Weston Beecroft @Westoncb Twitter profile

Pinned Tweet

Weston Beecroft

3 months

The repo for ConceptScript and a couple other LLM languages is now up (tutorial vid included): Here's some introductory info:

1

8

64

Last Seen Profiles

@Planet__Claire

@stwmaniax

@GQThailand

@221EXhonmono

@daddyskitty6

@ethicologies

@moon_star321

@Puchinatsu

@Sexyrohit2

@delicious_horny

@lunataur

@kevinklop

@GuildBabu

@Preserve1031

@riyadhkw

@Tan__p0

@chrisxcastle

@CBSSportsNet

@Alpeshrph

@Brendahodg76309

@monamjour

@tarrehyung

@yuyu_tiny

@Gaming_SoMe_

@nickmarxxshow

@benedicthx60872

@Dat_boi437

@FlightradarCAT

@_BigChubbz

@yoojisminie

@Agnlaf92

@HullCity

@PUPPYGOTER

@h_streich

@LincolnMullen

@PompeyWomen

Weston Beecroft

@Westoncb

3 months

After getting interesting results from ascii art gen with Sonnet 3.5 I decided to try applying same approach to pixel shaders We were speaking Lovecraft and math; this only emerged after next day manually 'unwrapping' the original with a non-euclidean transform. Story 👇

21

62

643

Weston Beecroft

@Westoncb

3 years

I was surprised yesterday to see so many people thrilled at the concept of viewing program state snapshots with Graphviz So here's a thread on two additions to the idea: 1. capture time-evolution of state; make it explorable 2. simplify integration with running programs

8

90

518

Weston Beecroft

@Westoncb

3 months

Inspired by some of the Claude ascii art I've seen people doing, I did some experiments this morning. The one that struck me most was uploading a sketch I did and requesting an ascii projection of it:

4

23

291

Weston Beecroft

@Westoncb

4 months

@kylebrussell Casually astral projecting in a park lol

2

0

196

Weston Beecroft

@Westoncb

3 months

I worked on a program (pixel shader) with Claude yesterday trying to capture some of the thought and aesthetics of William Blake and Frank Zappa yesterday: I've spent the past couple weeks doing a bunch of experiments giving LLMs loose directions and

8

11

119

Weston Beecroft

@Westoncb

8 months

Been trying to reconcile a bunch of different aspects of NN behavior and cognition by playing with formulations of latent space at a level (also) suited for reasoning about everyday cognitive phenomena In part it's led to a compact, diagrammable geometric system

8

6

113

Weston Beecroft

@Westoncb

6 years

Circular treemaps for directory preview icons. (From: )

2

22

105

Weston Beecroft

@Westoncb

5 months

I'd like to share more concretely how I'm actually using ConceptScript. The heart of it is: evolving conceptual models through LLM conversation On the right here is a Concept for the "semantic fiber bundle tower" I posted about. It's largely generated from those tweets

8

7

103

Weston Beecroft

@Westoncb

3 months

After 'unwrapping' it with a hyperbolic space transform and finding what i've been calling the 'angel core' from the first tweet, I tweaked another single param and the 'demon core' emerged:

8

7

94

Weston Beecroft

@Westoncb

5 years

@kenpex My on-again/off-again project is building an interactive/reusable program state visual debugger:

3

9

90

Weston Beecroft

@Westoncb

3 months

Also requested a geometric interpretation of latent space, then a 'reverse Rorschach test' version of same, and a couple that were attempts to be open-ended and let Claude pick

4

6

86

Weston Beecroft

@Westoncb

3 months

Claude generated this w/o modification: imo a pareidolic masterpiece folding together varied artists/styles from 4-dimensional hyperbolic geometry into some kind of organic cathedral tribute to the complex multifaceted nature of identity

5

12

85

Weston Beecroft

@Westoncb

1 month

new LLM context management utility I’m working on

4

84

Weston Beecroft

@Westoncb

9 months

So it turns out the “quilt-pinching” of latent space metaphor I was dwelling on ended up being fruitful I’m going to do some more sketches to clarify in a minute, but for now here’s the image i started with yesterday and part of an abstract on these “semantic fiber bundles”

9

6

81

Weston Beecroft

@Westoncb

3 months

I have links to all three shaders, the conversation with Sonnet 3.5 that led to the original, and details on how I found the other two here:

2

3

81

Weston Beecroft

@Westoncb

3 months

This was the original Claude generated. I found it so strange and fascinating I saved it as wtf.html to look more closely at next day

4

7

79

Weston Beecroft

@Westoncb

5 months

As an alternative to LLM summarizing, I've been getting very interesting results doing something like: Academic paper -> "functional ontology" If you consider the resulting FO as an intermediate rep, an entire paper can become a single (extremely detailed) "thought" for an LLM

7

3

74

Weston Beecroft

@Westoncb

7 months

Haven’t been excited about a new book for a little while

7

1

59

Weston Beecroft

@Westoncb

5 months

These past few months I've been doing a bunch of experiments with languages designed for LLMs. This is a thread about what got me started on it and some general principles and hypotheses I've arrived at along the way, including a kind of LLM Sapir-Whorf w practical implications

3

5

59

Weston Beecroft

@Westoncb

5 months

I’m working on this “latent space navigator” app where you can progressively break down concepts by following their parts into further decompositions thinking of calling the central operation “delve”

6

0

59

Weston Beecroft

@Westoncb

3 years

found at goodwill today

4

1

57

Weston Beecroft

@Westoncb

3 months

My baseless intuition is when OpenAI drops their next thing it’ll be a little shocking cause it’s profoundly good in some way, but will also have some kinda bad aura to it that’ll get more people asking, “what are we doing?”

6

0

55

Weston Beecroft

@Westoncb

9 months

big IDE has been tricking us into splitting our source into more than 1 file for too long

2

3

52

Weston Beecroft

@Westoncb

1 month

I went deep into this (structural text editing) without knowing what I was doing back in 2011-2012—but I think it increasingly makes sense for LLM-backed editors to have character-level operations be secondary to structure level, nontrivial design challenges notwithstanding

3

1

51

Weston Beecroft

@Westoncb

3 years

The more I use an app the further it should move up in the hierarchy of incarnation-stability 1. browser tab 2. native app 3. physical cube that only runs app

4

51

Weston Beecroft

@Westoncb

2 months

I'd mentioned only testing my shader generation approach with Sonnet 3.5 before and being curious about Opus and gpt4o. Here's a sample of what I found with Opus

3

51

Weston Beecroft

@Westoncb

6 years

@amasad @startuployalist There’s just enough people waving the idealism banner for selfish purposes that it’s been tainted with hypocrisy, in spite of the legitimate exceptions. Not every iPhone app is gonna improve the world.

0

45

Weston Beecroft

@Westoncb

3 years

Started a new weekend project: a Chrome extension that will add a spaceship to any page which can fly around and blast DOM elements First step is rendering the DOM in 3D of course. Here's a look at my first fumbling steps

1

6

48

Weston Beecroft

@Westoncb

4 years

The "Cube" language

3

9

45

Weston Beecroft

@Westoncb

3 years

Visualizations of structured data *evolving* are still rare, in part because it still takes too much work to produce them—but much of that work could be automated like it has been for the static scenario

1

2

42

Weston Beecroft

@Westoncb

4 years

Here's a thread on the evolution of my thinking on non-traditional program editor concepts, ending with a new concept for something like a "linear AST editor". (This does skip over my project that originally got me into this: —shown below too)

2

10

41

Weston Beecroft

@Westoncb

5 months

Yesterday I was describing translating papers into "functional ontologies" via ConceptScript; one cool thing you can do with the distilled structure is generate "instructive" versions: If you were curious about @VictorTaelin 's HVM2 or interaction combinators (full texts below):

4

1

41

Weston Beecroft

@Westoncb

2 years

watching an Alan Kay talk, got a nice image in my head of computation/computing machines as deep mysterious facet of the universe, more deserving of our curiosity interesting how certain things like that can only be conveyed seeing someone else's emotional relation to them

2

39

Weston Beecroft

@Westoncb

5 years

After years of mystery, I've finally written an article on the origin story and rationale of my Lucidity project: (And there's a new video illustrating more realistic usecases than I've shown previously)

3

4

36

Weston Beecroft

@Westoncb

3 years

Making progress on the DOM exploder game/Chrome extension :)

2

1

36

Weston Beecroft

@Westoncb

3 years

From the prologue to "Visual Differential Geometry and Forms"

4

7

34

Weston Beecroft

@Westoncb

4 months

App that just sits and reads your Twitter all day, connects every tweet to your interests and projects, conjectures on impact*relevance, sends notification if high enough

6

1

34

Weston Beecroft

@Westoncb

5 months

There is both a neuroscience and a psychology of LLMs; few people seem to recognize the depth to the psychology side, at most have it collapsed into something like "prompt engineering"

3

7

34

Weston Beecroft

@Westoncb

3 years

@nikitabier where do I get a list of these risky companies?

3

0

30

Weston Beecroft

@Westoncb

2 months

This is amazing. This is what I’ve been seeing in the shaders Claude generates and my whole reason for investigating interference/optics/metamateriels etc lately in connection with NN structure/behavior

Bij

@BijanFakhri

2 months

EM waves resonate within and bounce off of these patches, creating natural boundaries in our world simulation. These patches may even account for the boundaries of the self!

8

16

136

2

3

29

Weston Beecroft

@Westoncb

3 years

Seeing single snapshots of program state can be extremely useful on its own, but many times—in order to get quick sense of how an algorithm is acting upon some data—it's the sequence of program states that's needed.

2

1

28

Weston Beecroft

@Westoncb

3 months

This was a wild experience using Artifacts to build something. Definitely going to be experimenting more with it. You can go well beyond toys by working on multiple Artifacts independently and then re-combining

1

28

Weston Beecroft

@Westoncb

3 months

@algekalipso Less and less. Or maybe more and more, can’t tell

1

0

27

Weston Beecroft

@Westoncb

4 years

@RobertHaisfield I did an experiment on these lines where one control stick cycled through language constructs, and the other could've moved up/down abstraction levels (though I used buttons for this). Seems like it'd work for natural language too: abstraction levels are eg para, sentence, phrase

2

4

23

Weston Beecroft

@Westoncb

3 years

@mechadense @dm_0ney @spiralganglion @neurocy @disconcision More info on the project here:

1

5

24

Weston Beecroft

@Westoncb

3 months

teaching claude some magic tricks

2

1

24

Weston Beecroft

@Westoncb

3 years

Thinking about a possible near-future evolution for books the idea is they'll always come packaged with a language model that can answer questions as you read, and that this will be so valuable it'll overtake traditional solitary reading

3

4

24

Weston Beecroft

@Westoncb

6 years

Hypothesis: the only way for an already mature reader to significantly increase their reading speed—without also increasing their reading superficiality—is to thoroughly learn general concepts likely to underpin what they'd like to read quickly.

3

1

22

Weston Beecroft

@Westoncb

4 years

Okay I'll bite. What was it about Flash that made it so much better than JavaScript (+libs) for making funky creative games, art projects etc.? (Asking for real—I often hear it praised in this regard, but have never used it)

6

4

21

Weston Beecroft

@Westoncb

3 months

A better example: I distill the paper "A sheaf-theoretic construction of shape space" into a Concept, then request a concrete novel application Concept; I hit retry 3 times getting: NeuroDegenerationTracker EvolutionaryMorphologyAnalyzer TopologicalProteinDrugInteraction

3

2

21

Weston Beecroft

@Westoncb

5 months

I’ll try to be more concrete: people rarely see the full intelligence of models like Claude because prompts are lens-like wrt to it and it’s largely left unfocused There are techniques for bringing it into focus which can be combined with techniques for granting autonomy

3

0

22

Weston Beecroft

@Westoncb

2 years

A nice page from the original tree map paper ("Tree-Maps: A Space-Filling Approach to the Visualization of Hierarchical Information Structures"):

0

3

21

Weston Beecroft

@Westoncb

5 years

I'd be interested in a repository of "progressions" for building up to reading various works For instance, let's say I want to read @conal 's "Compiling to Categories" I'd like to enter the title and get people's curated, rated/reviewed reading progressions for it.

1

3

21

Weston Beecroft

@Westoncb

3 years

A problem that shows up when attempting this in practice is dealing with large volumes of data. To help: 1. users may select only relevant subsets to monitor 2. users unfold/explore data as needed in the viz 1. is pretty straightforward; 2. is an appealing space for innovation

1

0

20

Weston Beecroft

@Westoncb

5 months

The theory evolves: there are contraction languages and expansion languages for LLMs Another way of saying it: abstraction languages and concretization languages In terms of latent space nav: up langs and down langs

4

0

20

Weston Beecroft

@Westoncb

5 months

I feel a big relief having finally posted what I’ve been working on. Now I am going to get a pizza and clean my apartment, maybe do some reading

2

0

20

Weston Beecroft

@Westoncb

5 years

Unpopular opinion: your brain is already measuring this stuff and guiding future behavior via emotional responses. And it does a far better job of taking into account myriad complexly interrelated objectives, and not flattening your life into a dull chart of goal progress.

Sahil Lavingia

@shl

5 years

You can't make what you don't measure. Want to read more books, lose weight, save money, call your parents more frequently, become a better writer? Start measuring.

29

220

1K

1

20

Weston Beecroft

@Westoncb

5 years

New UI for: 1) Indicating location in hierarchy (i.e. filesystem path) 2) Previewing folder contents on hover

3

2

20

Weston Beecroft

@Westoncb

4 years

@thingskatedid

1

3

19

Weston Beecroft

@Westoncb

4 years

This is a file system view intended to make easy to find large files. The bigger the rectangle the more space it takes. White rectangles are folders, darker ones are files. I'm curious whether people can make sense of what's going on without more of a tutorial than that.

2

19

Weston Beecroft

@Westoncb

3 years

Bugs in the mapping from state -> visual can be seriously derailing, and as your program changes you often need to update the mapping, creating possibilities for new errors: The ideal simplifies to something like: monitor(state, filter/query)

1

19

Weston Beecroft

@Westoncb

2 years

@Meaningness @gwern if i'm not mistaken, Crungus already holds that honored title

Professor Bison Sexhorn

@Brainmage

2 years

Well I REALLY don't like how similar all these pictures of "Crungus", a made up word I made up. Why are they all the same man? Is the Crungus real? Have I discovered a secret cryptid?

662

2K

12K

4

3

19

Weston Beecroft

@Westoncb

3 years

secrets to programming: 1. break problem into small pieces in text file 2. sleep enough

1

0

19

Weston Beecroft

@Westoncb

4 months

Testing out 3.5 Sonnet :D

3

2

18

Weston Beecroft

@Westoncb

4 years

Thinking about "visualization-assisted algorithm development" again. Just idle, non-specific thoughts: visual feedback is nice, an 'immediate-mode' lib for this would be great, how many areas distant from e.g. game dev could use this that don't know about it? etc.

0

2

19

Weston Beecroft

@Westoncb

4 months

Progress update: finally recorded the video on llm languages and ConceptScript yesterday, editing now and assembling examples etc for repo

4

0

17

Weston Beecroft

@Westoncb

6 months

The reason there’s so much alpha in prompt construction is it’s literally a natural language interface to tensor network configurations—but most things we say aren’t laser-focused enough for the value of this metaphor to be apparent

1

18

Weston Beecroft

@Westoncb

2 years

making progress on the local Stable Diffusion desktop frontend (GenerationQ)—it now actually processes its queue (just showing fake terminal output atm though)

1

18

Weston Beecroft

@Westoncb

5 months

Church

4

0

18

Weston Beecroft

@Westoncb

3 years

Another lesson I learned while visualizing program state: while ultimately it often pays off to put non-trivial effort into setting up state viz for some program, the error-proneness of the manual work presents a real problem

1

18

Weston Beecroft

@Westoncb

3 years

Weird to me how consistent the tendency is for open source == little attention to usability/aesthetics Anyone have a good explanation for this?

13

2

18

Weston Beecroft

@Westoncb

3 years

I think there may be a legitimate upside to having a million unfinished projects more projects = higher probability one of them has an ideal bite-sized/appealing problem to tackle with morning coffee as a warmup

2

0

18

Weston Beecroft

@Westoncb

1 month

I basically always want to go into ai coding session with the LLM pre-oriented in this specific way Prompt preamble: 1 - project design/goals/intentions 2 - how existing code relates to/implements the above (1 we have to provide/maintain; 2 should be computed) —now we can talk

4

0

17

Weston Beecroft

@Westoncb

5 months

I wonder how many people know what I’m talking about.. it’s in this region: tree search + artifact construction + “memory-augmented prompts” + rich self-context

1

0

17

Weston Beecroft

@Westoncb

8 months

At the heart of it is a search space structured as a "semantic fiber bundle tower", dynamically constrained by "possibility cones" attached to the static fibers; fibers connect elements (tensors with 'free variables') to spaces of their possible configurations at layers below

1

0

17

Weston Beecroft

@Westoncb

3 years

This is as far as I'm going with the DOM destruction game I was working on source:

2

1

17

Weston Beecroft

@Westoncb

5 years

@SteveStuWill @41Strange What's the fish's experience of this? Assuming there's no top-down knowledge of the project he's undertaking, he must undergo a weird 'mood' for a week where he just finds strong appeal in brushing up against sand piles in a certain way...

0

15

Weston Beecroft

@Westoncb

3 months

I've actually tried teaching Claude and asking what's going on with this locked/wrapped up shader stuff; this was the result of the one time there seemed to be some understanding and the code included an explicit unwrap function itself—no work from me

2

1

16

Weston Beecroft

@Westoncb

3 years

Here is the thread that prompted me to write this. Hopefully @thingskatedid , @cfbolz , and/or @khinsen find the above interesting.

CF Bolz-Tereick

@cfbolz

3 years

This is an incredibly cool and important thread about programs visualizing their own state particularly using graphviz. @pypyproject uses this approach a lot and it's a real superpower and I wanted to show some examples to encourage people to do this more often!

5

29

122

2

0

17

Weston Beecroft

@Westoncb

2 months

Working on a language/tool (CLI-ish) for rapidly defining LLM Operations and Processes as the basic primitives/actions for an IDE that generates and evolves artifacts in various ways Been jumping around and prototyping a lot lately but feels like things are coming together now

0

16

Weston Beecroft

@Westoncb

10 months

@ShakyBarnes @Austen yeah that’s a better proof imo

0

16

Weston Beecroft

@Westoncb

2 years

i wonder if this will shift society's typical way of thinking/valuing toward the abstract: "particular representations are generated, not essential"

1

14

Weston Beecroft

@Westoncb

4 years

Thinking about 'interpersonal annealing': after a conflict, those involved—if it's resolved in a non-terminal way—often become closer than if there were no conflict This looks a lot like one of these processes (e.g. annealing) where: high-energy state, cooling -> reorganization

2

3

16

Weston Beecroft

@Westoncb

2 years

in the same way sublime text-inspired minimap scrollbars give quick visual context for where you are in a document, i'd love context for where that document/module is in relation to a larger program graph but... where do you put it? that seems to be the main issue 🤣

3

0

16

Weston Beecroft

@Westoncb

4 years

Concluding lines from Richard Gabriel’s Patterns of Software

1

2

16

Weston Beecroft

@Westoncb

3 years

How about: select a single variable in source and— 1. tool generates a (visually) weighted graph of other pieces of data whose values may affect it 2. press play and passively observe this focused+abstracted computation determining vals for your selected variable

3

0

16

Weston Beecroft

@Westoncb

5 months

I think once I get the ConceptScript stuff up I’m gonna switch to working on building a research tool on top of it It’ll let you do structured exploration of new or existing theoretical structures, using ConceptScript primarily under the hood, richer mixed media reps for user

2

0

15

Weston Beecroft

@Westoncb

3 months

Working on a quick app to render Concepts that'll do something like syntax highlighting but with blocks since the grammar isn't exact. First test building something real with Artifacts/3.5 too and it's awesome

4

0

15

Weston Beecroft

@Westoncb

3 months

@Nominus9 Thanks man. This is all pretty perplexing to me btw, but I will say when the first one showed up first thing I thought of was some things I’d seen you posting and was intrigued by. Curious where this all comes from, and if the structures are related

1

15

Weston Beecroft

@Westoncb

3 months

Having worked on an AGI project before that had an “artifact network” at its core, Anthropic’s projects/artifacts stuff feels like an intentionally inverted AGI system empowering people instead. Hope that works out

1

15

Weston Beecroft

@Westoncb

3 years

Very cool, and "white box" is a highly apt name I understood better after watching:

WhiteBox DevLog 2 - GUI, data trees and graphing values over time

This is a quick explanation of the new GUI features for WhiteBox.- an actual GUI!- familiar debugger-like data tree- graph and tree-based exploration of vari...

www.youtube.com

WhiteBox

@whitebox_sys

4 years

What's that? You're still looking at debug data a single value at a time? Here's an early look at WhiteBox's timeline: - see value changes in the context of other data - zoom in or mouseover for more detail - as ever, it updates instantly with code changes

1

8

39

1

5

15

Weston Beecroft

@Westoncb

3 years

Today’s theoretically appealing but probably practically misguided idea: what if wiki software added something like plug-ins that recognized patterns in topology + node-types of linked articles; when pattern is found, default behavior/presentation logic of wiki could be altered

1

0

15

Weston Beecroft

@Westoncb

3 months

I would like to chat with this pdf

2

0

15

Weston Beecroft

@Westoncb

4 years

Lawyer / self-taught programmer gets interested in morphogenesis during his wife's pregnancy, ten years later has built this and posts a Show HN ():

0

4

15

Weston Beecroft

@Westoncb

3 years

When people say structure editors are solving a non-essential problem

/Fay-lee-nuh/

@Felienne

3 years

10

75

535

0

2

15

Weston Beecroft

@Westoncb

3 months

@Shoalst0ne Even weirder: I'd generated two during the 'reverse Rorschach test'—didn't occur to me at the time but I definitely should've had Claude continue printing them out because these look unfinished. This is the 2nd

1

0

13

Weston Beecroft

@Westoncb

2 months

Nice old school page on optical diffraction and FFT:

1

0

14

Weston Beecroft

@Westoncb

4 months

I’m about to make the hardest pivot from knowledge representation to.. removing all the spiders from my apartment

4

0

14

Weston Beecroft

@Westoncb

4 years

If we achieved the commonly expressed “finished” form of computer programming where you basically tell an ai what you’d like built in a cycle of refinements, you’d basically be a manager at that point. People are already programming this way, they just aren’t programmers.

1

14