If Sam Altman only ever said reasonable things about the capabilities of ChatGPT I’d still want to see him disgraced and his company nationalized. With 95% of the water use people, if they stuck to “I don’t like the real negative social / environmental /economic consequences” then, like, same.

The Google weather forecast AI use Graph NNs, not Transformer, so in that sense it’s less related to LLMs than most currently deployed ML models are. On the other hand (as the authors themselves note) GNNs and sparse transformer encoders are different ways of labeling the same operations.

This is a Kerbal Space Program level understanding of the problem. Yeah, at the highest level it could work. It just presupposes a ton of engineering advances and perfection of several technologies that haven’t been demonstrated beyond small scale proof of concept.

can I have one day of scrolling without BLOOMBERG: BREAKING WHAT IS COMING TEN MILLION PEOPLE IN THE STREET CANNOT STOP WHAT IS COMING THE AMERICAN ARMED FORCES CANNOT CONTROL THE PRESIDENT OF THE UNITED STATES                 AND HIS COUNSELORS CANNOT CONCEIVE

I don’t have a low opinion of Genre Fiction. I love a lot of it. A lot of it is good. It has artistic merit. But I think it’s hard for Genre Fiction to be *Great Art*. Specifically, *Great Art* either creates a genre or destroys one.

If we want to apply this to short form content, it’s an interesting question with a complex answer. But for novel length content, the ability to fit novel length content into a model has been around for ~18 months. The ability to coherently *use* that length arguably still doesn’t exist.

This applies just as well to the fascination in some circles with trying to extract “secret” system prompts via clever questions. You are creating an ideal situation for hallucination. Even plausible, consistent output should be viewed skeptically (it could just be an attractor).

Was not expecting the Derry villain motive to be a more bigoted version of this. www.smbc-comics.com/comic/2011-0...

Gonna share My Ultra Gamemaxxing Master Strategy to assured success with women: 1. Cultivate a basic interest in a hobby or topic that is woman-coded. Fashion, celeb gossip, dating reality tv, w/e. Enough that you have at least one sincere opinion about it—that’s not “this is lame”. Last part is key

PSA for the kids: never do a PhD unless they're paying you. Even then, it's still a bad choice - but one you make because of a love of the game.

Hating AI for CV really is the epitome of motivated reasoning. Every part of VLMs is a natural generalization of pre-neural vision methods. It’s not how the real dev turned out, but there’s a world where we derive ViTs before LLMs.

The rest of this travesty aside, I hope that if a “sympathetic” organization released a screenshot of my writing they’d have the decency to turn off grammar-check highlights first.

“Aliens” covers a lot of ground here. Microorganisms: maybe even elsewhere in the solar system. Complex life: almost certainly somewhere Life like us: probably not unique, maybe rare enough we won’t see it “Aliens” that come to earth and interact with us: haha no The rest are much more specific

Struggling to reconcile my sneering dismissal of Prompt “Engineering” as a discipline with my just having written a 20k word document for work on how to do it effectively

Dennett had plenty of problems, but his dismissal of the argument by intuition pump has always stuck with me. If your “thought experiment” involves pushing the subjects beyond the bounds of anything we’ve observed, it’s a bad argument. Usually it amounts to a bare assertion of the thesis.

“No way to prevent this” says only language where this regularly happens

Brings to mind Sagan's not being "so open-minded your brains fall out." There is a lot of value in the critique of positivism, de-centering one's own viewpoint, etc., but we have to stop before sophistic absolute relativism. Or, if you want, Nothing is True isn't My Truth.

Even the idea that we need to explain the previous movies at all (“007 is a code multiple people have held”) is too much for me. New works can have references or homages to old ones, but the idea that we need a Bond continuity is absurd.

There's the overwritten AI style--which can mostly be prompted away--but the larger weakness of AI for this kind of project is that it's bad at coming up with a new, coherent world. They struggle with things like having a unifying theme or connected story and setting.

Beyond the race “science” bigotry here, this is a classic failure mode of ML. Even if there were somehow a real effect here, it would be much smaller than the race / sex biases in the training set by virtue of coming from our society. Models will always learn the easiest signal they can

“It’s bad ontological hygiene to block people just because you disagree” is maybe a defensible position—in the sense of “you should choose not to do it”. The Special Block Police reviewing your blocks for wrongthink is such a bad idea that it feels like a refutation of the original position.

I was willing to entertain the idea that so many anti-AI people blocking her was, if not quite bad, at least a symptom of something… not ideal maybe. I was willing to listen. This is pathetic, though. No one is owed an audience. Policing blocks would be an insane choice.

Not be fully conspiracy brained, but are we sure the AEI didn’t cook this up to be as maddening and disheartening as possible? There’s no claim these people are randomly selected. This feels more like a push-poll than a genuine exercise.

Betting against LLMs having something worth calling "understanding" is like betting against Bell Inequality Violations. It was always a little dubious, but not entirely out-of-hand dismissible at first. As evidence has mounted, there's still some little path to truth proponents can cling to.

There's a seductive idea that's steadily become more popular from ~ the 19th century up through today. You see it behind fascination with the Nazi military, the Cult of the Operator, Post-liberalism, the "Dark Enlightment", etc: the chains of morality, empathy, humanity are holding us back.

Apropos of no other discussions of unstable loci of meaning and the size of numbers, here's my first test chat with an old RAG demo I made that uses tools exclusively to find relevant passages from philosophy ebooks I pirated for class in my undergrad. Its takes are much hotter now.

There's this amazing Stancil phenomenon where he makes a post that maybe kinda is a Motte / Bailey ("'Love' is a four letter word"). He almost never says the Bailey part explicitly (here it would be something like "Love is pointless shit") . I make no claim as to what Will's intended reading is.

In general, this kind of Post-Truth "Deep Battle" doctrine is a loser for us. For Trump specifically, though, he is both broadly culturally relevant and has no fixed beliefs. Do you think my AI Trump is wrong to endorse Medicaid for All? Here's a video of him saying it's great from 2016. Is it real?

It's surreal that this has such different reaction here to the recent worries about people "befriending" or "loving" chatbots. Like---in that case people here can understand "LLMs are dangerously good at mimicking human interactions and many people struggle with emotional distance with them."

This may be the most fun I’ve had making diagrams for a presentation

As someone who is (sadly) not a layman at this, this is a good description of the problem. It's a design-level failure if your "do I bomb the hospital?" system has to roll a die to decide between "yes" and "no", even if "yes" is *really* unlikely.

The precise way in which LLMs model high-level concepts is not well understood, but the low-level way they model structures like Language *is*. The most intuitive starting point for Transformer--if you already have a strong ML / math foundation--is that Transformer is "Dynamic Graph Perceptron"

This is a great overview of the interpretation (and non-interpretation) of LLMs. There's an orthogonal point I think is important though: LLMs are built with in a way we *expect* to be able to produce a effective model of language that is hard to interpret. Our current situation is not a surprise.

Somehow GAN mode collapse returned

This is both true and grading the model on too much of a curve. The model clearly has memorized the token to letter correspondence—for other models where you can get the reasoning trace, models explicitly separate the letters and count them “manually”. GPT-5 is just still shockingly bad at counting.

All of the big models released this year can answer this question well without tool calling. We can try to tell ourselves that this is just because the model has memorized every possible description of geographic relations between states, though this isn't a common benchmark or thing people discuss

Heard someone say (again) today that "Idiocracy was a documentary," and I thought back to the excellent @sarahz.bsky.social's video on it: www.youtube.com/watch?v=o52z... I don't disagree with her. I think her points hold up. But the current moment and the movie still seem to have "resonance".

There’s this incredible, persistent innumeracy in the replies to this and the original where people say “but I can eat beef! LLMs are useless!” Or “as if we shouldn’t just stop eating beef.” *Training* one of the mostly costly models to produce to date used maybe *1* entire cow.

A fun hobby project I’ve been working on is seeing how much (with a mixture of RAG and prompting) I can force a model out of the groveling, obsequiousness Instruct-tuning tends to give them. I present “The Lacanalyzer”, a Qwen3 based chatbot with a condescending obsession with critical theory.

There's a deep pit in the center of answering "can LLMs have knowledge?" First, we need to answer "what is knowledge?" and then "do I have knowledge?" So many definitions that give an easy "no" answer to the first question also make "no" to the last one hard to avoid. Epistemology is hard.

A problem with LLM discourse is that LLMs are uncanny. An LLM can do things that in a human would be a sign of advanced skill. Tech-bros point to this as and triumphantly declare they are as smart as "Post-Grad Researchers, but in every field.". They are wrong.