Assistant Professor at Chapman University; Fellow, MIT IDE and Stanford DEL; Co-host of Justified Posteriors Podcast https://nitter.cf/t.co/BJbcIOiwuv
๐ฅฏ-to-โ๏ธ-to-โพ๐ฅ๐งฆ-to-๐
Joined January 2014
- Tweets4.7K
- Following1.5K
- Followers1.2K
- Likes9.1K
Pinned Tweet
Weโre launching the Justified Posteriors Podcast, co-hosted by Andrey Fradkin and me - Seth Benzell.
Thanks to Tyler for todayโs shoutout -- heโs a key inspiration.
We evaluate research on the Economics of AI and Tech, with a good helping of wit and light hearted irreverence.
Seth Benzell - the LA-atic Stranger retweeted
WTF?
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! ๐งต
Blog: rikiyat.github.io/blog/weighโฆ
Seth Benzell - the LA-atic Stranger retweeted
I made some plots to help visualize why coalition formation in Israel is so difficult right now
We are witnessing an extremely rare situation that political scientists call a "Condorcet cycle" - a demonstration of Arrow's impossibility theorem in the wild!
Seth Benzell - the LA-atic Stranger retweeted
My thought on the role of growth in solving fiscal problems:
1. I'm in favor of dynamic scoring for both tax & spending plans--you should take into account the best estimate of impact on budget.
2. Dynamic scoring needs to be done by CBO/JCT, they're generally unbiased and reasonably efficient. As a result they often disappoint advocates.
3. Ideally CBO would also display other information than just growth, the sole goal of every policy is not necessarily growth, it might also be about providing health insurance or increasing national security.
4. Dynamic scoring will usually be small relative to the first order effects of policies on spending and taxes.
5. I wouldn't overthink the second-order political incentives and impacts, if you have good information you'll generally (but not always) have better policy.
Seth Benzell - the LA-atic Stranger retweeted
ืฉืืืข ืืื.
ืืกืคืจ ืืืจืื ืืื ื ืืืืื, ืืืื ืืืืืืชืื ืืืฉืจืืืื, ืืงืจืืช ืืืืื ืืงืจืืืื:
ืกืคืืจืช ืืงืืืืช ืืกืชืืืื ืืืืืืจืืช ื ืืืจื. ืืื ืืจื ืื ืืกืชืืืื ืขื ืืืกืคืจืื, ืจืื ืืืจืื ืืฉืจืื, ืืืคืจืฉ ืฉื ืืืชืจ ืืืฆื ืืืืืื ืงืืืืช, ืืฆืืืขื ื ืื ืืืฉืืช ืืฉืืขื ืืืืงืืืืจ ืืืืฉืชืืืืช, ืืืขื ืฉืื ืื - ืจืื ืืงืจื ืืื ืืืจืื ืืฉืจืื, ืืจืื ืืงืจื ืืฆืืืืจ ืืฆืืื ื ืืืฉืจืช.
ืืืืจืชื ืืช ืืืืืจืืช ืืืื ืืฉื ื ืืืืจืื ื ืืืืจืืืืช ืืืืชืจ ืืชืืืืืช ืืฉืจืื ืืชืืฆืืืชืืื ืืจืืจืืช, ืืฃ ืฉืืฉ ืื ืฉืคืืขื ืืืืื ืืื ืืืคื. ืืืืืชื ืืืืจ ืืืืฉ ืืช ืืฉืจืื ืืืฉืื ืกืืฃ ืืชืจืืืช ืืฉืงืจ ืฉืืืืื ื ืชื ืืื ืืคืืจืงื ืืืชื ื ืืืคื ืื. ืืืืืชื ืฉืชืืช ืื ืืืชื, ืฆื ืืณ ืื ืืืจืืข ืืช ืฆื ืืณ, ืฉืชื ืขืืื ืดืื ืื ืืด ืืดืืืด. ืืืืจ ืืืืืืช ืืื ืืืฉืื ืืคืขืื ืื ืืฉืืืขืืช ืืงืจืืืื.
ืืืงืืื, ืืืืื ืชื ืืคืขืื ืืืืืืฉ ืืจืื ืืคืจืืื ืืจื ืืืืืืง ืืืืคื ืืืืื, ืืืืืช ืืื ืืืจืื ืืฉืจืื. ืืคืขืื ืืืงืืช ืืขืืช ืืงืืจื ืืืืืชืืช, ืงืืืื ืืืืช ืฉืืจืืช ืืื ืืคืขืืื ืืืืืืช ืืืืืื ืืืืืืง ืืืงืฃ ืืืฉืจืชืื ืืฆื"ื, ืชืืืื ืืืืงืจื ืจืืืื ืืืฉืจืชืื, ืืืืื ืืืงื ืืืคืืื ืืืฉืืจืืช ืืขืื. ืื ืชืชื ื ืืืฃ ืืื ืืืืขืืช ืืชืื: ืืืืื ืช ืืฉืจืื ืืื ืกืช ืืื ืฉื ืืชื ืช ืืช ืืืื ื ืืืืฉืื, ืืื ืืืคื.
ืืืืื ืงืืคืืื ืืืืืจืืช ืืฆืืชื ืืืจื ืืืงืื ืืืฉืื ืฆืืื ืืช ืืืืืชืืช ืขื ืืกืืื ืืืืืืช ืจืืื ืืื ืืืคืฉืจ ืฉืชืืื ืืจืืคืื ืืฉืื ืื ืืกืืื ืืืืืื ืืืื ืืฉืจืื ืืืืืช.
ืืฆืืชื ืืืืคื ืขืงืื 4 ืขืงืจืื ืืช ืฉืืืืงื ืืช ืืฉืจืื, ืืื ืฉืืชืคืืช:
1.โ โ ืืืจื ืืืฉืจืื ืืืืื ื ืืืืืืช, ืืืืงืจืืืช, ืืืืจืืืช ืืฉืืืืื ืืช.
2.โ โ ืืืืืืืช ืืขืจืื ืืืืืช ืืขืฆืืืืช.
3.โ โ ืืืืช ืฉืืจืืช ืืืื ืืืจืื ืืฉืจืื - ืฆืืื, ืืืืื ืืืืจืื.
4.โ โ ืืืจื ืืืืืก ืืืจืืื ืืจืืจ ืืืืืื ืืืฉืืืช ืืืื ืืฆืืื-ืฉืืืื ื ืฉืื ืืจืฆืืขืช ืขืื.
ืขื ืืกืืก ืขืงืจืื ืืช ืืื, ืืืืื ืชื ืืืงืื ืืืฉืื ืฆืืื ืืช ืืืืืืชืืช, ืจืืื ืืื ืืืคืฉืจ. ืืฆืืจื ืื ืืคืขื ืืชืืืื ืขื ืฉืืชืคืื ืืืคืขื ืืืจืืื ืืช ืืฉืืชืคืืช, ืขื ืื ืช ืืืืฉ ืืช ืจืฆืื ืืืืืจืื ืืืฉืคืจ ืืืืคื ืืืืื ืืช ืืฆืื ืืคื ืืื ืืืืื ืืืืื ืฉื ืืืื ืช ืืฉืจืื.
ืื ืืืจ ืื ืกืช ืฉืืืื ืืืชืื ืขื ืขืงืจืื ืืช ืืื, ืืื ืฉืืชืฃ ืืคืฉืจื ืืืจื ืืืื ืื ืืืฉืื ืืืฉื ืืืืื ืืืฉืจืื. ืืืช ืืืืฉืื ืฉื ืงืื ืชืืืื ืคืชืืื ืืื ื ืืืจ ืฆืืืืจ ืฉืืงืื ืืช ืขืงืจืื ืืช ืืื ืื ืืื ืืืชืขืืืช ืืืืื ืืฉืขื.
ืืจืืจ ืื ืืืืื ืืืฆืืขื ืืืืืืื, ืืื ืื ืืืชืืจืื ืฉืขื ืืคืจืง, ืืืืืืื ืืืืฅ ืขืืืื ืขื ืื ืช ืืืืื ืข ืืืืืจืืช ื ืืกืคืืช ืืื ืืกื ืืืฉืืจ ืคืืืืื ืืชืืฉื ื ืืกืฃ ืืื ืืจื. ื-27.10 ืืฆืืืืจ ืืืจ ืืช ืืืจื ืืืืืืื ืขื ืืจืืื ืฉื ืืืฉืืช ืืฉืจืื ืืืื ืืื ืฉื ืืืจืื ืืฉืจืื ืืืื, ืืื ืืขืืจืืืช ืืื ืืคืืืฆืื - ืื ืจืืื ืฉืืืื. ืืขืช ืืืืืช ืขืืื ื, ื ืืืจื ืืฆืืืืจ, ืืืืื ืืขืฉืืช ืืื ืขื ืื ืช ืืชืจืื ืืช ืจืฆืื ืืขื.
ืืืื ืืฉืืขื ืืืืงืืืืจ ืืืืืฉืื ืืืฆืืืืช ืืืืืืื ืืช, ืืคื ืืืืช, ืืืืืืืช ืืืืื ืืืืืืช ืืืืืืื ืืืืืืช ืืืืฆืืช ืืืืงืื. ืืคืขื ืืื ืืจื ืืืืื ืืกืืื ืืืืืืช ืจืืื ืืื ืืืคืฉืจ ืขื ืื ืช ืฉื ืืื ืืืชืืื ืืชืืงืื ืืฉืืงืื.
ืืกืืื, ืืืืืื ื ืื ืืืืืื ื ืื ืชืื. ืืืจืืช ืืืืืื ืื ืืืฉืื. ืขื ืืืงืืช ืืืืฉืื ืืชื ืืืืื ืืืืืืื ืฆืืืืื ืจืง ืืืืื ืืืืื ืืืืืื ืืฉืืจืื ืืืืืืืื ืื ืื ืืฉืจืื ืฉืื ืืืืืื ืืืืชืื ืขื ืืืฉืืขืช ืืืืฉืื.
ืื ื ืฉืื ืืช ืขืฆืื ื. ืื ื ืืคื ื ืฉืืืขืืช ืขืืืกืื, ืืืจืืืื ืืขืชืืจื ืกืคืื ืื ืืฉืงืจืื.ย
ืืืืื ืชื ืืคืขืื ืืืืืฅ ืืืชืื ืืืืืืืช ืืืืื ืืืืืจ ืืืืืื ื ืขื ืื ืช ืืืืฉ ืืช ืจืฆืื ื - ืืฉืงื ืืืืฆืืจืชืืืช.
ื ืขืฉื ืื ืฆืืื, ืืืขื ืืืื ื ืืืืขื ืืืื ืช ืืฉืจืื. ืื ื ืืชืืืื.
Seth Benzell - the LA-atic Stranger retweeted
Spec Ops the line totally works though because the player is the cosmic horror
Seth Benzell - the LA-atic Stranger retweeted
@SBenzell is one of the sharpest and most entertaining explainer of the economics of AI in the biz. In addition to being a great co-host of the brilliant Justified Posteriors podcast, turns out he's also a great guest on The Economics of Work:) Check out our conversation!
reveliolabs.com/podcasts/theโฆ
Seth Benzell - the LA-atic Stranger retweeted
Erin can do a Charlton Heston impression. Erin will not do a Charlton Heston impression.
(sub)Text ยท Civilizational Discontents in โPlanet of the Apesโ (1968)
Seth Benzell - the LA-atic Stranger retweeted
Japan just released the unclassified version of their annual defense strategy anyways here's the actual cover page of it
1/ Are you above or below the panopticon? ๐งต
On a new Justified Posteriors, @AndreyFradkin and I talk AI + hiring, +peer review & + religion with @SarahHBana of @LSEnews and @DigEconLab
Sarah's research: ~4M job applications, all screened by one vendor's video games (example: balloon-popping to test risk aversion). When every employer buys the same arbitrary screen, the same applicants get rejected everywhere. Algorithmic monoculture.
Seth Benzell - the LA-atic Stranger retweeted
is nathan fielder actually playing a character? said iโd run nathan fielderโs voice through soma, so here it is:
45 clips from 2008 to now. about half in character (nathan for you, the rehearsal, talk shows), half as himself (podcasts, festival q&as)
i expected the deadpan to show up somewhere in his voice. it basically doesnโt. 13 of 14 voice measures canโt tell persona nathan from real nathan. same pace too, 177 vs 175 words a minute. his reply timing doesnโt give it away either
the one thing that changes is his words. in his shows he says โumโ and โlikeโ about a third as often as on podcasts. but thatโs mostly the edit. live on talk shows, still in character, heโs almost back to normal
heโs also flatter than 70 of the 80 public figures weโve measured. and he and elizabeth holmes land on the same voice print, bamboo. close on melody, different on almost everything else, so not twins
holmesโ voice changed depending on who was in the room. his doesnโt. maybe the character is just him with the ums edited out
in honor of the doc of a generational releasing today, we obsessed over @ElizabethHolmes voice
a lot can be said about her, but her voice is irrevocably distinct andโฆpracticed
we pulled 16 recordings of her from 2014 to 2017, weโve pulled them from ted talks, charlie rose, cnn, mad money, a bunch of conference panels, and 3 full days of her deposition. about 4 hours of her talking total
we've been running her voice through @try_soma before the documentary drops and honestly the deposition sounds like a different person:
- hedging (โi thinkโ, โi believeโ, โsort ofโ) went from 3.8 to 19.5 per 1,000 words, about 5x. every deposition day was higher than every public appearance
- stumbles went from 29 to 50 per 1,000 words
- her sentences got simpler, 2.3 clauses per sentence down to 1.8
- her pitch dropped too, 152 hz in public to 136 hz under oath
the part i didnโt expect: if you score her against normal population benchmarks she looks totally fine. good to excellent in almost every recording, deposition included. her hedging under oath (17 to 23 per 1,000 words) is still technically โnormalโ for people in general
itโs only when you compare her to HER that the shift shows up. against her own baseline itโs super obvious
which is kind of the whole point of what weโre building. a score that compares you to everyone else wouldโve totally missed this
also fun fact, sheโs at her most polished on stage. slowest pace (129 words per minute), cleanest speech (3 filler words per 1,000). on tv sheโs way faster, 177 wpm. so even her โpublicโ voice isnโt one voice
maybe weโll run it again with @nathanfielder 's You Can See Everything, since thatโll be the first new audio of her since 2017?
Seth Benzell - the LA-atic Stranger retweeted
I don't think anyone has any idea what to make of this (from @I_Am_NickBloom survey)
In next Monday's episode of Justified Posteriors, we have the great Sarah Bana of LSE to talk how AI is changing hiring, academic publishing and religion.
Sarah used a cool metaphor for how AI may change labor productivity: AI as Steroids vs. Shoes vs. Coach.
Seth Benzell - the LA-atic Stranger retweeted
My frequent co-author Chiara Farronato, who's not on Twitter, just wrote a nice HBR article summarizing a lot of recent research (including ours) on digital platforms. Do check it out!
hbr.org/2026/11/the-new-platโฆ
