Economics postdoc researching info frictions + ext validity in technology adoption at @HarvardHBS. PhD @NorthwesternU. All opinions are wrong, some are useful.
Joined November 2008
- Tweets806
- Following177
- Followers200
- Likes1.3K
Related to this point nitter.gq/evavivalt/status/21216…
Maybe in the future alongside a paper submission I'll include a repo with an AI-written FAQ that points out why some things flagged by LLMs are not actually issues, so the reviewer's LLM has that.
Like an AI online appendix that dispenses with the false positives while keeping the draft clean and human-readable.
1
25
It would be nice to, alongside a paper, submit a Refine report and just say "here is me addressing each technical point. let's now converge to referees judging responses and otherwise focusing on non-technical issues."
1
2
25
Running the same paper for review via @RefineDotInk vs Claude only makes the AI referee issue more annoying. Refine doesn't hallucinate non-issues into R&R. Perhaps instead of journals using @RefineDotInk at end of process, use it before sent to referees? CC @ben_golub
1
1
33
I was thinking the same thing recently after having an llm review paper pre-submission. I wish there was a formal way to submit this.
2
244
That seems a bit too aggressive to attribute upside to him, no? Like if he simply retired, the counterfactual CEO in that scenario would probably be able to tame it much more effectively than him having an affair with a spy and putting NVIDIA in crosshairs.
159
10% is lower than we think as a cap. Treasury rates are low not only bc lack credit risk but also liquidity and other premia. The more relevant benchmark for opportunity cost of investment is maybe zero beta rate of @KurlatPablo et al. Frequently >10%! tinyurl.com/dt3ayujy
Let’s cap credit card rates at 10%. Take action on affordable housing. Protect rural hospitals. End price gouging.
Democrats have the bills ready.
@SenWarren says now it’s up to Trump to work with us and get things done.
1
1
25
(Happy to share examples via DM eg from process leading to arxiv.org/pdf/2611.01226 if of interest)
1
1
30
Yup. 100% unless I add edits outside AI flow, because all the writing is done by the AI, no matter how much you're iterating with it on ideation or at what granularity. Pangram partly picks up ticks, but it also picks up consistency; iterating with AI maintains consistency.
1
1
31
I do this since I have joint issues and dictating a conversation with AI is much more efficient. But I suspect ESL researchers have even stronger incentive to adopt a similar workflow where they work with AI to carefully tune writing details, but it will still be flagged AI.
1
1
38
I both strongly agree with you but also think this isn't incompatible with AI writing. I often work with AI to do sentence-level edits to refine / clarify a point. But if I put that paragraph through pangram, I suspect it will still score it 100% AI.
1
2
94
One theory says this means a collapse of the model. But I think the alternate take is that journal editors/conf chairs now have a much more important role in science and we need to design academia more around these positions.
151
Perhaps the most annoying thing about this in the AI era is that, given they're running every specification anyways, report the full distribution! Knowing the distribution of results is useful & eg anyone with minimax preferences can decide for themselves to value the worst case
1
1
217
(Mind you, lots of good reasons to pay teachers and nurses more! Just not sure about the same reasoning as PhDs is the best approach.)
15
I think of PhD student funding as early stage VC funding/incubators. You are paying for output (research) but they're early stage so they spend a lot of time on training (incubation) so you pay everyone a very small initial round (seed funding) in hopes of a few unicorns.
1
79
They could have picked the 'we prefer saving shrimp than humans' stuff but instead they went with the most reasonable of EA takes.
14
618
I think this is because empirics involves more tool calls. Though also I find it speeds me up in theory more than empirics. With code, I have to carefully monitor implementation as it obfuscates making weird empirical decisions. In theory, the chat is pretty self-contained.
1
4
174
We need just a equivalent tiered set of journals publishing descriptive stats. That way referees to those journals can focus on quality of the descriptive evidence, and whether worthy of the tier, instead of whether they care about the descriptive work.
1
35