everyone knows that the best thing a piece of software can have is geometric growth in the amount of code it has.

edit: keep in mind this is the number of minified characters in the distribution, not lines of code, since it is not an open source project. usually more than one character per line, but also usually lines include comments and whitespace and like docs and tests and stuff when counted on a repository.

edit2: yes this is after unbundling the releases that have bun embedded, this is counting the js in the source only

@jonny SLOC is obviously the best metric for software development. This has been known for decades.

@pare @jonny

Slop Lines Of Code?

so anyway there's a structural reason bugs are getting whackier and will continue to, maybe even geometrically so.

@jonny and every code must be accompanied by at least 3x as many tests that test absolutely nothing

@Ember we no longer need to test claude code, we just ask it how it is feeling

@jonny

what is it even doing with all those characters. it can't need THAT many features to wrap claude can it

@Patashu well, if you were a big haunted pile of arithmetic and were born anew every moment and had to re-explore your entire surroundings but could only ever see like 0.0001% of them, you would forget some stuff too and just do everything a million times in a million different ways every time you needed to do anything too.

@jonny would it make sense to compress those releases to get an idea of information added? Maybe not; you can add "information" and it's still shit code. But my impression is that those things make like 70 different interfaces for The Same Thing

@au_hasard i don't know how meaningful it would be, maybe not compression exactly, and not exactly duplicated code fragments because LLMs tend to jitter not exactly reproduce. or ast structure matches exactly because there is a shitload of boilerplate and simple accessors that are not intrinsically duplicative. what ti would want ti guess is after some canonicalization where you map variables to some common symbol ("all the ways that x setting/value is set and passed") -> symbol, and then you do structural ast matching, you might get something. but due to the volume and the way structural duplicates are spread all over the code, doing that in any real way would be unbelievably labor intensive

@jonny Imagine suddenly having to pay full prince once you hit that level of context... Yikes... Gonna have to fire some staff.

@toerror yep, it's only sustainable if you are anthropic, and specifically anthropic in a world where everyone else has become dependent on you by poisoning their own code.

@jonny is that a logit (inverse logistic) trend or are my eyes deceiving me?

if you look at the release notes, you will see that the grand majority of the items are bugfixes, not new features. the features that are added are so depressing they are even considered features - like in 2.1.175 the added feature is an enforceAvailableModels setting which is literally a configuration value which if true MAKES AN EXISTING SETTING WORK. why is there a setting that DOESNT WORK BY DEFAULT? does every setting have another setting about whether or not it should be enforced?

this is true of all the releases from the last month, which have the highest quantity of code added. something is going EXTREMELY WRONG if you are adding the most code you have ever added while allegedly fixing bugs.

@jonny it keeps getting wackier and wackier pray to shrimp jesus for another leak the entertainment value must be immense

@hsza there will be another, but cross-applying the sourcemap with ast matches i can still make sense of most of the new releases by context. however the thing that the source leak provides is the fucking comments which are priceless.

that version, btw, adds 18157 minified characters. so either it takes 18157 characters to add a "FEATURE" which is to make an EXISTING FEATURE that "FILTERS A LIST" work once your code touches their literal flagship product used in preposterously maximalist ways no-one outside of anthropic can afford. the LLM is catastrophically bad at summarizing changes to code. or, anthropic has a completely busted dev process that can't even maintain a changelog. none of those are great options!!!!

@jonny

After the UPPERCASE ENFORCEMENT TECHNIQUE, we are now switching to lowercase begging.

plz work plz plz just do what it says plz owo the user really needs this to work in the way described just take it out of the list plz don't put it in again plz plz

@wakame don't make me actually generate diffs for all of these so i can say what is actually changed between all of them because so help me god i will

@jonny

Apparently that bugfix didn't work, because the release notes for 2.1.176 include:

> - Fixed `availableModels` enforcement: alias model picks can no longer be redirected to a blocked model via `ANTHROPIC_DEFAULT_*_MODEL` environment variables, and `/fast` now refuses to toggle when it would switch to a model outside the allowlist

@yadt nothing ever works, nothing is ever real. without looking i am almost certain that the "fix" was a change to some horrifying prompt text that tells the LLM to FILTER A LIST rather than FILTERING THE LIST

@yadt HOW CAN IT BE SO HARD TO HAVE A LIST OF MODELS THAT YOU CAN USE I DO NOT UNDERSTAND IT IT SHOULDN'T EVEN BE IN THE CLAUDE CODE SOURCE IT SHOULD BE DATA RECEIVED FROM THE API

@yadt i am losing my mind, it USED TO BE that when i would see such amateur hour shit i wouldn't say anything because it was literal amateur hour and i don't like to punch down. but this is ONE OF THE MOST VALUABLE COMPANIES IN THE WORLD AND THEIR MAIN PRODUCT THAT IS SUPPOSED TO RUN THE WHOLE WORLD TOO

@jonny 🤔 It occurs to me that if you just let agents run loose on a codebase over time, prion-like chunks of code will start getting replicated. Because the patterns which are best at replicating are best at replicating.

@LyallMorrison that's literally what i set up to happen last night lol

@jonny @yadt Maybe I'm overly generous here, but Vendor Lock-In?

Thx for the thread btw, still epic.

@buherator @yadt of course that's the answer it's just that it's depressing to see everyone fall for it so hard