iiinova — where your digital self keeps evolving

Reposted by Mir Henglin

Demetri @phdemetri.bsky.social · 15/10/2026

Despite AI being the end of mathematics as we know it, AI lacks a certain style to writing mathematics. I find I have to boot strap it a bit by writing a few sections, but once I have then it certainly picks up what I'm putting down

131 0

Reposted by Mir Henglin

Mark Copelovitch @mcopelov.bsky.social · 08/10/2026

Just like the US News new formula, this “earnings after four years” is a stupid metric of the worth of college. If folks were serious about all this, they’d stop pretending that people don’t go to grad school & do other things in their 20s while not earning high salaries: bsky.app/profile/mcop...

510020 0

Reposted by Mir Henglin

Danielle Navarro @djnavarro.net · 05/10/2026

In which the author disagrees with Mr Wolfers and discusses what it *really* looks like when you sit down and have a conversation with your ideological opponents. When you do not have the protections of the ivory tower it turns out to be a much darker experience than he imagines it to be.

blog.djnavarro.net

Dining with wolves – Notes from a data witch

Polite discussion over lunch is not a reasonable option when you are on the menu

37735 0

Reposted by Mir Henglin

Dare Obasanjo @carnage4life.bsky.social · 04/10/2026

When I say AI coding tools mean execution is essentially “free”, I typically get pushback yet this is reality now. OpenAI’s Codex team plans to ship either a feature or a rewrite a day over the next month. This sort of velocity is impossible without AI. Tech debt is considered a manageable cost.

6272 0

Reposted by Mir Henglin

Will Stancil @whstancil.bsky.social · 29/09/2026

A better way to think of groups of people: a particle system, lots of little individual units all with their own characteristics and moving their own directions, but being acted on by forces that affect them all in related ways.

2494 0

Reposted by Mir Henglin

Colin @colin-fraser.net · 28/09/2026

I think something not emphasized enough is that OpenAI is subsidizing this. It’s not like they create the agent and then it’s a self-sufficient guy who runs amok. Every time the agent’s little brain says “it’s cyberattackin’ time!” OpenAI pays like a few cents and a drop of water.

18214 0

Reposted by Mir Henglin

hailey @hailey.at · 27/09/2026

this is posting

630925 0

Reposted by Mir Henglin

Colin @colin-fraser.net · 27/09/2026

Honestly this is not my intuition about the word “rogue” but I can be swayed by this. Has the dog gone rogue?

10272 0

Reposted by Mir Henglin

Boxo McFoxo @boxobark.ing · 27/09/2026

It is a complex system so it does things that were not anticipated by the people deploying it (especially when the engineers building it have internalised flawed ontological priors about what they're building lol) but that isn't the same as going rogue.

061 0

Reposted by Mir Henglin

Kiran @kirancodes.me · 27/09/2026

you crack your egg and then you are baby turtle

i am a baby turtle where is mama, google search response

Your mama is already gone. She laid your eggs in the sand and went back to the ocean a long time ago.

Growing up alone.
You do not need your mom.
You can live on your own from day one.
Your instincts tell you what to do.

2473 0

Reposted by Mir Henglin

Eryk Salvaggio @eryk.bsky.social · 26/09/2026

This, from Dennett, I wish I'd taken to heart earlier in my life. www.themarginalian.org/2014/03/28/d...

How to compose a successful critical commentary:

1 You should attempt to re-express your target’s position so clearly, vividly, and fairly that your target says, “Thanks, I wish I’d thought of putting it that way.”
2 You should list any points of agreement (especially if they are not matters of general or widespread agreement).
3 You should mention anything you have learned from your target.
4 Only then are you permitted to say so much as a word of rebuttal or criticism.

520236 0

Reposted by Mir Henglin

Colin @colin-fraser.net · 27/09/2026

> "AI goes rogue" because of what it suggests about the capabilities of AI. Yeah, in my opinion this is true: it suggests capabilities that don’t exist. It makes it seem like OpenAI could not have easily prevented this or stopped it at any time if they were paying a minuscule amount of attention.

1111 0

Reposted by Mir Henglin

conputer dipshit @davidcrespo.bsky.social · 27/09/2026

always interesting when someone makes a new kind of thing and shows there’s demand for the category and it becomes clear there wasn’t really a technical barrier to inventing it. it was just knowing people wanted that thing (we still don’t know if Jev’s alleged RLCD special sauce matters that much)

5816 0

Reposted by Mir Henglin

Hilary Mason @hilarymason.com · 26/09/2026

Overnight I trained a Jev-like model to output probability distributions over a set of 40 emoji, which runs in RAM at ~150ms per query. This was super easy and cost ~$10. It works pretty well on a weak training set, and with a day or two of work I could make this sing.

A terminal showing the output of the emoji jev-like model:

We launched the project!
🎉 0.41 👏 0.23 🔥 0.15 ✨ 0.08 💪 0.07
valence 3.92/4 intensity 1.98/4 | 154 ms

My cat is so cute and hates me right now
😂 0.15 😏 0.13 🥹 0.12 ✨ 0.06 😆 0.05
valence 2.34/4 intensity 1.85/4 | 140 ms

you stand in front of the dragon, triumphant
✨ 0.27 🔥 0.17 🎉 0.16 😍 0.09 😮 0.06
valence 3.65/4 intensity 2.19/4 | 144 ms

why would you do this with emoji anyway?
🤔 0.54 😏 0.17 🙄 0.06 😐 0.04 🧐 0.03
valence 1.86/4 intensity 1.30/4 | 138 ms

101329 0

Reposted by Mir Henglin

Samuel Moore @samuelmoore.org · 27/09/2026

One of the worst things about the current moment is the complete lack of imagination associated with "can AI do it?" Why not make peer review a social process that's less arbitrary and gatekeeping but generative and conversational?

chronicle.com

Opinion | Could AI Be Better at Peer Review Than Humans?

A journal editor weighs the pros and cons.

2288 0

Reposted by Mir Henglin

Dare Obasanjo @carnage4life.bsky.social · 26/09/2026

After a few months of living the AI native lifestyle, I now believe the default state is that it makes you worse at your job overtime as you understand less and hand off more thinking to AI. You’re more productive but it takes a lot of effort for your work quality not to drop precipitously.

1114913 0

Reposted by Mir Henglin

ǝʌɐp @dvshkn.com · 26/09/2026

Analogies aren't perfect, but it seems very similar to living a car-centric lifestyle. You can cover great distances, but you also need to separately exercise to prevent your body from falling apart.

2101 0

Reposted by Mir Henglin

conputer dipshit @davidcrespo.bsky.social · 24/09/2026

whether you agree with this or not, it’s important to understand that this is why they say sandboxing is not enough and we need to “solve alignment” (annoying SV language, I concede). most real-world agent tasks use the internet

3774 0

Reposted by Mir Henglin

Dare Obasanjo @carnage4life.bsky.social · 25/09/2026

Mark Zuckerberg made a recent comment that building user trust is more important than which model has the most benchmark beating capabilities when it comes to personal AI agents. I think he has a point.

7214 0

Reposted by Mir Henglin

Ted Underwood @tedunderwood.com · 23/09/2026

Something that worries me more every day: I suspect most fields have 5-25 breakthrough ideas latent in existing data and literature. Say you’re a midsize AI company looking for cred. Why not hire a *small* team of domain experts + 1000 agents … and promote your models as the ones that broke Econ?

12797 0

Reposted by Mir Henglin

conputer dipshit @davidcrespo.bsky.social · 24/09/2026

would love to hear what you think about the methods described in the preprint. haven't seen a single opinion yet from someone who knows something

291 0

Reposted by Mir Henglin

Adam Kucharski @adamjkucharski.bsky.social · 21/09/2026

I suspect we’ll increasing see a return to “you can only submit if you know an insider” systems, like the old-school PNAS ‘contributed’ track. It’s a bit like we’re already seeing with the new ArXiv ‘endorsement’ system...

3269 0

Reposted by Mir Henglin

Colin @colin-fraser.net · 23/09/2026

you don't need to appeal to metaphysics to see what whatever LLMs do is different from human thought. You can just look at the concrete differences in behavior. For example, LLMs are able to solve the Navier-Stokes Millennium Prize problem whereas humans are not.

41329 0

Reposted by Mir Henglin

Dare Obasanjo @carnage4life.bsky.social · 20/09/2026

Jev is the AI model for the token-pinching era. Instead of generating text, it makes decisions: yes/no, where to route this request, whether to escalate this issue, etc. The key idea is that not every AI task needs an LLM generating text. Nearly 13% of Vercel’s paid AI teams tried it in 24 hours.

4433 0

Reposted by Mir Henglin

Ethan Mollick @emollick.bsky.social · 20/09/2026

It is ironic that the thing that is now most annoying about long-running agentic tasks with Large Language Models isn't coding or errors or hallucinations, but the fact that their language gets worse due to drift & cross-agent talk as a task goes on Its in your the name! Just write better already!

5695 0

Reposted by Mir Henglin

Dare Obasanjo @carnage4life.bsky.social · 19/09/2026

CLAUDE.md is dead. Long live AGENTS.md

5575 0

Reposted by Mir Henglin

Dan Quintana @dsquintana.bsky.social · 19/09/2026

New preprint! 🎉 I analysed 1660 papers from 4 psychology journals and found materials sharing went from 9% of papers in 2015 to 82% in 2025, and these materials *do* get downloaded — a median of 135 times each. BUT shared code is often hard to run. doi.org/10.31234/osf... Let's walk through it 🧵

Two-panel figure. Panel a is a flow diagram tracking 1,611 empirical psychology articles from publication year (545 in 2015, 552 in 2020, 514 in 2025) to repository-link type: 785 link an OSF project, 85 link another platform, and 741 link no repository. Of those with an OSF link, download counts were retrieved for 670 and not retrieved for 115. Panel b is a line chart of the share of empirical papers linking OSF across 2015, 2020 and 2025. The overall rate, shown as a dashed black line, rises from 9% to 57% to 82%. All four journals rise steeply and end close together: Psychological Science 92%, JESP 89%, JML 83%, Cognition 76%, with Psychological Science highest throughout.

Three-panel figure. Panel a: ridgeline plot of downloads per file by material type on a log scale, with the percentage never downloaded labelled for each — archive 16% of 545 files, documents 20% of 2,970, code 12% of 3,471, other 17% of 1,646, data 21% of 10,051, media 35% of 2,120, images 35% of 4,102. Most files cluster between 1 and 10 downloads, with long right tails past 100. Panel b: ridgeline plot of downloads per paper by journal, log scale, with dashed median lines; Psychological Science is highest, then JESP, JML and Cognition. Panel c: stacked bars showing, for documents, data and code separately, the share of papers by download band (0, 1–10, 11–100, more than 100) in 2015, 2020 and 2025. The share exceeding 100 downloads falls sharply over time in all three types, from roughly two-thirds in 2015 to a quarter or less in 2025, as the 1–10 band grows.

Four-panel figure. Panel a: statistical languages detected among 333 papers with retrievable code — R 88%, SPSS 12%, Stata 4%, SAS 1%. Panel b: code red flags among those 333 papers — 34% hard-code an absolute path, 40% reference a missing file — above documentation among 672 OSF-linked papers — 21% have a README, 30% are documented by README, description or wiki. Panel c: among 562 Elsevier papers with no repository link, 44% (245) host at least one journal supplementary file but only 14% (77) host data, code or an archive. Panel d: composition of those 448 hosted files — documents 52%, data 21%, media 9%, other 7%, archive 6%, code 3%, images 1%. The code panels are green, the journal-supplement panels blue.

Coefficient plot (download predictors)

Dot-and-whisker plot of three standardised predictors of OSF download volume, each with a 95% confidence interval. Repository size (number of files) has the largest effect at about 0.76, citations about 0.38, and altmetric attention about 0.12. All three intervals sit entirely above zero, so each predicts more downloads, with repository size roughly twice the effect of citations and six times that of attention. X-axis: standardised effect, −0.2 to 1.0.

4269115 0

Reposted by Mir Henglin

(Un)Dead Carl 🎃 @deadcarl.bsky.social · 14/09/2026

Load-bearing post

21158289 0

Reposted by Mir Henglin

mallow's hyperrealistic blood @bajablastenjoyer.bsky.social · 14/09/2026

[on my 6th consecutive hour of doomscrolling and arguing on the internet without even an eyestrain break, 1,200mg of caffeine in my system, no solid food consumed since yesterday afternoon] why do i feel weird today?

922831 0

Reposted by Mir Henglin

Jordan Nafa @ajordannafa.com · 11/09/2026

I handed Fable 5.1 a 300-line Markdown file outlining a general vision, some architecture details, and some strongly held opinions as a prompt, and it returned a 2,500-line spec that was almost exactly what I had in mind. It required a bit of tweaking, but not much.

1142 0

Reposted by Mir Henglin

Ruben C. Arslan @ruben.the100.ci · 11/09/2026

An interesting thing is happening to me at the moment. I have a standing bug bounty policy for my first authored scientific work. That means, I've committed to pay people who report bugs in my code, up to 750€ for severe errors. Since GPT-6 Astra came out, I'm now getting several bug reports daily.

39334 0

Reposted by Mir Henglin

Daniel Litt @littmath.bsky.social · 11/09/2026

I gave a talk at CMSA on AIxMath on Wednesday. The title ended up a bit unrelated to the content--I rewrote the talk at the last minute for obvious reasons, and talked a bit more about my views on the future of the profession than I expected to. Link: www.youtube.com/watch?v=0wL8...

youtube.com

Daniel Litt | Working with LLMs to do high quality math

YouTube video by Harvard CMSA

5334 0

Reposted by Mir Henglin

conputer dipshit @davidcrespo.bsky.social · 08/09/2026

scott alexander being absurdly correct in 2019 slatestarcodex.com/2019/02/19/g...

Now imagine you prompted it with “P != NP”. This time give it near-infinite training data and computational resources. Its near-infinite training data will contain many proofs; using its near-infinite computational resources it will come up with a model that is very very good at predicting the next step in any proof you give it. The simplest model that can do this is probably the one isomorphic to the structure of mathematics itself (or to the brains of the sorts of mathematicians who write proofs, which themselves contain a model of mathematics). Then you give it the prompt P != NP and it uses the model to “predict” what the next step in the proof will be until it has a proof, the same way GPT-2 predicts the next word in the LotR fanfiction until it has a fanfiction.

The version that proves P != NP will still just be a brute-force pattern-matcher blending things it’s seen and regurgitating them in a different pattern. The proof won’t reveal that the AI’s not doing that; it will just reveal that once you reach a rarefied enough level of that kind of thing, that’s what intelligence is. I’m not trying to play up GPT-2 or say it’s doing anything more than anyone else thinks it’s doing. I’m trying to play down humans. We’re not that great. GPT-2-like processes are closer to the sorts of things we do than we would like to think.

511412 0

Reposted by Mir Henglin

Ethan Mollick @emollick.bsky.social · 08/09/2026

The breakthrough that solved Navier-Stokes in 88 hours used 130 billion output tokens. It seems clear that other breakthroughs will be possible, especially as models get better. However much compute we have, it is increasingly likely that we are going to need more compute.

716010 0

Reposted by Mir Henglin

John Burn-Murdoch @jburnmurdoch.ft.com · 08/09/2026

Two extraordinary charts from today’s PISA test results: 1) School test scores continue to collapse internationally, underscoring how this is no longer a Covid effect but sustained decline. Those falls in reading and maths are equivalent to about two years of lost schooling.

41856344 0

Reposted by Mir Henglin

Richard McElreath 🐈‍⬛ @rmcelreath.bsky.social · 07/09/2026

For Americans reading the German election news this morning, it might help to know that Saxony-Anhalt is like the West Virginia of Germany. Population similar size and age, GDP similar, high out-migration, old dying industry. No less worrying, but good for comparison.

3644 0

Reposted by Mir Henglin

Ethan Mollick @emollick.bsky.social · 06/09/2026

This is both an interesting experiment and a sign of a tsunami coming for academia. AIs retroactively reading the research and finding both opportunities and issues with published papers, then sharing those judgements publicly. d3jhl7jsny6f2h.cloudfront.net

610821 0

Reposted by Mir Henglin

Ben Harrap @bharrap.bsky.social · 05/09/2026

Why not both? (I'm disgusted that this works)

76411 0

Reposted by Mir Henglin

Yihui Xie @yihui.org · 05/09/2026

When I finished typing f() and realized I needed to assign this to a variable, I'd have to do: > data |> + f() -> x Or just run the code and then x = .Last.value

021 0

Reposted by Mir Henglin

Ethan Mollick @emollick.bsky.social · 05/09/2026

This impressed me: GPT-5.6 Astra turned Zork (the classic 1977 text adventure) into a full 3D playable action-adventure game It kept the original plot & puzzles, added fight scenes, and built all of the characters & environments directly in Threejs code. Play: zork-underground-empire.netlify.app

2929943 0

Reposted by Mir Henglin

Dare Obasanjo @carnage4life.bsky.social · 03/09/2026

GPT-6 Astra saturated ARC-AGI-3, scoring 99.9% on a benchmark designed to test whether AI can learn how to solve unfamiliar interactive problems. The notable part is the same model scored 62.7% with the standard harness. A great example of how the harness increasingly matters as much as the model

3363 0

Reposted by Mir Henglin

Dare Obasanjo @carnage4life.bsky.social · 04/09/2026

OpenAI’s GPT-6 Astra ad shows why CEOs love AI so much. Giving vague commands that get turned into real results by their staff is how they work. Now with AI they don’t even need people. However if you’re the one doing work with AI, you have to deal with “You’re absolutely right, I messed up” daily.

77919 0

Reposted by Mir Henglin

conputer dipshit @davidcrespo.bsky.social · 21/03/2026

in other words, the usefulness *is* the problem. the problem is not people's belief that there's a little guy in there

2435 0

Reposted by Mir Henglin

conputer dipshit @davidcrespo.bsky.social · 21/03/2026

put this way, the stochastic parrot framing seems aimed at people who, when they see an LLM output, immediately believe there's a little guy in there. for me the problem is: even if you guarantee this error has been thoroughly debunked in every mind, we still have all the same problems we have today

3451 0

Reposted by Mir Henglin

David Crawshaw @crawshaw.io · 02/08/2026

A surprising conclusion I have come to from working with agents. blog.exe.dev/devtools-mus...

blog.exe.dev

Devtools must be open source - exe.dev blog

We have entered the age of personalized software.

88211 0

Reposted by Mir Henglin

Demetri @phdemetri.bsky.social · 26/08/2026

Bayes doesn't make experiments run faster. Stop saying this dpananos.github.io/posts/2026-0...

dpananos.github.io

Bayes Does Not Make Experiments Faster On Its Own – Demetri Pananos Ph.D

5226 0

Reposted by Mir Henglin

Colin @colin-fraser.net · 26/08/2026

The thing is when you say that you think there can be such a thing as a slur for the computer, it makes it sound a little bit like you think that oppressing people by using slurs is the same as oppressing the computer. That's what a lot of people are going to hear even if that's not how you feel.

1164 0

Reposted by Mir Henglin

Sasha Gusev @sashagusevposts.bsky.social · 24/08/2026

It is good to keep in mind that Trump & co's attacks on these disciplines have faced their biggest challenge in courts, typically on 1st amendment, academic freedom, and labor law grounds (see examples below). So I would like to see the left move towards, not away, from defending such principles.

11211 0