Daniel Graviet

Writing

All posts

Old Ideas, New Machines

The same patterns keep reappearing in new disguises — branch prediction becomes speculative decoding, virtual memory becomes PagedAttention.

My Two Favorite & Greatest Rebrands

Japan and Darius Rucker don't have much in common, but both pulled off some of the most remarkable rebrands in modern history.

Agent-Aware Snapshots with Daytona

A look at using agent-aware snapshots to create and restore the development environments AI agents need more efficiently.

Kubernetes Is the Wrong Primitive for Diverse Workloads

Kubernetes excels at scaling stable services, but short-lived AI workloads often need many distinct, prepared environments that can be launched on demand.

Why I Am Studying Classic Reinforcement Learning Now

Modern AI is rediscovering old reinforcement learning problems: reward design, credit assignment, exploration, evaluation, and learning from feedback.

The Hidden Infrastructure Tax in Coding-Agent RL

Coding-agent RL depends on real software environments, so provisioning and command latency can become a significant part of training cost and rollout throughput.

GRPO: Learning From the Other Answers in the Room

A simple breakdown of GRPO, a reinforcement learning method that removes the need for a separate value model. Full of analogies, and images to make the concept more approachable.

What Longer-Timeline Intuitions About RL Progress Missed

An argument for why AI progress did not slow in the RL regime as much as some longer-timeline intuitions expected.

PPO Explain for beginners

A beginner-friendly breakdown of Proximal Policy Optimization — the RL algorithm that turned raw base models into useful AI assistants.

Hello World

My first blog post — a quick intro to what I'll be writing about.