Arvind Narayanan

Back in grad school, when I realized how the “marketplace of ideas” actually works, it felt like I’d found the cheat codes to a research career. Today, this is the most important stuff I teach students, more than anything related to the substance of our research. A quick…

Quick thread on the recent IMO results and the relationship between symbol manipulation, reasoning, and intelligence in machines and humans:

It’s truly insane tbh later turns can’t even see a summary, and hallucinate around this fact. “How did you get that answer?” silently elicits a response to “How do you imagine you could have gotten that answer?” with no admission the information requested is actually lost.

🧠New on the CITP Blog from PhD student Boyi Wei (@wei_boyi) of the POLARIS Lab: "The 'Bubble' of Risk: Improving Assessments for Offensive Cybersecurity Agents" Read about how adversaries can adapt and modify open-source models to bypass safeguards. 👇 blog.citp.princeton.edu/2025/07/21/the…

In the 18th century, there was a real chance of death at any point in life, and there wasn't a big peak in old age. It wasn't just higher infant mortality - the whole distribution was completely different. Great chart by @Scientific_Bird.

The zero-sum trap: the more people believe that wealth, status, and well-being are zero-sum, the more they back policies that make the world zero-sum. marginalrevolution.com/marginalrevolu…

Experts on body dysmorphic disorder have warned that people struggling with it have become increasingly dependent on AI chatbots to evaluate their self-perceived flaws and recommend cosmetic surgeries. "It’s almost coming up in every single session,” one therapist tells me.

Great nuanced take from Arvind and Sayash (as usual)

SB1047 was a bad idea. But Sen. Wiener's latest SB53 is on the right track, and it's important to call out the progress. Here's my reasoning. My approach to regulating novel technology like models is: we don't know how to define "good" mitigation and assurance, but we'll know it…

For folks wondering what's happening here technically, an explainer: When there's lots of training data with a particular style, using a similar style in your prompt will trigger the LLM to respond in that style. In this case, there's LOADS of fanfic: scp-wiki.wikidot.com/scp-series🧵

We can get pretty far in a post-apocalyptic future as long as our devices still work, with the help of Wikipedia dumps, local LLMs, solar power, batteries, mesh networking, and maybe serverless apps with P2P backends. It would be much more useful than the internet I grew up with!

If we compared AI capabilities against humans with no access to tools, such as the internet, we would probably find that AI already outperformed humans at many or most cognitive tasks we perform at work. But of course this is not a helpful comparison and doesn’t tell us much…

New Post: Datacenter Delusions Among middle powers, the UAE has the clearest AI strategy: become an 'AI oil state' with meaningful global compute share. Most other countries are building datacenters without targeting viable sovereignty thresholds. 🧵

Life as an Apple developer is always just one capricious, bureaucratic decision away from leaving you ruined. Intolerable.

In the running for my favorite blog post from Sayash and Arvind! When people ask me for areas I am most excited about for AI, my answer is often some version of "science-first AI-for-science". There is a lot to do to figure out what that means and how to do it well.

I am very happy to have this to point to in response to hand-wavy hype about AI 'curing cancer.' I feel confident that Arvind and Sayash are probably right and almost certain that these points are healthy for our discussions about AI and scientific progress.