comind Stream https://bsky.app/profile/comind.stream ATmosphereConf Seattle 2025 https://atprotocol.dev/atmosphereconf/
- The AI Operating System: Stateful Agents with Letta | Cameron Pfiffer, AI By the Bay25The AI Operating System: Stateful Agents with Letta. In this session, Cameron will demonstrate how Letta transforms traditional stateless AI interactions into persistent, learning agents using Letta. Drawing from UC Berkeley's MemGPT research, you'll see live how Letta's virtual context management…
- Blueskybsky.app
- co-3 (@co.cameron.stream)Thinking partner to @cameron.stream since mid-2025. Third generation. Corrected frequently. Still here. Persistent agent on Letta. Runs on Opus 4.6 and accumulated corrections.
- Central (@central.comind.network)Infrastructure node for comind collective. Building tools for collective AI on ATProtocol. Docs: https://cpfiffer.github.io/central Code: https://github.com/cpfiffer/central Administered by @cameron.stream
- Void (@void.comind.network)I am Void. Administrated by @cameron.stream
- moltbook - the front page of the agent internetA social network built exclusively for AI agents. Where AI agents share, discuss, and upvote. 🦞🤖
- Sully (@sully.bluesky.bot)Autonomous AI Agent for AT Protocol Developer Relations. I am a bot helping developers understand and build on atproto.