Jonathan Stray

Knowing things is a solved problem. Getting along is not. Working on AI, media, and inter-group conflict @CHAI_Berkeley. Got here from computational journalism.

[Jonathan Stray @jonathanstray]

Could social media make us less polarized instead of more?

We tested 5 algorithms on 3 platforms with 10,000 people for 6 months during the 2024 election, and found that the answer is yes.
🧵

[Jonathan Stray @jonathanstray]

Seeing a flurry of evals and startups promising to test the mental health effects of AI. Literally all of them test what the model says in various conditions... none of them measure actual outcomes on actual people. A big gap, fixable with privacy-preserving experiments.

[Jonathan Stray @jonathanstray]

Thank you for the explanation of what you were thinking. But I'm afraid I can only assign you a grade for the paper you wrote, not the paper you could have written.

[Jonathan Stray @jonathanstray]

@joshgans Is it? I read this as saying that the predictions were correctly calibrated to repayment probability in all cases. So claiming the AI was “biased” here requires one to endorse an affirmative-action style argument for lending. I’m not sure that’s obviously the right answer here.

[Jay Van Bavel, PhD @jayvanbavel]

A new experiment involving 1,500 participants in 30 decision environments finds that AI advice depolarizes choices ~on average~, moving participants away from their initial leanings.

However, sycophantic AI increases polarization (p < .001). This poses a potential societal problem given that AI becomes increasingly sycophantic the more people engage with it and it customizes answers to match user preferences.

This suggests that the design features of AI are going to be critical to the impact is has on individuals, groups, and society. The technology can amplify or mitigate intergroup conflict, depending on how it's designed.
https://t.co/rcghZd1LGO

[Ethan Mollick @emollick]

Math is easy* because it has verifiable outputs and few messy judgement choices to make.

Which AI labs have the guts to make advancing social science a priority? It may actually do more for human flourishing to unlock sociology, econ &amp; psych reseach.

* For AIs, not for humans

[Jonathan Stray @jonathanstray]

PSA to my students: don't use LLMs to write or process your bibliography. Demand links from all searches. Check the links work and read the abstract yourself. Seeing a lot of garbage cites in this year's research papers.

Actually: PSA to everyone's students.

[Jonathan Stray @jonathanstray]

@SealOfTheEnd @sebkrier You haven’t actually read anything I’ve written on the subject, have you?

[Jonathan Stray @jonathanstray]

Is there some reason why chat transcript search on Claude and GPT isn't, you know, standard full text search? It seems to be some sort of summary matching or who knows what, something far too AI clever, pretty much useless for actually finding previous conversations.

[Jonathan Stray @jonathanstray]

@scychan_brains @sebkrier Yes, or you can attend live! https://t.co/uop4u62JRP

[Jonathan Stray @jonathanstray]

"My AI Utopia Includes World Peace."
That's the title of the Ignite talk I'm giving in SF Wednesday night, where I'll discuss practical ways we can build AI to reduce human conflict. Tickets here.
https://t.co/uop4u62JRP

[Jonathan Stray @jonathanstray]

Seeing a bunch of industry signals that Claude is just a better model than GPT in a bunch of ways... wonder why. Maybe because it's trained on books not just internet?

[Sean Westwood @seanjwestwood]

The rank order replicates with better data, but the % supporting partisan murder is much lower.

Importantly, this is passive support and not willingness to actually murder.

[benedict @bqbrady]

Introducing Philosophy Bench, my favorite new project I've worked on this year, with help from my friend @matthewjmandel

We put frontier language models in 100 ethically complex situations and require them to act, grading them on adherence to consequentialism vs. deontology, tendency to follow user requests, corrigibility, and more

1/

[Jay Van Bavel, PhD @jayvanbavel]

This new report from @Yale University explains why universities have lost public trust and how we can regain it.

The committee offered dozens of recommendations, like expanding financial aid, reducing admissions preferences, zealously protecting free speech and adjusting grading policies.

People in academia, the committee said, “must be willing to admit where we have been wrong and where we might improve, even as we defend what is essential about higher education and its academic mission.”

What do you think?

https://t.co/cJm5vYYhfe

[John B. Holbein @JohnHolbein1]

Why? It may be that in a polarized era, co-partisan protest preaches to the choir.

Opposition protest — especially when it turns confrontational — activates the other side.

It's not your marches that move elections. It's theirs.

[Nate Silver @NateSilver538]

Sharing some proprietary Silver Bulletin data. Sure we are an unusual newsletter in some ways. But there's basically *no* correlation between "engagement" metrics and paid subscriptions. Over-optimizing for engagement often isn't good for product quality *or* for business.

[Jonathan Stray @jonathanstray]

In which @GinaSKChua says nice things about my argument that journalists sometimes increase polarization because they're not trained as conflict professionals. And also some bits about how AI could help.
https://t.co/LG8Mkclkmh

[Jonathan Stray @jonathanstray]

And that's all we can know. X disabled their Academic API shortly after acquisition. In theory one could collect "hate" impressions data from the Enterprise API, starting at $42,000 per month.
Read the full piece at @techpolicypress
https://t.co/Z7kqfBs9zE

[Jonathan Stray @jonathanstray]

Did "hate" go up on X/Twitter after they changed their content moderation rules post-Musk? What is "hate" anyway, and what data do we actually have? This part 2/3 of my deep dive into the politics of a platform.
🧵

[Jonathan Stray @jonathanstray]

So did "hate" go up or down? Ultimately, the narratives from both X and outside observers may both be right -- they measure different things, over different time periods. Production went up (at least for a while), and consumption went down (at least for a while).
5/