ArkType is a project I've been following for a while that scratches all my itches. Hyper focus on type safety, performance and covering all of the edge cases you'd expect from a first class type system.

Congrats David Blass. Excited to give it a whirl.

Taelin 'Feels like Python, scales like CUDA' is an awesome slogan

Ground breaking stuff. A massively parallel, high level language with a pythonic syntax and an algebraic type system.

Ideal for the compute intensive future. Imagine building a simulation this way! Just write it naturally and Bend will run it natively on the GPU, in parallel.

RELEASE DAY

After almost 10 years of hard work, tireless research, and a dive deep into the kernels of computer science, I finally realized a dream: running a high-level language on GPUs. And I'm giving it to the world!

Bend compiles modern programming features, including:

-

Never call your product 'quickstart' or 'blueprints'.

I only associate it with 'low quality', 'cheap', and 'more trouble than it's worth'.

In the future you won’t even have to remember your kid’s name or what they’re in to!

Finally, I’ll have my time back to really live my life.

Thanks, Gemini

What I learned from trying CDK with Python:

> JSII was unnecessary.

Just use TypeScript for the CDK config even if you are building primarily in another language.

> like you even have that many

This fkin guy. He's right, but still. 🤣

Dang it, why does this diverge from standard CDK practices?

Builder pattern is a no-no in TypeScript & CDK.

Wrapping cdk.Stack is a bigger no-no. Construct libraries should not be Stacks.

Doesn't extend eks.Cluster == hard to migrate.

github.com/aws-quickstart…

I'd like to use the opportunity to show off an instance of #Rust pattern matching that I'm quite fond of. IMHO, it demonstrates the strengths of this feature a bit better:

AI idea - filter paid sponsorships out of videos. Ad block plus plus.

This is declarative programming and reminds me very much of the benefits of managed services.

It’s almost configuration. A spec, even.

Tell your program what to do and let it figure out how.

This made something click for me.

Have an LLM improve itself by generating new samples with multiple turns.

Then feed those refined samples back into training.

Repeat.

This is what it means to “train on Q* outputs” btw

This is why I’m bearish on opensource fine tunes.

They don’t know how to generate the data to create recursive self-improvement

“I promise” - I see what you did there 😉

Once you work in the Spark ecosystem, it becomes obvious how little is being contributed back to open source by Databricks.

So many pitfalls. Compatibility and configuration hell. Rot.

All held for ransom so you pay them to set it up.

Open Source+Ray will eat their lunch 😋