iiinova — where your digital self keeps evolving

Carl T. Bergstrom @carlbergstrom.com · 14h

Reposted by Antoine Vernet

I have absolutely wondered about this. Amazing.

301644300

Antoine Vernet @antoinevernet.com · 11h

But do you really want a pay cut?

Captain Haddock @realcaptainhaddock.bsky.social · 15h

Reposted by Antoine Vernet

Microsoft branding nightmare re AI risks to humanity

724443

Antoine Vernet @antoinevernet.com · 13h

Try programmatically matching headings on a lot of documents… then you'll stop asking why I have no hair 😭

020

Reposted by Antoine Vernet

This skeet will self destruct @pardoguerra.bsky.social · 20h

It reached 109 degrees in San Diego today, October 2. And “we haven’t passed the 1.5C climate threshold“.

171

Reposted by Antoine Vernet

Petter Holme @pholme.bsky.social · 18h

Network of the day. From: JL Moreno's Psychodrama (1946).

082

Oh god, I still have a paper from my PhD I want to finish. I think the first draft is from 2012 👀

020

Berna Devezer @devezer.bsky.social · 02/10/2026

I would love to be able to write two good papers a year!

120

Reposted by Antoine Vernet

i cannot wrap my head around that 2 papers a week (socarxiv) or a month (arxiv) should constitute a meaningful cap on preprint submissions. 2 papers a year (no one) though? that i'd get.

9434

I have feelings about estate agents starting to use AI generated pictures in listings that showcase what the room could look like… Often they don't disclose it is AI, second, I am really buying what it looks like now

Peter Tennant @pwgtennant.bsky.social · 02/10/2026

Reposted by Antoine Vernet

"Not always towards the null: reconsidering the direction of overadjustment bias" - read our new paper in the @intjepidemiol.bsky.social led by Frerik Smit! Overadjustment for a mediator is often thought to 'bias towards the null', but the bias can be in any direction - as the paper explains!

academic.oup.com

Not always towards the null: reconsidering the direction of overadjustment bias

Abstract. Overadjustment bias occurs when adjusting for a variable increases rather than decreases bias. It commonly arises in studies estimating the total

15717

I switched everything to quarto and really likes it. Slides in reactjs like you do, and for manuscripts, word export is now mature enough that I use it even when co-authors use Word.

I redid my website in quarto (it was using RMarkdown and Blogdown) a year or so ago and really like it. How do you find using quarto for slides and manuscripts?

I do like the new feature in Claude of finding a reasonable place to stop when your 5 hour window runs out without charging you. That probably saves me a little each time, and also means I do not have to worry about the limit as much as a job will partially complete rather than fail.

This annoys me so much, yet I keep falling for this trap...

I have no idea what you might be referring to… 🫠

Jekyll and GitHub pages are great, but I think for ease of use, Quarto is now a very strong contender to replace Jekyll (plus, it is not just websites and you can basically all your writing with it)

120

I am sure there are plenty of interesting use cases that will emerge as people have more time to experiment.

000

Not extensively. I have tried it but only with toy examples. On a project I am part of, a colleague is testing using it to validate tasks as finished. Currently, we use an LLM to do that. But it is too early for me to tell you if it works, and if yes, how well.

Samuel Moore @samuelmoore.org · 02/10/2026

Entirely based on vibes, but Haiku 4.5 seems to fail more often than Luna 6

100

Reposted by Antoine Vernet

"The University of Cambridge refused to sign plagiarism detection software Turnitin’s new licence agreement, which could have allowed students’ work to be used to train artificial intelligence (AI) models."

varsity.co.uk

Uni opposes Turnitin AI training plans

The University is now exploring alternatives to the plagiarism detection service, according to the SU

220160

Also, Sonnet draws down my Anthropic usage faster than Luna draws my OpenAI one. Anyway, I am excited to see how good Haiku 5.5 is when it comes out… if it is close to Sonnet but cheaper, this will shift more of my tasks toward Anthropic

100

It could be that—despite what I have them do being pretty simple—Luna fails at a much greater rate than Sonnet and the feeling is entirely due to the retries taking time...

100

Gonçalo Pina @alvespina.bsky.social · 01/10/2026

Sonnet 5.5 feels much faster than Luna 6, is that just me?

220

Reposted by Antoine Vernet

José Mourinho, your Department Chair

08517

Johannes Haushofer @johanneshaushofer.com · 01/10/2026

10 movies that made history. Number 8 will surprise you!

020

Reposted by Antoine Vernet

Five years after founding @malengo.org, it’s time to let it fly! Today I’m stepping down as CEO, and Chad Sterbenz is taking over. He brings decades of experience in related roles, and I’m grateful that he is throwing his talents behind our mission. 1/3

1283

I am not AGI, I ask myself to do a breakthrough every morning and it isn't working

Is it because she gushes about "the strength of your analytical voice"? I'd hand over all my money right there after such a compliment

000

Telling the model to stop dilly-dallying… Do a breakthrough already 😂

120

Jeremy Berg @jeremymberg.bsky.social · 01/10/2026

Is it me or is Luna kind of slow? Or are my subagents failing and therefore I should upgrade them to Terra or Sol?

000

Reposted by Antoine Vernet

From a discussion of updating a department website...

A photo with baby Yoda and adult Yoda with the caption "Could it be time for a new photo?"

934670

Reposted by Antoine Vernet

Kieran Healy @kjhealy.co · 01/10/2026

RIP to everyone’s favorite p<0.05 factory

120737

Reposted by Antoine Vernet

Per Engzell @pengzell.bsky.social · 01/10/2026

CRIS/Sciences Po (Paris) and IAB (Grenoble) are recruiting two postdoctoral *or* doctoral researchers to join a project on how pollution and other environmental exposures affect children’s development. Deadline Nov 13. www.sciencespo.fr/cris/en/news...

089

Julia Silge @juliasilge.com · 30/09/2026

It is crazy how much Sol 6.1 overthinks the docs...

000

Reposted by Antoine Vernet

I'm super excited to join the folks from @motherduck.com at Data Outpost in November! I'll be speaking about why the investments that make SQL comfortable for humans turn out to be the same ones that make it tractable for an agent 🤖 Check out the details at www.dataoutpost.ai

"THE SQL AGENT HARNESS WE DIDN’T MEAN TO BUILD" talk at Data Outpost, MotherDuck's conference on Nov 5, 2026

0295

Antoine Vernet @antoinevernet.com · 30/09/2026

Thought regex was the right tool for the job 🤪