Tero Parviainen

We’re announcing something new today: Motif - music that moves with your yoga class. Instead of static playlists, the music follows the structure and energy of a class and adapts on the fly.

In-context learning in LLMs

"The line I use is that software development is the language of our time. Giving everybody the ability to speak that language should be very empowering. It’s similar to @drmichaellevin's idea of translating between a human and their liver."

@matdryhurst Only took that planarium 400 million years to make it to Venice. Life goals.

Strange Rules in Ocula ocula.com/magazine/conve…

we’re thinking about doing an ML music group show w/ @sound_obsessed called “consented corpus” where artists pick a different artist to train on & they release an artefact together — reply if interested to hear more

Anthropic launches Claude Monet Painters are cooked

> be github > invent copilot > you are literally the first one > you are literally the only one > you literally have access to all the code in the world > get mogged by literally every single agentic bs that came out in the past few years this level of fumble should be studied

Claude's first day at Dunder Mifflin

models throwing shade at each other is an under-appreciated genre

“why do we still need humans in the loop for the design of software” Because the cost function of bad architecture is measured in months and years, and it’s hard to measure that in an RL environment/benchmark like SWE bench multilingual that is measures a 5-10 minute “can we write the code and get the tests passing” Spoke about this at @AIEMiami - been Obsessed with this after a convo I had with @calvinfo back in April youtube.com/live/6IxSbMhT7…

shape rotators made by shape rotators

The Helmholtz decomposition is one of the fundamental results of vector calculus. It says any well-behaved vector field can be split into two parts, one capturing sources and sinks through divergence, and one capturing rotation through curl.

The quantity of code that devs ship has roughly 10xed. But net developer productivity (value created by unit of time) is only up by a bit, if at all. Part of it is that the additional code is solving more incremental problems. A bigger part is that the new code is creating problems of its own.

cowork is amazing

Vera Molnar in 1981: "Assisted by this diabolical instrument, one can afford the luxury of exploring regions that would otherwise have remained uncharted." Vera Molnar for Télévision Française 1, 1981 (dubbed to english).

claude enjoying some autechre

I built a cymatics-inspired sand simulation where sound shapes grains into patterns. Watching chaos organize itself never gets old.

every youtube essay critiquing ai music assumes only one framework (commercial prompt-based services where the user selects from finished outcomes) exists. I don't blame them, that's the framework currently blasting them with advertising. But dig deeper. Think weirder.

this is Krea 2. our first foundation model, built completely from scratch for aesthetic diversity and stylistic control. learn more and get early access 👇

> When the automation company says don't automate things people are understandably frustrated I think this gets to the root of it. How are they an "automation company" if they don't even let you use it in a non-interactive way? I'm making heavy use of non-interactive Claude in cloud containers. And I'm sure many people took the same intellectual journey as I did. It went something like this: > Try and do agentic work using Anthropic's Messages API with tools. > Quickly discover it's completely useless and unreliable, even if you do loads of harness engineering. > Start using the agent SDK and realizing how much more robust it is, but understand that it needs to be isolated for security purposes and scale. > Make the realization that the Agent SDK is basically just spinning up a Claude Code process in the background. > Move to using Claude Code CLI directly on a cloud containers on a schedule/events with claude -p. > Realizing it's possible to capture many paid workflows into the max subscription with async model > Realizing that their moat now is broken if I change a few characters in the cron script i.e. `claude` -> `codex` So I kind of get it. They're definitely missing out on a fair bit of revenue when people design distributed asynchronous agentic architectures. But this is only hobbyist projects. This doesn't scale in any meaningful way and it's a shit move from Anthropic. It's also important to point out that it's extraordinarily difficult to automate anything sophisticated using even the best frontier models. As we covered in great detail in our recent METR interview, any notion of autonomy is basically a myth, and even automation is only possible with thousands of iterations and constant engineering. This is a bit like bash shell scripting on steroids. If people are automating anything sophisticated with Claude, it's because they are extremely talented engineers who understand their domain well and have iterated many, many times. x.com/slimer48484/st…