Joel Levin @residualdeviant

Behavioral scientist at @UCSD @RadySchool. Working on expertise, human-algorithm collaboration, methods + applications in medical decision making. 🐙/2 San Diego Joined June 2009

Tweets

737
Followers

203
Following

815
Likes

1K

Joel Levin @residualdeviant

a month ago

@xsteenbrugge tbh I think gpt 4 would easily do this if prompted

0 0 0 12 0

View Details

Great thread, but likely addressable with a simple skill or even claude.md note -- e.g., to have independent agents read all claims and evaluate for plausibility. One-shot AGI is far off. Well prompted AGI is basically here.

lilly sharples @lillysharples

a month ago

My favorite Fable 5 benchmark is this equity research report. They gave a junior analyst a Claude Code subscription and told him to try and rebuild an institutional research shop:

18 4 768 290K 788

0 0 1 31 0

View Details

Joel Levin @residualdeviant

a month ago

@stillnesshum @alexolegimas @dioscuri Earlier models seemed to really degrade over iterations of adversarial self-review. Opus was pretty good. Fable is good good, to my eye. Can't wait to see what the next generation can do. Hard to imagine it not being AGI (to this layperson.)

0 0 0 15 0

View Details

Joel Levin @residualdeviant

a month ago

@stillnesshum @alexolegimas @dioscuri I essentially never ask for any important output without some kind of 'adversarial review by multiple independent agents' type workflow, which I think can often (but not always) result in good kinds of pushback.

1 0 1 8 0

View Details

Joel Levin @residualdeviant

2 months ago

@lakens @PublicThinkOrg Also a risk of false positives. A generally quite careful researcher who I know was recently accused of related behavior because they (slightly) mis-cited a paper (despite the passage being written pre genAI.)

0 0 0 9 0

View Details

Joachim Voth @joachim_voth

7 months ago

How did people in 1913 see the world? How did they think about the future? We trained LLMs exclusively on pre-1913 texts—no Wikipedia, no 20/20. The model literally doesn't know WWI happened. Announcing the Ranke-4B family of models. Coming soon: github.com/DGoettlich/his…

197 597 6K 563K 3K

View Details

Ahmad Lashkaripour @ALashkaripour

12 months ago

A 15% tariff "without" EU retaliation would likely increase "overall" US welfare per our projections: alashkar.pages.iu.edu/MAGA_July2025.… But this doesn't mean US consumers benefit. It's a massive redistribution of income from US consumers (and the EU) to US corporations and the Treasury.

13 83 284 125K 140

View Details

Joel Levin @residualdeviant

a year ago

@waltbrook @History__Speaks That's the article reporting on the survey, not the survey itself.

0 0 0 9 0

View Details

Joel Levin @residualdeviant

a year ago

@ryancbriggs @alexolegimas Very interesting. I also heard that o3 is great at math, but imo this problem is mostly not-math. Claude 4 Sonnet gave an answer that seems correct at first glance. It's possible even that some of the "traits" that make a model good at math could be maladaptive here

0 0 0 43 0

View Details

Joel Levin @residualdeviant

a year ago

@History__Speaks @waltbrook Those are the survey questions, but there are no results (it doesn't say how people answered the questions) Do you have access to a document like that, too?

1 0 2 94 0

View Details

Rachel Wetts @RachelWetts

4 years ago

🚨New WP!🚨 "Antiracism and its Discontents: The Prevalence and Political Influence of Opposition to Antiracism among White Americans" w/ @RobbWiller We find opposition to antiracism is a widespread ideology profoundly shaping US racial politics 🧵 osf.io/preprints/soca…

11 70 170 0 73

View Details

Joel Levin @residualdeviant

2 years ago

There are DOZENS of us

0 0 1 71 0

View Details

Joel Levin @residualdeviant

2 years ago

@KellyScaletta Exempting tips, overtime, and social security? (I hate them as policies, but would guess that they're relatively progressive on net)

2 0 0 33 0

View Details

Joel Levin @residualdeviant

2 years ago

Trump & Kamala have proposed a range of narrow tax cuts recently (tips; overtime; social security; now interest on car loans.) I'd be the last to note that these are horrible economic policies... but it's also notable how psychologically well-informed they are. Very clever.

0 0 0 57 0

View Details

Joel Levin @residualdeviant

2 years ago

@cboyack @elonmusk Now make a picture about externalities

0 0 0 28 0

View Details

Joel Levin @residualdeviant

2 years ago

💯 Having your paper retracted is not necessarily something to be ashamed of. But impeding the process almost always is. We've been fortunate to see many good examples of how to handle retraction recently, along with the bad

Oleg Urminsky @OlegUrminsky

2 years ago

A quick thought on a difficult topic: The purpose of retractions is to correct the academic record, and to remove conclusions that we now know are not supported by the evidence from the literature. 1/3

1 4 31 7K 8

0 0 2 115 0

View Details

Joel Levin @residualdeviant

2 years ago

Unfortunate situation, thoughtful response. We turn the ship slowly, largely through the voluntary efforts of individual researchers. (in case you're wondering what that looks like) Thanks @J_R_Schroeder

0 1 1 1K 0

View Details

Joel Levin @residualdeviant

2 years ago

The "one rough hour" bit is a more unambiguous endorsement of police brutality than I thought I'd ever hear from a US president. Believe him when he says that he admires the authoritarians.

0 0 1 90 0

View Details

Joel Levin @residualdeviant

2 years ago

@andre_quentin is spot on as usual. But I think there can be value in preregistering Hs if we care about controlling error at the level of the theory (rather than the test.) Hs are priors, and the posterior P that a theory is true is higher when it's consistent with priors.

Quentin André @andre_quentin

2 years ago

Many of these debates could be avoided if people stopped pre-registering hypotheses. If the goal of a pre-registration is to control Type I errors at the nominal level, then the researcher's pet theory doesn't need to be a part of it. Just describe the procedure and the test.