Finally MCP can become actually viable. I've been enabling max one or two per project ever since I found about how many tokens it used to cost. This should massively improve usability. Let's see. github.com/anthropics/c...

Hang this on the office wall.

Fun fact: The engine here is Hammerspoon, which is powered by Lua. Lua was developed at PUC-Rio in 🇧🇷 and is ideal for this because it’s super lightweight and designed to be embedded into other apps. www.hammerspoon.org

One angle I like about AI coding is how it enables you to build custom tools that augment your workflows. Ideally, software should fit you like a craftsman's tool fits their hand. Now that code is relatively cheaper, we can finally do more of that. Thats why I built this github.com/iurysza/ultra

Two things I do when using LLMs for researching a topic: - As you research select the most useful content it generates and organize it on a canvas. - Ask for ASCII Art to illustrate a concept (people have no idea how far you can get with pure text )

Been a Fragmented listener since before I got my first dev job. Surreal to now be joining @kau.sh as co-host for this new Al focused season! Looking forward to getting into the ins and outs of Al coding! What actually works, what doesn't, and how to think about all of it. Let's do this! 🎙️🤘

I knew this was a thing but I didn't care much, but I'm now planning to setup something to prevent bots from scraping my blog. Since 24 Google Analytics shows I have a huge number of users coming from Singapore and China. Like 5000% more than usual. 🙃 developers.cloudflare.com/bots/get-sta...

Running under -9° today ❄️❄️❄️

Flutter dropped a GenUI SDK last month. You register a widget library, give an LLM the component definitions, and it decides what to render. Server-driven UI, but the server is now an LLM. Not revolutionary as a concept, but it's a signal. Something's cooking. iurysouza.dev/non-determin...

That's why I think the "file over app philosophy" is so good. Even more so with LLMs. share.google/O0Xn32BIBC3h...

Still on the topic of who's reviewing all this code: From @kentbeck.com article > Code review made sense when humans wrote code at human speed. Now I'm on a beach with a genie, generating faster than anyone could review. The old model broke. What replaces it? tidyfirst.substack.com/p/party-of-o...

The grief is real. I think the reason I stay motivated is that I look AI coding as a another problem space I can learn. But I deeply get the feeling of that post. I used to say that my company could be selling bananas or apartments I would barely notice the diff. That's not the case anymore.

Another criminally underrated feature is the `AskUserQuestion` tool. This makes the UX so much better for many things. Here's an example of how I use it for something simple like generating an image. First image is part of my gemini-image-gen skill. The second is what happens when I use.

vibe coder aesthetics update Man, I think I gotta upgrade my outfit game or I'll be out of the market next year.

I have this generated for every UNIT of the project I want to use AI for (unit can mean a lot of things 👀). Then, I have a `rule` that causes that context to be loaded when certain files are touched. This makes sure that whenever an agent has to touch something important, the context is loaded.

New tech, new problems. Tech companies have our data, but that mostly affects the content we're served. For LLMs the side effects of that accumulated context are massive. Not to mention how it eventually locks you to a provider. Memory portability should be a thing iurysouza.dev/memory-porta...

I shared some thoughts on "agentic coding" over the holidays and got a lot of nice feedback! I've now added this image to better illustrate the context engineering section. I have to say that these things do compound nicely.

One of the biggest “wow” moments with LLMs was when GPT-4 came out. Now this open-source model with only 2.6B params, that can run locally on your phone... is BETTER than the original GPT-4. Not to toot my own horn, but this is exactly the case I was making when I gave this talk back in March.

You can't just let an agent loose on a repo and hope for the best. Coding is personal, so you'll likely have to build your preferences into the agent before you can trust it, and that means a lot of trial and error. Getting this right requires actual effort. iurysouza.dev/agentic-codi...

I've got this memory-service working. Looks promising, but I've no idea if it will scale well. I'm glad I found this because I was about to start building it xD

Anyone building or using memory abstractions for LLMs? I wonder what's the state of art and good usecases of the open source tools doing this right now.

Finally took the time to replace my old Disqus comment section with bluesky. Now, if you reply to a post where I share an article it shows up on the article's page. :cheffkiss:

After a loooong pause... I'm back! Are you folks still out there? 👀

I’ve avoided this topic because of the noise and heated debates, but I think we recently crossed a line. The polarization is still there but it doesn't change the outcome anymore: this isn't a fad. It's the new baseline. iurysouza.dev/agentic-codi...

Using the Default Hierarchy Template in KMP? You probably shouldn't 👀 It creates intermediate source sets you might not need — with real build cost. ⚠️ New article on the hidden costs and when to use a custom hierarchy instead 👇 dev.to/rsicarelli/t...

Sideloading apps, installing custom roms and overclocking cheap phones is how I got into this in the first place. Android survived so far without removing sideloading, it's not like it would die otherwise. This is a mistake. f-droid.org/en/2025/10/2...

Matt Levine on the AMD x OpenAI deal

Just got out of @zsmb.co Droidcon talk on Compose's Hot Reload, and I'm really impressed. It now looks and feels like a solid product. Also, super interesting to see some of the work that happens under the hood to make it seamless. Shows a lot of the platform's commitment to DEVEX.

Has anyone ever heard of tombstone files? I can't believe I never heard about this before after so many years. Also, I wish I didn't.

Gmail MCP is OP Been using it for one day and already hooked

Still buzzing from my time at Droidcon Lisbon! Thanks to everyone who came to my talk and for the great conversations and feedback. Massive thanks to Bruno and @cafonsomota.bsky.social and all the organizers. Fantastic conference! Slides are here: iurysouza.dev/speaking/dro...

🥹🥹🥹

Is this a test?

Made it to Lisbon! But as we took off at 7am I had a brilliant idea: Why not add another 30 slides to make a point? Improv time! See you at 14:10 on track 1!

I was supposed to arrive in Lisbon yesterday, but missed a connection due to a delayed flight. Leaving Paris now. Long day ahead 😅

I'm happy to share I'll be speaking at Droidcon Lisbon next month! If you're interested about on-device Gen-AI you'll like this one. The plan is to dig into how LLMs actually work and bridge to the tools that enable on‑device inference. Also down for a pastel de nata + coffee if you’re around 🇵🇹

Spending some time in Brazil with a proper home office setup 😌

Seems to be a lot of talk about Haze and Liquid Glass… At its core, Haze is a background content transform library, and the ‘visual effect’ (blurring, etc) is just an implementation detail. Liquid Glass will be a runtime shader like others which Haze already uses.

Apple sitting out of the whole AI assistant wars is nuts.

machinelearning.apple.com/research/int... Apple's new on-device solution is kinda like Google's AI Edge SDK, but as usual, less fragmented. Now, WTH are they comparing their model to Gemma 1 when Gemma 3 is out since March? I'm guessing their model is still quite basic. Lets see.

yes, but

Wireless debugging is actually happening this time!

AS new Agent Mode can understand Gradle structure and use IDE features. This will really set AS apart from alternatives which have no idea about kotlin/gradle context. Really good to see investment on that front! (There's also a ton more stuff. Huge release!) www.youtube.com/watch?v=KXKP...

Union types + Rich Errors Still not clear to me how union types will work (typescript PSTD), but Im all in for rich errors

Bold move 👀 github.com/Kotlin/kotli...

Google I/O update: New Gemma 3n, a 4B* model getting close to Sonnet 3.7 is both mindblowing and inevitable. Edge AI++

This was released less than 6months ago

New blog post up on UI layer architecture for persistent UI elements like navigation bars, navigation rails, and floating action buttons! www.tunjid.com/articles/ui-...

PSA: The Android Show (pre-io android only event) happens today! www.youtube.com/live/l3yDd3C...