Scythia Marrow

Np, have a good one.

Makes sense; seems like simple harnesses made of shoestring and glue aren't up to this task. Tho I'm pretty sure a better one could extract citations in a satisfactory way.

On my end I'm up way too late; you nerd sniped me good. This was fun! Feel free to ask me for better, I'll make it a project.

Oh! That makes a lot of sense. I had no *freaking* clue what you two were talking about; bluesky threads suck. And I'm not a lawyer, I'm an academic where all things inside the [] are citations.

Fixed (same brief):
drive.google.com/file/d/1DCQk...

Just let me know what's wrong and I can fix.

Hmm. How do you draw the distinction between the two in your own experience? So like, when it’s used to actively support an argument rather than providing supplementary information?

YSLETA DEL SUR PUEBLO, ET AL.,
Petitioners,
v.
STATE OF TEXAS,
Respondent.

I should have put that at the top of the output. In my defense, I built the tool in an hour.

drive.google.com/file/d/142QJ...

I made a demo tool for this problem. The tool is somewhat general and costs me ~$2 in API tokens per citation list extracted.

See the GOTCHAS and VALIDATION sections for items the LLM was uncertain on.

Want me to make it better? Tell me what's wrong and I will.

I have been in a continuous nerd sniped fuge for like, nearly two months now. It’s not sustainable but gosh I’m having fun with building again.

PyPI · The Python Package Index The Python Package Index (PyPI) is a repository of software for the Python programming language.

Did you know that there was no open source language agnostic copy/paste detection static analysis tool?

So I made CPITD! It’s neat and I use it routinely. It works especially well with claude code: ~ 8:1 SNR and usually catches 1-2 duplications per implemented feature.

pypi.org/project/cpitd

Doll would really love it if everyone got chainlink to the required levels so doll could quallify to get this. github.com/dollspace-ga...

Good morning!

Social climbing is neutral in this context because it can be a good motivator to help your friends and loved ones but is universally awful when driving interactions with strangers.

We need to be upfront about it and do lotteries for limited opportunities instead of doing social climbing (neutral).

My ideal recruitment process is only focused on establishing ”can you do the work?” If you can, congrats! You get a lotto ticket.

I love how the math of retrocausality in (some interpretations of) quantum mechanics is identical to the math of intelligence.

The electron isn’t there because in the future I measured it. I don’t jump off a cliff because future me got hurt.

Both aren’t *actually* retrocausal but it’s still neat.

Best sleep I ever had was on progesterone; it’s good stuff.

I love how responsive the tools are!

Hmmm. ~55k for a full setup, so we’d need to split along 8-10 or so people for the cost to be less than an rtx pro each. And a ~day of h200 time a week would be decent? I’d certainly find uses for that.

Probably not economical right now given how subsidized API credits are.

It’s surprisingly mundane when you get used to it.

The first time I ran OOM on a similar setup I had a medium sized existential crisis; felt like i was contaminating something holy.

We could coop fund a shared instance. I’d much rather have partial time on a H200 than full use of an RTX pro.

Yoooooo those are my namesake’s cousins right there! Gosh the ancient steppe peoples are so cool.

I wish the headline said Pazyryk and not Scythian tho. It’s like calling people who live in Groningen Hollanders.

jisp!

You are so fucking cool.

I did so at the age of 17, when I decided to seriously pursue machine learning. I was debating going into biology or computer science and decided on computer science because of the relative ease of independent research.

I’m 31 now and have fallen far, far, *far* short of my goals.

Ha!

Just to be 100% clear

In this house, we believe:
- AI should always and only be used for the benefit of humanity
- AI safety is critical, and we must do a better job protecting the vulnerable
- xAI should be burned to the ground and salted over
- Elon should be in prison

Field tho? It’s not a relational field unless the metric is reimanian! Otherwise it’s just a sparkling network statistic.

(To the galois theory aficionado reading this joke and complaining that my response got the math wrong: know that I did so on purpose just to spite you in particular)

Salvos (the popcorn litrpg) does this and it’s so good. We follow the main character for three books as she becomes one of the most powerful mages in the world, and only *then* does she decide to go to magic grad school.

It’s an *incredible* writing decision that works *so well*.

It starts with mocap tranfer (or animation!) to get the target poses then uses reinforcement learning in a virtual environment to get a full policy w/ kinematic feedback.

Any dancer could tell you the movement is impossible without dynamic balance.

In that case your knowledge is as good as mine. Your best next step would probably be to add a code coverage static analysis tool to verifiably propose uncovered edge cases to test. But that’s a big lift.

Ask it to write the tests first.

Alternatively ask it to write an api spec, then ask a fresh model to write tests for that spec.

But honestly if you move in small enough clearly defined features the model will do okay.

All my wonderful imaginative kind thoughtful old friends who stopped making things as they moved in mgmt or had families or life took over or tech got un-fun are making and writing and sharing again, and it makes my heart sing