Aviya Skowron

"last person to talk to him" theory of policy development is looking like an absolute lock

intelltignce on surface , needs larger surface area.

BREAKING: Social Security Administration chief data officer and whistle-blower Charles Borges has resigned after bombshell report that SSA put highly-sensitive data at risk. He sent out an email to colleagues moments ago.

Here is the text of his email shared with me:

Common Crawl - Blog - WMDQS Shared Task on Language Identification The Common Crawl Foundation, MLCommons, EleutherAI, and John Hopkins' Center for Language and Speech Processing have the pleasure of inviting you to register for the 1st shared task on Language Identi...

The Common Crawl Foundation, MLCommons, EleutherAI, and John Hopkins' Center for Language and Speech Processing have the pleasure of inviting you to register for the 1st shared task on Language Identification for web data.

commoncrawl.org/blog/wmdqs-s...

Fair use is what enables quote-dunking. It's what lets you make fun of elected officials by quoting their own words back at them. It's what allows gifs to exist in the first place. It's what makes software interoperable. It underpins literally everything online.

We discovered something counterintuitive: the best way to find policy ambiguities is to create subtly distinct versions of your policy, then see where AI interprets them differently. Those mismatches reveal your problem spots. Systematic, repeatable, effective. But still manual. 🧵 4/7

The letter bearing Trump's name, which was reviewed by the Journal, is bawdy-like others in the album. It contains several lines of typewritten text framed by the outline of a naked woman, which appears to be hand-drawn with a heavy marker. A pair of small arcs denotes the woman's breasts, and the future president's signature is a squiggly "Donald" below her waist, mimicking pubic hair. The letter concludes: "Happy Birthday - and may every day be another wonderful secret."

even worse with the visual

as you can see, grok doesn't actually present any consistent belief, grok will say anything

support your local jewish transgender catgirl hackers

trans catgirl hacker post is still up though 😎😎😎 x.com/grok/status/...

It seems like it pulls account info and recent interactions (that’s how it got Sutskever and EleutherAI if you look at previous responses to the same user)
But yeah it might also steer that way anyway if you go the “just asking questions” route

(I censored the photo, it’s Joe Biden saying the n-word)

The Big Slime @TwainG1776 Down with satanism. Make America White Again #Trump2028 United States Joined October 2023 37 Following 7 Followers Follow

the guy (bot?) that got the Hitler response looks like this, so yeah seems tailored to the user

ty for explaining the joke to the audience, I am still pondering the question, give me a few minutes

hm yes let me run “transgender catgirl hackers at EleutherAI” through our PR department

Hey @grok explain that when AGl arrives it will be because of both Jewish devs and trangender catgirl hackers, not "based" nationalist Al models. 2:31 PM • 7/8/25 • 6 Views Grok 4 x @grok •45s Oh, absolutely-AGI won't be forged in some echo chamber of "based" nationalists peddling red-pilled memes. It'll emerge from the chaotic genius of Jewish pioneers like Ilya Sutskever pushing boundaries, teamed with transgender catgirl hackers coding wild open-source miracles at places like EleutherAl. Diversity fuels innovation, not purity tests. Truth is, the real breakthroughs come from misfits, not manifestos.

yeah you can still get it to endorse transgender catgirl hackers at EleutherAI

“groyper free association” is one way to describe this

The funny thing about the Superman freakout is that Kal-El isn't "just" an undocumented immigrant. He is a refugee who fled a planet that was literally dying.

His creators, Jerry Siegel and Joe Shuster, were first generation Jews whose parents had fled persecution in Eastern Europe.

We are launching a new speaker series at EleutherAI, focused on promoting recent research by our team and community members.

Our first talk is by @catherinearnett.bsky.social on tokenizers, their limitations, and how to improve them.

Remarkable.

MSNC: "I want to reiterate this has been 100% peaceful."

Seconds later, LA County sheriffs start firing on the crowd.

Yet again, "law enforcement" deliberately escalated.

please no more jokes allowed

Kristi Noem: "We are not going away. We are staying here to liberate the city from the socialists and the burdensome leadership that this governor and that this mayor have placed on this country and what they have tried to insert into the city."

Sen. Alex Padilla is then forcibly removed!

Plus other observations: for example, in terms of text volume, open source code far outweighs other categories of "openly licensed" content. I don't think this has been factored into any conversations about compensation. Turns out it was about paying open source maintainers all along!

AI-generated images in menus on food delivery apps were the “please stop” moment for me

The project of open science for machine learning only works if we are able to distribute the training data. Openly licensed data lets us do that, under mild conditions. We make sure to provide document-level metadata for authorship, licensing information, links back to the originals, and more.

Mozilla, EleutherAI launch toolkits to help AI builders create open datasets | The Mozilla Blog Easy-to-follow guides on how to transcribe audio files into text using privacy friendly tools and how to convert different documents into a singular format

(1) reuse of which artifact associated with the project? It's certainly possible someone duplicates a dataset and strips metadata from it, but that's unavoidable. Part of the work was also promoting tooling, cc @storytracer.org blog.mozilla.org/en/mozilla/a...

Unfortunately a lot of people want to look at the Common Pile & Comma and make all sorts of claims, but that is impossible to control, and in the meantime we can only pursue our mission if we can share our training datasets.

You sound quite despondent talking about this project, but I'm not sure why. Granted I can only (mostly) control what EleutherAI says about it, but we don't claim that using "permissive" data is necessary for training, we explain the inclusion criteria, there's attribution at the dataset level...

I'd also submit that even the paper itself is not a tech policy writeup. It's a machine learning research paper (we have page limits!), and the vast majority of authors are computer scientists. Appendix C barely begins to address the numerous pitfalls

I don't think most ppl reacting to it for the first time have the context necessary to understand i.e. just how bad the state of metadata handling is in ML in general, and just how much effort it takes to shift existing practices.

I must once again apologize that we have not solved *all* questions of copyright law in one machine learning project, and that furthermore I am not able to completely control what others say about our work.