Gregor Schubert

"Family firms" talks about various non-monetary benefits that owners can derive from firms - which presumably mean they don't optimize prices if those conflict with their other objectives. I guess Holmstrom multi-tasking + multiple objectives imply similar things?

Vibe-coding’s complement is unit testing. If the writing of code is commodified, the validation of output becomes scarce.

Randomized trial AI for legal work finds Reasoning models are a big deal:

Law students using o1-preview had the quality of work on most tasks increase (up to 28%) & time savings of 12-28%

There were a few hallucinations, but a RAG-based AI with access to legal material reduced those to human level

These findings have policy implications - groups that have a greater affinity for homeownership will be more likely to be benefited / harmed by credit or housing policy changes. Heterogeneity in take-up can result from group affinity!

Link to paper is here: papers.ssrn.com/sol3/papers....

#Econsky

We find non-RE wealth increases of HHs with high HO affinity limited to those in the right housing markets at the right time & who happen to see high price increases, NOT a general consequence of homeownership - so housing policy is limited if housing booms are not guaranteed!

We obtain restricted HRS data to see if affinity for HO impacts the portfolios of foreign-born retirees: as expected, they are more likely to own a home & own more RE in their portfolios. Total non-RE retirement wealth is also higher for those who have high HO affinity in origin country! But why?...

We show that AFFINITY matters for housing cycles and effects of credit supply shocks. High HO affinity households enter more into homeownership during 2000s housing boom, default less during the GFC - see paper for causal evidence on greater response to credit supply shocks.

It's hard to find exogenous changes in homeownership (HO) to study effects on HH finance. We build on literature looking at role of experiences/origins driving financial choices and show that HO in origin countries (HOCO) drives HO of foreign-born in the US! (15% passthrough)

🚨 New working paper with Caitlin Gorback!

We ask what happens when households are more likely to WANT to own a home for cultural reasons? We find homeownership increases, they're more responsive to credit supply shocks, and more of their retirement portfolios are in real estate. 🧵

Managers and leaders will need to update their beliefs for what Al can do, and how well it can do it, given these new Al models. Rather than assuming they can only do low-level work, we will need to consider the ways in which Al can serve as a genuine intellectual partner. These models can now tackle complex analytical tasks, creative work, and even research-level problems with surprising sophistication. The examples I've shared - from creating interactive 3D visualizations of academic concepts to performing PhD-level analysis - demonstrate that we're moving beyond simple automation into the realm of Al-powered knowledge work. These systems are still far from flawless, nor do they beat human experts consistently across a wide range of tasks, but they are very impressive.

A new generation of AIs: Claude 3.7 and Grok 3 https://www.oneusefulthing.org/p/a-new-generation-of-ais-claude-37 #AI #Claude #Grok

Text Shot: I love this as an example of an expert using LLMs as tools to help further elevate their work. I'd love to read more examples like this one from experts in other fields.

A professional workflow for translation using LLMs - Simon Willison. https://simonwillison.net/2025/Feb/2/workflow-for-translation/

A Cross-Sectional Study of GPT-4–Based Plain Language Translation of Clinical Notes to Improve Patient Comprehension of Disease Course and Management The 21st Century Cures Act provides patients with access to their clinical notes, but most patients, particularly those with limited health literacy, have difficulty understanding and utilizing the...

🧪 A @ai.nejm.org study found that GPT-4 translation of discharge notes significantly improved patient understanding, especially for Black, Hispanic, and older patients. AI-driven readability could help close health literacy gaps.
🩺💻
tagging: @downing.bsky.social @thelizarmy.bsky.social

I am worried LLM researchers sometimes bury the lede with regard to "should we trust these systems". Framing below is: LLMs are failing to "earn human trust". But it turns out it's the humans who cannot be trusted - even seeing the LLM's answer, the humans do worse than the LLM!

I think one way of describing these results to our future LLM overlords is “inference time compute scaling leads to thinking more but not to thinking better - in humans”

Generative AI has flaws and biases, and there is a tendency for academics to fix on that (85% of equity LLM papers focus on harms)…

…yet in many ways LLMs are uniquely powerful among new technologies for helping people equitably in education and healthcare. We need an urgent focus on how to do that

To me this is one reason that good UX for LLM-based applications is important - users need to be guided by the designer to be able to quickly figure out "what is this good at" and "what is this not good at" - there is no time to validate all use cases for each chatbot encountered in the wild!

This thread was triggered by this great paper by @keyonv.bsky.social , Ashesh Rambachan, and @sendhil.bsky.social about how humans become overoptimistic about model capabilities after seeing performance on a small number of tasks.

That is, it is equally problematic when users (including researchers!) overestimate how trustworthy these models are on new tasks, as it is to not trust them ever.

It's necessary to constantly validate the quality of model output at different stages - "we asked the model to do X" is not enough!

Only after repeated use and exploration for where the weaknesses and pitfalls lie, and in which cases the LLM output can (with guardrails!) be trusted, does the user's expectation for LLM capabilities reach the "Plateau of Pragmatism".

BOTH the "Valley" and the "Mountain" are problematic places!

...but after some experimenting (@emollick.bsky.social suggests 10+ hours), many people find amazing abilities in some areas where LLMs exceed humans, reaching bedazzlement on "Magic Box Mountain".

However, their "jagged frontier" nature means that LLMs would fail on many other "easy" use cases.

Let me try to formalize some thoughts about Gen AI adoption that I have had, which I will call "The Bedazzlement Curve".

Most people still underestimate how useful Gen AI tools would be if they tried to find use cases and overestimate the issues - they're in "The Valley of Stochastic Parrots".

Reporting on AI adoption rates tends to show the importance of having priors.

In many examples of people actually implementing Gen. AI-based workflows, building the automation requires experience with the task at hand - suggesting that there might be upskilling / demand for experienced workers in those areas at least in the short-medium term, rather than simple “replacement”

The new Deep Research feature from Google feels like one of the most appropriately "Google-y" uses of AI to date, and is quite impressive.

I've had access for a bit and it does very good initial reports on almost any topic. The paywalls around academic sources puts some limits around it, though.

With regard to whether our firm-level Generative AI exposure measure predicts ACTUAL adoption - my forthcoming research shows it does!

See below - our exposure measure with 2022 data strongly predicts whether firms mention Gen AI skills in their job postings in 2024. 2/2

Was very surprised to stumble across a graph from my own research in a presentation by Benedict Evans today!

He makes the fair point that predicting technology effects is hard! Although I prefer to call our analysis "bottom-up" as it builds from microdata to a firm level exposure measure. 1/2

It's incredibly encouraging that even models for analytical purposes, like o1, can recite Shakespeare.

This means that there are still many "storage" parameters not fine-tuned for analytics and means that distillation can get large performance improvements at smaller model size.

I had to tell a Ph.D. student today what o1 is.

"The future is already here, it's just not evenly distributed"

The Gen AI Bridge to the Future Generative AI is the bridge to the next computing paradigm of wearables, just like the Internet bridged the gap from PCs to smartphones.

I found this Stratechery framework for how technologies evolve by the interplay of hardware, input modes, and applications quite thought-provoking - with innovations in one triggering innovations in the other dimensions.

Science seems to progress in similar ways.

stratechery.com/2024/the-gen...

I think firms worrying about AI hallucination should consider some questions:
1) How vital is 100% accuracy on a task?
2) How accurate is AI?
3) How accurate is the human who would do it?
4) How do you know 2 & 3?
5) How do you deal with the fact that humans are not 100%?
Not all tasks are the same.