Ted Suzman retweeted
BREAKING 🚨: Hubble and James Webb telescopes teamed up to drop new images of two spiral galaxies colliding
C.S. Lewis:
"It is a serious thing to live in a society of possible gods and goddesses, to remember that the dullest and most uninteresting person you can talk to may one day be a creature which, if you saw it now, you would be strongly tempted to worship, or else a horror and a corruption such as you now meet, if at all, only in a nightmare.
All day long we are, in some degree, helping each other to one or other of these destinations.
It is in the light of these overwhelming possibilities, it is with awe and the circumspection proper to them, that we should conduct all our dealings with one another, all friendships, all loves, all play, all politics."
Ted Suzman retweeted
Community Notes is setting speed records for how fast informative context can be added to potentially misleading posts at scale on the internet. Not only that, notes are found helpful across the political spectrum, making them effective and trusted.
Introducing Lightning Notes ⚡ People want fast notes. So we re-architected the system to score notes faster. Since rolling this out, we’ve seen notes go live in as little as 14m33s after being written, and 18m20s after the post itself was written. Ludicrously fast? Now reality!
Think about what needs to happen for useful context to be added to a post:
• Post has to be seen
• Post has to be evaluated — is it accurate? is the image real? is the video from today or 5 years ago? etc
• Sources need to be found — reverse image search? original video? raw data? helpful article or analysis?
• Proposed context needs to be written
• Proposed context needs to be reviewed by people from different perspectives for helpfulness, accuracy, etc
Thanks to Community Notes’ 800k+ global contributors and our new scoring architecture, it is now possible for this entire process to happen in under 20 minutes.
Contributors can also write notes on specific images or videos. These media notes show automatically on new posts with matching media, without any additional work from contributors.
Speed matters. External research made possible by notes’ open source data has shown that notes reduce sharing of posts by ~60%, increase deletion of posts by ~80%, and are highly accurate. On top of that, notes are found helpful by people from different perspectives, and people from across the political spectrum find notes significantly more trustworthy than traditional misinformation flags. See the studies: teapawt.net/CommunityNotes/status/…
In the past two weeks:
• Over 400 notes started showing <1hr after being written. They’re automatically showing on over 8000 posts and have been seen over 200M times.
• Over 60 notes starting showing <1hr after the post itself went live. They’re automatically showing on over 1000 posts and have been seen over 21M times.
Now, when you see a note that goes live in under 60 minutes, you’ll see a lightning indicator on it, recognizing the incredible work of contributors who make this happen:
Want to help?
• Everyone: If you see a post you think would benefit from a note, Request a Community Note
• Contributors: Check your Needs Your Help tab to see notes that could go live quickly with your ratings: teapawt.net/i/communitynotes
• Top Writers: Check your Note Requests tab to see posts that people think could use your help: teapawt.net/i/communitynotes/recen…
ALT Screenshot of the X app, showing a Community Note and a note details page for the note. The note has a new lightning indicator, and the note details page has a lightning indicator and says "Rated helpful and shown within an hour of the post"
Today's US minimum wage can purchase 8x more calories per hour than basic labor back in medieval England.
(Given all the growth and agricultural advancements we've had, one might've hoped to see >8x difference!)
Source: comparing the price of bread in "Standards of living in the later Middle Ages" to the price of bulk peanut butter at Walmart.
Ted Suzman retweeted
🚨🚨🚨 VERY excited to share our new paper on how AI can facilitate democratic deliberation, published today in @ScienceMagazine! Together with @mhtessler, @summerfieldlab, and other amazing collaborators at @GoogleDeepMind we've been building the "Habermas Machine"
Let's say the parents of a digital bit are all the bits that directly calculated it, e.g. the opcode and other operands. What are the oldest ancestors of the bits on my personal computer?
Can we trace back through early text editors (and the text editors used to write those text editors), until... what? punchcards? Or was there a total break in the lineage that came after punchcards?
Some people don’t like the idea of institutions.
But I think most such people are mainly against the sorts of poorly-functioning institutions that stick out in their minds (e.g. the DMV). Good institutions often become invisible — just part of the fabric of reality — the way things “are” / “should be”.
For example, prior to courts adjudicating disputes, many people were trapped in an honor culture in which they had to react with intense aggression to any slight in order to maintain their reputation as someone who can’t be messed with. This hair-trigger dynamic caused a lot of tension and violence. It's thanks to institutions that things don't work that way anymore for many of us, and you don’t see people complaining about it. That’s what functioning institutions are like!
That feeling of relaxation when the author is a multi lenser and you can let down your dogma defenses a bit.
from amazon.com/Lost-History-Endu…
I wonder what the story here is —
(a) founders are the ones with the internal world model that made the thing successful to begin with, and this world model is difficult to transfer, so the founder needs to be involved in decisions at many levels?
(b) execs the founders hire are self-serving, and principal-agent problems destroy the co unless the founders is very involved at many levels?
(c) the founder is actually just way more competent/agentic than the people they hire, and so more of their involvement works well?
(d) the founder has much higher standards than the people they hire?
(e) it helps a lot if a single person that both fully groks the whole "mechanism" of the product/business, and has the power to push hard to maintain/expand/evolve this mechanism? (letting execs be autonomous could result in losing conceptual integrity / strategic coherence)
If history repeats itself because old failure modes pass out of living memory, maybe we should keep our memories fresh via a civ-like game where to win you have to avoid common failure modes of the past.
Kinda like Sim City except instead of earthquakes and fires you gotta deal with rising authoritarianism, zealotry, dark ages, caste strife, hyperinflation, decadence, and whatever other things tends to wreck civilizations.
Maybe citizens in the game are LLM-powered and have twitter 😂 and you gotta help guide the discourse/pass the right laws/etc.
Ted Suzman retweeted
I model this like so: each type of job has (a) a lifespan (time until the type of work is obslete) and (b) a training time (how long it takes to train for). As long as job training times are << job lifespans, things should be okay.
But if automation drops the average lifespan of jobs significantly (i.e. we get quicker at automating away new types of jobs), then at some point people are just spending all their time training for jobs that then would quickly go away... Is it worthwhile to pay the training costs in this case? Possibly not. At that point I think we hit true structural unemployment from automation.
Someone should chart the average lifespan and training requirements of jobs across history!
Ted Suzman retweeted
Bat signal has been a top request from both contributors (who want to know where help is needed) and the rest of X (who want to see a note on a post they think would benefit from one).
If you think people would find added context on a post helpful, request a note!
This YouTube channel is great youtube.com/@asianometry
Has lots of analyses of how technology and politics interplay — both historical and modern examples
Looks like a good benchmark! It's a shame though that such evals lose reliability after they become public. (They end up in training data, which makes them unreliable for evaluating future models.)
I wish benchmark creators only publicly released a portion of their tasks — the remaining private portion could be used to evaluate each model as it comes out, which would let us graph how AI is improving with time and better anticipate how it might affect us.
