jaybaxter

X | Machine Learning Engineer - Community Notes | SF, SJ, Seattle, LA, NYC | All Levels

The role: build scalable production machine learning systems and infrastructure, and improve our open source algorithm: https://github.com/twitter/communitynotes

Community Notes has been discussed here on HN a number of times e.g. here: https://news.ycombinator.com/item?id=37292041

Apply here: https://twitter.wd5.myworkdayjobs.com/X/job/San-Francisco-CA...

There are many other positions at the company too, see: https://careers.x.com/en

I wouldn't say they were being childishly subversive -- it is very unclear that paper bags are better than plastic bags for the environment.

"Manufacturing a paper bag takes about four times as much energy as it takes to produce a plastic bag"

"Studies have shown that, for a paper bag to neutralize its environmental impact compared to plastic, it would have to be used anywhere from three to 43 times."

https://education.nationalgeographic.org/resource/sustainabl...

It is very confusing to me how this got painted as pressure solely from the right. Ackman seemed like the driving force behind it, and he has mostly donated to democrats: "Since 1998, with few exceptions to select Republicans, Ackman has mainly donated to Democratic election candidates. Ackman has donated more than $600,000 to Democratic campaigns"

https://www.influencewatch.org/person/bill-ackman/#:~:text=2....

There is a type of advertising called "brand advertising" where the goal is more about brand awareness (e.g. think of a Coca-Cola superbowl ad) than "direct response advertising". Such advertisers don't really care if you click on ads, but they still want to reach high-value users.

The code is open source: https://github.com/twitter/communitynotes

The notes and rating data is released to the public every day: https://twitter.com/i/communitynotes/download-data

Feel free to run the matrix factorization code on the data, and then try to interpret the resulting latent dimension it finds for yourself! And also, you can read the code to verify that it really is running a matrix factorization rather than hardcoding a particular left/right split.

Indeed, it does this!

>For a note to achieve a high intercept term (which is the note’s helpfulness score), it must be rated helpful by raters with a diversity of viewpoints (factor embeddings)

https://communitynotes.twitter.com/guide/en/under-the-hood/r...

How come? Just long hours or something else?

It started rolling out to everyone in the US in early October, and first version launched at smaller scale back in January 2021

https://twitter.com/birdwatch/status/1578004575990202370?s=4...

Hi! Birdwatch ML Engineer here-- these context notes are from Birdwatch. They are written and rated by users, and the notes are only added as context to Tweets if they are rated highly enough by multiple raters who have tended to disagree in the past. The core algorithm is open source as well as all of the data, and there is lots of public documentation about it too:

https://twitter.github.io/birdwatch/

https://twitter.com/birdwatch/status/1585794012052611076?s=2...

https://arxiv.org/pdf/2210.15723.pdf

>but you’ve got to think about the idea that the folks who conducted this study know as well how hard science is, and for them to publish this anyway means they (and their reviewers) did a pretty good job dealing with this issue (and many others).

What makes you say that about this paper specifically? Certainly in general there are career incentives for authors to publish papers which are not very conclusive, and we know that relatively useless observational studies do get published.

Sounds like it might be The Last Psychiatrist? https://thelastpsychiatrist.com/

Could you please link to the technical reviews you mentioned? I would be very interested to read them because as an outsider, Flash Boys seemed overly simplistic, but I haven't been able to figure out what the real story is yet.

Test questions like the SAT are designed to have wrong answers that look like they are right at first glance, so people who think they have a better than average chance by guessing often don't because they were baited by one of the wrong answers without thinking it all the way through.

When there is a penalty for wrong guesses, students who have a pretty good idea that they know the answer, but aren't certain, must waste time determining whether it is in their favor to answer the question.

Evaluating whether it is worth it to take a guess is a test-taking skill, and the SAT is trying to shift away from test-taking skills.

Once you have the VM up, you can try running demo scripts, located at ~/bayesdb/examples//.py after you've checked out and pulled master from ~/bayesdb and ~/crosscat.

Yes, the install process is a pain in the current release (sorry!), but the next release (almost ready) will be much more friendly and granular to install.

Hi, sorry for the trouble, and thanks for the bug report! If you git checkout master and git pull both the crosscat and BayesDB repos (at ~/crosscat and ~/bayesdb on the VM), this issue will be fixed. We are working on pushing a new VM where the proper commits will be checked out already.