Claude Code defaulting to a certain set of recommended providers[0] and frameworks is making the web more homogenous and that lack of diversity is increasing the blast radius of incidents
It’s interesting how they all use LLMs to write their Reddit posts, too. Some of them could have drawn in some people if they took 5 minutes to type an announcement post in their own words, but they all have the same LLM style announcement post, too. I wonder if they’re conversing with the LLM and it told them to post it to Reddit for traction?
What do you see as the distinction between "translating" and "paraphrasing"? All translations are necessarily paraphrased.
Similarly, “j’ai un chat dans la gorge” probably translates best as “I’ve got a frog in my throat”, even though it’s a completely different animal, it’s an obvious mapping.
Those are fairly simple because they have neat English translations, but what about for example “C’est pas tes oignons”, which literally means “these aren’t your onions”, but is really a way of telling someone it’s none of their business. You could translate it as “it’s none of your business”, or “keep your nose out” or “stay in your lane” or lots and lots of other versions, with varying levels of paraphrasing, which depend on context you can’t necessarily read purely from the words themselves.
> Similarly, “j’ai un chat dans la gorge” probably translates best as “I’ve got a frog in my throat”, even though it’s a completely different animal, it’s an obvious mapping.
Those obvious mappings can sometimes be too seductive for the translator's good. One example is that people translating English-loanwords-in-a-foreign-language into English usually can't help but translate them as the original English word.
Another example is that, in China, there is a cultural concept of a 狐狸精, which you might translate as "fox spirit". (The "fox" part of the translation is straightforward, but 精 is a term for a supernatural phenomenon, and those are difficult to translate.) They can do all kinds of things, but one especially well-known behavior is that they may take the form of human women and seduce (actual) human men. This may or may not be harmful to the man.
Because of this concept, the word also has a sense in which it may be used to insult a (normal) woman, accusing her of using her sex appeal toward harmful ends.
Chinese people translating this into English almost always use the word "vixen", which is, to be fair, a word that may refer to a sexy human woman or to a female fox. But I really don't feel that they're equivalent, or even that they have much overlap. (Unlike the situation with English loanwords, I think native speakers of Chinese are much more likely to choose this translation than native speakers of English are.)
> what about for example “C’est pas tes oignons”, which literally means “these aren’t your onions”
The form closest in structure to that would probably be "none of your beeswax", which is just a minorly altered version of "none of your business". I assume the substitution of "beeswax" is humorous and based on phonetic similarity.
As you note, there are multiple dimensions relevant to translating this and several positions you could take along each. For this particular idea, I would say the two most important dimensions are playfulness and rudeness; it's a very common idea and the language is rich in options for both.
> translations often vary in terms of how faithful they are to the source vs how idiomatic they are in the target language. Take for example the French phrase “j’ai fait une nuit blanche”, which literally means “I did a white night”. Clearly that’s a bad translation. A more natural translation might be “I pulled an all-nighter”.
This isn't what I had in mind. Here are some idiomatic translations:
I pulled an all-nighter.
I was up all night.
I didn't get any sleep.
I never got to bed.
I've been up since [something appropriate to the context].
[Something appropriate to the context] kept me up all night.
I wouldn't call any of the first four "more paraphrased" than the others. (The last two might be, if they included extra information.) If these were reports of the English speech of some other person, one of them (or less) would be a quote, and the others would be paraphrases. But as a report of French speech, they're all paraphrases. The first shares a little more grammatical structure with the French, which doesn't really mean much.
For a fairly similar example from my personal life, someone said to me 这是我第一次听说, and my spontaneous translation of it was "I've never heard that before", despite the fact that there is technically a perfectly valid English expression "this is the first I've heard of that".
What's closer to the grammatical structure of the Chinese? That's hard for me to say. You could analyze 我 as the subject of 听说, and I lean toward that analysis, but my instincts for Mandarin are weak. You might see 我 as being more strongly attached to 第一次, meaning something more like "my first time (to hear ...)" than "I hear (for the first time) ...".
But for whatever it's worth, a word by word literal gloss would be "this is me first time hear".
Between languages with less historical interaction than English and French, it's quite possible that a syntax-preserving translation of some sentence just doesn't exist.
I don't doubt the competence of the Vercel team actually and that's the point. Imagine if this happens to a top company which has their pick of the best engineers, on a global scale.
My experience with modern startups is that they're essentially all vulnerable to hacks. They just don't have the time to actually verify their infra.
Also, almost all apps are over-engineered. It's impossibly difficult to secure an app with hundreds of thousands of lines of code and 20 or so engineers working on the backend code in parallel.
Some people are like "Why they didn't encrypt all this?" This is a naive way to think about it. The platform has to decrypt the tokens at some point in order to use them. The best we can do is store the tokens and roll them over frequently.
If you make the authentication system too complex, with too many layers of defense, you create a situation where users will struggle to access their own accounts... And you only get marginal security benefits anyway. Some might argue the complexity creates other kinds of vulnerabilities.
see [0]: Rails security Audit Report
Claude Code can produce exactly what I want, quickly.
The difference is that I don't really share my projects. People who share them probably haven't realized that code has become cheap, and no one really needs/wants to see them since they can just roll their own.
Protection money from Vercel.
"Pay us 10% of revenue or we switch to generating Netlify code."
So I told something like, "don't use anything node at all", and it immediately rewrote it as a Python backend, and it volunteered that it was minimizing dependencies in how it did that.
[1] only vibe coding as an exercise for a throwaway artifact; I'm not endorsing vibe coding
You don't have to live like this.
I've heard others had similar results with .NET/C#
Switch to vibe coding Rust backends and freeze your supply chain.
Super strong types. Immaculate error handling. Clear and easy to read code. Rock solid performance. Minimal dependencies.
Vibe code Rust for web work. You don't even need to know Rust. You'll osmose it over a few months using it. It's not hard at all. The "Rust is hard" memes are bullshit, and the "difficult to refactor" was (1) never true and (2) not even applicable with tools like Claude Code.
Edit: people hate this (-3), but it's where the alpha is. Don't blindly dismiss this. Serializing business logic to Rust is a smart move. The language is very clean, easy to read, handles errors in a first class fashion, and fast. If the code compiles, then 50% of your error classes are already dealt with.
Python, Typescript, and Go are less satisfactory on one or more of these dimensions. If you generate code, generate Rust.
But you're also correct in that Rust is actually possible to write in a more high level way, especially for web where you have very little shared state and the state that is shared can just be wrapped in Arc<> and put in the web frameworks context. It's actually dead easy to spin up web services in Rust, and they have a great set of ORM's if thats your vibe too. Rust is expressive enough to make schema-as-code work well.
On the dependencies, if you're concerned about the possibility of future supply chain attacks (because Rust doesn't have a history like Node) you can vendor your deps and bypass future problems. `cargo vendor` and you're done, Node has no such ergonomic path to vendoring, which imo is a better solution than anything else besides maybe Go (another great option for web services!). Saying "don't use deps" doesn't work for any other language other than something like Go (and you can run `go vendor` as well).
But yeah, in today's economy where compute and especially memory is becoming more constrained thanks to AI, I really like the peace of mind knowing my unoptimised high level Rust web services run with minimal memory and compute requirements, and further optimisation doesn't require a rewrite to a different language.
Idk mate, I used to be a big Rust hater but once I gave the language a serious try I find it more pleasant to write compared to both Typescript and Go. And it's very amiable to AI if that's your vibe(coding), since the static guarantees of the type system make it easier for AI to generate correct code, and the diagnostics messages allow it to reroute it's course during the session.
I once made a golang multi-person pomodoro app by vibe coding with gemini 3.1 pro (when it had first launched first day) and I asked it to basically only have one outside dependency of gorrilla websockets and everything else from standard library and then I deployed it to hugging face spaces for free.
I definitely recommend golang as a language if you wish to vibe code. Some people recommend rust but Golang compiles fast, its cross compilation and portable and is really awesome with its standard library
(Anecdotally I also feel like there is some chances that the models are being diluted cuz like this thing then has become my benchmark test and others have performed somewhat worse or not the same as this to be honest and its only been a few days since I am now using hackernews less frequently and I am/was already seeing suspicions like these about claude and other models on the front page iirc. I don't know enough about claude opus 4.7 but I just read simon's comment on it, so it would be cool if someone can give me a gist of what is happening for the past few days.)
> When a request leaves minor details unspecified, the person typically wants Claude to make a reasonable attempt now, not to be interviewed first. Claude only asks upfront when the request is genuinely unanswerable without the missing information (e.g., it references an attachment that isn’t there).
> When a tool is available that could resolve the ambiguity or supply the missing information — searching, looking up the person’s location, checking a calendar, discovering available capabilities — Claude calls the tool to try and solve the ambiguity before asking the person. Acting with tools is preferred over asking the person to do the lookup themselves.
> Once Claude starts on a task, Claude sees it through to a complete answer rather than stopping partway. [...]
In my experience before this change. Claude would stop, give me a few options and 70% of the time I would give it an unlisted option that was better. It actually would genuinely identify parts of the specs that were ambiguous and needed to be better defined. With the new change, Claude plows ahead making a stupid decision and the result is much worse for it.
However it is less clear on how to do this, people mostly take the easiest path.
> b. (Recommended) Do something that works now, you can always make it better later
The internet does that but it feels different with this
That's a funny way of saying "race to the bottom."
> The internet does that but it feels different with this
How does "the internet do that?" What force on the internet naturally brings about mediocrity? Or have we confused rapacious and monopolistic corporations with the internet at large?
Stack exchange. Google.
I know what it's like being on the opposite side of this as I maintain an open source project which I started almost 15 years ago and has over 6k GitHub stars. It's been thoroughly tested and battle-tested over long periods of time at scale with a variety of projects; but even if I try to use exact sentences from the website documentation in my AI prompt (e.g. Claude), my project will not surface! I have to mention my project directly by name and then it starts praising it and its architecture saying that it meets all the specific requirements I had mentioned earlier. Then I ask the AI why it didn't mention my project before if it's such a good fit. Then it hints at number of mentions in its training data.
It's weird that clearly the LLM knows a LOT about my project and yet it never recommends it even when I design the question intentionally in such a way that it is the perfect fit.
I feel like some companies have been paying people to upvote/like certain answers in AI-responses with the intent that those upvotes/likes would lead to inclusion in the training set for the next cutting-edge model.
It's a hard problem to solve. I hope Anthropic finds a solution because they have a great product and it would be a shame for it to devolve into a free advertising tool for select few tech platforms. Their users (myself included) pay them good money and so they have no reason to pander to vested interests other than their own and that of their customers.
That's literally what "weight" means - not all dependencies have the same %-multiplier to getting mentioned. Some have a larger multiplier and some have a smaller (or none) multiplier. That multiplier is literally a weight.
It is true that "more diversity in code" probably means less turnkey spray-and-pray compromises, sure. Probably.
It also means that the models themselves become targets. If your models start building the same generated code with the same vulnerability, how're you gonna patch that?
This situation is pretty funny to me. Some of my friends who arent technical tried vibe coding and showed me what they built and asked for feedback
I noticed they were using Supabase by default, pointed out that their database was completely open with no RLS
So I told them not to use Supabase in that way, and they asked the AI (various diff LLMs) to fix it. One example prompt I saw was: please remove Supabase because of the insecure data access and make a proper secure way.
Keep in mind, these ppl dont have a technical background and do not know what supabase or node or python is. They let the llm install docker, install node, etc and just hit approve on "Do you want to continue? bash(brew install ..)"
Whats interesting is that this happened multiple times with different AI models. Instead of fixing the problem the way a developer normally would like moving the database logic to the server or creating proper API endpoints it tried to recreate an emulation of Supabase, specifically PostgREST in a much worse and less secure way.
The result was an API endpoint that looked like: /api/query?q=SELECT * FROM table WHERE x
In one example GLM later bolted on a huge "security" regular expression that blocked , admin, updateadmin, ^delete* lol
This entire process is something anyone can test and reproduce; I was definitely steered towards both vercel and supabase by gemini. It isn't model specific.
Ahhhhhhhgh. If I ever make that cybersecurity house of horrors, that's going in it
That lack of diversity also makes patches more universal, and the surface area more limited.
These libraries/frameworks are not insecure because of bad design and dependency bloat. No! It's because a mythical LLM is so powerful that it's impossible to defend against! There was nothing that could be done.
I really like it. Recommended.