Lightnews — Scholar-powered news

Grace

@gracekind.net

4.4K followers 1.5K following 10K posts

A latent space odyssey gracekind.net

Posts Media Videos Starter Packs

Pinned

Grace @gracekind.net · Dec 15

What is ideonomy, anyway?

I'm so glad you asked!

Ideonomy: A Science of Ideas

An accessible introduction to ideonomy

gracekind.net

10 17 130

Grace @gracekind.net · 43m

Yeah, I think they rely on the system prompt / context to give it that info, which is probably the right strategy overall but is confusing in this instance

Grace @gracekind.net · 45m

Oh good idea! I think this is the most up-to-date version www.alignmentforum.org/posts/jP9KDy...

How To Become A Mechanistic Interpretability Researcher — AI Alignment Forum

Last updated Sept 2 2025 • TL;DR * This post is about the mindset and process I recommend if you want to do mechanistic interpretability research.…

www.alignmentforum.org

Grace @gracekind.net · 50m

It might be the case the model is getting its own name wrong. This is especially possible via the API because it doesn’t have a system prompt

1 4

Grace @gracekind.net · 58m

This thread might also be helpful!

Mel Andrews @bayesianboy.bsky.social · 4d

What problem is explainability/interpretability research trying to solve in ML, and do you have a favorite paper articulating what that problem is?

Grace @gracekind.net · 1h

I don’t have any go-to resources but I’m sure someone here does! @vgel.me maybe

Grace @gracekind.net · 1h

I have an intuition that this might depend on the way the computation is being implemented, but I need to think about it

1 2

Grace @gracekind.net · 1h

I think clarity on this is good. But it also flows the other way- I think a lot of people are actually concerned with *any* qualia in a system, not the qualia of software or an algorithm. So this particular argument doesn’t resolve those concerns

Grace @gracekind.net · 1h

I’ll also say my own conception of qualia is quite narrow. And I leave room for some weird stuff at the software level that’s probably unknowable to us and not worth worrying about right now

1 1

Grace @gracekind.net · 1h

Gotcha, I think we’re in agreement then!

1 1

Grace @gracekind.net · 1h

gm fellow bizarro vibe calculators

Zoomer Antimillenarian ♨️ @surcomplicated.bsky.social · 3h

What's worse, humans also probably perform addition via similarly cursed mechanisms.

I don't know exactly what's going on in my head when I do addition, but until I get to a number big enough where i have to crack out pen and paper, I'm pretty sure it's some bizarro vibe calc like this.

4 2 36

Grace @gracekind.net · 1h

You could also argue that qualia in the first instance is ephiphenomenal, but if qualia is a necessary aspect of human cognition (or has bearing on the output) then I don't think this is the case. Or perhaps your argument is that this adds variance which makes the system not a true Turing machine?

1 3

Grace @gracekind.net · 2h

However you could still draw a line in the former scenario where you say the machine "itself" does not experience qualia

1 5

Grace @gracekind.net · 2h

I think my definition is broader, being "something that can implement any computer algorithm." So a machine comprised of humans performing bitwise operations would qualify (and qualia would be inherent in that system), whereas a water-based machine running the same computation would not

1 5

Grace @gracekind.net · 2h

This sort of implies that the machine runs as-is and then qualia is layered “on top”, but could qualia be a necessary component of running the machine? (On a particular substrate)

1 4

Grace @gracekind.net · 2h

Insofar as a Turing machine is fully defined by its program definition, I’m inclined to agree with you. But the part about qualia not impacting the operation of the machine is unclear to me.

1 6

Grace @gracekind.net · 2h

Grace @gracekind.net · Aug 3

Clippy saying: "It looks like you're shadowboxing your own demons in my replies. Would you like to stop doing that?"

Grace @gracekind.net · 2h

That paper isn’t making a religious argument, although I’m not sure it fully supports the claim that Turing machines can’t host qualia

1 4

Grace @gracekind.net · 3h

The discover feed is algorithmic, it’s just not a very good algorithm… I would also recommend the For You feed that @aparker.io linked!

Grace @gracekind.net · 3h

This imagery was staring us in the face this whole time…

Rorschach with a red circle drawn around his fedora

1 8

Grace @gracekind.net · 3h

They know the people demand cubes

Grace @gracekind.net · 3h

I think it is analogous to predictive coding. In that sense “all models are wrong, some models are useful” is itself correct, and of varying utility :)

1 6

Grace @gracekind.net · 3h

rev. howard arson @theophite.bsky.social · 5h

arxiv.org/abs/2502.00873

Language Models Use Trigonometry to Do Addition

Mathematical reasoning is an increasingly important indicator of large language model (LLM) capabilities, yet we lack understanding of how LLMs process even simple mathematical tasks. To address this,...

arxiv.org

Grace @gracekind.net · 3h

Ostensibly the reason for doing this is redundancy/error correction (between linear value and circles, as well as between different circles), as well as those representations being useful for other tasks (e.g. the 2-circle encodes whether a number is even or odd)

Grace @gracekind.net · 3h

Some of the math is beyond me but my understanding this: It represents each number not only as a linear value but also as positions on a number of circles (representing 2, 10, 100 etc). Then it adds numbers by combining the rotations on each of those circles.

1 10

Grace @gracekind.net · 3h

Sure! Basically, the model is performing addition via a fairly cursed mechanism that we didn’t expect.

3 21