⚔️
When you make a selection it cannot be changed
i.i.d.
IID
96 votes5 hours left
2
1
Follow
Click to Follow shortstein
Thomas Steinke
@shortstein
Computer scientist interested in (differential) privacy & related topics, e.g., generalization. 
Opinions are mine ©. 
Opinions are mine ©. Thomas Steinke’s Tweets
When *writing* a paper, do you primarily look at the TeX source code or the compiled PDF?
E.g., if you are making a pass over what your coauthor wrote, do you read the PDF first or go straight to the source?
When you make a selection it cannot be changed
Source
PDF
254 votes2 hours left
1
1
1
Math papers can be incredibly terse, but then they have random filler text like "We now state the next statement." or "This completes our proof. Q.E.D. □"
5
2
41
As usual, the reason academics do something weird is because of some reason that stopped being relevant decades ago:
Quote Tweet
Replying to @ysbhalgat and @adam_golinski
The ICML proceedings (we used to get a bound book when we showed up to the meeting!) has "global" page numbers (like pp321–328). The paper template had no numbers so they could all be added consistently later. Old proceedings on ACM have these numbers, but they've been forgotten.
1
5
Show this thread
When *writing* a paper, do you primarily look at the TeX source code or the compiled PDF?
E.g., if you are making a pass over what your coauthor wrote, do you read the PDF first or go straight to the source?
When you make a selection it cannot be changed
Source
PDF
254 votes2 hours left
1
1
1
FYI, you can add page numbers to the ICML template with
\icmltitlerunning{[title]\hfill\thepage}
6
16
114
Show this thread
A formally stronger looking version of this is a corollary of a Markov argument used in studying uniform integrability. It says that if X is integrable then
lim_{x -> ∞} E[X* I[X≥x]]= 0.
Quote Tweet
Here's a slight strengthening of Markov's inequality:
Let X be a random variable with a well-defined finite expectation.
Then
lim_{x -> ∞} x P[X≥x] = 0.
Markov's inequality would imply
limsup_{x -> ∞} x P[X≥x] ≤ E[X].
Show this thread
2
1
6
Show this thread
Here's a slight strengthening of Markov's inequality:
Let X be a random variable with a well-defined finite expectation.
Then
lim_{x -> ∞} x P[X≥x] = 0.
Markov's inequality would imply
limsup_{x -> ∞} x P[X≥x] ≤ E[X].
1
6
Show this thread
Markov's inequality is tight in the sense that for all x>0 there exists a random X≥0 such that P[X≥x]=E[X]/x.
However, it's not tight simultaneously for all x. This has always bothered me, as it means integrating Markov's inequality usually gives you the "wrong" answer. E.g.:
![For any random $X$ supported on the interval $[0,c]$, we have
\begin{align*}
\ex{}{X} &= \int_0^c \pr{}{X \ge x} \mathrm{d} x\\
&\le \int_0^c \min\left\{ 1 , \frac{\ex{}{X}}{x} \right\} \mathrm{d}x \tag{Markov's inequality} \\
&= \int_0^{\ex{}{X}} 1 \mathrm{d}x + \int_{\ex{}{X}}^c \frac{\ex{}{X}}{x} \mathrm{d}x\\
&= \ex{}{X} \cdot \left( 1 + \log(c) - \log\left(\ex{}{X}\right) \right).
\end{align*}](https://pbs.twimg.com/media/Fmwc5f7aYAA5944?format=png&name=large)
read image description
ALT
5
4
95
Show this thread
Let’s have a fun poll. Imagine you found a paper on arXiv and quickly skimmed through it. No obvious red flags found. What’s your prior that paper results are true?
- <20%7.4%
- 20-50%28.3%
- 50-80%43.5%
- >80%20.8%
660 votesFinal results
9
5
8
Show this thread
I went to Starbucks today and the barista noticed my hoodie and started asking me about ChatGPT. 🤔
12
8
339
I hear Googol’s LLM has 10^100 parameters. Beat that.
6
3
59
Please write a short (1-2 sentence) statement about why you think peer review is important to the advancement of science.
Wrong answers only.
Quote Tweet
Replying to @peter_richtarik and @icmlconf
Why are they asking this? Here’s what I wrote, which is very cheesy.
read image description
ALT
7
1
9
I'm more concerned about the possibility of using AI tools for writing reviews. If AI-assisted reviewing becomes commonplace, it means we've given up on writing papers that humans actually want to read.
Quote Tweet
And so it begins ...
Researchers created a peer-review dataset & review comment generation model: arxiv.org/abs/2212.04972
"[...] if you thought Human Reviewer 2 was hard to reason with, just wait until reviewer 2 is a language model!" -- via @jackclarkSF
1
7
Show this thread
Of course, AI tools can also be used to help write high-quality submissions. So I understand why people oppose banning them. But I also understand why people are worried that, on balance, AI tools will do more harm than good.
Quote Tweet
You are reviewing 35 submissions for #NeurIPS2025 Papers can now be written in a day or two thanks to AI writing tools. Even high schoolers are submitting papers written with the confidence of seasoned researchers. You decide to use AI to “draft” your reviews.
Show this thread
2
4
Show this thread
I don't see how using AI tools like ChatGPT to help write papers is inherently unethical. However, it makes it easier to spam conferences with low-quality submissions, which consume time & effort of reviewers.
Quote Tweet
ICML ethics guideline (see below) is shortsighted imo. ChatGPT and variants are part of the future. Banning is definitely not the answer.
"Papers that include text generated from LLM s.a. ChatGPT are prohibited unless produced text is part of the paper’s experimental analysis." twitter.com/icmlconf/statu…
4
1
16
Show this thread
Reminder: write all your preliminary math definitions in one big section at the beginning of the paper instead of close to where they are used.
Readers love memorizing definitions, and having to move back and forth in the paper helps them with the flow.
13
16
221
And so it begins ...
Researchers created a peer-review dataset & review comment generation model: arxiv.org/abs/2212.04972
"[...] if you thought Human Reviewer 2 was hard to reason with, just wait until reviewer 2 is a language model!" -- via
9
52
268
I'm not the first to say this, but I'll repeat: an interesting thing about differential privacy is it needs theorems even in practice. You can implement heuristic algs that are fast on your data, but 'heuristic privacy' doesn't exist. A mechanism isn't private without a theorem.
5
36
357
Show this thread
This seems like a battle Elon will lose. The value of Twitter is that you can find everything here. Trying to cut it off from other networks only decreases the value of Twitter.
1
12
Show this thread
All links to other social media are now banned (even if you spell out "dot" and "at" so that it doesn't parse as a link).
How long until arxiv is banned?
help.twitter.com/en/rules-and-p
1
1
5
Show this thread
All links to other social media are now banned (even if you spell out "dot" and "at" so that it doesn't parse as a link).
How long until arxiv is banned?
help.twitter.com/en/rules-and-p
1
1
5
Show this thread
Running a social media platform requires difficult choices around content moderation. We should cut Elon some slack, just like he has been respectful towards the decisions made by the prior management. ...oh, wait.
1
4
Show this thread
...hilarious. I'm guessing the filter only applies to things that get detected as links.
2
6
Show this thread
Let's see if this works: @steinke@mastodon.social (It might not be recognized as a link.)
2
13
Show this thread
I think Elon is about to learn a hard lesson in how free speech actually works, starting with the Streisand effect.
3
16
Show this thread
Twitter is now blocking *all* links to mastodon. I got this error message when attempting to tweet a link to my mastodon profile.
Elon has every right to do this on his personal website, but he cannot claim to be promoting free speech.
3
9
124
Show this thread
Amazing levels of commitment to free speech.
67
544
2,022
Show this thread
You are reviewing 35 submissions for #NeurIPS2025 Papers can now be written in a day or two thanks to AI writing tools. Even high schoolers are submitting papers written with the confidence of seasoned researchers. You decide to use AI to “draft” your reviews.
1
4
23
Show this thread
You are reviewing 35 submissions for #NeurIPS2025 Papers can now be written in a day or two thanks to AI writing tools. Even high schoolers are submitting papers written with the confidence of seasoned researchers. You decide to use AI to “draft” your reviews.
1
4
23
Show this thread
The year is 2025 and you have been on the phone for hours. You suspect that all of the customer service representatives you have spoken with are actually AI chatbots designed to frustrate you into giving up on getting a refund. Your suspicions are correct.
5
7
82
Show this thread
Replying to
The problem permits exact computation of the minimax risk (including constant). The argument proceeds with the prior Beta(a,b), a=b=sqrt(n)/2 (instead of Uniform). It appears in Hodges & Lehmann 1950 ("Some Problems in Minimax Point Estimation") who attribute it to Herman Rubin.
1
3
21
Oops: the 2 from k^2 got cut off in the margin.
3
Show this thread
It's a well-known fact™ that estimating the mean of a Bernoulli distribution from n independent samples must have mean squared error Ω(1/n).
But it's surprisingly difficult to find a clean proof or citation for this fact. Here's my attempt. 😅 Is there something cleaner/better?
read image description
ALT
16
51
379
Show this thread
Seriously, we should stop treating publishing and peer review as synonymous.
This made sense back when peer review was a gate keeper for print publishing, but the internet has disrupted publishing. You can now publish without peer review and vice versa.
Quote Tweet
A paper that is only on arxiv has been published but not peer reviewed.
A paper that is only behind the paywall of a conference/journal has been peer reviewed but not published.
3
1
30
It would be much more efficient if my kids got sick in parallel, rather than sequentially.
3
45
If professional societies like ACM/IEEE/SIAM didn't already exist, would anyone propose that we create them?
1
2
19
A paper that is only on arxiv has been published but not peer reviewed.
A paper that is only behind the paywall of a conference/journal has been peer reviewed but not published.
9
20
228