iiinova — where your digital self keeps evolving

OpenAI's most cited mathematician is OpenAI. Yesterday OpenAI published 722 math papers. I explored the citation network underneath them. appliedoverthinking.substack.com/p/openais-mo...

OpenAI’s most cited mathematician is OpenAI.

OpenAI released 722 math papers yesterday. What can we learn from their citation network?

192

There's lots of interesting structure to analyze in the network representing these papers and their citations. I wrote up the findings, with plots and code, in my latest post.

000

These observations don't necessarily mean that the proofs or logic in these papers are faulty, but they raise questions about how citation and publication norms may change.

100

OpenAI is a serial self-citer: more than half of the papers cite another OpenAI-authored work. I also found many instances of citation "cycles," where for example paper A cites to B, B cites to C, and C cites back to A.

100

OpenAI's most cited mathematician is OpenAI. Yesterday OpenAI published 722 math papers. I explored the citation network underneath them. appliedoverthinking.substack.com/p/openais-mo...

OpenAI’s most cited mathematician is OpenAI.

OpenAI released 722 math papers yesterday. What can we learn from their citation network?

192

Or is there some other reason for the gap between the interest and attention this topic is getting and its appearance in these journals?

000

Perhaps the scientific foundations for this area haven’t fully formed, so currently discussion is more conducive to news and editorial pieces, which use the term in greater numbers.

100

For example, perhaps AI safety researchers are simply publishing elsewhere rather than submitting to these legacy journals. Perhaps review timelines are longer than the time that has gone by since this topic became a major research focus.

100

This is obviously an imperfect way of measuring the amount of research on AI safety, as there is tons of it. But I wonder why it isn’t appearing in general-interest science journals.

100

From my attempt to measure this last week, I found that my collaborators and I wrote the only research article published in Nature, Science, or PNAS with “AI safety” in the title. (Our paper was in PNAS.) It is 1 of only 9 research articles across these journals to use the term anywhere in the text.

Bar chart showing the distribution of article types in Nature, Science, and PNAS that mention the term "AI safety" anywhere in the text. There are 28 new strikes, 13 editorials, and 9 research publications that include this term.

100

I became curious how many research articles in general-interest science journals mention “AI safety.”

100

They landed on a similar angle “two o clock” though not exactly the same. A hypothesis I have that isn’t exactly substantiated by this data is it’s good to face the hole in the direction of the side of traffic, since turns are sharper and quicker in that direction. I discuss this in the post.

100

This exercise also raised larger questions about what AI does and doesn’t make easier to study. AI was useful for finding data and building simulations. But even for “simple” questions, simulation can only tell us so much, and empirical measurement remains an important bottleneck.

000

This became the latest installment of Applied Overthinking, my attempt to use computation and AI to investigate everyday questions. (Read and subscribe, it’s free!) appliedoverthinking.substack.com/p/which-way-...

Which Way Should the Hole in Your Coffee Lid Face?

Using real driving data and a sloshing simulation, the answer seems to be "Two O'Clock."

120

More generally, the simulations suggest that a diagonal orientation toward the side of the car performs better than pointing the opening directly forward or backward. Sipping the coffee down a bit helps quite a lot, too.

So, with help from an AI coding agent, I built a simulation of liquid sloshing in a to-go coffee cup and combined it with 5.3 hours of real vehicle-motion data from Boston and Singapore. The answer, at least according to the model: point the opening roughly toward two o’clock.

Scientific figure showing simulation results which indicate the optimal direction to direct the opening of a coffee cup lid is 2 o’clock, given a 90% full cup. Generally pointing sideways and slightly askew rather than forward or back seem to reduce spill rate.

220

LSE Methodology @lsemethodology.bsky.social · 21/08/2026

Every time I put a hot coffee in a car cupholder, I make a small decision: which way should the hole in the lid face? Forward seems risky when you brake. Backward seems risky when you accelerate. Sideways seems plausible. But I wondered what the optimal direction is, and whether we can find it.

Image of a to-go coffee cup with lid opening oriented backwards.

140

Reposted by Benjamin Laufer

💡 We're hiring for an Assistant Professor in Computational Social Science Read more and apply before 11 October 👉️ www.lse.ac.uk/methodology/...

LSE Department of Methodology hiring — Assistant Professor in Computational Social Science; applications close 11 October 2026; apply at www.jobs.lse.ac.uk and follow @lsemethodology.

01211

Reposted by Benjamin Laufer

Kenny Peng @kennypeng.bsky.social · 18/08/2026

New York City 8th graders choose from 900+ high schools to apply to, in a process that’s spawned Facebook groups and dozens of expensive consulting services. Our new paper shows how application behavior leads to disparities, and how to effectively intervene. 🧵 www.nature.com/articles/s44...

1329

Yes! I’m definitely standing on their shoulders. I believe I have something new to contribute to the conversation. I also list a number of “open” directions and conjectures. :)

010

As I do it, I plan to share what I find. I’ve created a blog for these projects called Applied Overthinking. If you’re interested, subscribe. appliedoverthinking.substack.com/p/on-the-hor...

My slice at the mathematics of onion chopping

On the Horizontal Cut

210

AI makes it increasingly possible to take one of those curiosities and quickly build the model, dataset, simulation, or experiment needed to investigate it. I’m hoping to start doing more of that: throwing AI at these questions and seeing what I can learn.

100

Beyond the onion result, what I’m more excited about is the process. There are countless questions that are too trivial for a research project and too specific to Google.

100

I also found that for a typical onion, a second horizontal cut isn’t worth doing. It also feels a bit more dangerous, in my experience.

100

A horizontal cut halfway up the onion, by contrast, is usually much worse. Aquino’s article analyzes cuts halfway up the onion.

100

I found that if you’re going to make the horizontal cut, make it low, roughly 20% of the way up the onion. At that height, the cut attacks the few giant pieces that vertical cuts tend to leave behind. In the simulations, it can dramatically reduce the range of piece sizes.

Intuition showing that not doing a horizontal cut leaves giant pieces of onion.

100

With a lot of help from OpenAI's Codex, I modeled an onion as concentric layers, simulated different cutting strategies, and measured the area of the resulting pieces.

100

J. Kenji López-Alt, who has also written on this topic, supports the horizontal onion cut, but has not really resolved why it might help. So I did the natural thing and built an onion-chopping simulation.

100

There’s a surprisingly rich internet literature on this question. In an article in @pudding.cool, @aqandrew.com and his collaborators found that horizontally cutting an onion half, parallel to the cutting board, actually makes pieces less uniform.

100

University of Washington Information School @ischool.uw.edu · 07/08/2026

As a researcher, every day I think about questions I don’t have the answer to. It used to be that most of them just weren’t worth answering because they’d take too much time and resources. I think AI is changing that. Here’s one example: what is the best way to cut an onion?

Diagram showing the model of an onion as concentric layers, and showing intuition for horizontally cutting the half-onion at about 20% of its height.

150

Reposted by Benjamin Laufer

Incoming Assistant Professor Ben Laufer recently published a new paper in PNAS. Laufer, Jon Kleinberg & Hoda Heidari explore incentive effects of AI safety regulation using a mathematical model. (1/2) Check out the study: www.pnas.org/doi/10.1073/... @pnas.org @laufer.bsky.social

"The backfiring effect of weak AI safety regulation." Screenshot of article published by the Proceedings of the National Academy of Sciences (PNAS), a peer reviewed journal of the National Academy of Sciences (NAS). July 20, 2026. Full text available at https://www.pnas.org/doi/10.1073/pnas.2509768123.

161

Reposted by Benjamin Laufer

University of Washington Information School @ischool.uw.edu · 05/08/2026

We look forward to welcoming you to our faculty!

081

I'm extremely grateful to my mentors, friends, and collaborators who have supported me along the way, especially my advisors Jon Kleinberg and Helen Nissenbaum.

020

Before moving to Seattle in 2027, I’ll be spending some time in San Francisco. Please reach out if you’re around!

120

I'm looking for people who are excited by ambitious interdisciplinary work that connects technical systems to their social and institutional context. Please reach out or list me in your application if this sounds like work you might want to pursue.

- How regulation and market designs can influence the incentives of developers and users - How we should evaluate AI systems for fairness and safety.

120

I'm recruiting PhD students to begin in Fall 2027. I would be especially excited to work with students interested in questions such as: - How model attributes and risks propagate through AI supply chains

My research studies AI ecosystems: networks of models, developers, platforms, and users that collectively shape AI and its effects on society. My group will use empirical, theoretical, and normative approaches to understand these systems, and to develop better ways of evaluating and governing them.

I am delighted to share that I'll be joining the University of Washington Information School as an Assistant Professor!

2442

This was my first op-ed, and I’m grateful to @maxuf.bsky.social and Fast Company for the opportunity.

010

I also discuss a more encouraging possibility: well-designed requirements across both layers can help companies coordinate and make stronger safety investments.

100

AI regulation can create similar shifts in responsibility. Drawing on my recent research with Jon Kleinberg and Hoda Heidari, I discuss how a requirement placed only on downstream applications can sometimes lead upstream model developers to invest less in safety.

100

Imagine requiring restaurants to inspect every ingredient they serve while placing no obligations on large food suppliers. Restaurants should be responsible for what reaches the plate, but such a rule might also give suppliers reason to relax their own quality controls.

100

That framing treats AI regulation as if it occurs in a vacuum. It leaves out another important question: When policymakers place a safety obligation on one company, how does that change the behavior of other companies in the AI development chain?

100

These developments are often framed as part of a debate over whether safety requirements will slow innovation or place companies and countries at a competitive disadvantage. I argue that debate is incomplete.

100

In June, a U.S. government directive led Anthropic to temporarily suspend access to its newest models while it worked to comply with restrictions involving foreign nationals. More recently, the Chinese company Moonshot AI released Kimi K3 and published its full model weights.

100