Alex Lyman

If big companies have our data, can they use in conjunction with LLMs to persuade people on hot-button political issues? Our new study, published in PNAS, puts LLMs to the test persuading people on hot-button political topics. 🧵[1/6]

Bottom line: AI can shift opinions at scale, but personalization and chat didn’t add much. If you’re curious about the ability of AI to persuade-and its limits-check out our full open-access paper: pnas.org/doi/10.1073/pn… [6/6]

While AI is able to move people's attitudes towards the center, this does not increase "democratic reciprocity", a willingness to engage with people who hold opposing views. This suggests that merely changing policy opinions isn't enough.[5/6]

We tested four AI-driven persuasion strategies: 1️⃣ Generic message 2️⃣ Microtargeted message 3️⃣ Interactive direct arguments 4️⃣ Motivational interviewing All were aimed at shifting people away from their starting views, but which worked best? 🤔 [2/6]

Spaced repetition is a lie. Why don't I remember `tar -xvzf` after 10+ years?

No one is ready for this. (1908)

Synthetic data is something else. (Taken from Cosmopedia v2)

We just caught our first "BANANA!"

I just had work on Language Models for quantifying eyewitness confidence statements in Psychological Science. You can read the news brief here: law.upenn.edu/live/news/1650… Or the full paper on SSRN: papers.ssrn.com/sol3/papers.cf…

When I first saw this graph, the main thing that stood out to me was that Japan has the world's least spicy food. Even Brits eat spicier food.

>Code fails to compile under CUDA Toolkit 12.1 >Upgrades CUDA Toolkit to 12.2 >Code fails to compile earlier than before

One of our key sources of human data is no longer fully “human"! We estimate that 33-46% of crowd workers on MTurk used large language models (LLMs) in a text production task - which may increase as ChatGPT and the like become more popular and powerful. arxiv.org/abs/2306.07899

@pmddomingos argues that #AI regulation is misguided, and more on this podcast: spotifyanchor-web.app.link/e/9Y3Ywme1eAb #artificialintelligence #generativeai #regulation #aiact #europe #podcast

“AI Safety” seams to be a euphemism for censorship and I don’t think big corps should be bragging about it quite so much.

I see someone on this website has come up with a specific number of training FLOPS above which a language model becomes AGI and needs to be regulated. Very glad I logged in this morning.

advanced math in deep learning papers

Had a great time shoring up my reinforcement learning with @huggingface's course. Thanks @ThomasSimonini!