Notes on Rohit Krishnan's LLMEnron study: the main lesson is that agent systems fail less from raw model capability than from poor state structure. Explicit thread state, shared boards, and actor identity matter more than giant scratchpads or simply adding more agents.
Karpathy's autoresearch pattern — an AI agent that iterates on a model while a fixed eval script keeps score — turns out to work surprisingly well for Medicare fraud detection. Here's what happened when we pointed it at 90M claims records.
Sequoia partner Julien Bek argues the next $1T company will be a software company masquerading as a services firm. If you sell the work instead of the tool, every model improvement makes you faster and cheaper — not obsolete.
AI browser agents can now interact with web apps in real-time. This changes how we build — apps need to serve two users: humans clicking buttons, and AI agents calling tools.
We don't need more apps. We need something that connects the ones we already have. AI isn't replacing your tools — it's becoming the layer that finally makes them work together.
Forget SaaS. Forget multi-tenant. The future of enterprise AI is bespoke applications deployed inside the client's own infrastructure. Here's the business model — and why it works.
CMS publishes incredible healthcare data — if you can wrangle it. Here's what I learned building a pipeline to ingest NPPES, Open Payments, and Medicare utilization data into DuckDB.
How do you know that 'Dr. Jane Smith' in CMS data, 'Jane A. Smith MD' on Google, and 'Smith, Jane' on a practice website are the same person? This is the matching problem—and solving it unlocks massive value in healthcare data.
I gave an AI agent access to my web app. Then I rebuilt the app to actually talk back. Here's what happened — and why every web developer should be paying attention.
The most powerful AI agent I've used isn't a chatbot, a copilot, or an app. It's a terminal session with tools. Here's how OpenClaw works under the hood — and why 'bash is all you need' isn't a joke.
Your doctor's office can't reliably tell which John Smith you are across two different systems. That's not a tech problem — it's a 50-year-old design choice we never fixed.
Peter Steinberger (creator of OpenClaw) explains why 80% of apps will disappear in the agent era, why fighting agent access makes you less useful than your competitors, and what agent-friendly companies should do right now.
Pi is a while loop + 4 tools. That's it. This 'minimal by design' agent harness powers OpenClaw and proves that the most elegant agent architecture might just be bash.
WebMCP is like responsive design for AI — annotate your website once, and agents can discover and use your tools automatically. No separate MCP server needed.
You can build an AI app in a weekend. But Azure's portal makes your eyes glaze over. This guide maps Azure's concepts to what you actually need.
Why the SaaS model you grew up on is already obsolete — and what replaces it.