---
title: "Duane Nielsen (@duanerocks) | Hashnode"
description: "Obsessed with RL. Read the latest articles by Duane Nielsen on Hashnode."
url: "https://hashnode.com/@duanerocks"
---

@duanerocks

# Duane Nielsen

@duanerocks · Joined September 2025

Obsessed with RL

### [The Building Blocks of an Agent Memory System](https://proximal.hashnode.dev/the-building-blocks-of-an-agent-memory-system?utm_source=hashnode&utm_medium=feed)

The Building Blocks of an Agent Memory System Most agent "memory" systems retrieve too much. They paste the last N turns into the context window, or they dump the top-K results from a vector search, and they hope the model finds the signal. The model...

14 min read · [proximal.hashnode.dev](https://proximal.hashnode.dev/the-building-blocks-of-an-agent-memory-system?utm_source=hashnode&utm_medium=feed)

[Apr 26](https://hashnode.com/posts/the-building-blocks-of-an-agent-memory-system/69eda982024dd2c7df9eb04f)

### [Smaller is Better: Replacing GPT-4o-mini with a 7B Local Judge](https://proximal.hashnode.dev/smaller-is-better-replacing-gpt-4o-mini-with-a-7b-local-judge?utm_source=hashnode&utm_medium=feed)

I expected the 30B model to be the better judge. It wasn't. When I set out to replace OpenAI's GPT-4o-mini as the judge for the Oolong benchmark, my plan was simple: use the biggest local model I had. Qwen3-coder at 30B parameters seemed like the obv...

4 min read · [proximal.hashnode.dev](https://proximal.hashnode.dev/smaller-is-better-replacing-gpt-4o-mini-with-a-7b-local-judge?utm_source=hashnode&utm_medium=feed)

[Feb 5](https://hashnode.com/posts/smaller-is-better-replacing-gpt-4o-mini-with-a-7b-local-judge/69843519bca6d7ede506eb08)

### [How InfoNCE Creates Exploration: The Hidden Engine of Contrastive RL](https://proximal.hashnode.dev/how-infonce-creates-exploration-the-hidden-engine-of-contrastive-rl?utm_source=hashnode&utm_medium=feed)

A personal exploration of the mechanisms behind emergent exploration in goal-conditioned reinforcement learning Contrastive RL made a huge splash at NIPS 2025, with "1000 Layer Networks for Self-Supervised RL: Scaling Depth Can Enable New Goal-Reach...

11 min read · [proximal.hashnode.dev](https://proximal.hashnode.dev/how-infonce-creates-exploration-the-hidden-engine-of-contrastive-rl?utm_source=hashnode&utm_medium=feed)

[Jan 18](https://hashnode.com/posts/how-infonce-creates-exploration-the-hidden-engine-of-contrastive-rl/696c22e116bc8fe44370b78f)

### [Contrastive RL: A Step-by-Step Guide to Learning Reachability](https://proximal.hashnode.dev/contrastive-rl-step-by-step-guide-to-learning-reachability?utm_source=hashnode&utm_medium=feed)

The paper "1000 Layer Networks for Self-Supervised RL" won Best Paper at NeurIPS 2025, and for good reason. It demonstrates that goal-conditioned RL can scale to 1000-layer networks—something previously thought impractical. But the real insight isn't...

8 min read · [proximal.hashnode.dev](https://proximal.hashnode.dev/contrastive-rl-step-by-step-guide-to-learning-reachability?utm_source=hashnode&utm_medium=feed)

[Jan 9](https://hashnode.com/posts/contrastive-rl-step-by-step-guide-to-learning-reachability/69617f939c2177a405fd615e)

### [How wp.ScopedTimer Found My 12x Speedup](https://proximal.hashnode.dev/how-wp-scopedtimer-found-my-12x-speedup?utm_source=hashnode&utm_medium=feed)

I was benchmarking a gridworld RL environment built on NVIDIA Warp. The native Warp version hit 8.4 million world-steps per second on small grids - impressive. But when I wrapped it with JAX for compatibility with standard RL training pipelines, perf...

5 min read · [proximal.hashnode.dev](https://proximal.hashnode.dev/how-wp-scopedtimer-found-my-12x-speedup?utm_source=hashnode&utm_medium=feed)

[Jan 6](https://hashnode.com/posts/how-wp-scopedtimer-found-my-12x-speedup/695c68fdda9ef94189726736)
