Matthew Hindman

Want to understand RT, the Russian outlet indicted for funneling $10 million to right-wing influencers, including @DaveRubin and @Timcast and @bennyjohnson? Good news! Our new article on RT (by @yangyunkang, @mccabe_s & I) is just out in IJPP. Just in time for #APSA2024

A novel and very promising approach: specialized guardian models that are as resistant to abuse as big frontier models, but much smaller and more efficient.

Some people find politics interesting. Others do not. In a new paper, I show that appealing to MEANING increases political interest. In 6 experiments, connecting what people find meaningful in their lives to politics increases political interest. Link: osf.io/preprints/soca…

ML researchers just built a new ensemble technique. It even outperforms XGBoost, CatBoost, and LightGBM. Here's a complete breakdown (explained visually):

@d_e_holliday Welcome! We are thrilled to have you.

It's official -- I'll be taking my talents to Foggy Bottom this Fall! Excited to announce I've accepted an offer to join the Department of Political Science at the George Washington University as an Assistant Professor of American Politics!

Striking chart from recent @pewresearch work on how US Americans think about the news. Just 7% say that the news they get often makes them feel empowered - a fraction of the many more who say it makes them angry, sad, scared, and/or confused. #what-americans-think-news-is-and-is-not" target="_blank" rel="nofollow noopener">pewresearch.org/journalism/202…

Levitsky, Way, and Ziblatt: "No One Has Ever Defeated Autocracy From the Sidelines" (gift link in thread) Key points: 1. We are now living under competitive authoritarianism. 2. We need ALL of civil society to resist the temptation of appeasement and defend democracy.

Happy to see our paper out on @SciReports! We introduce a statistical method for detecting coordination and apply it to Facebook, using 11.2 million link posts shared by 16k pages that discussed US politics in 2021. nature.com/articles/s4159…

Thrilled to see this paper out! Thanks to everyone who has given us feedback over the years, including the anonymous reviewers and handling editors.

@ScottNover @washingtonpost Congrats Scott! Well deserved. I’m just sorry that this opportunity comes at such a slow time for the media beat. Should be a sleepy six months—good luck finding something to report on!

Some personal news! Today is my first day covering media for the @washingtonpost. I'll be here for the next six months writing about news organizations, press freedom, and how people get their information. Send tips, press releases, and friendly barbs to scott.nover@washpost.com

🚨 Job Alert: 2 PhD Positions Available at 📍 @FU_Berlin 🚨 I recently started a new position as Prof of Communication @FU_Berlin, honored to lead the Media Effects Research Group. I have 2 PhD positions available—if you're passionate about media effects, political communication

🚨 🚨 🚨 🚨 🚨🔥 ❗️❗️❗️ @HouseGOP Republicans are launching an unprecedented attack on the residents of DC. This is the most significant Congressional interference that I’ve seen in my 22 years living in DC, & it is being done maliciously to harm our city. (1/8)

NEWS: For the first time, @DCAttorneyGen is suing three non-D.C. drivers for tens of thousands of dollars worth of unpaid traffic camera tickets racked up over a decade. The three drivers collectively have 226 tickets worth more than $90,000 in fines.

This paper addresses the issue of inefficient inference when using long contexts in transformer models. Introduces a method to reduce inference costs. It attends only to the most important tokens at each step using a top-k selection, enhancing efficiency for long contexts. ----- 📌 By offloading key/value cache to CPU and using top-k selection, this method drastically reduces GPU memory needs. It enables million-token context inference on commodity GPUs. 📌 This work empirically validates inherent attention sparsity in LLMs. Top-k attention effectively exploits this, focusing on crucial tokens while maintaining performance with minimal overhead. 📌 Architecturally, top-k attention smartly shifts key/value cache to CPU, using approximate nearest neighbor search. This decouples context length scaling from GPU memory limitations for practical deployment. ---------- Methods Explored in this Paper 🔧: → It selects only the most relevant key-value pairs for attention computation at each layer. → Key and value vectors are stored in CPU memory within a vector database. → Approximate k-Nearest Neighbor search is used to retrieve top-k keys from the CPU cache for each query during decoding. → The retrieved keys and values are moved to GPU for attention computation. → This method reduces computation and memory overhead by focusing on crucial tokens. → The value of k can be adjusted per layer to optimize performance and efficiency. ----- Key Insights 💡: → Modern LLMs exhibit sparse attention patterns. → Only a small fraction of tokens significantly contribute to the attention mechanism. → Top-k attention effectively exploits this sparsity. → Models can maintain performance even when attending to a very small percentage of input tokens. → This sparsity is observed across different model sizes, architectures and training types. -----

🚨After eliminating parking requirements last year, Cambridge, Massachusetts has passed one of the more ambitious rezoning efforts in America tonight allowing 6-stories citywide by an 8-1 vote. Prior to this update, the City estimated only 350 new units would be built by 2040.

New open source reasoning model! Huginn-3.5B reasons implicitly in latent space 🧠 Unlike O1 and R1, latent reasoning doesn’t need special chain-of-thought training data, and doesn't produce extra CoT tokens at test time. We trained on 800B tokens 👇

An underappreciated reason why Harris lost was flat real disposable income growth in the election year. RDI was above trend, but voters care about growth in the election year itself. Between Jan. and Nov. 2024, nothing happened.

@ML_Burn Congratulations! Well deserved, and a great spot to continue your work.

This Fall I'm joining the Texas A&M political science department as an assistant professor! Thrilled to be joining such a fantastic program and can't wait to get to work.

My thoughts on China, export controls and two possible futures darioamodei.com/on-deepseek-an…

everyone comparing deepseek-r1 to o1 and forgetting about Gemini 2 Flash Thinking which is better than r1 on every cost and performance metric

We have two (2!) jobs at GW, one in Political Science and one at SMPA. Both will be affiliated with @GWIDDP. SMPA job: gwu.jobs/postings/116060. Poli Sci job: gwu.jobs/postings/116106. Let me know if you've got questions!