Keep the posts
you pass.

What if the future of software development isn’t about maintaining code, but regenerating it?

Dusan will take the stage at AI2FUTURE to explore this idea in his talk:

“AI code will be regenerated, not maintained”

📅 October 29, 9:30 - 10:00 
📍Zagreb, AI2FUTURE

He’ll share what we’re learning at *codeplain, where regenerative software approach already works and what this could mean for the future of software

Two talks from Granola's design engineering night that I'm thinking about 👇

1. Maggie Appleton (GitHubMaggie Appleton (GitHub Next) on speculative design.
I've been following Maggie's work since her writing on digital gardens, where ideas grow wiki-style instead of being published as finished posts in chronological order.

She talked about her work at the R&D labs of GitHub, noticing weak signals, asking what

1.5 days. That's how long it took from my first Figma screen to Forgettable running on my iPhone.
Forgettable is for the thought that ruins my mornings: did I lock the door? Is the oven off? Did I switch off the lights?
You can't check from the office, so the doubt follows you all day.
I started from one belief. A reminder only works if it comes by itself. If you have to remember to open a reminder app, you've

If you are building skills for the developer community, please budget very experienced team member time to build evals and garden the set of skills. As models continue to improve, it is beneficial to user outcomes and token efficiency to prune either portions or entire skills that the base model and harness are freshly capable of doing better. This is similar to long-established good habits of tending our code bases

What used to take weeks in C++ now takes one hour or so in Rust with a half-burned ChatGPT Pro plan and much less code. Plus a real-time viz for fun. Most of what I did was to review and agree on some architecture SVGs with the agent before letting it loose.

The idea is that a performant single-threaded high-bandwidth data recorder owns buffer lifecycles for reuse and performs memory-to-storage DMAs via io_uring as

Every time I have used automatic differentiation, I felt like some Dark Magic was taking place. With a lot of hand-waving, I understand why it should work, thanks to the chain rule, but how it performs the magic exactly was never clear to me. So I embarked on a side quest: try and implement it naively from scratch, in F#, to see what lessons I might learn along the way, and hopefully finally figure out how the magic

🤗 transformers serve now speaks 🔊

We just added an OpenAI-compatible `/v1/audio/speech` endpoint to transformers serve, and we couldn't have picked a better model to launch it with than Phonon, Gradium brand-new on-device TTS: tiny, fast and multilingual.

Test it with:
transformers serve gradium/phonon-40m-en

Phonon isn't alone. transformers serve supports:
• voice cloning from a reference clip
• streaming, to

Very excited to share that ContextCov has been accepted to #ICSE2027! 🎉

Coding agents can read AGENTS.md, but do they follow it? ContextCov turns natural-language repository instructions into executable checks for coding conventions, tooling restrictions, and architectural boundaries.

This one means a lot to me as a senior PhD student. I carried the project independently, from the initial idea through

AMD has several presentations accepted/presented at this year's PyTorch Conf about our effort to create a lower-level DSL than Triton which allows model developers to take better advantage of the underlying hardware while remaining in Python.

In keeping with our open source ecosystem strategy, efforts are under way to integrate FlyDSL as a back-end at several different model developer APIs, including PyTorch

Feels so cool to be part of Garaža's Frontier Hackathon! 😎

It takes place November 7-8 in Belgrade and is open for teams 2-4, under the age of 26.

We're joining as sponsors and technology providers and if you're thinking about signing up, make sure to check out what we're building at *codeplain and how we approach software regeneration:
lnkd.in/dYtVzQ_v