iiinova — where your digital self keeps evolving

...are clearly *productive* in the classic linguistic/philosophical sense, i.e. capable of producing infinitely many *new* - previously unseen - samples of text.

010

Piotr Wilkin @ilintar.org · 7h

...the quasi-conversational "user-assistant" format that end-users are used to seeing. Also, the one thing problematic with the parrot metaphor is that "parroting" assumes LLMs are only capable of outputting the exact patterns that they have been trained on - and this is not the case, i.e. LLMs...

110

Piotr Wilkin @ilintar.org · 7h

* models *are* trained with "conveying thoughts to listeners" - first, because such discourse is present in their raw texts, second, because they are trained not only with raw text data, but also post-trained with so-called "instruction training" which embeds the capability to generate text in...

110

Piotr Wilkin @ilintar.org · 7h

...well *how* they do it; going away from LLMs for a bit, would we really assume eg. that the chess-playing Stockfish's NNUE (the neural network used by the engine to evaluate final positions, i.e. positions at the end of depth of its evaluation) *does not* have a model of chess positions?

110

Piotr Wilkin @ilintar.org · 7h

To be clear, as someone who has worked with the internals of LLMs for quite some time now, some of these assumptions are either false or tentative: * LLMs can form internal models during their training because that's what neural networks do during training - although we don't understand...

130

Piotr Wilkin @ilintar.org · 22h

"Aren't you being hysterical? Why are you saying things like 'MAGA are being overtaken by Nazis'? Why are you calling today's mainstream right parties Nazi? Why would you make such outrageous claims?" www.sacurrent.com/news/politic... Oh.

sacurrent.com

Karl Rove says Republican Party is being taken over by 'Nazis'

The Bush-era GOP political operative also said Republicans would be winning the battle for Texas' U.S. Senate seat if John Cornyn was the nominee.

120

Piotr Wilkin @ilintar.org · 22h

Sure, like she did when overruling Roe vs Wade?

130

Piotr Wilkin @ilintar.org · 29/09/2026

For all the fans of holding Catholic officials accountable for their actions, let's name a proper analogy that they'll recognize: this is akin to a justice saying that state-mandated forced abortions are fine and can proceed and we'll hear the case in December. I wonder what the reaction would be.

020

Piotr Wilkin @ilintar.org · 29/09/2026

Can ACB stop calling herself Catholic please? This stance has nothing to do with Catholic teaching. Morally, this is akin to a justice stating that the government is allowed to orchestrate state-demanded abortions.

1110

Piotr Wilkin @ilintar.org · 29/09/2026

The opinion is very civil and dry, but in the details it's extremely scathing. The Circuit is basically saying the Trump lawyers failed to raise *any* substantial points.

2721

Piotr Wilkin @ilintar.org · 29/09/2026

TL;DR: the government case is so weak that it loses on all the possible points.

000

Piotr Wilkin @ilintar.org · 29/09/2026

Nie sądzisz może, że czas porozmawiać o powszechnym dostępie do broni? ;)

0100

Piotr Wilkin @ilintar.org · 29/09/2026

Rao and Katsas should also be there explaining their bullshit order that caused this in the first place.

090

Piotr Wilkin @ilintar.org · 28/09/2026

Raczej braunista jednak.

010

Piotr Wilkin @ilintar.org · 25/09/2026

In other words: the relentless probe of the defense finally forced the lady between a rock and a hard place, so she had to resign because the DoJ was about to throw her under the bus. If they defended her properly, she still wouldn't have done anything. Did I get that right?

010

Piotr Wilkin @ilintar.org · 25/09/2026

How many more of those stupid Katsas-Rao panels? This is insane...

090

Piotr Wilkin @ilintar.org · 25/09/2026

Nie, kontakt osobisty jest niehigieniczny. Wyłącznie rażenie prądem.

020

Piotr Wilkin @ilintar.org · 24/09/2026

Thanks for proving my point.

111

Piotr Wilkin @ilintar.org · 24/09/2026

I haven't seen her engage here - on BlueSky - in any meaningful discussion with critics that doesn't include disparaging remarks or disdain for opponents who don't share her views. If you can point me to a discussion that includes *substantial* disagreements, I'd be grateful.

110

Piotr Wilkin @ilintar.org · 24/09/2026

If you want to preach to someone else than the choir, then tackling what LLMs *actually can produce* (instead of claiming, against a rising pile of evidence, that they cannot produce any non-trivial artifacts) is the first step, but it requires getting out of that comfy space.

010

Piotr Wilkin @ilintar.org · 24/09/2026

I know. But she's also pigeon-holing herself into a group of yes-men and cutting off any critics. That's not a smart intellectual strategy. The "stochastic parrot" idea was cute for a while, but it's quite obvious that the text-generation capabilities of LLMs extend past pure parroting.

220

Piotr Wilkin @ilintar.org · 24/09/2026

...do you start introducing randomness, by adding the extra computational mechanism known as "samplers" that manipulate the logits, sometimes with a pseudorandom mechanism fueled by a random seed.

020

Piotr Wilkin @ilintar.org · 24/09/2026

...weights or probabilities that the model would assign to the next token. Under so-called "greedy decoding", you can just pick the highest-rated token and discard the rest and the model works fine under that mode of operation. Only if you decide to not always take the highest probability item...

120

Piotr Wilkin @ilintar.org · 24/09/2026

Tokenization is a deterministic translation from natural language to sequences of tokens. Then tokens are represented as integers and fed as an input tensor into the LLM. What happens next is a series of operations that end up in an output tensor that represents something called "logits" - the...

120

Piotr Wilkin @ilintar.org · 24/09/2026

Fun fact, because I just saw a proposal by Emily Bender for people to use "probabilistic automaton" instead of "artificial intelligence" - LLMs aren't, at their core, "probabilistic". Yes, you can add randomness to introduce variety to the models, but the core algorithm is *deterministic*.

220

Piotr Wilkin @ilintar.org · 24/09/2026

I would agree with most of those, but thinking that people will actually use "probabilistic automaton" instead of "artificial intelligence" is pretty wild.

110

Piotr Wilkin @ilintar.org · 24/09/2026

So that's basically in an hour+, right?

000

Piotr Wilkin @ilintar.org · 24/09/2026

Can we please have some contempt finally?

0211

Piotr Wilkin @ilintar.org · 24/09/2026

US crony capitalism in one picture: (full text: theintercept.com/2026/09/23/d...)

062

Piotr Wilkin @ilintar.org · 23/09/2026

If I were the Dem leadership right now, I'd make sure news of every Rep candidate who distances themselves from Trump reaches Trump 😃

010

Piotr Wilkin @ilintar.org · 22/09/2026

Yeah. Those are the guys that will never, ever take on the big business. When forced to make a choice: tackle big business or embrace a full-on mafia state - they will always chose the latter - because their position in the latter is high enough that they will not be harmed.

000

Piotr Wilkin @ilintar.org · 22/09/2026

Waiting for Asmongold, though he might not get out of his basement often enough...

000

Piotr Wilkin @ilintar.org · 22/09/2026

I wish him a successful lifetime career alongside his masculinity influencer friends, Andrew and Tristan Tate.

000

Piotr Wilkin @ilintar.org · 20/09/2026

No dobra, to się Tacie Maty udało.

0133

Piotr Wilkin @ilintar.org · 20/09/2026

Raczej: żeby ta druga ruska swołocz się utrzymała pod progiem.

100

Piotr Wilkin @ilintar.org · 20/09/2026

I was just about to say, this is the level that my 9-year-old was at about a year-and-a-half ago, he's already outgrown it since :)

000

Piotr Wilkin @ilintar.org · 20/09/2026

Nie, właśnie dlatego cieszy mnie każdy cios w imperium Putina, który przybliża koniec tej wojny, a co?

100

Piotr Wilkin @ilintar.org · 20/09/2026

Gotta admit, there's few things that warms the Polish heart as much as waking up and seeing Moscow look like this :)

1112

Piotr Wilkin @ilintar.org · 19/09/2026

WTF. Of all times, how do you go over to the side of Bari Weiss *right now*?

030

Piotr Wilkin @ilintar.org · 19/09/2026

Fakt: czas na wygenerowanie tych wysrywów jest dłuższy niż czas potrzebny na wyniesienie posegregowanych śmieci.

030

Piotr Wilkin @ilintar.org · 18/09/2026

...just plowed right through them, naming their attacks as bullshit culture war fearmongering. I think him surviving all the stuff thrown at him by Orban's cronies has left him immune to such manipulations. Democratic politicians should take note - you do not beat thugs by backing down.

000

Piotr Wilkin @ilintar.org · 18/09/2026

I love how Magyar, who was probably submitted to more bullshit propaganda attacks than any US politician, casually dismisses the right's culture war points. They already tried with trans rights and gay adoption and in both cases (though he's generally a conservative) instead of walking back he...

110

Piotr Wilkin @ilintar.org · 18/09/2026

To jest ta katolicka prawica, która struktury grzechu widzi wszędzie, tylko nie we własnych szeregach. A potem zdziwko.

040

Piotr Wilkin @ilintar.org · 18/09/2026

There are no "legitimate local concerns". It's that simple.

020

Piotr Wilkin @ilintar.org · 17/09/2026

...to be evaluated in the most charitable circumstances possible. The only way you can *actually* make the mentioned claim work is if you insert one more premise: that the differences between the global poor and global rich are *not* incidental. And you know what that is? You guessed it: racism.

050

Piotr Wilkin @ilintar.org · 17/09/2026

In case of people from rich countries, most of their potential is realized already, i.e. if they didn't invent a genius cure for cancer by now (and they're not in a team doing so now), there's very little chance they will do so in their lifetime, simply because their potential has had a chance...

130