Normally, it's:
1) write a paper & submit
3) get reviews (~3 months)
4) revise paper & resubmit
5) wait for response (~3 months)
...what if we could simulate this process in minutes?

Could we fix issues?
Anticipate misconceptions?
Get ideas for new analyses/experiments?
1/ To do this, I created a Python notebook (link at the end) that simulates the process. It lets you upload a PDF & simulate steps 3, 4 + 5 (and so can you, with your own paper):

I created referee personas: 2/

Gave a talk today to @cepr_org thanks invite from @akorinek ; great discussion by @Afinetheorem here are my slides in PNG form b/c of the link this 1/N

The "there should be a rule/law/training/office" impetus after every bad event is both a) understandable and b) how you get the kinds of organizations most people hate to be in For the specific thing I'm sub-tweeting - who exactly do ppl think would provide the oversight that would prevent this? Almost all the peculiarities of research orgs are a reaction to how difficulty the assessment of research is in practice; some putative office of research non-fakery staffed w/ non-experts faces the same problem but w/o a chance of solving it

Posters I thought were interesting

I wanted to do a thread about how I've been thinking about Gen AI & production. So, imagine a job that's a sequence of tasks to be done 1/n One of these tasks *might* be doable with AI, so we give a shot by asking nicely:

.@sama laying out an optimistic and compelling vision for AI at @MIT

https://twitter.com/johnjhorton/status/1786104916521468192

A few paraphrased thoughts: 1) views the acceleration of scientific discovery as highest potential pay-off from AI

I'll do a longer thread---and this little toy example just scratches the surface of what's possible---but edsl (pip install edsl) is an open source python package & domain specific language for asking LLMs questions

https://twitter.com/expectedparrot/status/1780935480386695672

You can ask questions of various types (free text, multiple choice, numerical), combine them into surveys, add skip logic, parameterize w/ data (to create data labeling coding flows) & then administer to or model models. It abstracts away from the model-specific details

New working paper by @BenSManning & @Kehang_Zhu that & I’m super excited about benjaminmanning.io/files/rs.pdf
What we do, in a nutshell, is let the LLM propose hypotheses, design experiments, RUN those experiments insulation and then have it estimate the results

Just wanted to give a personal update - despite my prior skepticism, I'm now bullish on blockchain as a solution to a very real problem we all now have no idea how it's supposed to work or it if will, but the problem of deciding what's real or not will become acute

So I'm working on a little system to create a personal academic webpage from a google sheet automatically. It takes a public google sheet as input (docs.google.com/spreadsheets/d…) and makes a GitHub page (johnjosephhorton.github.io) automatically Background: So I'm a bit prone to common nerd flaw (willing to spend lots of time with tech to save little bits of time in the future), but this might benefit others once it's actually built

I’d guess probably 10-15 major generals in total. If they’ve already killed 3 of them, it sure seems like Ukrainians are effectively targeting Russian commanders and this isn’t just luck

https://twitter.com/larisamlbrown/status/1502296821204500481

A caveat is I’m not sure if same ratio of troops to generals holds in Russian army. This is my guess using approximate US ratio.

This article describes a style of fighting that will be very hard for the Russians to effectively counter nytimes.com/2022/03/05/wor… Small, nimble groups of soldiers using light but highly lethal weapons (sniper rifles, modern anti tank missiles, armed drones, etc) with excellent intelligence would be a nightmare for that convey leader

hard to know exactly what's happening of course, but I *think* what we're seeing is that with modern anti-tank missiles like the javelin, tanks/APCs are toast w/o extensive infantry support. Just rolling down a highway in a column ---> 💀

https://twitter.com/johnjhorton/status/1486327562871746564

The latest javelin has a ~4KM range. It's hard to even see infantry at that distance, nevermind engage with them effectively. But they sure can see you and your hot tank! And the javelin has such a low signature, they're already running away before you even notice the rocket

A theory for while many people are misinformed about academia: Imagine you went to a pediatrician until you were 18, then at 19, you switched to an "adult doctor," who did the same stuff---gave you checkups, advised you on health issues, prescribed medicines etc. But if you asked, you'd learn that unlike pediatricians, adult doctors had this side hustle--making water color painings for other adult doctors. Many adult doctors viewed painting as their "real" work; that they got into adult doctoring to paint water colors.

Thread: Want to include links to presentations, videos, etc. but (a) want to be able to control what that URL points to in the future, (b) don't like the look of very long, cryptic URLs and (c) worried that URL shorters come & go (bye, goo.gl)? Here's a trick - create an HTML file with the name you want e.g., "<cool_project)_slides.html then, inside put a snippet of code for an immediate redirect: (nb: I'm missing a ">" in my screenshot - that's not part of the code)

A thread on R&Rs - so I did not appreciate in grad school how much time I'd spend revising papers for re-submission, nor did I learn much about how to do it efficiently. This is seemingly commonplace: I by no means have this nut cracked - in fact, I had one failed R&R in grad school (never re-submitted to Journal of Law & Economics despite what I now recognize wasn't a hard revision) & nearly de-railed another one though a poor revision practices.