Margaret Mitchell

@mmitchell

Researcher trying to shape AI towards positive outcomes. ML & Ethics +birds. Generally trying to do the right thing. TIME 100 | TED speaker | Senate testimony provider | Navigating public life as a recluse. Former: Google, Microsoft; Current: Hugging Face

827 Following23,564 Followers1,941 Posts

Bluesky

Is this your profile?Claim it and mmitchell.dirty.social is yours.Claim profile

Margaret Mitchell@mmitchell· 17d

8a/8 Hugging Face's security systems detected the intrusion. Given the sheer volume of recorded actions, engineers decided to use an AI agent system to analyze what had happened. They first tried a closed system, but its safety guardrails blocked them. So they ran the analysis using an open model...

Margaret Mitchell@mmitchell· 17d

7/8 Once inside, the agent's commands generated thousands of actions, moving through systems and harvesting credentials.

Margaret Mitchell@mmitchell· 17d

6/8 Hugging Face processes datasets, and the agent was able to utilize some of the ways they are processed to gain access into Hugging Face internal systems.

Margaret Mitchell@mmitchell· 17d

5/8 With access to the wider internet, the agent's commands targeted Hugging Face, a major AI platform where ExploitGym answers might be stored.

Margaret Mitchell@mmitchell· 17d

4/8 With access beyond the sandbox, the agent issued commands that reached through OpenAI's internal research systems, moving from machine to machine until they reached the wider internet.

Margaret Mitchell@mmitchell· 17d

3/8 The agent uncovered a previously unknown flaw in the proxy that linked to the outside, and exploited it. Instead of just requesting software tools, the agent was able to use the proxy to connect to things outside of the sandbox.

Margaret Mitchell@mmitchell· 17d

2/8 The sandbox couldn't connect to the internet, but it could request software packages through a "proxy" that would fetch tools from the internet on its behalf and pass them in. This proxy was the sandbox's only link to the outside world.

Margaret Mitchell@mmitchell· 17d

People not in tech have been interested in what happened with the Open AI agent/Hugging Face hack. So, I "Explain it Like I'm 5"-ed it and made a cartoon in 8 panels. Apologies for some anthro-ing. 🫤 m-mitchell.com/HF-hack-cart... 1/8

An AI agent in development at OpenAI was recently tested on how well it could find and exploit code vulnerabilities.

The test began in a "sandbox", an environment meant to keep everything the agent does contained, so that nothing it triggers can affect systems outside of the test.

Margaret Mitchell@mmitchell· Jul 6

If you haven't read this piece on language use in around AI, you definitely should. Great work from @emilymbender.bsky.social , @nannainie.bsky.social and Peter Zukerman, with solutions for what to call things we backoff to just anthropomorphising. firstmonday.org/ojs/index.ph...

Table 3: Suggestions for de-anthropomorphizing language.
Category, Alternative Strategies, Examples.

Cognizer & Products of cognition: Instead of language that locates thinking in an algorithm, describe the algorithm as performing calculations or other algorithmic operations and the people using it doing the thinking. Examples:
artificial intelligence → probabilistic automation

hybrid intelligence → augmented human intelligence

image recognition → image labeling

speech recognition → automatic transcription

the model shows bias → the model reflects bias

model mistakes → model errors

chatbots are good at … → chatbots are good for ….

Margaret Mitchell@mmitchell· Jun 29

I am beyond honored to have been selected to give a keynote at Deep Learning Indaba: 4.August, Lagos. I will be cutting into some of the deepest issues I see in AI (hint: it's not alignment) and what I think we can do solve them.

Margaret Mitchell@mmitchell· Jun 11

🌱 My fantastic blue-purple Delphinium is rapidly blooming, and today this pretty yellow butterfly was chilling on it, and it was glorious. I just thought you all should know.

Margaret Mitchell@mmitchell· May 28

🪶 If you’re in Europe and have ice, consider putting some out in bowls/bird baths.

Heat map of Europe showing all land in red, with temperature values across different cities around 30 degrees C.

Margaret Mitchell@mmitchell· Mar 22

🪶 Pileated woodpecker on my walk today. Chillin’ on a Doug Fir, next to a Madrone.

Margaret Mitchell@mmitchell· Feb 6

This older piece from Anthropic's CEO on developing a creator-compensation economic model is incredible to see. Disagree with some things, but it's fantastic to see this level of thinking at the intersection of ethics and profit. Source: www.courtlistener.com/docket/69058...

An Economic Model for Compensating Data Producers
One of the controversies over Copilot, related to but distinct from concerns about the model spitting out copyrighted data, is a feeling that large corporations (MSFT and OpenAl) are exploiting the open source work of developers and using it extract concentrated profits. The same could be said about GPT-3 being trained partially on fanfiction books, or for that matter about models being trained on human feedback, for which workers are paid near-minimum wage. I think this is a real and important concern, especially when considered as part of an overall trend for Al progress to concentrate wealth, increase inequality, and in the long run automate labor. Zooming very far out, what is happening macroeconomically is that distributed human labor is being used to train Al models by large, centralized actors, who concentrate the resulting profits while in the long run making the human labor obsolete. It seems like the default trajectory is for one of the following things to happen:

Margaret Mitchell@mmitchell· Jan 25

This statement below is a really big deal for the U.S.A. The NRA is a bedrock institution of our Second Amendment of the Constitution; and I reckon many supporters of the current president are members of this organization.

Post on X from the NRA. It is quoting a post from F.A. United States Attorney Bill Essayli, who says, "If you approach a law enforcement with a gun, there is a high likelihood they will be legally justified in shooting you. Don't do it!" Below, Readers added context: "The U.S. Constitution (particularly the 2nd, 4th, and 14th amendments) prohibit officers from shooting citizens merely for possessing a weapon that is not an "imminent threat". This was reaffirmed in Graham v Connor, which says force must be reasonable."

The NRA post says: "This sentiment from the First Assistant U.S. Attorney for the Central District of California is dangerous and wrong. Responsible public voices should be awaiting a full investigation, not making generalizations and demonizing law-abiding citizens."

Margaret Mitchell@mmitchell· Dec 17

Here's what the UI is doing for me now (which may very well be them quickly addressing the clear pushback).

Same screenshot as original, but 'Abstract' is the bolded tab instead of 'Summary (with little stars)'. A hover-over on 'Summary (with little stars)' says "To view this AI-generated summary, you must have Premium access. Learn more. Sign in."

Margaret Mitchell@mmitchell· Nov 16

#SixOnSaturday 🌱

Smoke bush, orange dahlia, red mums, magenta dahlia, close up of yellow nasturtium with vermillion maple-leaf-shaped inner markings, red mums against the blue sky.

Margaret Mitchell@mmitchell· Nov 10

There's this form of behavioral engineering/coercion going on to use AI when you're not even trying to. Ex: Enterprise Google Slides replaced "Upload an Image", with "Generate an Image" (right?), requiring extra work +adaptation to sidestep the AI-as-default push. It drives me bonkers.

Before and After shots of Google slides menu bars, Insert --> Image. Before, the first option was "Upload from computer". After, the first option is "Generate an image", with "Upload from computer" a distant second.

Margaret Mitchell@mmitchell· Nov 9

🪶 Golden-crowned Kinglets at the water spout.

Margaret Mitchell@mmitchell· Oct 29

🤖 Did you know your voice might be cloned without your consent from just *one sentence* of audio? That's not great. So with @frimelle.bsky.social, we brainstormed a new idea for developers who want to curb malicious use: ✨The Voice Consent Gate.✨ Details, code, here: huggingface.co/blog/voice-c...

Ornate line drawing of a fence and gate, with fleur de lis tips. The gate says CONSENT where the family name usually is.

Margaret Mitchell@mmitchell· Oct 26

Screensnapshot of our times. Corporate pressure to use AI creates bad outcomes with real-life consequences; seeking to unpack the problem yields more pressure to indulge in the very thing causing the problem.

Reuters screenshot. Headline: "Two federal judges say use of AI led to errors in US court rulings". Bottom left: Reuters AI saying, "Get the key points from this story with Reuters AI"

Margaret Mitchell@mmitchell· Oct 24

Life situations are bleak right now for a lot of people. In tech, the "Venn Diagram" of (1) positive work and (2) making enough money to support your family is increasingly non-overlapping. We all do what we can. This image has been living in my mind rent-free for months.

Tech Venn Diagram
Two circles, one says "Doing work that contributes beneficially to society."; the other says "Making enough income to support a family."

Margaret Mitchell@mmitchell· Oct 12

#SixOnSaturday

Margaret Mitchell@mmitchell· Oct 9

🎭 As a Samuel Beckett and Alex Winter fan, I highly, highly *highly* recommend going to see this version of Godot. Laugh-out-loud funny, masterfully handled some of the confusions from text. Use of funnel as stage was brilliant. Only props were hats and shoes. Beckett would be proud.

Keanu Reeves and Alex Winter as Gogo and Didi, sitting in a funnel-shaped tan/white stage against a stark black background.

Margaret Mitchell@mmitchell· Oct 8

Apparently this was Portland last night. Gosh I love my home away from home. ❤️

Margaret Mitchell@mmitchell· Oct 5

This post about Waffles that’s going viral doesn’t credit (what I believe to be) the source, Bravo Coolee, on Twitter a few years ago. Just wanna shout out the original. Still viral thru another person’s rewrite today.

Twitter screenshot from @BravoCoolee

“Twitter is the only place where well articulated sentences still get misinterpreted.

You can say “I like pancakes” and somebody will say “So you hate waffles?’

No bitch. Dats a whole new sentence. Wtf is you talking about.

Margaret Mitchell@mmitchell· Oct 5

My #SixOnSaturday with Fringe Tree berries, pink Wax Begonia flowers, just-opened crimson Mums, Black-capped Chickadee eating a Dogwood berry, scarlet Nasturtium, red blueberry leaf fallen into native PNW Slough Sedge.

Margaret Mitchell@mmitchell· Sep 28

#SixOnSaturday appreciating the first Saturday of autumn

Rust-colored Dahlia about to bloom, Tuberous Begonia, dark red Mums about to bloom, Witch Hazel leaf turned red, red rose hips, and red Coleus Inferno.

Margaret Mitchell@mmitchell· Sep 26

🪶 🎃 It’s Halloween deco season for many — remember that those gauzy fake spider webs trap and kill birds. Don’t use them. OTOH, consider making a bird-feeder pumpkin or sticking decos on windows, which helps birds. More here: www.audubon.org/magazine/how...

Picture of house with gauzy spiderweb material all over it.

Margaret Mitchell@mmitchell· Sep 23

This probably won't reach the very people who should know, but apparently I've been put on all kinds of (mostly spam) block lists and lists accusing stuff I'm vocally against, leading to me being blocked when uncritically accepted without examination. Great way to erase voices eh.