Anchi Xia

love is a doing word

[Anchi Xia @bryantxia22]

RT @pmbruera: Excited to share our paper with @ojblanchard1 and @GonHuertas on sovereign debt and multiple equilibria!

https://t.co/7k6AkM…

[Anchi Xia @bryantxia22]

@aledinola RIP Aiyagari. Gone too soon. The working paper version of his classic '94 paper is a thing of beauty (https://t.co/0oyZsa0KN1). The published version at QJE did not do this masterpiece justice...

[Anchi Xia @bryantxia22]

@juanescolar captain 🫡

[Anchi Xia @bryantxia22]

next to flo balogun and tyrese maxey, this list has to be a joke...

[Adrien Bilal @AdrienBilal]

Truly honored and grateful to be included in the #TIME100Next list alongside so many inspiring people. Thank you @TIME. This recognition belongs just as much to my amazing coauthor @drkaenzig. The work continues!

[Josh Hendrickson @RebelEconProf]

Why would I take part in this charade?

[Tom Holden @t_holden]

In endogenous growth models with both research and entry margins, restricting entry actually accelerates growth (by cutting research duplication), and leads to fat profits for incumbents.

If Dario really wanted to slow AI development, then he would be encouraging entry (e.g. Chinese models). AI prices get driven to near marginal serving costs, research incentives evaporate, and OpenAI and Anthropic probably declare bankruptcy.

[Anchi Xia @bryantxia22]

@akorinek @ChadJonesEcon @SzymonSacher @PeterMcCrory It's a cool model with A LOT of ingredients, but how well does it perform? Does all the richness allow it match historical episodes of technological change given the right inputs?

[Anchi Xia @bryantxia22]

@BasilHalperin If AI were to trigger a wave of sovereign debt crises, where is capital going? The market is questioning the safe asset status of US treasuries, but with no consensus for an alternative. Would a portion of the funds paradoxically flow back into financing AI somehow?

[Anchi Xia @bryantxia22]

@Acyn https://t.co/xR8UR9eJhO 😨

[Anchi Xia @bryantxia22]

but credit constraints could also bite the other way. suppose chinese labs wish to undercut but have limited capacity, then making models open weight is a way of recruiting others to do its bidding. though characterizing when this is optimal will be trickier. will dig into this further...

[Anchi Xia @bryantxia22]

Kimi K3 was recently released, with benchmark results rivaling those of Claude Fable and GPT-5.6. Yet the model is open-weight—something neither Anthropic nor OpenAI has done with a flagship model.

Unless Moonshot AI and others are simply giving away the store, what gives? Can a profit-maximizing lab ever find it optimal to release its model weights? Jan Morgan (@JanTMorgan) and I argue that the answer is yes. A thread. [1/10]

[Anchi Xia @bryantxia22]

@andrewjkoh @arena thanks Andrew! agreed credits constraints may be a big deal. it could be that chinese labs know they can count on the gov't for funding, so releasing weights hurts rival labs' capacity to raise but has limited backlash on themselves.

[Anchi Xia @bryantxia22]

This thread came out of a lunch time discussion with @JanTMorgan. We’re posting since we haven’t seen this particular angle on Econ Twitter. Tagging a few people who may find this interesting: @andrewjkoh, @pmbruera, @alexolegimas, @pawtrammell, @ChadJonesEcon, @dwarkesh_sp, @f_paternollo, @gaoooric. We have benefited from reading many thoughtful AI related discussions on Econ twitter and from conversations with @Jean_Tirole. Hope to continue the discourse! [10/10]

[Anchi Xia @bryantxia22]

Finally, what does all this mean for welfare? On the one hand, open-weight models reduce incentives to invest in frontier research. On the other, they drive down LLM prices by increasing competition. Which force dominates in equilibrium? A lagging lab’s incentive may be misaligned from the societal perspective, creating a welfare externality. Whether that externality is positive or negative depends on how consumers value improvements in inference quality relative to lower prices—and, more provocatively, on whether slower frontier progress carries safety benefits. [9/10]