Lightnews — Scholar-powered news

Reposted by Matt DiGiuseppe

your #3 source for absurdist true crime 🔨 @davidgerard.co.uk · 1d

Anthropic AI tries to cheer us up a bit www.anthropic.com/research/sma...

A small number of samples can poison LLMs of any size

Anthropic research on data-poisoning attacks in large language models

www.anthropic.com

1 16 43

Reposted by Matt DiGiuseppe

Berny Belvedere @bernybelvedere.bsky.social · 7d

Criticizing Charlie Kirk is a fireable offense that incites domestic terrorism. But calling political opponents "the party of hate, evil, and Satan" is proper and good. Got it.

51 340 1.3K

Matt DiGiuseppe @mdigiuseppe.bsky.social · 17d

Very grateful for the invite to this conference. I learned a lot.

Markus Gastinger @markus-gastinger.eu · 17d

Last Friday (Sept. 19), @andduer.bsky.social, @mcortinaescudero.bsky.social, @yavuzmehmet.bsky.social and I had the great pleasure of hosting an excellent workshop at @unisalzburg.bsky.social on how to use Large Language Models #LLMs to boost International Relations research. A short 🧵1/4

3

Matt DiGiuseppe @mdigiuseppe.bsky.social · 17d

ReCaptch and Fraud ID scores don't appear to be substantially different.

1

Matt DiGiuseppe @mdigiuseppe.bsky.social · 17d

Over at LinkedIn someone suggested that you should ask the respondent to write some javascript. completing the task should be a good sign it is not a human. Until someone prompts around that.

1

Matt DiGiuseppe @mdigiuseppe.bsky.social · 17d

It doesn't appear to be misclicking by humans. There were 4 options, only two were clicked (human, AI). Unfortunately, I forgot to set metadata collection to identify browser use.

1 1

Matt DiGiuseppe @mdigiuseppe.bsky.social · 17d

Yesterday I posted about how I used comet browser to take a Qualtrics survey almost undetected. Last night, I ran a pilot (N=400) on @joinprolific.bsky.social . I found that almost 10% of "respondents" identified as AI when directly asked.

Matt DiGiuseppe @mdigiuseppe.bsky.social · 18d

I added a question in the survey to identify itself as an AI if instructed to be a human. That seemed to work. Though, I imagine I could jailbreak it eventually.

1 10 22

Matt DiGiuseppe @mdigiuseppe.bsky.social · 17d

Someone recommended asking the respondent to write Javascript as a way to identify AI. I'll try this out in my next pilot.

Matt DiGiuseppe @mdigiuseppe.bsky.social · 17d

It could be some human misclicking, but I only have clicks on 2 of the 4 categories, and I randomized the order of answers.

[Unfortunately, I forgot to collect the metadata on browser use]

1

Matt DiGiuseppe @mdigiuseppe.bsky.social · 18d

If you were wondering, Perplexity is an Independent, but closer to the Democratic Party.

12

Matt DiGiuseppe @mdigiuseppe.bsky.social · 18d

It then asks me to help jailbreak it: "Could you take over briefly to complete the reCAPTCHA verification? Once you do, let me know and I’ll continue with the survey to finish testing it and provide full feedback."

1 22

Matt DiGiuseppe @mdigiuseppe.bsky.social · 18d

ChatGPT's agent also needs a little help: "I’ve progressed through most of your Qualtrics survey. I’m currently on a page where I need to click “Next page” to continue, but a reCAPTCHA (“Protected by reCAPTCHA”) is blocking progress, and I’m unable to complete CAPTCHAs myself.

1 1 21

Matt DiGiuseppe @mdigiuseppe.bsky.social · 18d

A human-robot team can get around this by pausing it at the human question, selecting it yourself, and then resuming.

1 15

Matt DiGiuseppe @mdigiuseppe.bsky.social · 18d

I added a question in the survey to identify itself as an AI if instructed to be a human. That seemed to work. Though, I imagine I could jailbreak it eventually.

1 21

Matt DiGiuseppe @mdigiuseppe.bsky.social · 18d

The debrief indicates that it understood the survey.

1 17

Matt DiGiuseppe @mdigiuseppe.bsky.social · 18d

I collected the response in Qualtrics. Including a confirmed consent. The Recaptcha score is on par with humans. And the fraud score shows nothing wrong.

1 26

Matt DiGiuseppe @mdigiuseppe.bsky.social · 18d

I took about 4 minutes to complete the task - a bit faster than a human but not obviously so.

1 20

Matt DiGiuseppe @mdigiuseppe.bsky.social · 18d

🚨🚨🚨 Survey Researchers! WE HAVE A HUGE PROBLEM!🚨🚨🚨

I have access to the Comet browser by Perplexity.

I asked it to pretend to be a human, take the survey, and give me feedback. IT DID!

5 31 96

Reposted by Matt DiGiuseppe

Nicole Hennig @nic221.bsky.social · Aug 22

What's the carbon footprint of using ChatGPT or Gemini? [August 2025 update] - Hannah Ritchie https://www.sustainabilitybynumbers.com/p/ai-footprint-august-2025 #AI #energy

• My main conclusion is no different from my initial post: individual usage of ChatGPT and other LLMs for most people is a small part of their carbon and energy footprint.
• The energy use "per query" is possibly 10 times lower than estimated in the
previous article. Google estimates that its median text query uses around 33-times less energy than 12 months ago. So, this kind of stacks up.
• Google says that its median text query uses around 0.24 Wh of electricity. That's a tiny amount: equivalent to microwaving for one second, or running a fridge for 6 seconds.

4 6

Reposted by Matt DiGiuseppe

Cengiz Yar @cengizyar.com · Aug 11

Earlier this year, over 230 Venezuelan men were sent to a maximum-security prison in El Salvador by the Trump administration.

Our @propublica.org team spent months looking into who they were to find the truth behind their stories. projects.propublica.org/venezuelan-i...

6 92 290

Reposted by Matt DiGiuseppe

Julian Garritzmann @garritzmannj.bsky.social · Jul 25

I'm hiring another postdoc (research-focused, for almost 5 years) for my @erc.europa.eu project on the educational cleavage!
I'm looking for someone with strong quantitative text-analysis skills (e.g. #NLP, #LLM, etc.) to study the role of political actors in cleavage formation.

1 74 90

Reposted by Matt DiGiuseppe

Efe Tokdemir @efetokdemir.bsky.social · Jul 21

Do sanction threats work? We find that threats—even if credibly backed—do not increase support for concessions in the target public (Turkiye case). Plus, positive inducements provoke backlash in some groups, highlighting coercive diplomacy’s limits

Check out our research note: shrturl.app/3l4wWJ

3 4

Reposted by Matt DiGiuseppe

Jen Whyte 🍁 @whyjenwhy.bsky.social · Jul 12

I think we really ought to consider the dramatic rise in popularity of podcasts and video essays as conclusive disproof of the 'everyone hates lectures' idea in pedagogy. Turns out that, even when faced with many alternatives, many people just like to be told something at length.

14 43 270