Sunil Mallya
Writing about LLMs, reinforcement learning, and building intelligent systems.
Writing
-
Jev Is an Implementation. The Control Loop Is Moving Back Into Code.
We've spent a few years moving control flow out of code and into inference. Typed semantic decisions, the "semantic if," suggest orchestration is moving back into code while the judgment inside each branch stays learned.
AI Agents Software Architecture -
AI Doesn't Replace Workers. It Rewires the Boundaries Between Them.
The real change isn't speed — it's that role boundaries are dissolving. The marketer is building. The engineer is running a fleet. The model is only the steam engine. The factory system forming around it is the story.
Future of Work AI Agents -
Models Eat Their Scaffolding
AI platforms mature by absorbing the scaffolding built around them. What starts as wrappers, retrieval layers, tool glue, and agent frameworks turns into model primitives or platform infrastructure.
AI Platforms AI Agents -
AgentDiff: git blame for vibe coding
When an AI writes your code, git can tell you what changed, but not why. AgentDiff captures every decision your coding agent makes: the prompt, the reasoning, and which requirement it maps to.
Developer Tools AI Agents -
The Great Compression
How AI is collapsing the enterprise stack. When models can reason, act, and interact, the layers built for human navigation of software start to hollow out from the middle. Five layers become three.
Enterprise AI AI Agents -
The Myth of RLVR
Why verifiable rewards aren't enough and what it takes to build open-world agents. The belief that correctness is the bottleneck is the myth. The real bottlenecks are information flow, credit assignment, and trajectory-level learning.
Reinforcement Learning AI Agents