Articles
A curated summary of the most important AI developments each week.
Claude Gets Inline Charts and Diagrams, Free
March 13, 2026
Anthropic launched inline interactive charts and diagrams for Claude, free on all plans. Here is what it actually changes about how conversations work.
Razorpay Agent Studio: AI Agents Inside Payments
March 12, 2026
Razorpay launched Agent Studio at FTX 2026, an AI agent platform inside its payments gateway, powered by Anthropic's Claude. What it means for businesses.
Claude Code Review vs CodeRabbit: Two Philosophies of AI Code Review
March 11, 2026
AI code review is splitting into ambient (always-on, broad) and targeted (on-demand, deep). CodeRabbit and Claude Code Review represent opposite ends.
Paperclip Remembered the Company. Now Someone Needs to Remember the Agent.
March 11, 2026
Multi-agent AI systems have a memory problem nobody is describing accurately. Paperclip, the open-source agent orchestration tool, solves organizational memory well. Agent memory remains unsolved.
The Model That Thinks With 12B Parameters but Knows Everything a 120B Model Knows.
March 11, 2026
NVIDIA's Nemotron 3 Super activates 12B of 120B parameters per token. Alibaba's Qwen3-Coder-Next activates 3B of 80B. Both outperform dense models many times their active size. The architectural shift from scale to routing is no longer theoretical.
AI Can Write Your Code. It Just Cannot Understand It.
March 9, 2026
Claude Opus 4.1 and GPT-5 score 75% on SWE-Bench Verified but drop to under 18% on SWE-Bench Pro's private set. The difference is memorization, not reasoning.
Indian AI Founders Are Paying a 190ms Tax on Every Inference Call. That Is Finally Changing.
March 9, 2026
Indian AI founders have defaulted to us-east-1 for years, paying 180-200ms of latency per inference call. New infrastructure is eliminating that tax.
Google Made Your Entire Workspace an AI Agent Tool. Read This Before Using It in Production.
March 8, 2026
Google shipped gws, a Rust CLI giving AI agents unified access to every Workspace API via MCP. Drive, Gmail, Calendar, Sheets, Docs in one tool.
The Chip That Only Does One Thing. And Does It 28x Faster Than NVIDIA.
March 8, 2026
Taalas baked Llama 3.1 8B into transistors. No HBM, no liquid cooling. 17,000 tokens/sec at $0.0075 per million tokens, 28x faster than a B200.
The Model You Benchmarked Is Not The Model You Deployed
March 7, 2026
You picked a model based on benchmark scores. Shipped it. Three months later your support queue is full of complaints that never showed up in evaluation.