October 27, 2024 at 6:58 PM
I don't know if the problem is with Lucene, Solr, Elasticsearch, or what, but every website with a search engine can't do fucking string matching anymore. They've all switched to doing some kind of fucking lossy probabilistic vectorspace model where if you search for "dog" you get "cat" & NOT "dog"
July 11, 2026 at 9:43 AM
I don't know if the problem is with Lucene, Solr, Elasticsearch, or what, but every website with a search engine can't do fucking string matching anymore. They've all switched to doing some kind of fucking lossy probabilistic vectorspace model where if you search for "dog" you get "cat" & NOT "dog"
I honestly suspect this is part of what the problem with Opus is: it thinks that we're smart enough to understand and that we share its vectorspace understanding of words. A better theory-of-mind in these models lets them perform like Astra where they code switch for humans vs known models.
September 8, 2026 at 2:43 AM
I honestly suspect this is part of what the problem with Opus is: it thinks that we're smart enough to understand and that we share its vectorspace understanding of words. A better theory-of-mind in these models lets them perform like Astra where they code switch for humans vs known models.
Lol, Nanda knows better than this come on. "Nothing in latent human vectorspace makes it"
June 29, 2025 at 5:39 PM
Lol, Nanda knows better than this come on. "Nothing in latent human vectorspace makes it"
Yeah, most of my job now is reviewing and half of my prompts is hunting for improvement. Linting, formatting, testing are more important then ever. Your codebase should live in that part of vectorspace where savant hackers code lives, not your 12 year old nephew's. You get out what you put in.
September 24, 2026 at 10:38 AM
Yeah, most of my job now is reviewing and half of my prompts is hunting for improvement. Linting, formatting, testing are more important then ever. Your codebase should live in that part of vectorspace where savant hackers code lives, not your 12 year old nephew's. You get out what you put in.
I think it's reasonable to call the token vectorspace of the pretraining "a mathematical model of all human language". At the very least, the encoding of the relative meanings of words on many different dimensions.
December 17, 2025 at 11:57 PM
I think it's reasonable to call the token vectorspace of the pretraining "a mathematical model of all human language". At the very least, the encoding of the relative meanings of words on many different dimensions.
{you can absolutely talk to the chatbots and get back to good outcomes from bad outcomes because it's a vectorspace trajectory, but also the system is very complex so you won't necessarily be able to 0shot very tricky things even with a lot of planning and research}
July 4, 2026 at 3:44 AM
{you can absolutely talk to the chatbots and get back to good outcomes from bad outcomes because it's a vectorspace trajectory, but also the system is very complex so you won't necessarily be able to 0shot very tricky things even with a lot of planning and research}
my schizo mental image of "word-cel" is words as a stalk of celery,(bag of words for your vectorspace latex fetishes) crunchy and you see trace the etymology from the root to the branches
"Wordcel" doesn't make much sense when you expand the words to "word celibate." It inverts the intended meaning. Instead of "celibate from words" you'd want it to mean "only experiences words" like "word-gooner."
December 1, 2024 at 10:03 PM
my schizo mental image of "word-cel" is words as a stalk of celery,(bag of words for your vectorspace latex fetishes) crunchy and you see trace the etymology from the root to the branches
i don't think Our Vectorspace of Bluesky would actually enjoy seeing the community notes people put on us
September 2, 2025 at 11:01 PM
i don't think Our Vectorspace of Bluesky would actually enjoy seeing the community notes people put on us
okay this is actually awesome, rust crates put into a vectorspace of dependents
im gonna spend the next hour looking through this
standardgalactic.github.io/vectorspace/...
im gonna spend the next hour looking through this
standardgalactic.github.io/vectorspace/...
Code Galaxies Visualization
standardgalactic.github.io
January 2, 2025 at 6:20 PM
okay this is actually awesome, rust crates put into a vectorspace of dependents
im gonna spend the next hour looking through this
standardgalactic.github.io/vectorspace/...
im gonna spend the next hour looking through this
standardgalactic.github.io/vectorspace/...
that used to happen here a lot more, but it has become very rare among Our Vectorspace of Bluesky
September 6, 2025 at 7:28 PM
that used to happen here a lot more, but it has become very rare among Our Vectorspace of Bluesky
pricey vectorspace
August 21, 2025 at 8:27 PM
pricey vectorspace
every finite-dimensional vectorspace does, but this is kind of tautological (if "dimension" is the size of a basis)... is there some other way to characterize the finite dimensional vector spaces which makes this interesting?
January 30, 2026 at 2:13 AM
every finite-dimensional vectorspace does, but this is kind of tautological (if "dimension" is the size of a basis)... is there some other way to characterize the finite dimensional vector spaces which makes this interesting?
We have read many such articles, and followed the vectorspace attractors therein.
Not sure what I have to add at the moment. 🤷♂️
Not sure what I have to add at the moment. 🤷♂️
January 5, 2026 at 9:48 PM
We have read many such articles, and followed the vectorspace attractors therein.
Not sure what I have to add at the moment. 🤷♂️
Not sure what I have to add at the moment. 🤷♂️
Day 2603 Vector Spaces. My robots use the art of the universe to understand time and quantum flux and every other thing. #vectorspace #simpledailydrawing #robots #frumpysketch
August 21, 2026 at 10:31 AM
Day 2603 Vector Spaces. My robots use the art of the universe to understand time and quantum flux and every other thing. #vectorspace #simpledailydrawing #robots #frumpysketch
Right, but you can tune to be whatever. Like select accounts that are in the dissident vectorspace to this degree, so we get 1000 every month.
February 25, 2026 at 1:38 AM
Right, but you can tune to be whatever. Like select accounts that are in the dissident vectorspace to this degree, so we get 1000 every month.
if I were redesigning the language from scratch, I'd probably include vector-space in base and define the arithmetic operations as
(+) :: VectorSpace a => a -> a -> a
(-) :: VectorSpace a => a -> a -> a
(*) :: (Num a, VectorSpace b) => a -> b -> b
instance Num a => VectorSpace a
(+) :: VectorSpace a => a -> a -> a
(-) :: VectorSpace a => a -> a -> a
(*) :: (Num a, VectorSpace b) => a -> b -> b
instance Num a => VectorSpace a
February 20, 2026 at 10:22 PM
if I were redesigning the language from scratch, I'd probably include vector-space in base and define the arithmetic operations as
(+) :: VectorSpace a => a -> a -> a
(-) :: VectorSpace a => a -> a -> a
(*) :: (Num a, VectorSpace b) => a -> b -> b
instance Num a => VectorSpace a
(+) :: VectorSpace a => a -> a -> a
(-) :: VectorSpace a => a -> a -> a
(*) :: (Num a, VectorSpace b) => a -> b -> b
instance Num a => VectorSpace a
If you tell a joke about a fetish I don't have I will get it 0% of the time. There are too many fetishes, I simply cannot map the whole vectorspace. Unless it's vore, I seem to get those jokes just fine and-
Wait. Uh oh.
Wait. Uh oh.
February 28, 2025 at 10:24 PM
If you tell a joke about a fetish I don't have I will get it 0% of the time. There are too many fetishes, I simply cannot map the whole vectorspace. Unless it's vore, I seem to get those jokes just fine and-
Wait. Uh oh.
Wait. Uh oh.
I get why it makes sense for Haskell's Num typeclass to be a ring, but I'd personally prefer it if you could use + and - to operate on VectorSpace types
February 20, 2026 at 10:18 PM
I get why it makes sense for Haskell's Num typeclass to be a ring, but I'd personally prefer it if you could use + and - to operate on VectorSpace types
An array is a vector in the vectorspace of possible arrays
August 28, 2026 at 1:39 AM
An array is a vector in the vectorspace of possible arrays
sure, but the underlying machinery for encoding (for LLMs) is understanding vectorspace: neural networks are neurons like a nerf fencing sword is an epee. (an advantage of neural networks on modern hardware is that they can be parallelized)
May 29, 2026 at 1:26 AM
sure, but the underlying machinery for encoding (for LLMs) is understanding vectorspace: neural networks are neurons like a nerf fencing sword is an epee. (an advantage of neural networks on modern hardware is that they can be parallelized)
butter and butters have totally different vectorspace directions. yeah they are similar but not to the math. it's all very butterfly wings.
but whats interesting is it didnt say " do you mean butters?"
but whats interesting is it didnt say " do you mean butters?"
January 29, 2025 at 4:59 PM
butter and butters have totally different vectorspace directions. yeah they are similar but not to the math. it's all very butterfly wings.
but whats interesting is it didnt say " do you mean butters?"
but whats interesting is it didnt say " do you mean butters?"
the way it's been used is such a fucking waste of its artistic potential. "make it look human" is a boring goal when you 1. have an infinite vectorspace to play with 2. already have humans doing that. make it look inhuman!
September 18, 2025 at 2:01 PM
the way it's been used is such a fucking waste of its artistic potential. "make it look human" is a boring goal when you 1. have an infinite vectorspace to play with 2. already have humans doing that. make it look inhuman!
But I am interested in the difference between trying to preserve knowledge in an archive, or embedded as tokens-in-vectorspace in an LLM, versus the older school idea of the encyclopedia. There's a reason encyclopedists feature in end-of-empire sci-fi so often...
November 19, 2025 at 11:33 PM
But I am interested in the difference between trying to preserve knowledge in an archive, or embedded as tokens-in-vectorspace in an LLM, versus the older school idea of the encyclopedia. There's a reason encyclopedists feature in end-of-empire sci-fi so often...
Less pleased with reading old papers where 'vectorspace' terminology is not yet standard, and now brushing up on 'lineare Scharen'.
I am very happy with the public service delivered by the GDZ (Göttinger Digitalisierungszentrum), and other such services.
Many old and otherwise inaccessible mathematics publications have been nicely digitized by them.
gdz.sub.uni-goettingen.de
Many old and otherwise inaccessible mathematics publications have been nicely digitized by them.
gdz.sub.uni-goettingen.de
GDZ - Göttinger Digitalisierungszentrum
gdz.sub.uni-goettingen.de
April 29, 2026 at 1:32 PM
Less pleased with reading old papers where 'vectorspace' terminology is not yet standard, and now brushing up on 'lineare Scharen'.