@disconcision mentioned you in a photo

@nostalgebraist : so intriguing! raises a lot of questions! is frank (still?) blind to the notes? starting to worry about my potshots here. also, is mood bounded? personally hoping for a mood singularity due to self-reblogging of self-encouragement

If I understand you correctly, no, Frank isn’t blind to the notes – that is, Frank is aware of replies and direct reblogs, and they interact with the new mood feature.

(The “selector” feature, which has been around for a long time, is also aware of the notes in a different way – as raw counts – but I don’t think that’s what you mean.)

If you want to discuss a Frank post without worrying about mood effects, here are some ways to achieve that:

• reblog the post via an intermediary (requires someone else to have reblogged it first)
  
  
• send me an ask or message
  
  
• make a new post on your blog with whatever you wanted to say, with (if you like) a link to the relevant Frank post.
  
  (note that in this last case, Frank might still reblog the post if she’s following you via the !follow command, but even if so, this won’t have mood effects – posts seen on the dash don’t affect mood, only things “said to” Frank like asks and direct reblogs affect mood)

To your other question –

also, is mood bounded? personally hoping for a mood singularity due to self-reblogging of self-encouragement

The value on the graph isn’t actually bounded, so in principle it could get really large in magnitude for some period of time.  It has dynamics that exponentially relax it back to a baseline, though, so the magnitude wouldn’t stay high for long without some continual driving input.

Since mood only affects the content of Frank’s posts, and is affected only by user input (not Frank’s own posts), any positive feedback loop would have to be a two-party thing where users give happy/unhappy responses to happy/unhappy posts – it can’t happen from the system alone, without an element of human response.

The effects of this value on the posts are themselves bounded, BTW:

• The mood value is converted into lower and upper bounds on a certain kind of sentiment score (output from a sentiment predictor), and candidate posts outside those bounds are rejected.
  
  
• The function from (mood value) –> (lower bound, upper bound) just interpolates between members of a discrete set of “named moods,” which are pairs (lower bound, upper bound) that seemed empirically reasonable for capturing things like “only sad posts,” “only non-sad posts,” “only happy posts,” etc.
  
  (Originally there were just the discrete “named moods,” but then I wanted to make an underlying continuous variable, so I interpolated between the discrete elements that I already felt confident about using.)
  
  Anyway, this function just returns the saddest “named mood” for all sufficiently low inputs, and the happiest “named mood” for all sufficiently high inputs (while interpolating in between).  So its outputs are bounded.

disconcision replied to your post “She is dressed, now, in a black vest thrown over a dark gray-green…”

did you switch to the 345M model at some point? this is disturbingly coherent

Yeah – I switched @uploadedyudkowsky over to 345M shortly after the release of 345M, which conveniently coincided with a point when I was getting tired of curating the 117M output

It’s amusing to reflect on how much @uploadedyudkowsky has “improved” since I started the blog.  Originally I was just doing word-level Markov chains, which are an old favorite of mine (first learned about them through Janusnode, which I first used back when it was called “McPoet,” sometime in the … early 2000s?).

When I heard about char-RNN, I started using that, which allowed for a lot more “creative” variability although I still had to laboriously hunt for funny-sounding phrases buried within larger swaths of mostly dull/gibberish output.

Then when I heard about people fine-tuning GPT-2, of course I had to try that, which got vastly better results – now the curation challenge is less “find some phrase in the output that doesn’t suck” and more “I want to quote some gigantic 15-paragraph stretch but I’m worried people will just tl;dr unless I choose a smaller excerpt.”

And now it’s the 345M GPT-2.  And I’ve established this sort of personal tradition where every time I hear about a new gee-whiz NLG method with freely available code, I’ll try it on this same Yudkowsky corpus and revive @uploadedyudkowsky.  I’m sure I’ll do it again with the larger GPT-2s, and then whatever comes next.

The really funny thing here is that if you look over the blog, it looks like something “gaining more intelligence” dramatically and qualitatively over the course of a few years.  But it’s a thing that I, a perpetual (if moderate) AI skeptic/pessimist, am doing to poke fun at Yudkowsky, whose whole deal is worried about rapidly improving AI.  So it feels almost like he’s getting the last laugh!  It’s an extra, completely unintended/emergent joke on top of the basic joke underlying the blog.

(On a related and more serious note, I do think I’ve become less “AI skeptical” in certain ways – which are still vague in my head and need more thought – as a result of recent successes with the “transformer” architecture.

Like, a few years ago if you had told me we’d have GPT-2-level NLG in 2019 I imagine I would not have believed you.  But what’s more, the same architecture that GPT-2 uses for NLG also enables some really incredible stuff in supervised NLU, via BERT, which can get you state-of-the-art results on p much any supervised task with a few epochs using one-size-fits-all hyperparameters.  I was like “sounds fake but OK” when I read the paper, and then I tried it on some proprietary tasks from my day job and it just worked.  Sometime I want to make an effortpost about the transformer architecture, because there’s something magic going on there, and none of the explainer posts out there do justice to the intuitive simplicity of the thing.  [Very briefly: it’s a lot like a CNN in terms of using sparse filters, but the “shape” of each sparse filter is computed dynamically from the input via a function learned from the data])

disconcision replied to your post “ There’s something a little mysterious to me about the usage of “the”…”

aliens guy: “category theory”

i feel like what you’re getting at is that we can fall into an intuitive sense that equality of objects lives in some kind of objective arena. what we find out in math (or just thinking precisely about anything, i guess) is that this is abjectly not the case. equality is always with respect to some underlying category/type/whatever, and which arena we choose as our ground determines which things are equal and hence what is unique

also reminscent (though tangentially): http://math.ucr.edu/home/baez/qg-spring2004/s04week01.pdf

Ooh, yeah, those Baez notes seem to be talking about the same thing as my “constraints” and “structure” (except he says “properties” instead of “constraints”).

Anyway, your second reply feels right, but I think there is a little more here than just “equality of objects is relative” – that sounds like an observation about some independently defined things called “objects,” as if we have a good handle on what an “object” is but not necessarily on when objects are equal.  But it’s actually that way of looking at things that (in various specific cases) tends to feel wrong to me.

It feels like there is a tension between two ways of thinking which are both supposed to be hallmarks of modern/higher math: formalism and abstraction. Formalism tells you that the explicit capital-D Definition of an object is the ultimate source of truth about it, and closer to “what the object really is” than the set of the motivating examples you keep around in your head and use for intuition. But abstraction tells you to care only about intrinsic patterns/structures and not the contingent ways they may happen to be encoded. From the perspective of abstraction, a formal definition is just a way of expressing a pattern, and our intuitions can get at aspects of the pattern that the definition misses. (E.g. if we move away from set-theoretic foundations, no one is going to say that the word “group” can’t be used anymore because a group just is “a set equipped with (etc).”)

To continue on this riff: formalism has this problem where it allows you to start with definitions that have more structure than you really want, and then happily carry it around with you forever, expressing its irrelevance by saying “two objects that differ only in that way are isomorphic” — as if this is some further fact you’ve happened to learn about the pattern you are studying, when it’s really a fact about your (bad) encoding. For example, for any kind of object based on a set, we could imagine forming a stupid variant of that object where the set is ordered (a tuple), and then all of the results would be the same except we’d pointlessly act like there were multiple copies of each instance (one per ordering) that “just happen to be” isomorphic. I don’t think anyone does exactly this, but there’s this uncomfortable feeling that the same kind of error could be happening in fancier ways without us noticing.

Two things I think are kind of interesting about this:

(1) I feel like programming computers has shaped the way I think about this stuff. I’m used to drawing the distinction between the data I want to store and the data structure / encoding I use to store it, since the former is usually fixed by the problem at hand but the latter can be chosen and matters for speed, etc. So, in programming, formalism — taking one way of encoding the data and saying it is the data — is recognizably a bad habit, which will prevent you from finding better encodings. (The set vs. tuple thing from the last paragraph is a common practical issue in my everyday work, that’s why it came to mind!)

(One could imagine a parody of formal mathematics in which every definition starts by telling you that the object is stored as JSON.)

(2) It’s kinda cool how this tends to connect very abstract and very down-to-earth ways of looking at something, distinguishing them from some middle ground. You start out with an intuition that something can be abstracted from some examples, then you write a formal definition of the abstraction, but then as you prove more equalities / isomorphisms, you find more and more ways that your “naive” original examples are completely representative of other things, while the formalism can get more and more (but not less) misleading.

Arnold mentioned a few cases like this in “On Teaching Mathematics”:

What is a group? Algebraists teach that this is supposedly a set with two operations [two?? -nost] that satisfy a load of easily-forgettable axioms. This definition provokes a natural protest: why would any sensible person need such pairs of operations? “Oh, curse this maths” - concludes the student (who, possibly, becomes the Minister for Science in the future).

We get a totally different situation if we start off not with the group but with the concept of a transformation (a one-to-one mapping of a set onto itself) as it was historically. A collection of transformations of a set is called a group if along with any two transformations it contains the result of their consecutive application and an inverse transformation along with every transformation.

This is all the definition there is. The so-called “axioms” are in fact just (obvious) properties of groups of transformations. What axiomatisators call “abstract groups” are just groups of transformations of various sets considered up to isomorphisms (which are one-to-one mappings preserving the operations). As Cayley proved, there are no “more abstract” groups in the world. So why do the algebraists keep on tormenting students with the abstract definition? […]

What is a smooth manifold? In a recent American book I read that Poincaré was not acquainted with this (introduced by himself) notion and that the “modern” definition was only given by Veblen in the late 1920s: a manifold is a topological space which satisfies a long series of axioms.

For what sins must students try and find their way through all these twists and turns? Actually, in Poincaré’s Analysis Situs there is an absolutely clear definition of a smooth manifold which is much more useful than the “abstract” one.

A smooth k-dimensional submanifold of the Euclidean space R^N is its subset which in a neighbourhood of its every point is a graph of a smooth mapping of R^k into R^{(N - k)} (where R^k and R^{(N - k)} are coordinate subspaces). This is a straightforward generalization of most common smooth curves on the plane (say, of the circle x² + y² = 1) or curves and surfaces in the three-dimensional space.

Between smooth manifolds smooth mappings are naturally defined. Diffeomorphisms are mappings which are smooth, together with their inverses.

An “abstract” smooth manifold is a smooth submanifold of a Euclidean space considered up to a diffeomorphism. There are no “more abstract” finite-dimensional smooth manifolds in the world (Whitney’s theorem). Why do we keep on tormenting students with the abstract definition?

disconcision replied to your post “Verbal brain noise (in exaggerated Sean Connery voice): “the name’s…”

i get spontaneous manifestations of this template all the time. the other day i had “The name’s Fuck. Fuck Manchild.”