Tessera
Folgen

Michael Ekstrand

@md.ekstrandom.net

Info sci prof @ Drexel, trying to keep the machines (esp. RecSys & IR) from learning bigotry and discrimination. ADHDS9. Usually self-propelled. Opinions those of the Vulcan Science Academy. 🐰x2. 🏡 https://md.ekstrandom.net 🧪 https://inertial.science

3053 Posts2255 Follower1014 Folgt
Michael Ekstrand@md.ekstrandom.net · yesterday

This is interesting work, and important question! It's also a question we touch on, briefly & much less formally, in our paper on LLM fairness: we argue that general model providers should have an affirmative responsibility to support downstream auditing. md.ekstrandom.net/pubs/acl-llm...

The Impossibility of Fair LLMs

md.ekstrandom.net

Benjamin Laufer@laufer.bsky.social · 2w ago

One intuition behind many AI policy proposals is that downstream AI applications -- the companies deploying AI in healthcare, finance, education, customer service, etc. -- should bear responsibility for ensuring safety. Our paper asks: What incentives does that create for the firms building AI?

Illustrative example of our game-theoretic model. This numeric instance of the game consists of one general-purpose producer and three domain-specialists. Each player has a different utility in performance-safety space which dictates the path of development. Within this setting, the no-regulation game (Upper Left) reveals the players’ investment efforts when no floor is imposed on safety. Regulating the domain-specialist alone (Upper Right) exhibits backfiring for all three domains, meaning the regulated safety level is lower than it would be without regulation. In this particular example, the same floor is assumed for all three domain-specialists. Regulating the generalist alone (Lower Left) improves the safety level slightly across all three domains, compared to no-regulation. Finally, a regime that targets both generalists and specialists with regulation (Lower Right) is able to 1) retain the improved safety performance from regulating the generalist, 2) improve the safety level of least-safe domain-specialist, while 3) avoiding backfiring. The purpose of this figure is to visualize the model’s incentive mechanisms; none of these panels represent real regulations.
Sarah Taber@sarahtaber.bsky.social · yesterday

I would like to propose the Taco Bell Principle When a person/institution Did A Problem, & IMMEDIATELY fesses up & takes proactive steps to solve it That's exactly what we want them to do? So how bout we let them live ¯\_(ツ)_/¯

Sarah Taber@sarahtaber.bsky.social · 3w ago

A quick word on Taco Bell: That's where we found the suspected lettuce, and it's not bc Taco Bell is gross. It's bc Taco Bell is REALLY GOOD at tracing product! Keeping track of exactly where they got their ingredients, & exactly what restaurants each shipment of ingredients went to!

Brooke Harrington@ebharrington.bsky.social · 2d ago

"don't make enough money" is a fig leaf All those degree programs--social work, art, etc.--have disproportionately female student bodies. This is designed to make it harder for women to attend college--college education being a key factor in women's autonomy from men & left-leaning voting pattern.

The New Republic@newrepublic.com · 2d ago

President Trump is banning students majoring in degrees that don’t make enough money from taking out college loans. Degrees for social work, art, religious studies, teaching aides, and music will be hit the hardest. trib.al/aaAEu6v

John Scalzi@scalzi.com · 2d ago

Bluesky is a "liberal bubble" because fascist chuds can't get traction without algorithmic assistance and because people here trained themselves to block "debate me" trybois the instant they roll up

The Verge@theverge.com · 2d ago

Bluesky's new CEO Toni Schneider on the platform's reputation for being a liberal bubble: "Yes, we definitely want that to change. It is already changing. It certainly wasn’t designed to attract one specific group of people."

Michael Ekstrand@md.ekstrandom.net · 2d ago

typst doesn't have pdfpages support yet, so I am assembling my tenure dossier PDF as $DEITY intended: emitting JSON and blank pages to indicate PDF inclusion, extracting them with pdftotext, parsing with Tcl, and using qpdf to inject the included documents in their correct positions.

Sonja Drimmer@sonjadrimmer.bsky.social · 2d ago

It requires expertise to be able to say, “I do not know the answer to that question, and that is becauae thus far scholarly consensus is that we do not have the knowledge needed at this time to answer that question.” Sometimes “I don’t know” is a far more rigorous answer than people might think.

Dare Obasanjo@carnage4life.bsky.social · 2d ago

People complain about employers wanting to replace them with AI then fall into the bad habit of responding to questions with “Claude said…” or sharing AI slop documents. They fail to realize that they’re making the case that they should be replaced with AI. Your coworkers can prompt Claude too.

Don't be a meat proxy

gruhn.me

GOLIKEHELLMACHINE@golikehellmachine.com · 3d ago

stripping OMB of the ability to dole out grants is an enormous win, this nukes the SAVE act, it prohibits further ICE bailouts, adding a "no, really, trump can't break already on-the-books laws regarding elections the federal government doesn't run" clause would be both unnecessary and unproductive

Murshed Zaheed@murshedz.bsky.social · 3d ago

Patty Murray and Chuck Schumer decide to strike a “bipartisan” deal with … checks notes … Susan Collins to fund the Trump regime through December which appears to not to any include any safeguards against Trump’s meddling with midterm elections.

Screenshots of posts from Burgess Everett on X: “The Senate's CR deal is out from Sens. Collins and Murray. Funds government through Dec. 11 Bill blocks a new rule that would allow the admin to cancel grants to, for example, blue states Collins says the bill "prevents the Office of Management and Budget (OMB)’s proposed rule regarding federal financial assistance from taking effect" Murray: "his bill blocks implementation of OMB’s corrupt new grants rule for the duration of the CR. The proposed rule would systematically politicize federal funding and allow Trump officials to cancel grants at any time for any reason." “Bill also extends surface transportation programs through Dec. 11 (was set to expire Sept. 30 Also has health care extenders Bottom line is this is going to kick a ton of stuff to Congress's lame duck”
Alexandra Olteanu@aolteanu.bsky.social · 3d ago

So what do AI evaluations do or should do? Wrote a brief reflection about prototypical questions evaluations often try to answer and why it is useful to have clarity about what exactly one is trying to learn about some phenomenon of interest. rigor-in-ai.leaflet.pub/3ms4ax4sojs2j

Back-to-basics: so what do evaluations do

TL;DR — Developing useful evaluations requires clarity about what exactly one is trying to learn about a phenomenon of interest.

rigor-in-ai.leaflet.pub

Shannon Vallor@shannonvallor.bsky.social · 4d ago

This is so frustrating to read from The New Yorker. Reward hacking *long* predates LLMs. The Coast Runners paper on reward hacking (in reinforcement learning, pre-LLMs) is 10 years old! Reward hacking is not new, and not a sign of a breakthrough intelligence, much less sentience. Did AI write this?

Bild
John Scalzi@scalzi.com · 4d ago

It's a small thing, I know, but there really is a difference between "palate cleanser" and "palette cleanser," and also, while we are on the subject, "pallet cleanser," and it is worth making the distinction

Dan Hon@danhon.com · 4d ago

st:tng s19e7 The Enterprise hosts the Rigellian Regional Bake-Off diplomatic event, in preparation for the sector-wide finals.

st:tng title card for the episode "Necessary Cookies Only"
thank u, hexed@lindsaykingmiller.bsky.social · 5d ago

Using lists of three is not an "AI tell." It's a common trait in human writing because our brains biologically want to make sure there's always one thing to marry, one thing to fuck, and one thing to murder.

Kate Starbird@katestarbird.bsky.social · 6d ago

Our relationship to visual evidence is being profoundly disrupted by GenAI tools, which will impact how we know, what we can know, and what we think can be known about the world. A massive epistemological shift is upon us.

Faine Greenwood@faineg.bsky.social · 6d ago

I used to do some work on the ethical use of satellite and drone imagery for aid and disaster response, and it is genuinely hard to express in words what a dangerous, stupid thing Google is doing by making it easy to use GenAI to create faked satellite imagery:

Michael Ekstrand@md.ekstrandom.net · 6d ago

I remember a paper from several years ago about classical RL agents learning to exploit bugs/limits in their training gyms to “solve” the problem. One was a little physics sim, the agent learned to exploit the sim’s floating-point error to fly across the world at impossible speed to meet the goal.

Timnit Gebru@timnitgebru.blacksky.app · 6d ago

We're in the era of incompetence & cybercrimes headlined as "unprecedented model capabilities gone rogue." So OpenAI & Anthropic are trying to one up each other with such incompetence because the "press release as a service" performing media & clueless politicians parrot pre IPO CEO talking points.