And I love that you can do that. And I love that you share it. I don't know what your skill in making images by hand is, and honestly? Who the fuck cares. You're bringing things to life with the help of technology. Things that are uniquely yours. Things that wouldn't otherwise exist without YOU.

I would be doing myself a disservice. I would be doing myself harm. This technology is a dream come true. And I understand that by openly talking about it, everything I make from this point on will raise the question: Did José make it? Or did the LLM?

On LLMs as a new medium: sbgeoaiphd.github.io/rotating_the...

this is about so much more than just the ability to generate fucking text this is an instrument for the perception and exploration of a different plane of existence Something which is very, very real but exists everywhere and nowhere It’s an extraordinary thing, no real precedent

#30DayMapChallenge Day 23: Process 🔧 Geospatial embedding datasets allow us to quickly do fairly sophisticated analyses quickly and at a large scale. To make this map, I used the AlphaEarth embeddings dataset in #GoogleEarthEngine to detect changes. Blog post in replies ⬇️

A key piece of received wisdom right now is that LLMs by definition produce "average & predictable" output and can only homogenize human culture. This is where it matters that people envision models working via search/retrieval and haven't thought about high-dimensional probability distributions. +

I am actually very optimistic about the long-run outcome of genAI for the humanities but the reaction *from* humanities specialists is going to delay the benefits to their fields. It's really good when ppl suddenly have a bunch of tools they interact with via effective use of *natural language*.

The way you communicate with the tool is through language in general: multi modal models accept voice as audio. But even besides them, the interface is text which can be produced by writing (well, typing) or voice via speech to text.

To get myself to work with LLMs better, I found it easier to use different kinds of analogies, such as coding like a surgeon, or coding like a tank. This is another one. I think how Geordi LaForge uses "Computer" to do engineering is more akin to how we'll do it in the near future.

"The LLM can generate a view you couldn’t have produced yourself—not because it’s smarter, but because it’s drawing on patterns from across its training distribution. You see that projection, react to it, redirect. The space of accessible projections expands."

People still want shared experiences in entertainment, although there's a market for personalized entertainment This essay points to another alternative: fiction becomes a landscape rather than a story, and the models allow us to shine a light and project our way through many stories

An attempt to express how I principally use LLMs. Rotating the Space: On LLMs as a Medium for Thought sbgeoaiphd.github.io/rotating_the...

"Teach it to me like I'm 5" is PCA and visualise only the first 3 dims on a 1024 dim embedding. Sure, you can do it but don't pretend its the a full expression of the full richness of the embedding space.

Curious if anyone else uses or has examples of using fiction as a thinking tool with llms as I do here. This is not an isolated example for me.

I use LLMs a lot for non technical things, as an interactive thinking space for want of a better description. It's hard to precisely explain what I mean by that because it means a lot of things, but let me give an example...

Somehow he only ended up watching the mirror universe episodes and he's really wishing for the Terran Empire.

GenAI is rotating the skills space such previously entangled skillets become separable.

I wonder if the heptapods could count the Rs in strawberry.

Consider LLMs as a first contact scenario. Most AI discourse is still in the "point weapons and demand it speak English" phase. The magic is in the "learn to draw circles" phase.

Behold the AlphaEarth Concept Menagerie! sbgeoaiphd.github.io/aef-concept-... (Or: "why you should always include linear probes when evaluating geo embeddings")

Here are its methods gist.github.com/dollspace-ga...

I haven't said it like this before I don't think, but I do believe it. Slop peddling is a human failure. It's bad ends from bad means, usually through inattention (underspecification -> hallucination), incuriosity (inability to discern good output from bad before releasing it), or malice.

I've spent a couple of days working on this one, and it didn't end up quite where I expected it to. But I think this is the clearest way I can describe where my head's at, why I approach LLMs in the way that I do, and why, while it's describing-not-prescribing, it's a little light prescribing too.

Human dialogue also does this. To be charitable to the ongoing blowup, it's providing examples of 2 in that there are (rare) aspects of the critique which are largely because I was clumsy in my expression.

The current (very early!) literature on the cognitive effects and learning outcomes of using GenAI is mixed. There are headline negative ones but also positive ones. If one wants to appeal to the literature in this state you need to provide a thesis on why it is mixed.

This is also true for writing.

Standard use will indeed pull towards received opinion but that doesn't mean they aren't still useful for exploring beyond, it just doesn't come from naive use. I fully expect some of Ted's more exploratory students to be using LLMs to poke around at the edges of recieved wisdom and push beyond.

A common pattern for me with LLMs is I start to unpack some idea, the LLM rephrases then unpacks further. Then one of a few things happens:...

I think this entire conversation is suffering from a narrow view of AI as the "essay writing and answers without thinking too hard machine". I think we have actually invented an entirely new medium with way more postures, afforances and uses than we yet realise.

Merry Christmas. Goodwill toward men. New things can indeed be produced by playing old things against each other in clever ways. Happy New Year!

Embeddings exist precisely to shove non-linearity upstream.

Very well put. This tracks to my experience.

If you follow the chain of quote posts far enough, it sounds very much like their model of AI use is without human intervention. Like "go find me something exciting", not "use AI as a tool to explore the space and find surprising things". Perhaps AI can do the former. It can definitely do the latter

You will never have a conversation with a human which is as unconstrained or lacking contextual guidance as when you hit "new chat" with an LLM chat bot. That alone makes it a distinctly different kind if interaction than humans are used to having.

Our communication skills as humans are the result of continuous practice through our lives and depend heavily on Theory of Mind and generally understanding how other people think. We also appreciate that one can learn to be a better communicator...

The Marginal Scapegoat Fallacy. Blaming the newest marginal user for a pre-existing systemic constraint. The Right: immigration The Left: AI water use

... and concluding therefore that knifes are useless.

Equivalent to "I asked a random person to do a really good job at making an index and they failed, therefore people in general can't make indexes". Just wait till LLMs erode folk's critical thinking skills still further...

When I see people show concern about the impact on critical thinking if AI while we have this insane discourse around the environmental impact which is exhibits such a lack of critical thinking, I can't decide between:...

I'm with Atticus here. I find LLMs to be an extremely powerful learning tool. But I don't think the "you can ask followup questions" framing really does justice to what's going on. At least not for me. It makes it sound like it's just being able to click links in a wiki article...

I think this is right—and also think it makes it increasingly important for us to figure out how to describe and convey the needed metacognitive skills.

@andymasley.bsky.social Sit down before reading: www.theguardian.com/technology/2...

Really refuse to let the worst people have ownership of techno-optimism

I'm not sure *how* it happened, but I encounter a lot of undergrads proposing remarkably thoughtful and well-informed indep. study distant-reading projects. It's not like those methods got incorporated in the curriculum! But students are somehow self-educating—possibly with LLM assistance?

As Ethan says in the follow-on post, this doesn't make roles useless. It actually maybe makes them even more useful. Roles are a powerful way of influencing interaction patterns and the mode(s) of interaction are critical to useful human-AI multi turn collaboration...

This is the equivalent of saying farms only produce high fructose corn syrup.

This idea that data centers are just for AI needs to die. They underpin the Internet and the entire digital economy...

Just as there are different ways of collaborating with people, there are different ways of collaborating with AI. And they have different implications for the style of work and the outcome.

They people who say you *have* to learn a lot of abstract maths before you can start learning, let alone working, with ML/AI should know their structure excludes a lot of ADHD folks who struggle when things are abstract but thrive then they are applied.

For me, the promise of language models is that we get to explore new ways of thinking. What if you could fit a million words in short-term memory? What if you could adjust the balance between brainstorming and critique? Are these better ways of thinking? Idk. But we never know that in advance. +