Madhavan Suresh·Aug 31I Tested Two Memory Frameworks for AI Agents. Both Scored 100% by Forgetting.A paper says stored memory makes models reason worse. I tested that on Mem0 and Graphiti. Both scored perfect, and then the logs showed me…
Madhavan Suresh·Aug 21The Weights Never Changed. The Score Doubled.An agent went from 43.5% to 93.0% on real web tasks with no retraining, and the part that did it is the part you already own.
Madhavan Suresh·Aug 18The Attack That Never Lies to Your AgentTwo papers from August 2026 break AI agents without a single malicious prompt. The method explains why your safety tests keep passing while…
Madhavan Suresh·Aug 17Stop Picking a Model. Start Picking a Router.Nvidia says the same work costs a third as much if you send each request to the cheapest model that can do it. Here is the mechanism, and…
Madhavan Suresh·Aug 14One AI Solved a Number Theory Problem. Another Messed Up a Refill.Both happened this week. The gap between them is not intelligence, and it is not model size. It is one design decision you can copy.
Madhavan Suresh·Dec 16, 2025The Future of SaaS Is a Download Button, Not a Login PageHere is an idea that struck me as a Aha! moment. These are solely my personal convictions.
Madhavan Suresh·Nov 21, 2024The AI vs Software Paradox: Lessons from Japan’s Software Development ApproachMy manager always said this story about his Japan counterparts during the initial days of software development. They always tested…
Madhavan Suresh·Mar 16, 2023The next global virus or the next transistor?As an engineer, I’ve always been looking for ways to improve my productivity and efficiency. That’s why I was excited to discover ChatGPT…