iiinova — where your digital self keeps evolving

Marinus Ferreira @notrolleys.bsky.social · 17h

I did, but it is questionable whether I succeeded in making any one else understand it. I did get my first publication by applying my framework to the case of offloading decisions to opaque AI: onlinelibrary.wiley.com/doi/abs/10.1...

010

Marinus Ferreira @notrolleys.bsky.social · 20h

Just like academia, the trades have priced in the ability to get skilled work for very little pay under the label of 'training'. Like grad school, the rate of completion of an apprenticeship is low partly because people towards the end of their training resent being worked like dogs.

020

Marinus Ferreira @notrolleys.bsky.social · 20h

If you think grad school involves a lot of badly paid, underappreciated work in exploitative settings, wait till you do a trades apprenticeship.

120

Reposted by Marinus Ferreira

ponder @ponder.ooo · 08/10/2026

lol arxiv.org/abs/2610.08144

paper that claims to have identified discrepancies between the lean formalization and natural language proof of navier stokes blowup

38216

Marinus Ferreira @notrolleys.bsky.social · 08/10/2026

You might be getting your wish: bsky.app/profile/pond...

010

Marinus Ferreira @notrolleys.bsky.social · 08/10/2026

I endorse this joke, but I will add that the AI-generated papers that I have been sent for review do things no professionally-authored paper I've read has done, like write 11000 words about a topic without ever advancing a thesis. Also, please stop sending me purely AI-generated papers for review.

150

Marinus Ferreira @notrolleys.bsky.social · 07/10/2026

Really, doing anything else shows a corrupt mind.

030

Marinus Ferreira @notrolleys.bsky.social · 05/10/2026

I didn't realise that the Guardian endorses Aristotle's view that we cannot assign truth values to future contingents.

020

Marinus Ferreira @notrolleys.bsky.social · 04/10/2026

Humans use it as well, of course (that's where LLMs learnt the construction), but LLMs overuse it to the extent that in RLHF they try to train the models out of it. On Bayesian grounds, seeing this construction makes it much more likely that an LLM wrote it.

000

Marinus Ferreira @notrolleys.bsky.social · 04/10/2026

Among the many concerning things in this slide is that the language in the heading is almost certainly AI-generated: LLMs use staged contrasts like 'Worker's slice shrinks, Profit's slice glows' much more than human authors.

100

Reposted by Marinus Ferreira

Lev Parikian @levparikian.bsky.social · 03/10/2026

Happy birthday Steve Reich appy birthday Steve Reich H ppy birthday Steve Reich Ha py birthday Steve Reich Hap y birthday Steve Reich Happ birthday Steve Reich Happy birthday Steve Reich Happy irthday Steve Reich Happy b rthday Steve Reich Happy bi thday Steve Reich Happy bir

14463124

Marinus Ferreira @notrolleys.bsky.social · 02/10/2026

As documented by Hobbes in 'The Leviathan', white people naturally co-existed in peace and harmony and the Leviathan of the title was the whale of a time they were having in commonwealth in 17th century England, sharing drinks at a place called the Civil Bar.

010

Marinus Ferreira @notrolleys.bsky.social · 02/10/2026

For #BandcampFriday this month I'm very excited by The Black Drumset, a one-man psych rock band who makes percussive music which is simultaneously intricate and hypnotic. I bought the whole discography: theblackdrumset.bandcamp.com/track/the-ic...

theblackdrumset.bandcamp.com

The Ice Melted, We Had to Swim, by The Black Drumset

from the album Friends in Dark Places

031

Reposted by Marinus Ferreira

Elana Levin @levin.bsky.social · 02/10/2026

Damn. Cronos from Venom just died! They are an extemely influential to modern heavy metal. Founded in 78 their 1982 album Black Metal is where the genre got its name. The singles on their 2026 album kick ass. That’s a legacy to be proud of. Rest in power chords, Conrad "Cronos" Lant 🤘🏻

youtu.be

Man & Beast

YouTube video by Venom - Topic

2448

Marinus Ferreira @notrolleys.bsky.social · 30/09/2026

I can tell Postdoc Appreciation Week is over because you're rubbing the fact that you know where you'll be working next year in our faces 😥

030

Marinus Ferreira @notrolleys.bsky.social · 30/09/2026

Capsule of what time? That time you had a glimpse of happiness that has since turned to ash between your fingers? [One song stops playing, the next one begins, it's 'Rose Parade']

021

Marinus Ferreira @notrolleys.bsky.social · 29/09/2026

On that point, it is really remarkable the extent to which even this referendum, which was specifically about negotiation with the ANC, did not mention the ANC, except when the opponents of the referendum said it was surrendering to them.

120

Marinus Ferreira @notrolleys.bsky.social · 29/09/2026

OK, you were agreeing with me. Yes, I specifically had the '92 election in mind. The self-conception pretty clearly was of (white) South Africans stewarding their future. Apartheid was no longer viable, if it ever had really been, and the National Party thought this was the best way forwards.

020

Marinus Ferreira @notrolleys.bsky.social · 29/09/2026

If the claim is the worries by Treurnich et al, well, that was only griping. The usual 'stab in the back' narrative.

110

Marinus Ferreira @notrolleys.bsky.social · 29/09/2026

I was there for that! I remember it well, though I was only a child. But it was not conceived of as a matter of how to manage foreign attitudes. It was a matter of how to South Africa should manage its own problems.

210

Marinus Ferreira @notrolleys.bsky.social · 29/09/2026

Among the many problems (Afrikanerdom views perseverance as its chief virtue), it is not accurate to describe apartheid as ending due to bullying by foreign powers. What is even meant to be the basis of that? Some foreign elites disapproved of apartheid, it ended, so their disapproval ended it?!

110

Marinus Ferreira @notrolleys.bsky.social · 28/09/2026

My 3-year-old still reflexively answers most questions 'no', even if he goes on to say 'yes' only two seconds later. My favourite is him saying 'no' to some food being offered as he was taking it into his hands and pushing it into his mouth.

020

Marinus Ferreira @notrolleys.bsky.social · 27/09/2026

A lot of this Springbok loss reminds me of our loss against the All Blacks in the first of the four matches last month. Not taking opportunities and not keeping their head under pressure. This time we don't have three more games to make up for it. Well done again to the Wallabies.

000

Marinus Ferreira @notrolleys.bsky.social · 27/09/2026

The Wallabies did something the All Blacks could not, disrupt the SA lineout. All-in-all, they showed how to beat SA, which in some ways is to use their tools against them: place SA under pressure, meaning that when they have the ball they need to do more with it, and sometimes they'll fail.

100

Marinus Ferreira @notrolleys.bsky.social · 27/09/2026

Not Ikitau, Suaalii. Getting my Wallaby 13s mixed up

100

Marinus Ferreira @notrolleys.bsky.social · 27/09/2026

Lots of people are unhappy with the ref showing 4 yellow cards to AUS in the last 12 minutes, but 3 of those are nailed on by the current rules, and the Ikitau one (killing an attack by playing offside) was also fair. The issue, if there is one, is using sending off as a penalty, which is unpopular

100

Marinus Ferreira @notrolleys.bsky.social · 27/09/2026

I think the score flatters SA, who only got close because AUS gave up a series of yellow cards late. Even so they flubbed their lines a few times even when up 15 men to 11 (!!!). The obstruction by Horn to end the game was very bad but in keeping with how SA played at the end: list their heads.

100

Marinus Ferreira @notrolleys.bsky.social · 27/09/2026

My thoughts on the ridiculous rugby match that just finished between Australia 42 - 37 South Africa, Australia won I think by denying SA territory and possession, mainly by winning the battle in the air and being as usual excellent at stealing balls at the breakdown. And SA kept dropping the ball.

100

Marinus Ferreira @notrolleys.bsky.social · 27/09/2026

I think they call it 'exhilarating rugby that will always keep you guessing'

010

Marinus Ferreira @notrolleys.bsky.social · 27/09/2026

Well done Wallabies, they had a plan and they executed it beautifully.

110

Marinus Ferreira @notrolleys.bsky.social · 27/09/2026

That's two games in a row without Kolisi on the field where SA lose their heads.

1191

Reposted by Marinus Ferreira

Australasian Association of Philosophy @aaphilosophy.bsky.social · 25/09/2026

Congratulations Bhanuraj Kashyap, Macquarie University, Winner of the 2026 AAP Media Prize for the SBS podcast series "Rethinking South Asian Philosophy" aap.org.au/Media-Prize sbs.com.au/audio/podcast/rethinking-south-asian-philosophy #aaphilosophy #philsky

0112

Reposted by Marinus Ferreira

Dr SpookyGoth @ghostingdani.bsky.social · 24/09/2026

Using generative AI means that you can't answer the most basic element of research, which is "where did you get this information?"

48933

Marinus Ferreira @notrolleys.bsky.social · 23/09/2026

I was thinking the same. Reminds me of how people say that physics shows that everything is caused and there's no space for free will, when ancient Greeks and Romans already cogently discussed what the causal closure of the physical would mean for free will.

010

Marinus Ferreira @notrolleys.bsky.social · 23/09/2026

It would be churlish for me to try and undercut each of the examples of evidence you're presenting. I acknowledge what you take as the best evidence. I disagree; I tagged you when I presented what I thought is the strongest. Perhaps you'll find it interesting. If not, that's your prerogative.

010

Marinus Ferreira @notrolleys.bsky.social · 23/09/2026

People hated LLMs refusing their corrections, so the purveyors of these models have trained them not to do it through RLHF. It reveals no deeper features about LLMs You can tell because, again, you can get it to hedge and acknowledge ignorance about even obviously true things.

000

Marinus Ferreira @notrolleys.bsky.social · 23/09/2026

It didn't know that what it said is false, it is responding to a prompt correcting it. You can trivially get an LLM to acknowledge being wrong about things that are obviously true by 'correcting' it. This is a behaviour introduced by RLHF; early versions of GPT-2 was very rude about being corrected

220

Marinus Ferreira @notrolleys.bsky.social · 23/09/2026

I would love it if he did an Anglo Saxon version of Romeo and Juliet which is like The Norseman as compared to adaptations of Shakespeare's Hamlet. Fewer apothecaries, more berserkers

020

Marinus Ferreira @notrolleys.bsky.social · 23/09/2026

I don't believe their line either. I think their approach makes LLMs parasitic on genuine language understanders and unable to stand on their own, even more so than speakers depend on their speech communities to imbue words with meaning. But their work is the stuff I worry about the most for my view

110

Marinus Ferreira @notrolleys.bsky.social · 23/09/2026

Link to that paper: philosophymindscience.org/index.php/ph... Millière & Buckner also have a Phil Compass article on LLMs in general: compass.onlinelibrary.wiley.com/doi/10.1111/...

philosophymindscience.org

The Vector Grounding Problem

120

Marinus Ferreira @notrolleys.bsky.social · 23/09/2026

As I said on their post, I disagree with @moultano.bsky.social about what is the best evidence here. For LLMs being able to display linguistic understanding, the best work I know is by Raphaël Millière and co-authors, e.g. Mollo & Millière, 'The Vector Grounding Problem', Phil. & the Mind Sciences

philosophymindscience.org

The Vector Grounding Problem

230

Marinus Ferreira @notrolleys.bsky.social · 23/09/2026

It is genuinely impressive that LLMs can produce fluent text! The culmination of decades of work in NLP. That it does so using embedding spaces is very interesting. Trying to track the processes it uses to do so is important. But this is not thinking.

120

Marinus Ferreira @notrolleys.bsky.social · 23/09/2026

Once you acknowledge that LLMs can produce fluent text that (to use some jargon) represents an embedding space, then this is not a surprise. Fiddling with the vectors in the model changes what is closest to what in the embedding space, so what output gets produced.

120

Marinus Ferreira @notrolleys.bsky.social · 23/09/2026

I'm sorry, but despite mechanistic interpretability being important work, but this is exactly no evidence of thinking. This is what you'd expect from fiddling with vectors in order to manipulate the results of linear algebra. What it shows is that LLM outputs don't fall out of the sky fully formed.

110

Marinus Ferreira @notrolleys.bsky.social · 23/09/2026

I don't think even Ayer would endorse Stancil's argument, because good science often involves postulating theories for which we only have indirect evidence, i.e. psychological processes that produce but are not identical with behaviour. LLMs obviously don't have some of these, e.g. personalities

020

Marinus Ferreira @notrolleys.bsky.social · 23/09/2026

I think his view is closest to Ayer-style behaviourism, where observations of behaviour are the gold standard of evidence and everything else is suspicious or outright bad. So, because LLMs often behave like thinking beings (display thinking behaviour), and there isn't another standard, they think

230

Marinus Ferreira @notrolleys.bsky.social · 22/09/2026

The one-post summary of Anderson's view is that valuable things are the products (top-level goods) of concrete processes (intermediary goods), and business economics by trying to maximise top level goods undermines the intermediary goods that produce them. It makes a virtue out of eating seed corn.

010

Marinus Ferreira @notrolleys.bsky.social · 22/09/2026

Not the most approachable treatment of the problem, since it is trying to speak both to economists and philosophers so it takes on two different technical frameworks (at least), but the best.

110