I'm not going to run a long thread on this because the paper is very detailed and quite enjoyable to read, but here's some of the things I loved 2/n
The central idea of the paper is beautifully laid out at the beginning: use an open/closed pair of modalities on types to treat *intension* as structure on *extension*. So cool! 3/n
These modalities are generated by a certain predicate ¶_E which classifies the extensional phase. This seemed odd to me at first, but then this example made it so intuitive: 4/n
The other key idea of calf is to use model computational cost as a side-effect. Hence 'cost structures' become effectively algebras of a writer monad for a monoid of cost (e.g. running-time would be Nat) 5/n
So basically the extension modality 'forgets' step^c: 6/n
In the paper you find a full specification of 'calf', some categorical model theory (to prove it is non-degenerate), and most importantly, examples of verifying the running time of a program (such as an implementation Euclid's algorithm)
And there's an Agda formalization too! 7/7
• • •
Missing some Tweet in this thread? You can try to
force a refresh
here's the deal: proofs are mainly artifacts (usually text, but figures count too!) mathematicians produce to convince other mathematicians of some fact about their *shared* imaginary world. without the *shared* part, they'd mean nothing. 1/n
specifically there is no such thing as a 'correct proof', there is only a consensus about which proofs are correct. correctness it's not an objective fact. 2/n
And now a very interesting concept/observation: some emergent effects have intrinsic significance since they feed back in the components of the system which gave rise to the them:
Hence this classification of emergence (paraphrasing Cruthfield below): 1. 'something new appears' 2. 1 + observer identifies a pattern 3. 2 + the observer is part of the system itself (strong 2nd order cybernetics vibes from this one)
@math3ma just gave a very interesting talk about this paper, with wonderful intuitions
My understanding of the situation is the following (and I hope she'll correct me if I'm wrong). At a first approximation, a lang model (LM) learns a Markov kernel π:X->X where X is a set of strings.
The question is, what structure shall we expect this kernel to have?
The idea of π is that given a string x:X, π(-|x) is the probability measuring the likelihood of a given string y:X to follow x. Hence π(y|x) is π(xy) up to normalization.
Idea: the structure of scientific revolutions identified by Kuhn is an instance of the more general features of evolutive/inferential dynamics. Available evidence provides the selective pressure for scientific theories.
For instance, lack of selective pressure produce adaptive radiation in evolution. en.wikipedia.org/wiki/Adaptive_…
This is analogous to the pre-paradigmatic phase of a science, where lack of evidence produces a plethora of alternative theories and models.
A 'revolution' would correspond to speciation/extinction, i.e. the strong selection of a few successful traits (revolution/crisis). Then for a long time these traits don't vary (paradigmatic periods), giving rise to punctuated equilibria. en.wikipedia.org/wiki/Punctuate…
David Spivak delivered one of the best motivational talks about ACT I've ever seen:
It's a replica of his NIST talk from last week, here's a few key points I personally vibed with 👇🧵
First: ACT is about better communication and better language *for SMEs* (Subject Matter Experts).
The corollary (this is me not David) is you shouldn't exact applications from applied category theorists.
It's not our job!
We provide the fishing cane, not the fish.
Second: mathematics as account systems. It's all in the slide. The example is great because is so easy to disregard $s as 'just numbers'.
Mathematicians (but CTists are somehow more sensible to this) know they must be parsimonious with structure.
👉🏼 Fibred categories are like woven fabric and doing a Grothendieck construction is a like weaving on a loom: a... thread about textile intuitions for fibrations 🧵👇🏼
The threads of woven fabric, when you look up close, are entertwined in a distinctive pattern: some of them run vertically (that's the *weft* or *woof*), and some run horizontally (that's the *warp*). The *bias* is the 'diagonal' direction, along which fabric is easy to stretch
Likewise, when a category E is fibred there is a factorization system that tells you how to decompose each morphism into a 'warp' part and a 'weft' part.
If we forget about the weft, we can project down our fabric on the selvedge. This projection is the fibration!