Matt Blackwell
@matt_blackwell
Followers
16K
Following
6K
Media
1K
Statuses
10K
data, causal inference, experiments, politics
Cambridge, MA
Joined January 2009
Reminder that I have a textbook on an introduction to mathematical statistics and regression, designed for first year PhD students in poli sci, but maybe useful to others. Let me know if you use it and have feedback! .
7
140
710
I think some introspection is needed by the industry about their own unrealistic narratives about AI. These are useful tools, of course, but when people constantly hear that we’re 3-6 months from some job being completely automated, we’re getting into boy cried wolf territory.
For some reason, a lot of journalists and academics have decided that AI is useless because it sometimes gives wrong answers. Fine if they don’t want to use such useful tools, but I feel sorry for “civilians” who take such claims at face value and get put off using AI too.
2
1
7
Nothing has made my want to quit social media more recently than seeing someone that has tweet >100k times in 5 years. That’s 55x a day on average….
3
1
10
It’s impressive that it can find it in 8 out of 100 iterations of the same setup, but the stochastic nature of the outcome renders it only useful to someone who can and is willing to verify the answer.
o3 for finding a security vulnerability in the Linux kernel:.
1
0
2
The final mission impossible is good but it’s hard to take a movie seriously when it uses the word “cyberspace” multiple times in dramatic scenes.
1
0
0
A Republic, if you can keep it.
At this point the TROs are basically an automatic judicial veto on all new policy, and then the courts of appeals decide whether the judicial veto should be suspended. Whatever form of government that is, let’s please not call it “democracy.”.
0
0
4
bouba / kiki.
1
0
1
It’s interesting to compare how people react to sending the same image through a diffusion model 1000x to how they react to two AI instances “talk” to each other. I think people see images as natural stochastic drift of the model where they see something more purposeful in text.
When Claude instances talk to each other, in ~90% of open-ended interactions they spiral into discussions of consciousness, then profuse gratitude, then abstract spiritual/poetic expressions with Sanskrit and emojis.
0
0
5
(whispering into my phone during the job talk) grok is that true?.
0
0
10
Sounds encouraging until you see the 5 books.
you’re 5 math books away from redirecting your entire life trajectory btw.
1
0
9
No it wouldn’t be cheating in the same way as using a simple calculator wouldn’t be cheating for a calculus exam. But using Mathematica probably would be cheating.
Microsoft Word first had a spell check feature in 1985. It had a grammar checker in 1997. Would you consider these to be "AI tools"? Would you consider it cheating if a student used these features to improve an essay that they wrote?.
2
0
2
This kind of thing is great, but you have to wonder how much of it depends on the reams of lecture notes out there explaining these older papers. I worry about the future incentives to produce that kind of important training (for humans and LLMs) material.
5
2
24
Thinking about saying “orthonormal” instead of “orthogonal” in normal conservation as a bit.
1
0
4
getting asked to explain the 5 diff-in-diff estimators you used in your job talk
0
1
26
The death of the URL (“gibberish”) is tragic imo.
A modern URL bar (in @diabrowser):. • Page Title not "/2025/12/seo-spam" gibberish .• Space on both sides of "/" for readability .• Hover to reveal & edit URL.• Emphasize domain for trust+security . Dia isn't just AI. It's refined browser basics too, @browsercompany style.
2
0
5
Mexico, truly the heart of Asia.
There are very strong regional differences in attitudes toward AI. Asian countries are generally quite positive about AI while Western countries are more nervous. @Ipsos
0
0
14
I think a lot of “gifted kid burnout” is just “burnout” combined with a poor theory of mind for kids that weren’t labeled as “gifted”.
I wonder if “gifted kid burnout” is just: You learn to navigate by the reward signal of achievement (easy, ppl tell you how to get it) so you never develop the more subtle ability to navigate by what interests you. Then when achievement gets sparse, you feel aimless & unmotivated.
0
0
5
Can polling be inaccurate? Of course, but making assumptions about how to map aggregate vote to individual characteristics is also fraught.
1
1
18
Pre-election polling on this shows richer people were less supportive of the minimum wage expansion than poorer people, which is the opposite conclusion you’d draw from this county-level analysis. Solving the ecological inference problem is hard!
In 2024, rich Californians voted to raise the minimum wage. Poor Californians voted against it. A wild chart below showing people voting against their own class interests.
3
20
128