iiinova — where your digital self keeps evolving

Henry🔸

@henryjosephson.com

103 followers 66 following 121 posts

I may look like a 15x15, but I'm actually two minis on each other's shoulders in a trench coat. Send me your favorite pdf! let’s elect democrats who know what’s going on www.henryjosephson.com

Posts Replies Media

Henry🔸 @henryjosephson.com · 13/01/2025

My friend Thomas wrote a piece on death (simplereflections.substack.com/p/on-death, meme unrelated)

100

Henry🔸 @henryjosephson.com · 13/01/2025

@moskov.goodventures.org does it count if i only made it to $500 with transaction fees included?

110

Henry🔸 @henryjosephson.com · 13/01/2025

The most interesting thing that I haven't seen other recaps touch on, though, is that the rule also embargos some model weights!!

010

Henry🔸 @henryjosephson.com · 13/01/2025

Quick takes on the new Interim Final Rule on Artificial Intelligence Diffusion: 🧵

120

Henry🔸 @henryjosephson.com · 12/01/2025

welp, i only went to the gym bc i hoped ai returns on robotics would lag far enough behind returns on intelligence that i'd still be useful as a thing-picker-upper in twenty years...

000

Henry🔸 @henryjosephson.com · 21/12/2024

new sunday record, too! adrian johnson, you make a hell of a grid

020

Henry🔸 @henryjosephson.com · 21/12/2024

fwiw, i smashed my friday record yesterday! i think this guy and i are just on the same wavelength

150

Henry🔸 @henryjosephson.com · 21/12/2024

@evavivalt.bsky.social posted -> i texted it to the gc -> gc emails to tyler -> marginal revolution post!

321

Henry🔸 @henryjosephson.com · 20/12/2024

It’s been a long time coming, but I’m excited to announce my NYT crossword debut! I hope you like it :) www.nytimes.com/crosswords/g...

7321

Henry🔸 @henryjosephson.com · 18/12/2024

why on earth would you ever manually insert line breaks into your webpage

110

Henry🔸 @henryjosephson.com · 18/12/2024

lebanon, new hampshire

100

Henry🔸 @henryjosephson.com · 18/12/2024

update: just got a paper back, professor is proud of me for reading things that aren't gpi working papers:

010

Henry🔸 @henryjosephson.com · 18/12/2024

...and found big differences between them (look at those y-axes!) why do claudes do so much better than geminis and gpt-4os? the authors have a few thoughts:

Alt text: A figure showing three line graphs labeled (a) Claude 3.5 Sonnet, (b) Gemini 1.5 Flash, and (c) GPT-4o. Each graph shows five different colored lines and one dotted black "Average" line tracking performance across 10 generations (x-axis). The y-axis shows "average final resources" with different scales: up to 6,000 for Claude, 600 for Gemini, and 40 for GPT-4o. Claude's lines show generally increasing trends with high variance, Gemini's lines show erratic behavior with eventual drops (labeled as "cooperation crash"), and GPT-4o's lines show gradual decline across generations. The figure caption explains these trends and notes the different y-axis scales.

110