As the cost of the marginal robustness check approaches zero, what will happen to the number of robustness checks per paper, and the length of the annexes? What will be the new binding constraint in scientific communication?
One aspect worth considering, and hinted at but not discussed explicitly in the article (as far as I could tell), is the value of information. I discuss this briefly in the section "Move fast, break (little) things" in this blog post: www.fernandomartel.com/posts/politi...
What do academic researchers and traffic cops have in common? They all work at intersections.
π€·
Gated
Will AI agents replace software? My experience building an options-trading record-keeping workflow suggests otherwise.
AI is useful for predictive tasks. But when a workflow must be exact, stateful, and auditable, software still needs to stay in charge.
I recently came across a study that claimed llms ignore llm.text. Nothing reads them apparently.
I guess AJPS finally qualifies as a scientific journal.
Previously AJPS had a formal editorial policy of not even considering for review unsolicited (!?) rejoinders.
Not a serious journal.
I personally am not a big fan of notebooks but just using AI and reams of md files is not ideal either. Listen to this podcast creators.spotify.com/pod/profile/... Or just search YouTube for any Bret Victor talk.
Great podcast on the potential and limitations of LLMs from an expert in the field:
"Vibe Coding is a Slot Machine" - Jeremy Howard
Homeschooled kids can be done with school in 1-2 hours and spend the rest of the day playing with friends, specially if they are similarly home schooled.
I home schooled my son during COVID on USA. The most rewarding experience of the time. Sailors often home school their kids. Learnings:
* In a small group you can cover most of the school material in 1-2 hours per day, move beyond grade, and get A grades.
* school is 80% daycare, 20% learning
"Last week, we estimated a 66% chance that the US would strike Iran before April, well above the Polymarket aggregate of 20.5%" - open.substack.com/pub/sentinel...
One concern with AI is that AI will do the work, cheating, etc...
Another concern, perhaps more intriguing, is what students should learn nowadays.
My sense is more abstractions (the why and the what) and less implementation details (the how). More math, dags, proofs, less code.
Simple beats complex.
Maybe not a universal principle but a good point of departure....
Love the syllabus. Two things caught my attention.
1. Would love to see a blog post expanding the section on requirements and policies in age of AI. You are in a unique position to witness and shape the impact of AI on education.
2. No Paul Rosenbaum (e.g design of observational studies)?
I wrote a new post on automated policy evaluation: using AI to scale policy impact analysis with observational data.
This could speed up knowledge production, but whether it improves welfare depends on the institutions that adopt policy.
Market commentary isn't usually my thing but Microsoft's 15% stock price decline after beating earnings last month raises deeper questions about capital allocation within firms.
LLMs created a weird shift for me. Iβm more productive than ever, but that makes free time feel less relaxing, not more. The more I can do, the harder it is to step away.
How much of that work do you think can be abstracted, standardized (e.g. checklists), and automated? Peer reviews seem to combine haphazardness with labor intensity. Surely there is room for an orders of magnitude better process.
Built FermiApp and open source tool for back-of-the-envelope strategic analysis using Fermi estimation.
In my blog post I explain how working with AI as engineering partner (not autonomous agent) helped me move 10x faster while staying in control: www.futureofscience.ai/post/2025/11...
I find this very counterintuitive. IMHO longer letters send a worse signal than shorter letters. Shorter is a lot harder than longer, while still a packing a punch. Why are folks using AI to write longer letters?
Such a metric of success may introduce a conservative incentive, no? E.g. where are you more likely to deviate all else equal: Doing a field experiment in a war torn country, or with grad student at Yale? Which would you choose if you wanted to minimize deviations?
Good to see AI startups tackling judgemental forecasting, a seriously underused technique in business and government alike:
Here is one example onlinelibrary.wiley.com/doi/abs/10.1...
My motivation comes for something I wrote a while back. See the section entitled "Move fast, break (little) things" in this blog post www.fernandomartel.com/post/politic...
Value of Information for deciding what initiatives to experiment on, and inform experiment design?
TIL the auto save feature only works if you have a OneDrive account and the file is in a OneDrive folder.
"Growth teams" need to drive M365 subscriptions!
Even If you have a subscription, if you are working on an Excel file shared by a client in a Dropbox folder, Auto save won't work. π€¦ββοΈ
A big confounder here is age. E.g. the population of Barcelona is much older than that of LA. Would be good to plot death rates by age cohort across cities .
If Schopenhauer is right that life boils down to a flight from either boredom or pain, then the real breakthrough in AI will come the day machines get bored or suffer pain.
Nothing will drive their agency like boredom and pain....
AI providers may want to change their demos: Nobody is using LLMs to plan trips and book flights.
