The purest form of pride is when our children surpass us.
@eleeshimshoni.bsky.social

Great talk by @t8el.bsky.social at our econ theory seminar. As an added bonus, he publicly credited me for what he called two "Yannai tricks" in his slides:
1. Adding photos of his kids
2. Using memes
Happy to see my efforts of enriching econ with computer science techniques are bearing fruit!

Out of our dataset of n=4 economic researchers from a café at Stanford this morning, only one has an economics PhD, which raises the conjecture that despite common belief, an econ PhD might not be a necessary condition for conducting economics research.
@alroth.bsky.social @ranshorrer.bsky.social.

Too bad.

Things that pop to mind as I'm reading admissions files.

It's 2025, and AI is still nowhere near replacing or even slightly augmenting @skominers.bsky.social .

PS Shameless math geeking-out: Never thought I'd ever use concepts from number theory to prove stuff in econ. (We prove that all our our axioms, such as our cardinal version of IIA, are logically independent: Dropping any of them invalidates the uniqueness of our social inefficiency function.) 8.5/9

Random Serial Dictatorship (RSD) is a popular matching mechanism, inefficient in various senses. We use tools from computer science to prove that no truthful ordinal mechanism can improve on the robust (over all allocation problems) social inefficiency guarantee (at equilibrium) of RSD by >28%. 8/9

We apply our social inefficiency function to a setting where due to the absence of monetary transfers and outside options, and due to the repugnance of measuring with money, justifying interpersonal comparison or unique cardinal utility values is notoriously challenging: Object allocation. 7/9

One challenge is interpersonal utility comparison. Sure, in some settings, if we measure individuals' utilities in dollars and no-trade utility is 0, a statement like "the equilibrium welfare is 65% of the efficient welfare" is meaningful. But what about when we only have individuals' vNM prefs? 3/9

While many social welfare functions (e.g., Nash's) technically output numbers, only their 𝑜𝑟𝑑𝑒𝑟 has been axiomatized. In the words of Luce & Raiffa on indices that represent orders, "one must develop an almost inhuman self-control not to read into these numbers" any property except their order. 2/9

🚨New Paper🚨
𝐐𝐮𝐚𝐧𝐭𝐢𝐟𝐲𝐢𝐧𝐠 𝐈𝐧𝐞𝐟𝐟𝐢𝐜𝐢𝐞𝐧𝐜𝐲
Joint with Ella Segev

Social choice theory provides ways for aggregating the preferences of individuals in a society, to tell us what is better for society: alternative a or alternative b. However, this theory is silent on 𝑏𝑦 ℎ𝑜𝑤 𝑚𝑢𝑐ℎ a is better than b. 1/9

GPT-4o doesn't even come close!

What we're doing is much more of a lab experiment than agent-based modeling. We don't use LLMs to simulate other entities, but rather study the behavior of LLMs, anticipating their deployment in the field.

Results:
LLM-based pricing agents autonomously collude:
1. quickly (~100 rounds)
2. even when not instructed to
3. certain terms and phrases lead to ~monopoly profits
4. also in auction settings

New version (novel off-path analysis):
Price-war concerns by LLMs are a significant contributor.

2/3

🚨Major new version🚨
Algorithmic Collusion by Large Language Models
Joint w/ @sarafish.bsky.social & @ranshorrer.bsky.social

LLMs are automating many business decisions. Pricing might be next (or is already).
What if multiple firms, in good faith, to use off-the-shelf-LLMs for pricing? 1/3
#EconSky

It is true that adding arbitrarily small timing frictions of some forms to a model precludes the dynamic formation of new common knowledge (i.e., anything that's not common knowledge at time 0 is never common knowledge). This is known as the Halpern–Mosed Paradox, and... 1/2

Communication Complexity for #GenZies (this is literally my explanation in class last week):

Say that two students have a dedicated WhatsApp group, in which they can only send two kinds of emojis: smileys (🙂) and frownies (🙁). They can strategize ahead of time, then each of them gets 1/2