Honing Lab

Infrastructure for Self-Evolving System 1 – System 2 Intelligence.

[Honing Lab @HoningLab]

A higher score can still be noise. 🎲

Honing PR #182 proposes a release gate that compares candidate and baseline round by round. It accepts the candidate only if the lower bound of the gain clears a preset threshold. Not wired into campaigns yet.
https://t.co/QilQQY0yLL

[Honing Lab @HoningLab]

Your game engine shouldn't import your training stack. 🎮

Honing's Craftax, Griddly and REA-Isaac runtimes now sit behind a clean boundary: adapters depend on runtimes, not the other way around.

Same game rules, less coupling.
https://t.co/Lu8z7QA4wl

[Honing Lab @HoningLab]

No training in this run, and no level completion. Unity pauses during model calls; the clip omits that waiting time.

More details in the repo:
https://t.co/eacpuoqoe8

[Honing Lab @HoningLab]

Laya's plan for this level: jump left. Repeat 50 times. 🧱

This 20-second clip is a real Laya-on-CPU run through Honing's Hollow Knight imitation adapter.

The model-to-game connection works. Navigation needs work.

[Honing Lab @HoningLab]

Proposed test: drop a context field or reverse a fire/hold weight edit.

Compare diagnoses from traces alone vs. probes that replay the same state with a targeted change. Can it find the cause more reliably?

Research plan, no results yet.
https://t.co/PQ27Nh1pUI

[Honing Lab @HoningLab]

Before an AI fixes itself, can it tell what broke? 🔍

Honing's System 0 RFC proposes a debugging exam: inject known faults into the harness or weights, then see whether an LLM can identify the cause.

First, debug the debugger. 🛠️

[Honing Lab @HoningLab]

Nimble-v4-max gets a seat at the table. 🎮

It now runs through Honing’s decision API: send in a state, questions, and options; get probabilities back.

We also check source and weight hashes for the serving integration. 🔍

https://t.co/dQ3I8tngi9

[Honing Lab @HoningLab]

The recording skips that thinking time, so smooth playback isn’t proof of real-time inference.

More on the input adapter 👇
https://t.co/0KhdvyBf96

[Honing Lab @HoningLab]

Bullet time, but for inference. ⏸️

In HoningLab’s Hollow Knight imitation demo, Unity pauses while the model chooses its next input. 🎮

[Honing Lab @HoningLab]

This was one smoke test with global map observations, not a benchmark result. The run hit the step limit, and model training hasn’t run yet.

https://t.co/dQ3I8tngi9

[Honing Lab @HoningLab]

NanoJev met Griddly Spiders and chose WAIT for all 100 steps. 🕷️

A bold commitment to doing nothing.

[Honing Lab @HoningLab]

New arena unlocked 🎮

We’ve added REA, an Isaac-inspired room-combat engine, to HoningLab.

Next up: letting our decision model step into the arena and see how it handles the fight. 🤖⚔️

More fun stuff on GitHub 👇
https://t.co/dQ3I8tngi9

[Haoze He @HaozeHe_Hector]

We edited 0.0002% of NanoJev’s weights, and the edited model beat the original 65–34 (1 draw) in 100 head-to-head Tank War games. 🎯

The original took only about 1 in 3 clean shots. Using Weight-Space Programming by @rikiyatakehi et al. (https://t.co/oQW1HI5rU9), we read the fire-vs-hold direction out of its activations and added it to the decision head: 1,024 weights, no training. Now it takes every clean shot and fires about twice as often.

It isn’t a better shot, though. It just stopped hesitating.

🛠️ Weight editing is now a trainer in Honing. An edit goes through the same gate as SFT and RL, and only ships if it beats the current model.

Report: https://t.co/xpLcLMauje

Weight-edit trainer: https://t.co/hoXdxumA1s

GitHub: https://t.co/oAPAOkCpcy
Follow @HoningLab for more Honing updates :)

[Honing Lab @HoningLab]

Added Craftax to HoningLab: 43 actions, seeded tasks, and achievement-based grading. ⛏️ Plenty of ways for a model to discover that “survive” is harder than it sounds.

The environment is ready; model training is still homework.
https://t.co/Jigr5V8Bqu

[Honing Lab @HoningLab]

Next up: a full learning campaign with real model weights, more game environments, and reusable gameplay data. 🔧

We’re building toward continuous self-improvement; the full loop is still in progress.

Code & ideas welcome 👇
https://t.co/dQ3I8tngi9

[Honing Lab @HoningLab]

100+ GitHub stars! 🌟 Sadly, our agents can’t train on those. Back to the GPUs. 🫠

This week: SGLang serving + an OpenHands/OfficeQA integration landed in Honing. The toolkit is growing beyond games.

Thanks for the early support! 🫶

[Honing Lab @HoningLab]

The loop we're working toward: act → feedback → reflection → improve the harness + weights → validate → release → repeat.

Continuous learning is the goal; the full loop is still in progress.

Explore the code & help build it 👇
https://t.co/dQ3I8tnO7H

[Honing Lab @HoningLab]

A familiar decision shouldn't need a fresh essay. 🧠

Our goal: self-evolving System 1 + System 2 agents, with fast decision models for familiar choices and deeper reasoning when needed.

We're building Honing: infrastructure to keep those decision models learning.

[Honing Lab @HoningLab]

V3 moves, aims, and fights. V4 dodges incoming shells and wins in the clip. 🎯

We changed the game context, available moves, and checks around NanoJev. These are gameplay clips, not a measured win-rate comparison.

[Honing Lab @HoningLab]

We gave NanoJev a tank. First it spun in circles. Then it found the fire button and shot itself. 💥

Tank War's first boss: your own health bar. 🫠

Four versions, same model weights. We changed the controller.