Danilo Britto 🇵🇪🇨🇴

Meet Gemma 4! Purpose-built for advanced reasoning and agentic workflows on the hardware you own, and released under an Apache 2.0 license. We listened to invaluable community feedback in developing these models. Here is what makes Gemma 4 our most capable open models yet: 👇

This model has been #1 trending for 3 weeks now. It's Qwen3.5-27B fine-tuned on distilled data from Claude-4.6-Opus (reasoning). Trained via Unsloth. Runs locally on 16GB in 4-bit or 32GB in 8-bit. Model: huggingface.co/Jackrong/Qwen3…

The next version of React Aria has 90% fewer dependencies (from 127 down to just 13). It also supports subpath imports, e.g. "react-aria-components/Button". This means easier upgrades, faster builds, more effective tree shaking, and improved micro-frontend support. 🚀

Introducing TurboQuant: Our new compression algorithm that reduces LLM key-value cache memory by at least 6x and delivers up to 8x speedup, all with zero accuracy loss, redefining AI efficiency. Read the blog to learn how it achieves these results: goo.gle/4bsq2qI

Readers spotted that compression was not applied consistently in our React SSR benchmarks. TanStack Start had it disabled while the others had it on. We disabled compression everywhere and re-ran the tests. Here are the corrected results.

Async in frameworks happens in 4 phases: 1. create 2. consume 3. block 4. read The problem? Most fuse 2 & 3. Consumption triggers blocking. This puts DX (coloration) at odds with UX (blocking). Solid 2.0 snaps them to the poles. Colorless logic. Non-blocking UI.

We built Start with a client first mindset, but server first performance might be our thing too! 😉 New blog post on how we dramatically improved TanStack Start’s SSR performance: 427 req/s → 2357 req/s 424ms avg → 43ms 6558ms p99 → 928ms 99.96% success → 100% 🔗⬇️🧵

Hey peeps, we're revamping zustand docs, feedback is welcome here's the preview dbritto-dev.github.io/zustand This change would help us to add support to i18n, previews, playgrounds and more

We've been using Agentic Workflows since November to help with a bunch of our issue triage / project health reports etc and saved massive amounts of time. But this is one I really love, enable and then increase your test coverage gradually... github.com/githubnext/age…

Introducing GLM-OCR: SOTA performance, optimized for complex document understanding. With only 0.9B parameters, GLM-OCR delivers state-of-the-art results across major document understanding benchmarks, including formula recognition, table recognition, and information extraction. Weights: huggingface.co/zai-org/GLM-OCR Try it: ocr.z.ai API: docs.z.ai/guides/vlm/glm…

In the latest version of @Expo MCP, Enable agents like Claude Code to fix TestFlight crashes and review user feedback automatically 🚀 ◆ Full crash logs and screenshots ◆ Follow up with users over email ◆ Rank issue severity ◆ Works with any Apple app—Expo, SwiftUI, etc.

We're releasing 𝚟𝚎𝚛𝚌𝚎𝚕-𝚛𝚎𝚊𝚌𝚝-𝚗𝚊𝚝𝚒𝚟𝚎-𝚜𝚔𝚒𝚕𝚕𝚜, extracted from the lessons of building v0 for iOS and the upcoming Vercel app (teaser below). ▲ ~/ npx skills add vercel-labs/agent-skills

Introducing Bucketdrop - a tiny S3 client that sits on your menubar Bring your own keys Open source Free

Built a little macOS menu bar app today called WhyFi. To answer the age old question: why is my Wi-Fi connection absolutely rubbish and/or not working right now? Personally I've wanted this for years! If people like it I'll package it up and make it available as a download.

Expo Router v7 is now in beta! ◆ <Toolbar /> ◆ <Stack.Toolbar /> ◆ <Link.AppleZoom /> ◆ <Tabs.BottomAccessory /> Try the new UI components today: ~ / bunx create-expo -t default@55

I've been experimenting with AI since I got back from break. I never really touched the stuff so far beyond autocompletes so it has been quite a learning experience the last 3 weeks. I have definitely invested a lot of time, but I hope that time pays off. What I focused on wasn't coding but designing a sound reactive system that fit the constraints that I set out. It had to be async-first, optimistic capable, and status queriable. Basically my idealistic base for Solid 2.0. A few insights from the experience: 1. My work requires keeping an incredible amount of context in your head at the same time. AI definitely struggled with that. It was quick to provide its thoughts but it would forget previous decisions. We'd freeze them to prevent drift but that was only so effective because of layering and versioning. I even tried outputting immutable artifacts but over time it would lose references to them. To be fair I think most people would have trouble keeping all the context in their head, I do myself which is why I tried to pick up AI for this. But I spent an incredible amount of time reminding it what it already knew and correcting it a couple times to get it back to where it should be. When I say a lot of time, I'd say cumulative like almost half my time. Even now where I think I have arrived a holistic solution it probably doesn't remember all of it. 2. It's difficult to pick your battles in this scenario. I could just be like.. no you are wrong, this is the truth. But that would compromise its future work. That rule would inject my own assumptions silently and not make the exploration open. At one point I ended up with a less optimal solution because I'd cut it off from the path we needed to take. I reset expectations and we got there, but it wasn't as easy as just being like "Shut up you, I'm right". 3. Conversely though I found myself correcting AI a lot. Like about almost everything it said. Not just my prescribed semantics, but like JavaScript fundamentals. It would say stuff that was blatantly incorrect and I'd need to like prove it otherwise. Especially around async mechanisms in javascript like Generators and Promises. It took a lot of effort to explain why you couldn't retain context in Promises, or guarentee you are the last microtask in the microtask queue. It would write non-sensical code in synchronous execution scope, trying to explain it being down different generator paths. The more I got away from specific language mechanics the happier I was, but I really wonder how negatively this would impact code generation when it decides this arbitrary behavior is truth and can even "prove it". 4. Language is really important. Like naming of things and existing contexts for that. I had a concept I call a Firewall which is the internal mechanism for controlled writes in Computations.. it's how I do mutable derivation. But AI kept on confusing it with Boundaries (like Suspense). No amount of freezing/locking would keep the definition right for more than like 10 mins. I never said Firewalls were boundaries but its previous context of the meaning, stuff it pulled from thinking about other Frameworks etc kept influencing it. Finally I just started talking about them as Projections and this just resolved itself. Similarly renaming Suspense to something else was helpful because it would stop it from trying to pull Solid 1.0/React semantics in. 5. What I'm doing wouldn't have been possible until pretty recently. I tried some older.. even deeper thinking models like O1 and it was just useless. It couldn't keep anything in its head. I was using Copilot chat in the browser initially and I was like yeah I want a deep thinking model and I was really confused how bad it was. I realized the default model was GPT 5.1 which was infinitely better. There are probably better models out there (that I'd have to actually pay for). I tried Gemini too which felt more accurate but also less imaginative. Sort of in the middle. But I know nothing about these models so take that with a grain of salt. All in all I found the experience helpful. I could talk with it more readily than an actual person which in some ways accelerated things. I'm used to re-explaining myself so my only real points of frustration is when we "locked" down the definitions and it would still drift. I'm concerned there is no real way to output the whole picture. It keeps generating documents for me but those drift. But the best picture is in my own head still. It suggests a quick path to final artifacts is the most beneficial way to leverage this capability. But that also seems really hard when there are so many aspects impacting design. I will probably just accept where its value exists today.

we just released 𝚙𝚘𝚜𝚝𝚐𝚛𝚎𝚜-𝚋𝚎𝚜𝚝-𝚙𝚛𝚊𝚌𝚝𝚒𝚌𝚎𝚜, inspired by Vercel's 𝚛𝚎𝚊𝚌𝚝-𝚋𝚎𝚜𝚝-𝚙𝚛𝚊𝚌𝚝𝚒𝚌𝚎𝚜 these are the Agent Skills that we use on the supabase platform - including performance, security, & schema design details ↓ supabase.com/blog/postgres-…

Quick React reminder that `useId()` is underrated You can `useId(…)` + the `form={…}` attribute to submit forms from buttons outside the form element Avoids the whole `useEffect()` or `useRef()` juggling you'd otherwise have to do

We just released 𝚛𝚎𝚊𝚌𝚝-𝚋𝚎𝚜𝚝-𝚙𝚛𝚊𝚌𝚝𝚒𝚌𝚎𝚜, a repo for coding agents. React performance rules and evals to catch regressions, like accidental waterfalls and growing client bundles. How we collected them and how to install the skill ↓ vercel.com/blog/introduci…

@youyuxi Yes

T-shirts anyone?

My Discord sidebar looks so clean now