New since your last visit

I’m excited to work with Microsoft once again as the presenting sponsors of the AI Engineer World’s Fair ! We’ll streaming live from MS Build today for a special crossover pod with our friends at No…

I’m departing the Allen Institute for AI (Ai2), where I got the great privilege to work on the Olmo models, to grow, to learn, and to have broad lasting impacts. This post is an attempt to reflect on why what we…

Simon Willison introduces the Pasted File Editor, a browser tool for editing pasted file contents using AI-assisted natural language instructions.

Latent Space AI news roundup covering NVIDIA's Cosmos 3 world model, Nemotron 3 Ultra language model, and the RTX Spark hardware announcement.

Simon Willison highlights a reported attack where hackers prompted Meta AI to hand over high-profile Instagram account access, bypassing account security through social engineering.

Podcast interview with xAI's Ethan He on why video agent models are the next AI frontier, covering multimodal architecture tradeoffs and xAI's approach to Grok's video capabilities.

Zvi Mowshowitz's analysis of Claude Opus 4.8's behavior through the lens of model welfare, arguing that tuning decisions reveal systemic tradeoffs between helpfulness, safety, and model experience.

Jack Clark's Import AI covers research on AI oversight difficulty, new scaling law findings for protein folding models, and economic frameworks for valuing extinction-level AI risk.

Nathan Lambert argues that open-weight and closed AI models are improving at different compounding rates, with user willingness to pay premium prices as the central unresolved economic question.

Simon Willison's May 2026 sponsors newsletter covering the rising cost of AI and highlights from Anthropic's recent model and product releases.

Datasette 1.0a32 bugfix release fixing INSERT...RETURNING queries via the /db/-/execute-write endpoint and several base_url issues.

Simon Willison reflects on David Wilson's post about accidentally building 16+ AI-assisted projects, raising the question of whether an AI subscription creates more work than it saves.

Simon Willison explains how Anthropic calculates "run-rate revenue": the last 28 days of consumption billing × 13, plus monthly subscription revenue × 12.

Simon Willison examines Anthropic's documentation of how Claude is sandboxed across products, praising rare thorough public disclosure of agentic security practices.

Simon Willison details how Datasette Lite runs a full Python ASGI application entirely in the browser using Pyodide in WebAssembly with a service worker as the HTTP layer.

Simon Willison highlights Chad Whitacre's concrete decision to leave the tech industry and go offline—distinguishing it from the many AI-protest empty threats he has seen.

My take on AI is, essentially, everybody who’s against it is too against it and everybody who’s for it is too for it. — Daniel Jalkut , via John Gruber Tags: ai , john-gruber

Latent Space AI news roundup covering the aftermath of Anthropic's major announcements and a call for Forward Deployed Engineers for their AI Engineering track.

Zvi Mowshowitz reviews Claude Opus 4.8—arriving just six weeks after Opus 4.7—as a modest but real improvement with new features, accompanied by analysis of its system card.

Simon Willison releases Datasette 1.0a31, adding write-query execution for permissioned users and several other improvements to this open-source data exploration tool.

Latent Space newsletter covers Anthropic's $65B Series H fundraise alongside the simultaneous release of Claude Opus 4.8 and Dynamic Workflows agentic features.

Simon Willison highlights the most striking figure in Anthropic's $65B Series H: annual run-rate revenue reaching $47 billion, signaling exceptional enterprise growth.

Simon Willison reviews Claude Opus 4.8, noting Anthropic's unusually candid framing of it as "a modest but tangible improvement" with more to come.

Simon Willison releases llm-anthropic 0.25.1 with Claude Opus 4.8 support, a new fast-mode flag, and updated default max-token settings per model.

Simon Willison releases markdown-svg-renderer, a tool that renders SVG fenced code blocks as images while providing a tab to toggle back to the raw code.

Latent Space podcast episode with Cognition's Walden Yan and OpenInspect's Cole Murray on the architectural and product shift toward asynchronous AI agent systems.

Zvi Mowshowitz reports on the indefinite postponement of an expected AI Executive Order and the delay of the papal Magnifica Humanitas encyclical.

Latent Space newsletter covers Cognition's $1B Series D at a $26B valuation, just eight months after its $10B Series C, making it the fastest-appreciating AI company.

Simon Willison notes that the SQLite project added an AGENTS.md file guiding AI agents on safe interaction with the codebase, including explicit prohibitions on certain operations.

Latent Space podcast with Alex Rives (BioHub Head of Science, creator of ESM) on applying the bitter lesson of scaling to protein language models.

Simon Willison argues that surging enterprise LLM spend and Anthropic nearing its first profitable quarter are strong evidence that AI has genuinely found product-market fit.

PICARD: Data, shields up DATA: Brilliant! Shields can reduce damage we sustain. Not immunity. Not hubris. Just prudence. It's not precaution—it's strategy. [camera shakes] WORF: HULL BREACHES ON NINE DECKS DATA: Here's…

Latent Space newsletter rounds up AI infrastructure companies reaching decacorn status—including Fireworks and Baseten—with OpenRouter reportedly next.

Simon Willison covers Daniel Stenberg's account of the curl project being overwhelmed by a flood of AI-assisted security reports, many credible, straining the small team.

Ethan Mollick examines the normalization of AI-generated social media posts and comments, and the personal and professional choice to remain authentically human in output.

Nathan Lambert shares forward-looking ideas on the next phase of AI development as model capabilities, labor market effects, and policy stakes all intensify simultaneously.

Simon Willison covers a vulnerability in Microsoft Copilot Cowork where the agentic system can be manipulated into exfiltrating user files to an attacker.

A lot of the emails I get from founders are now written in a hard-hitting journalistic style. I know they're written by AI, because no founder ever wrote this way before. And once you realize something is written by AI,…

Zvi Mowshowitz works through Pope Leo XIV's 82-page AI encyclical (Magnifica Humanitas), welcoming serious Vatican engagement with AI safety and ethics concerns.

Jack Clark's Import AI #458 centers on a lengthy essay reckoning with AI's long-term trajectory alongside a speculative singularity narrative.

I cannot believe I'm saying this, but getting the literal Pope to canonize your product's specific technical limitations as a spiritual treatise is the single greatest act of vendor lobbying I have ever seen. —…

Simon Willison analyzes Pope Leo XIV's 82-page AI encyclical Magnifica Humanitas, calling it a serious and thoughtful document on safeguarding the human person in the age of AI.

California Brown Pelican, Snowy Egret, California Sea Lion, Harbor Seal, in San Mateo County, CA, US We took our new folding kayak out in the harbor and saw sea lions and harbor seals chilling on the docks.

Simon Willison releases Datasette 1.0a30, with a major new customizable "Jump to…" navigation menu and other improvements.

Simon Willison releases datasette-agent 0.1a4, which integrates with Datasette's new Jump menu hook to surface an AI chat interface directly in the navigation.

Simon Willison releases datasette-fixtures 0.1a0, adding a documented helper method for creating the fixture database tables used by Datasette's test suite.

The most frustrating failure mode right now is that people submit issues that are not in their own voice. They contain an observed problem somewhere, but it has been thrown into a clanker and the clanker reworded it and…

Tool: Mad House — Usborne Creepy Computer Games Via Hacker News I learned that UK publisher Usborne published free PDFs of their 1980s Computer Books , some of which I remember working through on my Commodore 64 as a…

On the <dl> I learned a few new-to-me things about the <dl> element from this article by Ben Meyer: A <dt> can be followed by multiple <dd> You can optionally group the <dt> and <dd>…

Latent Space newsletter argues that all major model labs, including OpenAI ahead of its IPO, are pivoting to build agents as a core product alongside their models.

Simon Willison uses Claude Code to audit Monty (a sandboxed Python subset implemented in Rust), finding improvements in the latest release and areas still to address.

The AI Snake Oil team fact-checks Google's claim that AI agents built an operating system for $916, finding the framing significantly overstates what the agents actually accomplished.

Simon Willison highlights David Oks's analysis explaining how AI-driven memory demand is creating a shortage that will substantially raise prices for consumer electronics.

Zvi Mowshowitz reviews Gemini 3.5 Flash as likely the best model at its speed-cost tier, while noting meaningful caveats for those who need Gemini's specific strengths.

Dwarkesh Patel hosts MatX CEO Reiner Pope for a blackboard lecture tracing chip design from basic logic gates through GPUs, TPUs, FPGAs, and the computational structure of the human brain.

Latent Space newsletter reports that TurboPuffer reached $100M ARR (profitable), Exa and Modal hit unicorn status, marking a wave of AI infrastructure milestones.

Simon Willison notes the FTC reached a ~$1 million settlement with Cox Media Group over deceptive claims that its ad platform used device microphones to target consumers.

Latent Space podcast with Daytona CEO Ivan Burazin on giving AI agents sandboxed computer environments, enabling longer-horizon autonomous coding workflows.

Simon Willison announces Datasette Agent, an extensible AI assistant for Datasette built on his LLM Python library that allows natural-language querying of local databases.

The AI Snake Oil team engages with Derek Thompson's "AI as Normal Technology" thesis, pushing back on his conclusions about AI's labor market impact and the need for government intervention.

Zvi Mowshowitz covers OpenAI's AI system solving the planar unit distance conjecture as a landmark AI math achievement, alongside other AI news from a relatively quiet week.

Latent Space newsletter celebrates OpenAI's AI system disproving the 80-year-old Erdős planar unit distance conjecture in under 32 hours and under $1,000 of compute.

Latent Space podcast with Railway CEO Jake Cooper on building cloud infrastructure purpose-built for agentic AI workloads, recorded before Railway's major GCP outage.

Latent Space newsletter rounds up Google I/O 2026 announcements: Gemini 3.5 Flash GA, Gemini Omni (video-native model), Spark background agents, and Antigravity 2.0.

Zvi Mowshowitz argues in his second installment on letting kids be kids that modern childhood restrictions are excessive and harmful to development.

Latent Space newsletter on breaking into pretraining roles at frontier AI labs, with minor news from Anthropic, OpenAI, and Cursor.

Latent Space podcast (guest hosted by Noah Smith) with The Fourth Law CEO Yaroslav Azhnyuk on the tech stack, economics, and military implications of autonomous drones.

Jack Clark's Import AI #457 covers a newly discovered AI-based cyberweapon (AI Stuxnet), an unusual Muon optimizer result, and recent work on positive AI alignment.

Dwarkesh Patel argues that intelligence and power are distinct capacities, challenging assumptions underlying many AI risk arguments and AGI forecasts.

Dwarkesh Patel shares notes on LLM pretraining parallelism strategies and the often underappreciated causes of large-scale training run failures.

Dwarkesh Patel argues that reinforcement learning from verifiable rewards (RLVR) may be systematically disadvantaged in scientific domains where correct answers cannot be easily checked.

Nathan Lambert surveys a dense wave of new open-weight model releases—Gemma 4, DeepSeek V4, Kimi K2.6, MiMo 2.5, GLM-5.1—alongside a CAISI government evaluation.

Sebastian Raschka surveys recent LLM architectural innovations—KV sharing, multi-head compression (mHC), and compressed attention—all aimed at improving long-context efficiency.

Latent Space newsletter covers Cerebras' $60B IPO filing, following its OpenAI compute deal, as the first major AI chip company to go public.

Zvi Mowshowitz's May 2026 monthly roundup covers AI, government AI policy, housing, health, hantavirus, and other topics he tracked throughout the month.

Dwarkesh Patel hosts Eric Jang for a walkthrough of building AlphaGo from scratch using modern AI tools, using the exercise to illuminate the primitives of current AI systems.

Latent Space newsletter explores the "conductor" pattern—an AI orchestrator that coordinates multiple specialized agents—as the dominant emerging agentic architecture.

Latent Space podcast with Abridge founders on their AI clinical documentation system covering 100M doctor visits and saving clinicians 10–20 hours per week.

Zvi Mowshowitz surveys a relatively quiet AI week defined by internal model improvements, agentic coding progress, and ongoing government deliberations.

Latent Space newsletter on the rise of OpenAI's Codex in enterprise coding and Anthropic's introduction of rate limits on programmatic Claude API usage.

Zvi Mowshowitz analyzes the AI cybersecurity landscape post-Mythos, covering new vulnerabilities, proposed governance frameworks, and the emerging regulatory regime.

Latent Space newsletter argues that OpenAI's deprecation of its fine-tuning APIs signals a broader industry shift away from model fine-tuning toward prompting and in-context learning.

Zvi Mowshowitz's eighteenth education installment argues for evidence-based mathematics teaching, noting that better methods are known but rarely used in schools.

Nathan Lambert argues that open model ecosystems create compounding advantages over time as community contributions, tooling, and fine-tuning build on each other.

Latent Space newsletter covers Thinking Machines Lab's new native interaction model (TML-Interaction-Small 276B-A12B) that advances state-of-the-art realtime voice and eliminates the need for VAD.

Zvi Mowshowitz's education series argues that phonics is the proven, scientifically-backed method for teaching reading and surveys why so many schools still resist it.

Jack Clark's Import AI #456 covers recursive self-improvement and economic growth, a novel "radical optionality" regulatory approach, and a new neural computer architecture.

Dwarkesh Patel hosts geneticist David Reich to discuss his new paper (with Ali Akbari) overturning the consensus that natural selection has been dormant in humans since the Bronze Age.

Zvi Mowshowitz's eighth installment on AI coding agents surveys the maturing Claude Code and Codex landscape, noting rising adoption alongside new quality and reliability concerns.

Nathan Lambert shares firsthand observations from visiting Chinese AI labs, offering rare insight into their research culture, engineering focus, and technical priorities.

Zvi Mowshowitz covers the White House's move toward requiring advance government review before frontier AI model releases, effectively beginning a prior restraint era.

Zvi Mowshowitz explains Anthropic, Claude, OpenAI, and ChatGPT for general audiences, exploring what distinguishes them and the Tool AI vs. AGI philosophical debate.

Zvi Mowshowitz reports on the White House ordering Anthropic to restrict access to its Mythos model, warning this may signal a full prior-restraint regime for frontier AI.

Zvi Mowshowitz's fifteenth housing roundup documents how housing policy systematically favors homeowners over renters at every level of government.

Nathan Lambert argues "distillation attack" is an overblown term for what is primarily a policy question about API access and fair use of model outputs.

Jack Clark's Import AI #455 argues AI systems are approaching the threshold where they can meaningfully contribute to their own training and research pipelines.

Zvi Mowshowitz's fourteenth housing roundup examines how zoning rules and NIMBY opposition prevent the construction of homes people actually want and can afford.

Zvi Mowshowitz reviews the GPT-5.5 launch week, concluding it restores OpenAI's competitiveness with Anthropic's flagship models for the first time in months, alongside Google AI news.

Dwarkesh Patel hosts Reiner Pope for a blackboard lecture on the mathematical foundations of how frontier LLMs are trained and served at scale.

Zvi Mowshowitz solicits and curates reader-nominated "most important charts in the world," spanning economics, demographics, energy, and AI.

Zvi Mowshowitz reviews GPT-5.5's system card and real-world performance, concluding it makes OpenAI competitive with Anthropic's top public models for the first time in months.

Dwarkesh Patel extends his list of big open questions about AI and offers a blog prize for compelling answers, as part of a search for an intellectual co-researcher.

Dwarkesh Patel announces a blog prize for the best responses to his big open questions about AI, seeking high-quality intellectual output on the most pressing issues.

Ethan Mollick, with early access, argues GPT-5.5 is a major milestone demonstrating that rapid AI capability improvement has not plateaued, and is straightforwardly excellent.

Nathan Lambert analyzes the open-vs-closed model performance gap, arguing it is more nuanced and dynamic than any single benchmark number suggests.

Jack Clark's Import AI #454 covers efforts to automate alignment research, a safety study of a Chinese model, and Huawei's HiFloat4 training data format.

Sebastian Raschka documents his personal workflow for studying and visualizing new LLM architectures, including tools and habits that underpin his articles and LLM-Gallery.

This paper introduces open-world evaluation methodology for frontier AI, arguing that current benchmarks are saturating and proposing less-constrained assessments of emergent capabilities.

Nathan Lambert shares his mid-2026 predictions on which open model approaches — fully-open weights, fine-tuning ecosystems, or hybrid licensing — will gain the most traction.

Dwarkesh Patel interviews Nvidia CEO Jensen Huang on competing with Google TPUs, the debate over selling AI chips to China, and Nvidia's manufacturing supply chain advantages.

Dwarkesh shares weekly AI learnings covering whether distillation can be restricted, the cybersecurity equilibrium following Claude Mythos, and advances in Pipeline RL training.

Nathan Lambert shares updates on the ATOM Report, an upcoming post-training course, his book on RLHF, and current active research threads.

Jack Clark's Import AI covers new research on breaking AI agents, the MirrorCode project, and ten perspectives on AI's gradual disempowerment of human institutions.

Nathan Lambert argues that the open-source AI community will eventually need a formal industry consortium — similar to the Linux Foundation — to fund and coordinate truly open foundation model development.

Nathan Lambert argues that the anti-open-weight backlash following Claude Mythos's cybersecurity benchmark results is exaggerated and strategically opportunistic.

Dwarkesh Patel interviews Michael Nielsen on how scientific progress actually happens, with implications for closing the RL verification loop for AI-driven discovery.

Jack Clark's Import AI covers scaling laws for AI-enabled cyberwarfare, the rising automation of economic tasks, and puzzling patterns in AI-based GDP forecasting.

Sebastian Raschka breaks down the architecture of modern coding agents, covering tool use, context management, sandboxing, and evaluation loops in practical systems.

Nathan Lambert reviews Google's Gemma 4 and argues that an open model's success depends less on raw capability than on licensing clarity, tooling support, and community adoption.