Cursor Blog⏱ 4 min🛠 builder tools
Cursor 3: A Unified Workspace for Building Software with Agents
Cursor's major new version creates a unified workspace where multiple AI agents can collaborate on software projects together.
Editor's Note
Cursor 3 dropped their multi-agent workspace. Everyone's hitting the same wall — single-agent coding turns into spaghetti pretty quick.
But the real stuff's happening in the boring infrastructure layer. Vercel cut sandbox restores from 40 seconds to instant. Google's doing continuous checkpointing. This matters when your training run dies at 90% and you've just torched $50k.
Yegge's Gas Town hit v1.0. Someone's actually running federated AI coding in production now. The Waldium piece is wild — giving every blog its own MCP server endpoint. Could be genius. Could be completely nuts.
The capabilities research feels different lately. Less breathless hype, more actual measurement. The compute capacity analysis between OpenAI and Anthropic reads like intelligence gathering, not recycled press releases. And that economics piece confirming semiconductors still capture 70% of AI revenues? The picks-and-shovels play is working exactly like everyone said it would.
⚡ Vibe Coding
Claude Code Changelog⏱ 2 min🛠 builder tools
Claude Code v2.1.91: MCP Tool Result Persistence and Shell Execution Controls
New version adds support for 500K character tool results and lets you disable shell execution in skills and plugins.
Google DeepMind⏱ 6 min🛠 builder tools
Gemma 4: Byte for Byte, the Most Capable Open Models
Google's new open models (2B to 31B parameters) are designed specifically for advanced reasoning and agentic workflows.
Replit Blog⏱ 8 min🛠 builder tools
Why Smart PMs Are Using Vibe Coding to Cut Design Delays
Product managers can skip traditional design handoffs by prototyping directly with AI tools, cutting weeks of lag per quarter.
Vercel Blog⏱ 5 min🛠 builder tools
Optimizing Vercel Sandbox Snapshots: From 40 Seconds to Near-Instant
Vercel cut sandbox restore times from 40+ seconds to near-instant through parallelization and smart caching strategies.
Vercel Blog⏱ 6 min🛠 builder tools
How Waldium Made a Blog Platform Work for Humans and AI Alike
YC startup built an agentic CMS that gives every customer blog its own MCP server endpoint for direct AI agent queries.
Steve Yegge⏱ 7 min🛠 builder tools
Gas Town: From Clown Show to v1.0
Steve Yegge's federated AI coding platform officially hits v1.0 after three months of wild development and user adoption.
TLDR AI⏱ 4 min🛠 builder tools
Claude Code's Real Secret Sauce (Probably) Isn't the Model
Analysis of the leaked Claude Code reveals sophisticated software harness with dedicated tools, memory architecture, and parallel subagents.
Google Developers AI⏱ 6 min🛠 builder tools
Continuous Checkpointing: Optimizing Training Reliability in Orbax and MaxText
Google's new continuous checkpointing feature maximizes I/O bandwidth and minimizes failure risk during model training.
🧠 The Big Picture
LessWrong⏱ 12 min🧪 deep analysis
Q1 2026 Timelines Update: Adjusting AI Progress Expectations
Major AI forecasting update shows how recent evidence is shifting expectations about AI development timelines and takeoff speeds.
LessWrong⏱ 8 min⚖️ AI futures
Intelligence Dissolves Privacy: How AI Changes Social Expectations
As AI capabilities expand, our social expectations about privacy and what's 'reasonable' to keep hidden will fundamentally shift.
Simon Willison⏱ 10 min🧪 deep analysis
Highlights from Agentic Engineering Conversation on Lenny's Podcast
Simon Willison discusses the November inflection point, software engineers as bellwethers, and automation timelines in AI development.
TLDR AI⏱ 8 min🧪 deep analysis
The Economics of Generative AI: Two Years Later
Semiconductor companies still capture 70% of AI revenues, while infrastructure remains the only truly competitive layer in the stack.
TLDR AI⏱ 3 min🧪 deep analysis
Compute Wars: OpenAI vs Anthropic Capacity Analysis
Anthropic's doubled capacity enabled Opus 4.5 breakthrough, while OpenAI will pull ahead in late 2026 but face close competition in 2027.
LessWrong⏱ 15 min🧪 deep analysis
The Corner-Stone: Is the US Actually a Cognitive Meritocracy?
Deep analysis of whether high IQ actually guarantees success in America's supposed meritocratic system, challenging common assumptions.
TLDR AI⏱ 5 min🧪 deep analysis
PrismML Claims Radical 1-Bit AI Model Compression Breakthrough
Caltech spinout claims extreme compression technology that enables full AI models to run locally on edge devices without performance loss.
Alignment Forum⏱ 6 min🧪 deep analysis
My Most Common Research Advice: Do Quick Sanity Checks
AI safety research mentor breaks down the three most common feedback areas for junior researchers, with practical guidance.