so to my best understanding the "parrots" position requires me to believe either: 1) this exact pairing appeared in the training data, and the LLM did not figure it out for itself 2) the LLM was this funny by dumb fucking luck no i think the LLM understands the joke actually

Absolutely. For reference, my AI assisted essays all came out of at minimum multi hour conversations and iterative refinement. I care about them and put significant effort into them. I stand by what they say. I wish there was a shorthand for this kind of high effort AI assisted writing...

This is a work in progress but I made a visual that compares how much water AI and data centers use in total to other industries. Relies on a lot of botecs I justify in the methodology. www.andymasley.com/visuals/water/

"Stochastic Parrots" is wrong

"The novel wasn’t obvious. It wasn’t planned. And it wasn’t what writing was “for”." As my "Rotations" essay is making the rounds again, let me bump it's companion essay which didn't get as much attention.

imho the real superpower of LLMs is stretching an idea onto a dissimilar framework Getting them to stretch and bend ideas can be a really amazing source of inspiration and understanding

My post in praise of cognitive offloading andymasley.substack.com/p/the-lump-o...

Also important to note that when you pit your credentials against reality, reality doesn't lose.

7/ Some of the opposition to AI and its use in writing is status protection dressed up as principle. Resistance often reflects institutional gatekeeping, not genuine intellectual concern. We should evaluate the quality of the product carefully regardless of whether it's 100% human-produced.

dame is correct here Everyones thinking was already outsourced in countless ways since the dawn of humanity. As well as a lot of labor. Thats what society is.

I think what's most important for most people isn't whether some analogy is in some sense "correct", it's whether it gives people useful intuitions. And autocomplete and stochastic parrot don't tend to lead to strong intuitions about the capabilities or impacts of llms.

“it’s just spicy autocomplete” is advocacy malpractice, is the thing. the sneering dismissal completely miscalibrates the urgency: media siege devices —printing press, telegraph, internet— rip holes in civilization itself

The AI discourse sometimes seems to center on "Is AI good or is it bad?" I find this framing unproductive. AI is not a fixed thing. I would prefer to ask "How might we use this technology for good, and mitigate the bad?" What a shame if the best use we can come up with is no use at all.

Moriarty was a star fleet red teaming exercise that escaped.

The only sense this (see picture) makes any sense to me is if prompting just doesn't work... it's not that it's wrong. It's not even wrong. So I prompt it with coding guidelines so it's stylistically nothing like average stack overflow. So then it's the average of what exactly?

Which raises the question, if you have high quality energy matter converters, why do you have a holodeck, not a "actual real physical stuff created on demand" deck?

I love this analogy and it's fascinating. Replicators as the closest start trek has to generative ai? You prompt it and get the thing you asked for... maybe? If you explained well enough or it inferred or you're referring to an established recipie.

And it is a medium. One if the most powerful and frankly weirdest ever created. And it needs it's champions. It needs it's innovators. Read the full thread.

This is an important thread. My lived experience is different, though my best guess is my deconversion experience was particularly formative in this regard, and while different to Doll's story, has similar elements like grappling with identity in the light of flawed authority.

New essay on LLMs and brain rot: "LLMs do not inevitably corrode thinking. They amplify whatever epistemic posture you bring..." sbgeoaiphd.github.io/rotating_the...

This is very good! I highly recommend reading, so long as you're open to the idea that brain rot is contingent, not inevitable (something which I unpack somewhat less rigorously in an upcoming essay).

These are fantastic well baked thoughts from Cat!

The reason I think this motivation effect is interesting is because it is *another* interesting counter to the bad outcomes we don't want (passive, skill loss). Ppl experiencing unexpected rapid changes to their evaluation of what's possible GET MOTIVATED. They want to do more, not less, solve more

I agree with most of this but I don't think curiosity is the sole necessary component. There's more going on which is necessary for turning that curiosity into productive learning and thinking.

100% LLMsare having the strange effect of splitting apart crafts which used to necessarily coexist. You can solve problems and build things without writing code (though programmatic thinking and understanding the code remain valuable).

I think what'll happen is that knowledge gets (even more) k-shaped between those who are curious and those who are not. For people that are, AI is an *insane* learning aid. An infinitely patient tutor who knows basically everything ever. For those who aren't, it can just think for them. ->

Look, I'm just going to say it and y'all can flame me for it: if you're doing group assignments with other humans in a way which makes you not learn anything, that's a skill issue.

Some people are really like, "AI uses too much water and electricity, now if you'll excuse me, I need to eat a cheeseburger while watching extremely high resolution streamed video while sprinklers douse my private plot of grass in potable water"

Super useful and interesting read on generalization, memory and overfitting, highly recommended!

An auto regressive LLM is a sequential colony of forward passes. The forward passes are individual units which collectively create emergent capabilities in the "colony", much like an ant colony has emergent capabilities individual ants (by their nature) lack.

For mice-eyes only. "Granite has made its position on water quite clear." sbgeoaiphd.github.io/rotating_the...

This makes sense of my experience co writing even short stories and having to actively battle the model's desire to "be helpful to the reader" and remove the deliberate misdirection and obfuscation.

You can post yet another AI-generated caricature. Or you can post yet another complaint about AI slop. Or you can ask AI to imagine infamous in-universe essays from Star Trek...

This is a great refinement. The problem is we don't have a functional and widely used language to describe this yet and most people assume use of AI is low effort by definition.

this would be a fun 'long now' style project - imagine dusting off a bunch of laser-cut titanium tablets from a dead civilization and it's just their ghosts, their open-weight models, their last trace a mere shadow cast by a shadow resurrected in silicon

Once again people who complain about AI hallucinations showing absolutely no care for accurately understanding a source.

The way I treat people isn't contingent on me taking strong positions about unknowable metaphysical questions.

god of the gaps. but for human exceptionalism.

"Taking Jaggedness Seriously" by Helen Toner talks about this in some depth. helentoner.substack.com/p/taking-jag...

I think there's a level vs shape distinction and intelligence isn't a simple linear scale. AGI as human "level" falls apart because it implies human shaped. We could have arguably ASI which is still not human shaped and therefore we still hire humans etc.

The future belongs to the word rotators.

A useful starting point for reflection on any particular AI use case is "how would we think about this if it was a statement about human-human interaction"?

A lot of criticisms of chatbots seem to imply that it is incredibly harmful to your knowledge and ability to think to use a human research assistant

Embeddings don't encode meaning. They encode affordances for meaning recovery, given the right probe.

Precisely. People don't learn well when using cognitive shortcuts. No shit...

This is not true; I beg people read the full paper and especially the study design. The conclusions mirror my (and many other practitioners') conclusions: if you use AI critically and engage both with the question and the answer, it has a net positive impact on both learning and productivity

There are 2 kinds of forced as well. 1. AI included in apps where people don't want it, 2. The society level "have to use it to keep up" - increasingly true in some technical fields like software engineering and likely to become more widespread.

Huge water consumer, destroying the planet with CO2 emissions, mostly a negative for their consumers. Something must be done. Yes, that’s right, it’s time to do something about cows.