JavaScript is not available.

How

applies language models compositionally to make it easier to check models' work: 1 When language models extract info from or answer questions about papers in Elicit, users can quickly see the source - the part of the paper the model got its answer from.

0:19

10.2K views

201

I'm sure these are all great bands coming to the Fillmore in SF but this just looks like someone made a page of joke band names, right?

Feb 2

A little update here, this is a link so folks can read the paper without the paywall: rdcu.be/c4JlP

Quote Tweet

Nov 10, 2022

Big news: my paper "Identity effects in social media" with @LevMuchnik @madhavkumar2005 & @sinanaral was published in @NatureHumBehav It's an 89-week experiment with a novel intervention: on 5% of comments the viewer could not see the author's username. nature.com/articles/s4156

Show this thread

Feb 1

An AI that makes my slides look nicer without messing with my content.

Feb 1

Believe it or not it is really challenging to find a reason to discuss this slide with a causal directed ~acylcic~ graph for the Terminator.

Feb 1

Unlike the normal state of affairs, the less engagement this tweet gets, the more special I feel.

I wonder if there are other people around my age who read Popular Science as a kid and experienced the excitement of reading the mail-order ads for wacky technology and then eventually feeling really burned by finding out it was all quackery.

For arrays alone are so many design choices: - length vs. count - distinct vs unique - mean vs. average - how missing values are treated - indexing

I've had to learn a different set of built-in functions for data work numerous times: base-R, Tidyverse, Pandas, many SQL dialects, Stan, Excel and Google Sheets, etc. Who has done the absolute best job of creating a small, expressive, and intuitive set of functions like that?

IYKYK

Wish this button existed when chatting with some people.

Simpson's Paradox and Existential Terror

Vera Wilde

@vk_wilde

Jan 25

Finally wrote up something on Simpson's paradox, a widely misunderstood statistical phenomenon that doesn't really exist: wildetruth.substack.com/p/simpsons-par.

wildetruth.substack.com

There is no paradox -- "D'OH!"

406

Jan 26

Cautionary tales about why not to do this are arriving right on schedule.

I can't be the only one who sees cloud computing prices and thinks "how is this better than a bunch of Mac minis on my shelf?" Yes I know how wrong this instinct is.

131

Adding this to the list of mandatory reading materials for managers in tech. Such a clear explanation of an ever-present tradeoff, and one that's currently very under-articulated.

Quote Tweet

Sebastian Bensusan

@sebasbensu

Jan 23

High Variance Management: blog.sbensu.com/posts/2023-01- If you want a team to produce something outstanding, you need to manage variance, increasing it in some places, reducing it in others

Show this thread

My friend

sometimes uses the term *polycausal*, which I like -- there many phenomena that have property. It feels like there should be (but isn't) an equivalent term for the effect side: polyeffectual, a cause that has a variety of downstream consequences.

@DYtweetshere

Jan 23

We're releasing our

@scale_AI

hackathon 1st place project - "GPT is all you need for backend" with

@evanon0ping

@theappletucker

But let me first explain how it works:

102

614

2,300

Ok “one pour” was being generous, apparently.

When I came to pickup my belongings from the Lyft office (left since the start of COVID) after I resigned about a year ago, I regained possession of a bottle of Yamazaki 12 year with exactly one pour left in it. Today is the day I enjoy it.

Nonstick pans come and go but a cast iron skillet is forever.

Erik Brynjolfsson

@erikbryn

Jan 22

Some perspective on the tech layoffs: 1. California and the Bay Area are still adding jobs overall: 16,200 jobs in December in CA and 13,000 in the Bay Area alone. 2. The unemployment rate in Silicon Valley is just 2.3% 3. California is adding jobs faster than the rest of the US

352

2,146

I started working out again in July after a couple years off. I am making progress, but the workouts are often so *mentally* challenging for me. I dread when I see the workout on the board, I want to give up soon after it starts. I still don't know how to overcome these feelings.

Uri Shalit

@ShalitUri

Jan 16

Are you interested in causality, machine learning and healthcare? Come work with Mihaela van der Schaar (

@MihaelaVDS

) and me in a joint PhD or postdoc at Cambridge University, UK and the Technion, Israel Contact via email: shalit-lab AT technion ac il

127

David Holtz | @dholtz@zirk.us on mastodon

@daveholtz

Jan 17

🧵Tomorrow is the first day of a new MBA class I'm teaching at

@BerkeleyHaas

, "Online Marketplace and Platform Design." Interest is high! Despite being a new class taught by a new prof, the class is ~full. I wanted to share the reading list in case it is of interest to others:

460

What's the best current thinking on whether we're data-constrained or parameter-constrained for improving LLMs right now? Like if I could magically have 10X the data vs 10X the parameters, which of those would I choose?

Jan 12

See now my optimism is waning, I don't love this list of policy suggestions for housing affordability (compare to yimbyaction.org/2021/solutions/).

On the danger side, a lot of potential for abuse, and it would be very difficult to detect it. I'm sure that's well-discussed already. I just happen to be weirdly hopeful about it, like it will organize and align our thinking as a society a bit.

This could obviously be bad... but it could also be good? The suggestions could be a reasonable consensus over an extraordinarily large data set. It could convince a large number of people to, say, take effective steps to improve their lives (health, happiness, etc).

I wonder about these use cases for suggesting what people should do. If people follow the suggestions the AI makes, it creates a really interesting causal structure where that system has widespread influence over people's behavior.

I asked ChatGPT for policy suggestions for reducing homelessness. This is a good response IMO!

Show this thread

Description will go into a meta tag in

Rachel Cohen

@rmc031

Jan 12

German words win again

My former team at

just released a new Python package called `balance`, which automates the adjustment of biased samples (e.g. for surveys but many other applications). Exciting for me because the package was in its infancy when I left. Check it out!

import-balance.org

balance | balance

223

File photo

Hundreds of them in my garden in Oakland! Butterfly milkweed FTW.

Quote Tweet

The Progress Network

@progressntwrk

Jan 11

The number of western monarch butterflies counted in California: 2020 - 2,000 2021 - 247,000 2022 - 390,000 "It's really encouraging," said local volunteer coordinator Jessica Griffiths. "I'm personally so delighted by it." phys.org/news/2022-12-c

Show this thread

Jan 10

A lot of apps are bad at clearing notification state but Slack is consistently the worst and it's not even close. I wonder what's so hard about it.

What Does Apache Arrow Unlock for Analytics? (w/ Wes McKinney)

🎙️New

Analytics Engineering Podcast is LIVE!

@jthandy

and I sit down with THE

@wesmckinn

-- co-creator of Pandas, Arrow, and CTO of

@VoltronData

. Shaking 🧂🍸 it up and bringing some data science to the pod🤩

open.spotify.com

Listen to this episode from The Analytics Engineering Podcast on Spotify. Wes McKinney is the creator of pandas, co-creator of Apache Arrow, and now Co-founder/CTO at Voltron Data. In this conversa...

Barry McCardel

@barrald

Jan 4

Should we enable pie charts in

@_hex_tech

chart cells? We are undecided after several internal ballots so I turn this decision over to the masses

Yes
43.7%
No
56.3%

343 votesFinal results

At Motif we've noticed that our unit tests serve a few purposes: they describe system behavior, measure performance, as well as checking system correctness. Is there visual/interactive test reporting software that is searchable/browsable, and allows for annotations/modification?