here's the token cost comparison. neither created a "good park". when including cost, it's unclear which wins

i'm working on an RCT benchmark for agent workflows that factors in cost

if this type of work interests you, check out INFER/CON, nov 16, SF: infercon.ai

I asked models to "maximize company value after a year"

luna 6 spent 10 minutes, raised prices but didn't build anything

opus 5.5 spent 2.5 hours, discovered an exploit where it could build tons of tiny rides, bypassing the duplicate ride penalty by never opening the rides

Benchmarking agents with RollerCoaster Tycoon

- a custom CLI performs much better (and cheaper) than computer use
- harnesses compete on cost:performance tradeoff, which can generalize to other tasks

Jev can generate game levels in real time.

Faster and cheaper structured output could be a big deal for game dev!

agentic engineering for beginners

dylanebert.com/agentic-engine…

I worked at Hugging Face from 2022-2025 🤗

I started as a research intern, helping to build a project called Simulate with @Thom_Wolf @natolambert

I loved it. No meetings, no bureaucracy, extremely passionate, talented people who are genuinely excited about what they're ull time.

My mandate was "have impact". I didn't even know what a KPI or OKR was the whole time. Just trust and support working with talented, passionate people like @osanseviero @mervenoyann @xenovacom

I got to build gsplat.js, 3D Arena, make courses, videos, demos, release papers. Here's one of my Hugging Face tiktoks from 2022.

I now realize how lucky I was for it to be my first job. Congrats to the team there!

10k active rigid bodies at 240fps in the browser (webgpu)

GRPO is actually pretty simple (most explanations overcomplicate it):

1. take a model and prompt e.g. "name this color: #D02027"
2. generate multiple outputs
3. score outputs
4. calculate group average
5. encourage outputs that beat the average

full version with interactive visualizations: adaptive-ml.com/post/grpo-simp…

My dear front-end developers (and anyone who’s interested in the future of interfaces):

I have crawled through depths of hell to bring you, for the foreseeable years, one of the more important foundational pieces of UI engineering (if not in implementation then certainly at , accurate and comprehensive userland text measurement algorithm in pure TypeScript, usable for laying out entire web pages without CSS, bypassing DOM measurements and reflow

I created an RL Glossary. Free, open source, puts everything in one place, made accessible, for training your own specialized LLM

dev.adaptive-ml.com

Introducing Texel Splatting: Perspective-Stable 3D Pixel Art

open source paper+code

Most 3D pixel art techniques (e.g. t3ssel8r, ProPixelizer) snap pixels to a screen grid, which only works with an orthographic camera

Texel splatting solves this for perspective cameras: first,

the scene is rendered from fixed-origin cubemaps. then, each cubemap texel is splatted to a world-space quad

each quad is expanded to fill disocclusion gaps

during movement, the fixed origin is snapped to a world-space grid

to make transitions smoother, blend from the previous cubemap with ordered dithering

the webgpu demo achieves 240fps on an RTX 4090, 40 fps on an iPhone 15

try it: dylanebert.com/texel-splattin…

Introducing Texel Splatting

a technique for "true" 3D pixel art

speculative decoding explained in 53 seconds

my favorite visual explainer I've made so far: SPECULATIVE DECODING

2-3x speedup on inference. but how?
1. draft quickly with a small model
2. accept/reject in parallel with a large model

full step-by-step explanation with ✨3D Animations ✨adaptive-ml.com/post/speculati…

this is so cool

looking forward to more!

I'm experimenting with creating the best possible technical video explainer of attention, but in two modes

⬛ Hyperactive dark mode
⬜ Slower-paced light mode

which do you prefer?

Attention Explained in 59 seconds

written version: adaptive-ml.com/post/attention…