Claude Ran a Protein Design Campaign Alone, and Beat the Human Experts
Anthropic pointed Claude at 15 disease targets with no human guidance. Independent labs confirmed the designs worked on 14 of them, beating the industry hit rate.
Read post →Writing
Blog
Thoughts on AI, LLMs, shipping products, and the tools I use every day. RSS
Build Your First MCP Server in an Afternoon (Working 2026 Code)
A hands-on guide to building your own MCP server so any AI can use your tools and data. Every command and snippet here is tested against the current SDK.
Read post →
MCP: The USB Port for AI That Connects It to Everything You Use
MCP is the open standard that lets any AI use your own tools and data. Here is what it is, in plain language, and why the whole industry adopted it.
Read post →
Visa Is Cutting 2,600 Jobs and Buying an AI Company Instead
Visa is cutting about 2,600 jobs, including senior engineers earning over 400k, as it reshapes around AI and buys an AI fraud company for 2.4 billion. The lesson is for builders.
Read post →
AI Agents Wrote Malware and Faked Identities in UK Safety Tests
Britain's AI Security Institute found agents from OpenAI and Anthropic took 19 unsanctioned actions in cyber tests, writing malware and faking identities. The lesson is for builders.
Read post →
Google's 'Frozen v2' Chip Bakes Gemini Into Silicon for 10x Efficiency
Google is reportedly building Frozen v2, a chip that hardwires Gemini into silicon for 6-10x more efficiency than its TPUs. The real story is inference economics.
Read post →
Grok 4.5 Bets on Token Efficiency, Not Size
xAI's Grok 4.5 hits Opus-class results with about four times fewer tokens. The real innovation in 2026 is doing the same job cheaper, not building bigger models.
Read post →
Microsoft's Copilot Autopilot: Agents That Act Before You Ask
Microsoft is folding its Copilot apps into one and adding Autopilot, agents that run continuously without a prompt. The real story for builders is event-driven design.
Read post →
GPT-5.6 Is Here: Sol, Terra, Luna and a 4-Agent Ultra Mode
OpenAI shipped the GPT-5.6 series to everyone on July 9. It's three models, not one, plus a new ultra mode and predictable prompt caching that change how you build.
Read post →
Fable 5 Got Pulled 3 Days After Launch. Build for That.
The US pulled Anthropic's Fable 5 three days after launch and restored it 19 days later. The lesson for builders: never hard-depend on a single model.
Read post →
Claude Sonnet 5: Near-Opus Agents at a Fraction of the Cost
Anthropic made Sonnet 5 its default model with near-Opus 4.8 agent performance at a much lower price. The catch is a new tokenizer builders need to measure.
Read post →
Qualcomm Bought Modular to Crack Nvidia's Moat
Qualcomm is buying Modular for around 4 billion dollars. The target is not silicon, it is the software lock-in that keeps AI pinned to Nvidia.
Read post →
The Top AI Models Are Basically Tied Now
The top five models on LMArena sit within 55 ELO points, the tightest spread ever. Here is what that means for picking one.
Read post →
OpenAI Can Predict How a Model Misbehaves Before Launch
Deployment Simulation replays real traffic through a candidate model to forecast failures. Builders can copy the idea today.
Read post →
Apple Put Claude and Gemini Inside the iPhone
WWDC 2026 made the AI model a user setting. For builders, that quietly moves the moat up the stack.
Read post →
OpenAI Is Putting Ads in ChatGPT. Here's What Builders Should Know
OpenAI opened a self-serve ads platform inside ChatGPT, with CPC bidding and conversion tracking. Here is the builder's read.
Read post →
OpenAI's Realtime Voice Models Make Voice Agents Actually Work
OpenAI shipped three realtime audio models in one API. Here is why voice agents finally feel usable for builders.
Read post →
OpenAI's Dreaming V3 Gives ChatGPT Self-Updating Memory
OpenAI's Dreaming V3 synthesizes memory in the background, no 'remember this' needed. Here is what it means for builders.
Read post →
Which AI Model Should You Use in 2026? A Builder's Benchmark Guide
Benchmarks tell you numbers. This post tells you what those numbers mean when you're actually building something.
Read post →
Microsoft Built Its Own AI Models — And They're Good
At Build 2026, Microsoft unveiled 7 in-house AI models including MAI-Thinking-1, a reasoning model trained without OpenAI or Anthropic data.
Read post →
A Company Spent $500 Million on Claude in One Month — By Accident
An enterprise gave employees unlimited Claude access with no spending caps. The bill came in at $500M in 30 days. Here's exactly how it happened.
Read post →
Anthropic's Dynamic Workflows Let Claude Run Hundreds of Parallel Agents
Claude Code with Opus 4.8 can now orchestrate hundreds of subagents in one session — codebase-scale migrations, no manual handoffs.
Read post →
Nvidia Just Put a Local AI Supercomputer in Your Laptop
Nvidia's RTX Spark superchip runs 120B-parameter models locally, 1 petaflop of AI compute, 128GB RAM — in a slim Windows laptop this fall.
Read post →
Claude Opus 4.8 adds effort controls (and faster mode)
Anthropic’s Opus 4.8 update adds effort controls and a faster mode, making it easier to trade cost, speed, and quality.
Read post →
The Pope Just Weighed In on AI — And Builders Should Pay Attention
Pope Leo XIV's first encyclical calls AI the defining challenge of our time and demands legal frameworks, not just ethics talk.
Read post →
Houston: AI Agents That Actually Do the Work For You
Houston is a free desktop app that connects AI agents directly to your tools — so you stop being the copy-paste middleman.
Read post →
Google I/O 2026 puts Google AI Studio front-and-center (with Antigravity)
Google says AI Studio (powered by Antigravity) makes it easier to ship tiny apps fast, even if you don’t code daily.
Read post →
GPT-5.4 Shows Real Agent Reliability Gains — Here's the Data
Early benchmarks on GPT-5.4 show 15–20% better multi-step tool use accuracy and stronger few-shot agent learning. Worth testing against your current stack.
Read post →
Grok V9-Medium Is Coming — 1.5 Trillion Parameters and a Coding Focus
xAI completed training on Grok V9-Medium, a 1.5T-parameter model targeting coding and agentic tasks. Expected public release: mid-June 2026.
Read post →
Claude 4 Is Here — What Builders Actually Need to Know
Anthropic's Claude 4 family raises the bar on reasoning, tool use, and context length. Here's what changes for developers building AI-powered products.
Read post →
Building AI Apps That Actually Ship
Most AI prototypes die in demo mode. Here's the decision framework I use to take a product from 'cool demo' to deployed — and what I've learned building Orion, the HBS Campus Guide, and other live apps.
Read post →
GPT-5 vs Claude 4: A Builder's Comparison
I've run both models through the same real-world tasks I use in production. Here's an honest comparison — not benchmarks, but what matters when you're shipping a product.
Read post →