Misha looked at the rapid progress in using language models for biological tasks, and offers a forecast for when we will see AI generation of custom proteins.
arbresearch.com/files/gen_bio.
Arb
@ArbResearch
Arb’s Tweets
David returns to his roots and sharpens our ignorance about some important concepts people have not been treating well:
problem-solving ability vs goal-orientedness vs phenomenal consciousness vs self-consciousness vs moral patiency
1
3
Max Tabarrok on recent futurists. Incredible that they still weren't using probabilities in the 90s.
What accuracy should impress us? Well it's only n=12, but Szabo's 46% and Miller's 42% put them among the best. (Is that enough?)
Couple important caveats:
1. Asimov stands out but the other two don't. He still got lots of big things very wrong (mainly supply chains and overpopulation)
2. We looked at their nonfiction rather than their fiction because novelists have a complicated objective with truth last
Quote Tweet
Please stop using "science fiction" to mean "nonsense"!
1. Sci-fi is one of the few ways we have to debate the future we are building
2. Despite not being primarily about prediction, sci-fi is actually about as good as anything in predicting the future
cold-takes.com/the-track-reco
Show this thread
One aspect of expertise we've been thinking about since writing this is Heisenberg's line "an expert is someone who knows exactly which mistakes you must not make". Clipping the left tail often matters more than the mean.
Quote Tweet
With thanks to Sam Enright, Sam Harsimony, Sam Glover, Irish Sam Bowman, Alejandro Ortega, Patrick Atwater, Max Langenkamp, Christian Robshaw, Ozzie Gooen, Kristi Uustalu, Michael Story, Nick Whittaker, Philip Stephens, Adam Russell, Alec Stapp, Santi Ruiz, Nuño Sempere for input twitter.com/IFP/status/163…
2
This is great to see but indicative of our challenges in the forecasting community:
A) did the accuracy advantage matter for decisions?
B) was anyone making decisions based off these estimates?
progress.institute/can-policymake
1
6
27
(Note the separate analysis for the community predictions and the weighted "Metaculus predictions")
+ as always, this stuff is only possible because Metaculus are really open with sharing data.
5
Show this thread
Vasco, sometime data wizard at Arb, took a look at the average calibration of the Metaculus community.
Overall Brier is 0.13 (good), but 0.24 for questions about AI (near chance).
+ some nonsig. correlates with worse performance, incl. more predictions.
3
9
35
Show this thread
With thanks to Sam Enright, Sam Harsimony, Sam Glover, Irish Sam Bowman, Alejandro Ortega, Patrick Atwater, Max Langenkamp, Christian Robshaw, Ozzie Gooen, Kristi Uustalu, Michael Story, Nick Whittaker, Philip Stephens, Adam Russell, Alec Stapp, Santi Ruiz, Nuño Sempere for input
Quote Tweet
How should policymakers think about integrating forecasters, prediction markets, and expert judgment into decision-making?
@g_leech_ and @mishayagudin lay out a framework:
progress.institute/can-policymake
1
1
How should policymakers think about integrating forecasters, prediction markets, and expert judgment into decision-making?
and lay out a framework:
2
12
24
This is a nice overview of why use softmax (do good options often, worse options less often) rather than argmax (do the best option 100%) in many situations. Big issue for me is how to find optimal temperature schedule - but we can have individual ones!
2
4
51
Pair of posts about the political economy of emulated minds, from our in-house philosopher David. Messier than it seems!
forum.effectivealtruism.org/posts/imY2soAH
forum.effectivealtruism.org/posts/qLYFBpT2
Dan Luu has released a characteristically superhuman red teaming of futurism which includes Arb's study of the Big 3.
He covers more ground than did our team of 6 people.
1
2
33
big rewrite of the alignment wikipedia page is out!
1
12
82
It's easy to mock the old futurists - to assume they must've all suffered Jetsons or Star Trek naivete.
We investigated 3 non-cherrypicked examples - processing hundreds of their books to get a decent estimate of their accuracy. And they were ok!
1
10
45
I'm helping judge this extremely cool new contest for good new criticisms of effective altruism.
Go on, get it off your chest. Hit me with yr best shot. Give it to me straight, like an etc etc etc.
forum.effectivealtruism.org/posts/8hvmvrgc
3
7
25
A useful response from Good Judgment Inc:
4
Show this thread
New series about the EA response to covid and what it means for longtermist method.
1
1
6
Show this thread
Arb are featured in this week's ACX forecasting post!
1
More at the link, including three big tables of studies (forecasters vs experts, prediction markets vs pooled forecasts, and the inexpert vs the expert)
1
7
Show this thread
This has implications for the laudable drive to forecast important things like machine learning progress: it might be better to train ML experts in forecasting rather than vice versa.
2
4
21
Show this thread
Tracking down the original studies, we find mixed results and weak methods. The best study* found similar performance between forecasters and experts. Another decent one found a small (3%) forecaster advantage.
Compare the original 30% claim!
*
1
1
14
Show this thread
Superforecasting - that certain teams of people are reliably better than other prediction mechanisms like crowds and statistical rules - seems sound.
But serious interest in superforecasting stems from the reported triumph of forecaster generalists over non-forecaster experts.
1
6
Show this thread
Our first public work is out!
We review the idea that the very best generalist forecasters can beat experts at predicting events *in their own domain of expertise*.
1
37
105
Show this thread






