Umberto Raimondi

(gen-AI) x.com/HaibaraAi__APT…

This is basically performance art for the era of AI, not as sophisticated as the "company" that was cloning OSS but still a good one.

@ferologics Blessed gstack 🙏

the valuable cannot be collapsed into one production metric. unless it's gstack and we're talking LOCs

@alexisgallagher Yeah, being able to quickly translate well is probably the best feature here, adds a lot of variety to my TL

My timeline is full of auto-translated converations between Americans & Japanese, talking about food, tatoos, comics, history, hobbies, cherry blossoms, exercise, etc.. I can’t recall seeing such an efflorescence of cultural exchange on an individual level. It’s like a thousand mini semesters abroad. There’s a sense of real warmth and interest, and also maybe a bit that everyone is on their best behavior. Heartwarming! ☺️🇯🇵🇺🇸

Wait. Moltbook still exists?

@badlogicgames at least he didn't call base64 "encryption"

i had a cto once, gaming industry, ca. 2010ish. i was just a humble tech lead. he'd cite us into the meeting room to "tackle large asset sizes in mobile app bundles once and for all". he literally proposed base64. he called it asszip (i have witnesses). this is what it feels like seeing all the posts from former engineers turned VCs now getting clanker induced ai psychosis.

We made Muon run up to 2x faster for free! Introducing Gram Newton-Schulz: a mathematically equivalent but computationally faster Newton-Schulz algorithm for polar decomposition. Gram Newton-Schulz rewrites Newton-Schulz such that instead of iterating on the expensive rectangular X matrix, we iterate on the small, square, symmetric XX^T Gram matrix to reduce FLOPs. This allows us to make more use of fast symmetric GEMM kernels on Hopper and Blackwell, halving the FLOPs of each of those GEMMs. Gram Newton-Schulz is a drop-in replacement of Newton-Schulz for your Muon use case: we see validation perplexity preserved within 0.01, and share our (long!) journey stabilizing this algorithm and ensuring that training quality is preserved above all else. This was a super fun project with @noahamsel, @berlinchen, and @tri_dao that spanned theory, numerical analysis, and ML systems! Blog and codebase linked below 🧵

@mitsuhiko Random walk, reacts more to rumors than to news, often disconnected from the state of the economy 🤷‍♂️

How is the market this calm after such a massive disruption and continued escalation? This makes absolutely no sense to me.

People discover that knowledge is fluid and that you lose what you don't use.

Mistral just open-sourced a text-to-speech model that beats ElevenLabs. 3 GB of RAM. Runs locally. Free. The thing people were paying per-word for last year runs on your laptop now.

@andreasklinger I wonder if they expect sessions where I saved the agent from itself or something else :)

New question in the YCombinator application form 👀

Maybe it’s just me, but if I had to choose, I would prefer models with additional training for better code and architecture quality instead of models with a few more points in some wacky SWE benchmark. We can discuss about how to measure it.

People and companies need to realize this, with this new development paradigm the code does not come included with the developer anymore, there is a third party provider you’ll have to pay. It’s like having to pay for each key press. The provider decides the price. The provider will change the price.

Writing Hypervisor in Zig - hv.smallkirby.com/en/

The interesting thing here is that he used a dataset with just 3k reasoning traces, way smaller that I would expect.

Open Source Pangram is out now! We have released the datasets, code, and two models based on our EditLens work on quantifying the extent of AI editing in texts.

It’s curious how I ended up photographing the Kowloon Walled City, along with co-author Ian Lambot. We ended up making the most thorough (and most frequently referenced) record of the place. Why didn’t any HK photographers concentrate their efforts on documenting the place? One reason is that every parent told their kids to stay clear of the place. The other, and maybe more important reason, is that it wasn’t considered a worthy subject. Had the Walled City lasted just a few years longer (it was demolished in 1993), I have no doubt that young HK photographers and filmmakers would have been all over the place. But it was erased just moments before HK woke up to itself. A generational change. The culture generation now values HK in a way the previous generations didn’t. “City of Darkness Revisited” will be republished later this year. Stand by for updates on pre-ordering! 1. Walled City, southwest corner. 2. Facade along Tung Tau Chun Rd. 3. Rooftop kids. 4. Mahjong at home. Link to books in bio. greggirardpictures.com

THE APPLE APP STORE IS DROWNING IN AI SLOP people are treating the App Store like a Medium blog spitting out apps one after another. All with zero users and $0 revenue. Apple reviews that used to take hours are now stretching into WEEKS and even months > more than 550k apps were submitted just last year, highest in a decade.

@code_star Headache Fez, looks great.

So fez?