AlexKesling

E.g. why is Claude's active context not phrased in terms of memories, history, etc. _and_ vice-versa, why are knowledge bases not phrased in terms of contextual value?

I'm thinking more and more that the graduated graph of context should have better layers of abstraction that reduce the representational discrepancy across domains of consideration.

Is it odd that we keep talking about memory and context as separate things?

The harness is a subset of the domain.

Are people using it more generally for the "classical" programmic portion yet?

A skill on Claude.md is in many ways programming various Claude harness constructions. What do we call the holistic memetic system? Harness/Workflow/Whatever + "memes"?

Having a lot of fun building things that put us back in control of how our agents operate and what they do.

LLM case of the Mandela effect? vgel.me/posts/seahor...

I think "Slopware" is the stuff that comes out in codegen unbridled by taste.

"Thinkware" is what runs on/around the model. If we loosely consider an LLM a computer, a prompt is somewhat of a program. Thinkware is all that stuff that runs on / wrangles this "new kind of computer".

Mmm "Slopware"

"Thinkware Engineering"

There needs to be a name for the prompt and context that "runs" an Agent/LLM -- Promptware, Thoughtware, Vibeware?

We now have many different "applications" running on the same models; different Vibeware running on Claude or Llama = different behavior = a different application.

Billion dollar idea: stand up a marketplace on top of Replicated or something which lets people set up a "virtual cloud" that runs other people's jobs in their own environment as a SaaS offering to leverage non-transferable cloud credits.

"Aggregated cloud as a service as a service"

ACaaSaaS

Code review is more important now than ever. I keep seeing posts espousing a view to review AI generated code less, not more. Best of luck to those who wander down that dark path.

It's good to see thoughtful practice shared too --
diwank.space/field-notes-...

Software gaining more autonomy (AI agents, etc.) may very well make this problem worse. I'd be surprised if we didn't end up with per-task data puddles proliferating before we figure out systemic solutions.

If data puddles are tiny data lakes, do we need puddle houses?

How do people wrangle all their puddles of data now? Permissions management alone on hundreds of independently owned puddles sounds like home directories all over again.

How are you thinking about what this means for AI code generation? What new patterns are you applying to help mitigate the risk of your codebase getting stuck in the mud?

Feature creep is the norm. If a project is a watercolor portrait, things become muddy shades of brown. Each beautiful addition, vibrant on its own, detracting from the work as a whole.

We must keep composition in mind. Readability is about more than a local change, how does it fit into the whole?

It's a testament to the lifetime elision ergonomics that, even after six years of enjoying regularly writing Rust, I occasionally have to re-up on my borrow checker knowledge.

I'd recommend this oldie-but-goodie for anyone but the most grizzled rustacean: github.com/pretzelhamme...

Software engineering is building with sand. Nothing is permanent. Code is change.

Like buddhist mandalas, there's an ephemerality to the art that must be acknowledged to be accepted.

As AI codegen gets better, code review is becoming an even _more_ important "basic" skill for engineers still mucking in code-land.

If you can't immediately find what's wrong in generated code, you're going to struggle to keep up.

Get Weird w/ DataFusion · Luma Ever wanted to turn something into a database? What about write a crazy query language? Send SQL to your toaster? This night's for you! DataFusion is like…

Ever wanted to make a database out of something or query a DB with your own custom language? In SF March 19th?

I'm hosting a hackathon just for you.

Come get weird with DataFusion!

lu.ma/3nrr8n61

Alex Kesling on Apache Arrow DataFusion [PWL NYC] YouTube video by PapersWeLove

The Kesling Criteria (informally) --

A system is a database if:
1) You can ask questions of it
2) It returns data

As presented in my @paperswelove.org talk last week youtu.be/6A4vFRpSq3k

Alex Kesling on Apache Arrow DataFusion [PWL NYC] YouTube video by PapersWeLove

PWL NYC is pleased to present Alex Kesling (@alexkesling.bsky.social) on Apache Arrow DataFusion

February 5, 2025

📺 youtu.be/6A4vFRpSq3k
📜 andrew.nerdnetworks.org/other/SIGMOD...

#compsci #research #datafusion #database #sql

For Papers We Love last week,
@alexkesling.bsky.social gave a talk on DataFusion and live demoed an increasingly awesome set of Rust databases built with it. If you’re free tomorrow, you should head to Rust NYC where he’ll go deeper on the internals and take it further.

Come learn how to "build" a database with me!

I'm presenting on February 5th at Datadog HQ on the paper "Apache Arrow DataFusion: a Fast, Embeddable, Modular Analytic Query Engine" (andrew.nerdnetworks.org/other/SIGMOD...).

If you're in NYC, register at www.meetup.com/papers-we-lo...

Bazel works best when deps are defined to be fully immutable and consistent for all new builds (or fail clearly). This means CI reflects local, etc. in the common case. rules_rust branch selection is leaking Cargo's looser model a bit.

Sounds like a plan ;). If you ever wander back into Bazel-land, I'm happy to help.

Note that it's better form to use a rev instead of a branch as Bazel won't check the repo on every build and you probably want to be able to control when it pulls a new version. I used a branch here for simplicity of example.

If you just want to replace the crates.io package with a git branch, you should be able to alter the spec. If you want them to coexist, or this gist doesn't work, let me know.

gist.github.com/akesling/7bb...