Mastodawn

Stefan Frede 1d ago

Anthropic's Claude Code's full source code leaked. Claude is seen by many to be the best coding LLM on the market with Anthropic proudly stating that Claude Code itself is mostly written by the LLM.

Now this sounds good as long as nobody can see the code which is quite the trash fire. Detecting "code sentiment" via regular expressions, variable and functions names containing prompt parts trying to influence the bot, a completely intransparent mess of a control flow that makes actual maintenance and debugging functionally impossible and the prompts ... of the prompts. All the begging and pleading to the chatbot not to do this or not to do that or please to do this.

It is fascinating but it is as far away from actual engineering as drunkenly pissing your name in the snow. Dunno what you call the people prompting software at Anthropic but "engineer" is not it.

Now it is fun to look at the currently hyped product striped bare and showing its pathetic quality but that is the future of software if we let those companies continue to undermine every good practice software engineering has tried establishing.

The software we have to use will be bad, insecure, unmaintainable, expensive with nobody having the skills or resources to build something better. As I wrote a few months ago: LLM based software production is equivalent to saying that fast fashion should be the only way to produce clothing. A tragic degeneration of the quality of the artefacts we rely on build for maximum profit on the backs of people in countries from the global majority.

Rens van der Heijden 2d ago

@tante I often wonder how much of (especially proprietary) software already suffered from these issues before the whole LLM situation came about.

David Nash 2d ago

@Namnatulco @tante

This Reddit thread is a good example. A dev posts some visible problems with the leaked code — some problems more serious than others, but just about everything in there counts as dubious “code smell”, sketchy team/organization culture, or both, at a minimum.

The general response? Blowing it off. Stating that the person posting this doesn’t know what real software development is like. Reinforced by an AI-generated summary at the top of the thread, emphasizing the sentiment just described, and generally belittling the concerns raised here.

(TL;DR summary of my own experience with enterprise software development, going back almost 30 years: the quality of what’s in the Claude Code leak is, at the absolute best, definitely on the “very dysfunctional” side of “typical”. I’ve seen lots of dubious code, outdated code, tech debt time bombs as code, and just plain crap code; I have *not* routinely seen codebases for major applications that look quite like this.)

https://www.reddit.com/r/ClaudeAI/comments/1s8lkkm/i_dug_through_claude_codes_leaked_source_and/

@dpnash @Namnatulco @tante

re: AI top of thread summary

can't get much more dystopian than having AI's defend themselves using some unknown energy amount.

bad code actually costs energy.

whole bunch of thesis papers on it.

@Namnatulco @tante software engineers are in general excellent at criticism - the fast iterative work flow makes this skill more useful than in say surgery. But asking them about code their group wrote will get the answer that it is terrible. But IME the problems are more often with conditions than code itself. A lot of times management won't pay for the right way, and a lot of times years of tech debt is hard to maintain, but quality is higher than the people say.

Rupert V/4h ago

@Namnatulco @tante I maintain a codebase that I started 8 years ago. It works, but there are definite here be dragons bits that I dread ever having to change.
The difference being that I know to fear that, whereas an LLM will confidently produce a "fix" that will introduce subtle data corruption in certain cases.

Rens van der Heijden 1h ago

@rupert @tante I was thinking of a different perspective - a lack of appreciation for quality in IT overall.
Obviously there are many people out there wo maintain their projects with care, even in commercial settings. But thinking back to when I was a teenager, the "have you tried turning it off and on again" approach was already the norm. Society already accepted IT is a mess and this makes it hard to see how the impact of AI on quality is harmful, even as it leads to more systematic problems.

Rens van der Heijden 1h ago

@rupert @tante there's an obvious connection to a critique of capitalism here, but I don't think that's the whole story, since we're seeing lots of open source projects both using and doing LLM stuff. It's not really a clean cut between open source and closed source anymore (it probably never was, though..)

Not a Spring Onion 30m ago

@Namnatulco @rupert @tante

Taking Microsoft as an example (they are definitely not the only ones):

Most of the time, when something goes wrong, the user will be confronted with an impersonal message, often accompanied by a long error code or hex value.

So instead of a "Sorry you likely lost work", you get a passive-aggressive "Have fun googling this code. Or just give up."

Half the non-technical users I know instantly blame themselves. The other half shrugs or vents their frustration.

So yeah, software has trained people pretty well to just expect low quality.

Ein „Primat“2d ago

@tante might well be the event that will make the bubble-popping begin. especially because its so far away from the values that anthropic tries so hard to project.

Ein „Primat“2d ago

@tante der kaiser hat keine kleider after all

Cap Ybarra 2d ago

@lechimp @tante i hope so but this might be wishful thinking, there is no evidence to dissuade True Believers anymore because they have bet their skills and careers on slop. their ships have been burned

Morgan ⚧️2d ago

@cap_ybarra @lechimp @tante yeah :/ unfortunately I'm gonna have to agree with this here. If anyone hasn't heard enough to sway them against AI by now, they're probably not going to read the source code for Claude. I'm not going to either, because I've already made up my mind on the software (I hate it, tbc). So have the people using it, so why would they put more effort into knowing how bad the code is than I will?

@tante
I love this so very much I could cry.

Christian Schlieker 2d ago

@tante great timing only a couple of months before the plannend IPO of Anthropic.

@tante "drunkenly pissing [my] name in the snow" might be the closest thing I ever did to vibe coding

Я⊛ḶḶ model ☭1d ago

@qwertzuiop @tante

vibe leaking

cf. anthropic

Ken Brucker 2d ago

@tante I know devs that are saying if AI makes them 10-20% more productive it is worth it. And there’s a rejection of the ethical implications of the technology. I’m thankful I’m on the tail end of my career.

Svavar the Neurospicy 2d ago

@sigsegv @tante

There's a study that found that engineers report being 20% more productive but they're actually 20% less productive.

So it's delusional all the way down.

Steve Purcell 2d ago

@svavar @sigsegv @tante here are recent numbers: more activity, less getting shipped https://mastodon.cloud/@jasongorman/116327493190480229

Jason Gorman (@[email protected])

CircleCI's analysis of 28 million CI workflows confirms the same picture the DORA data shows. While feature branch activity's up significantly, the median impact on *main* (i.e. release) branch activity's net-negative 7%. Only the top 5% of teams saw significant gains. The top 10% flatlined at 1%. For the average team, AI slows them down overall. Told ya! https://www.linkedin.com/pulse/what-28-million-workflows-reveal-ai-codings-biggest-risk-circleci-j9syc/

mastodon.cloud

tranarchy_a_genderqueer

@tante but it's the AI Overlord god code it's not meant to be understood by meat sacks ... 🤣

Svavar the Neurospicy 2d ago

Salesforce has Agentforce Vibes and if you have access to it you can see a lot of the prompts they wrote to guide the LLM. There's very little to stop you from copying it all.

It's the same with a lot of their agentic offerings, you can look under the hood and customise how it works. All they're selling you is a prompt library and the ability to run the LLM on their infrastructure.

https://www.salesforce.com/agentforce/developers/vibe-coding/

Enterprise Vibe Coding - Agentforce Vibes

With Agentforce Vibes, leverage vibe coding to customize Salesforce faster and transform software development directly in your coding environment.

Salesforce

Openhuman 2d ago

@tante but why does it work better than the other machines then?

mx alex tax1a - 2020 (6)2d ago

@Openhuman @tante it doesn't. it's all psychological tricks to make you think it works.

none gender with left politics 2d ago

@tante @atax1a comparing LLM code to fast fashion is really an A+ comparison. I'm gonna use that one

@vikxin @atax1a I elaborated on that here https://tante.cc/2026/01/15/software-as-fast-fashion/

Software as Fast Fashion

Clothes have never been cheaper. These days a t-shirt is often cheaper than a decent cup of tea in a cafe. The wonders of capitalism. At least that is how it is often described. And when you point at the underpaid, gruesome labor that people in poorer regions of the planet have to do to […]

Smashing Frames

none gender with left politics 2d ago

@tante @atax1a I'll give this a read when I have a bit more time

@vikxin @tante @atax1a i wanted to note the same sentiment, great analogy. I'll make sure to read that post!

Paul Hebert 1d ago

@tante this is a good post.

It’s telling that the main selling point of AI coding tools is “building things faster” not “building better things”

Moe Lassus 2d ago

@tante instead of prompt engineer, they appear to be prompt therapists.

Skjeggtroll 2d ago

"Dunno what you call the people prompting software at Anthropic but "engineer" is not it."

'Software mendicants', perhaps?

@skjeggtroll @tante Technopriests.

@skjeggtroll @tante I've seen "prompt fondler" and "sloperator" and both terms have value.

Parth Parikh 2d ago

@tante like this

const negativePattern =
/\b(wtf|wth|ffs|omfg|shit(ty|tiest)?|dumbass|horrible|awful|piss(ed|ing)? off|piece of (shit|crap|junk)|what the (fuck|hell)|fucking? (broken|useless|terrible|awful|horrible)|fuck you|screw (this|you)|so frustrating|this sucks|damn it)\b/

😂

Adrian Chadd <verified.png>2d ago

@parikhparth23 @tante (linguistics doing parts of speech tagging enter the chat and all rub their hands together saying "YESSSSSSSS")

Parth Parikh 2d ago

@erikarn @tante You are right 🤣

GunChleoc 2d ago

@parikhparth23 @erikarn @tante I guess I'll have to swear in a language other than English then.

Tòn an t-Sàtain!

Parth Parikh 1d ago

@gunchleoc @erikarn @tante exactly !English

Marcos Dione 22h ago

@gunchleoc ohhh, I know curses in many languages, thanks for the tip!

@parikhparth23 @erikarn @tante

Parth Parikh 1d ago

@erikarn @tante exactly !English

Adrian Chadd <verified.png>1d ago

@parikhparth23 if only we had a way of classifying text as different language/dialects based on some sort of rules and bayesian statistical based processing, then a way to apply different subsets of regular expression tables based on this

(sorry as a programmer and a linguistics graduate, i can not resist)

Prema Marsik 2d ago

@parikhparth23 @tante sounds like me googling some new feature from windows update

Aslak Raanes 2d ago

@parikhparth23 @tante Things like this would work fantastically in languages and dialects with a richer swearing tradition

Parth Parikh 1d ago

@aslakr @tante exactly thats my point…swear in !English

Lord Caramac the Clueless, KSC 2d ago

@tante Ever since the Industrial Revolution, capitalism has been thriving on producing bad quality for cheap, and then selling that cheap disposable shit to everyone so they have to throw it out and replace it when it breaks. Before the Industrial Revolution, people didn't have many pieces of clothing, but their clothes lasted for decades. Then mechanical looms produced so much cloth for so cheap that weavers lost their livelihoods, even though their artisanal cloth was far superior. All of a sudden, many people could afford to buy a lot of new clothes, but then they had to keep buying them because the fabric didn't last as long as it used to.

Making things cheap, shoddy, disposable, has always been the general direction of industrial mass production. If it's cheap enough, who cares if it breaks after a while, just buy a new one.

First it was textiles. Then it was all kinds of consumer goods. After WW2, everything became increasingly disposable. Nowadays even entire washing machines are often made out of plastic so that they break after a decade. Automated factories don't care if their products are of worse quality than the things that came before, they just dump their cheap shit onto the marketplace until the competition crumbles.

Roxy ⭐ (she/her)2d ago

@LordCaramac @tante good analysis

Cap Ybarra 2d ago

@LordCaramac @tante this llm project is primarily about getting you used to paying the same price for worse service. they're selling you two chips in a king size bag and telling you it's inevitable

@cap_ybarra @LordCaramac @tante slopflation

🇨🇦 CowMan 🇪🇺 🇺🇦 🇲🇽2d ago

@LordCaramac @tante washing machines out of plastic Haven't seen that!

Though they do use things like alu or magnesium spiders with steel drums which are susceptible to galvanic corrosion. Or light gauge painted steel instead of hravy weight steel and enamel.

Lower weight saves significantly on shipping for those overseas factories, at the expense of longevity.

@tante Was it the full code base? I heard that some things were leaked which would prove useful for reverse engineering but that the code itself wasn't leaked.

Jim Jones 2d ago

@angiebaby @tante my understanding was the the code was leaked and then quickly reimplemented in another language using Claude Code in order to thwart takedown requests.

@angiebaby @tante

A javascript source map file for the claude code client was accidentally published and people were able to convert that back into source directory.

We don't know what's going on on their servers though.

@angiebaby it was the code

@tante I wouldn't claim that I expected something else, but I sure hoped for it. So, at least the harness is known now to be iteratively developed to adress the symptoms. Looks like LLMs are too complex even for the "real" experts, compared to an eternal GenAI beginner like me. They just have more optimism about how much control they have.

Bruce Simpson, Ph.D.2d ago

@tante I am currently significantly refactoring a medium sized C++ code base from the 00s. I have mostly only used LLMs to suggest sed substitutions to me. I will probably rely more on traditional AST-based refactoring e.g. clang-tidy initially, though I have an item to turnup Opencode.ai. I do not trust Cursor, Codex, GitHub Copilot or Claude Code.

Hugs4friends ♾🇺🇦 🇵🇸😷2d ago

@tante After reading this, very carefully - and reading replies with the same care - I now hope that I am still compos mentis when AI goes tits-up. The world will need the knowledge of the ancient ones...to rebuild.