Pinned Tweet
I’ve almost built a new web browsing tool 🕸(to build tool for hypertext notes & thoughts from it later). It is called Linkkraft.
2 major points: 1) it knows your web history & navigation trails is precious for you. 2) it makes offline copies for everything.
A great use case for Wasmer! 🚀
Introducing @Infrawrench CI: 2–5× faster than GitHub Actions.
Every job runs in its own WebAssembly sandbox over an in-memory snapshot of the commit. One clone, one install, then 64 jobs at once on one runner.
Our 897-job pipeline re-checks in 9.1 s when nothing changed. ↓
Gleb Arestov retweeted
After many attempts over years, I came up with a new and improved interpolating spline. I call it “S9”, try it in the new demo. gasiulis.name/interpolating-…
AI has made no system too small to attack. In Japan, attackers hit a tourist train's seat booking and a funeral company's flower orders.
Japan's guidance starts where defense should: know every component you run and patch what's already public.
socket.dev/blog/japan-cybera…
Gleb Arestov retweeted
We just settled the DOM-based vs WebGPU canvas debate.
This is our WebGPU canvas rendering ~7,000 real HTML/CSS frames with over 2 million elements, at 120 fps.
And it's not just static designs. Every frame runs your full app so you can click around etc.
We built our own HTML renderer that supports every single CSS property (blurs, shadows, gradients, 3D transforms, you name it). We have full parity with chrome's own renderer!
This is possible because we own the entire pipeline. Our custom bundler takes almost any codebase and splits out its HTML, CSS and JS that is optimized for our canvas.
We basically built a mini browser, in Rust, running inside your browser.
Two years of infra work compounding and coming together finally 🥲
wow opencode support built in. admired unreal engine from afar for a decade and it's mindblowing to have any connection to it
The Unreal Engine 5.9 MegaJam Preview is here!
It’s the first Unreal Engine release built primarily with agentic workflows and opens up a few new Experimental features: check out Unreal MCP’s expanded toolsets, work with AI coding agents directly in Editor, explore AI-assisted content creation, and run your projects in a browser.
This isn’t your typical Unreal Engine Preview, but it’s ready just in time for the Epic MegaJam.
Invite your agents and start experimenting: epic.gm/ue-5-9-preview
Gleb Arestov retweeted
Built a physical simulation of a record player with WebGL2 and Web Audio. It's *shockingly* accurate: vinyl-gl.vercel.app.
Instead of playing a song, it works by converting it into the physical shape of a groove. Then a simulated stylus rides those walls 384,000 times a second, and you hear the voltage that motion would induce in a cartridge.
So the crackle, the pops, the wow/flutter, the dull inner grooves are natural byproducts of the simulation. You can also "DJ" the record.
(The converted songs are in the public domain or CC-licensed. Code is linked for you to remix and re-deploy.)
Gleb Arestov retweeted
OpenCode 2 far outscores OpenCode 1.19: from 8th of 9 harnesses to tied for first on all three models.
It replaces 1.19 in the @huggingface Harness Leaderboard, which compares coding agents on the same models. Like most harnesses there, it now uses the Chat Completions API directly, so it's more comparable.
At high effort it solves 65.7% of tasks instead of 43.3% on GLM-5.3-Flash and 69.4% instead of 57.0% on DeepSeek-V4.1-Flash. On Flint, 22.9% instead of 7.1%.
Cost per solved task fell from $3.35 to $0.45.
Part of the old gap was the setup. Harbor's standard route for OpenCode 1.19 is OpenAI's Responses API, which I translated to Chat Completions so it works with Hugging Face Inference Providers.
It ran without errors but silently dropped the model's earlier reasoning. Thanks to @thdxr and @rekram11 from @opencode for spotting it!
Leaderboard: huggingface.co/spaces/joelni…
Gleb Arestov retweeted
Build domain-specific canvases where humans and LLMs collaborate through mixed-initiative interactions
Enabled by the exo-harness: each component declares how an agent should interact with it, while a lightweight core inspects, adapts to, and wraps the component
Gleb Arestov retweeted
Bringing Your Own Components to the Mixed Initiative Canvas
Demo of how Tutuca’s Meta-Component Protocol enables dynamic component loading and runtime capability enhancement through optional, namespaced metadata schemas
youtube.com/watch?v=to5SuDeS…
Gleb Arestov retweeted
I burned a shitload of compute so you don't have to. Releasing the Harness Leaderboard.
Which harness should you run your open model in? I ran 28,849 trials to find out.
The leaderboard compares nine coding-agent harnesses on open models. In each comparison only the harness changes: the model, provider and reasoning effort stay fixed, and every task gets three attempts.
The trials ran GLM-5.3-Flash, DeepSeek-V4.1-Flash and Flint on Terminal-Bench 2.1 and DeepSWE, for $4,746 of inference.
Averaged over the three models, mini-swe-agent (@KLieret), a bash-only agent, solves 51.4% of tasks and @openclaw 27.3%. Four others are statistically tied with mini-swe-agent.
One of them is Pi (@pidotdev), at 49.4% for $0.29 per solved task, about 45% less than mini-swe-agent's $0.53.
The best harness depends on the model: mini-swe-agent on GLM-5.3-Flash, Hermes Agent (@NousResearch) on DeepSeek-V4.1-Flash and Kimi Code (@Kimi_Moonshot) on Flint.
@opencode costs $3.35 per solved task, 11.7 times as much as Pi.
See Harness Hub's Harness Lens for what the nine harnesses do differently. It documents the prompts, tools and compaction of 34 coding-agent harnesses and links every claim to its source.
Leaderboard: huggingface.co/spaces/joelni…
Harness Lens: huggingface.co/spaces/harnes…
Gleb Arestov retweeted
Electrobun 2.0.6 just shipped
- New GPU UI with fine-grained reactivity from Bun, Zig, Rust, and Odin
- Your app updates now stream patch reconstruction using less disk space
- full changelog docs.electrobun.dev/electrob…
This demo is direct to gpu ui, no webview in sight.
Gleb Arestov retweeted
Medicine is a compute problem.
This model was able to predict your next health event.
Trained on 300M patients and 16B health encounters.
We overestimate the mystery of our bodies and are more predictable than we want to believe.
Cool seeing that illusion chipped away.
That’s because GPUI Kit has specifically optimized and improved performance for this scenario.
We also have use cases involving AI chat.
Replying to @anaisbetts @theo
Scrolling perf in GPUI isn't difficult. Agents will not get it correct most of the time though. It's not surprising vibed GPUI apps have perf issues.
Here's a GPUI app w/ gpui-kit scrolling with images, graphs, markdown, mermaid charts stuck at 120fps.
Gleb Arestov retweeted
Drag-and-drop makes the digital graspable, but building it well is fiddly enough that many don’t bother. Declarative Dragging (with Josh Horowitz & Jeff Heer, #uist2026 best paper award), lets you make an element draggable just by listing the states dragging it can reach.























