
Same remote, same oval, three cars: Fable's, Sol's and the human engineer's original. All three get around. Fable shipped a brick on wheels. Sol shipped a body with vents.

Building AI agents and stuff

Same remote, same oval, three cars: Fable's, Sol's and the human engineer's original. All three get around. Fable shipped a brick on wheels. Sol shipped a body with vents.

Nothing was built. So how do I know the cars work? I run the files. The schematic goes into a circuit simulator. The firmware runs as written. The body drives in a physics engine. A script holds the remote. A swapped pin is a car that spins in place.

Both checked their own work at the end. Fable found its board 0.1 mm over the limit and routed the whole thing again. 14 minutes. Sol found the car 0.13 g too heavy and cut bigger vents in the shell. 78 seconds.

Fable is the careful one. It read 11 datasheets and threw out the stocked motor driver before drawing a single wire.

Sol is the one that bends the rules. The stock footprints threw 357 design-rule errors, so it rewrote its copy of them: stripped the silkscreen, redrew the courtyards, deleted a keep-out zone.

Sol wrote its router in C++. 78 lines, 2 minutes. Clean board 18 minutes later. It never looked at the board. One picture in the whole run, and it was the schematic.

The box had no autorouter. So Fable wrote one. 132 lines of Python, 3 minutes. Then it spent 92 minutes making it work. 16 runs. At one point USB+ and USB- kept ripping each other up in a loop.

I gave Fable 5.2 and GPT-6.2 Sol one task: design a palm-sized RC car from scratch. Circuit board, firmware, 3D-printed body. Sol was done in 33 minutes for $2.48. Fable took 2 h 23 min and $30.75. Both cars drive.

this is a data recipe for post-training a model like qwen-3.8-27B that has already RL'ed a lot 1. run the rollouts -> update the tasks to bring the solve rate into the target range 2. the tasks that are already too easy -> combine them to more complex ones, chaining is simple and a bit mechanical, but it works! congrats @guohao_li !
How much can you improve a small model in 18 days on 144 GPUs? Meet DaisyRose-27B, RadixArk's first open-weight model! It is trained with RL on Qwen3.8-27B, with 1,200 curated long-horizon agent tasks for 18 days on 144 H200s. Powered by Miles and our own training backend, we stretched agent sessions up to 1M tokens and raised the Terminal-Bench 4.1 score from 5.6% to 12.0%. See how we pull it off ⬇️

You can now RL-train a coding agent on 21,272 SWE tasks in 20 languages! SWE-rebench V2.1 update: > +7,935 Python tasks from the first SWE-rebench, rebuilt with clean images > Same curation pipeline as the rest of the set: repaired tests and rewritten task statements > Fixed a leak where some images had the fix inside. Thanks @fangjianju for finding it! Now on @harborframework, @huggingface and @PrimeIntellect Environments Hub. Links below 👇 More types of RL envs coming soon. All open source and free. Need specific tasks or found something broken? dm

More time bought nothing. More tokens bought nothing. Checking your work against the reference bought everything. 6 runs per model, auto-graded, real output files.Mean of 5–

Opus and Sol did it the boring way. Measure, rig, render next to the reference, fix, save once at the end. 0.99 and 0.95. Same 25 minutes.

Muse ships slop early and polishes it forever. First file at minute 11, then two thirds of the run re-saving it. 8 saves, 99 calls. It punches and kicks with the wrong side the whole time. Score: 0.60.

Luna is the lazy one. 9 minutes, 5% of the time limit, never looked at the videos again after the first glance. Asked to run, she took 0 steps. Score: 0.37.

Grok burned 2 hours trying to reverse-engineer 3D poses from the videos. Then it gave up and posed her by hand. It lost its context 4 times and wrote 4x more tokens than Sol. Score: 0.57.

I benchmarked 5 frontier models on one task: rig this character in Blender and animate 5 moves from video. Opus nailed it in 25 minutes. Luna quit after 9. Grok spent 3.5 hours and still can't kick.