Hacker News Reader

Comment by stingraycharles | Hacker News Reader

Because they want it to be executed quickly and cheaply without blocking the workflow? Doesn’t seem very weird to me at all.

View on HN

They probably have statistics on it and saw that certain phrases happen over and over so why waste compute on inference.

The problem with regex is multi-language support and how big the regex will bloat if you to support even 10 languages.

Did you just complain about bloat, in anything using npm?

We're talking about Claude Code. If you're coding and not writing or thinking in English, the agents and people reading that code will have bigger problems than a regexp missing a swear word :).

I talk to it in non-English. But have rules to have everything in code and documentation in english. Only speaking with me should use my native language. Why would that be a problem?

Because 90% of training data was in English and therefore the model perform best in this language.

In my experience these models work fine using another language, if it’s a widely spoken one. For example, sometimes I prompt in Spanish, just to practice. It doesn’t seem to affect the quality of code generation.

They literally just have to subtract the vector for the source language and add the vector for the target.

It’s the original use case for LLMs.

Thank you. +1. There are obviously differences and things getting lost or slightly misaligned in the latent space, and these do cause degradation in reasoning quality, but the decline is very small in high resource languages.

It’s just a subjective observation.

It just can’t be a case simply because how ML works. In short, the more diverse and high quality texts with reasoning reach examples were in the training set, the better model performs on a given language.

So unless Spanish subset had much more quality-dense examples, to make up for volume, there is no way the quality of reasoning in Spanish is on par with English.

I apologise for the rambling explanation, I sure someone with ML expertise here can it explain it better.

The agent may be translating to English behind the scenes. There are all kinds of hacks layered on top of the underlying models. But however exactly it works, it’s clearly not the case that these things only make use of Spanish training data when you prompt them in Spanish.

Lets take a language with orders of magnitude less training data, like Nahuatl. The following is what happens if I ask Claude Code in Nahuatl to give me an example of a simple program that I can execute inside a web browser (avoiding any direct keyword prompts like 'Javascript' or 'Hello World'). Interestingly, it responds in Spanish. I doubt that it has any relevant training data in Nahuatl. (Apologies if the Nahuatl is wrong – I just used Google translate.)

    ¿Uelis techmakas se neskayotl tlen se programa tlen amo ouij tlen ueliskia nijchiuas ipan se navegador web?

    -----

    Entiendo tu pregunta en náhuatl. Aquí un ejemplo sencillo — una calculadora en HTML puro que puedes abrir directo en el navegador, sin servidor:

    <!DOCTYPE html>
    <html>
    <head>
        <meta charset="UTF-8">
        <title>Calculadora</title>
        <style>
        body { font-family: sans-serif; display: flex; justify-content: center; padding: 40px; }
        .calc { border: 1px solid #ccc; padding: 20px; border-radius: 8px; width: 200px; }
        input { width: 100%; margin-bottom: 10px; font-size: 1.2em; text-align: right; }
        button { width: 22%; margin: 2px; padding: 8px; font-size: 1em; cursor: pointer; }
        .wide { width: 47%; }
        </style>
    </head>
    <body>
    <div class="calc">
        <input type="text" id="pantalla" readonly>
        <div>
        <button onclick="tecla('7')">7</button>
        <button onclick="tecla('8')">8</button>
        <button onclick="tecla('9')">9</button>
        <button onclick="tecla('/')">÷</button>
    
        <button onclick="tecla('4')">4</button>
        <button onclick="tecla('5')">5</button>
        <button onclick="tecla('6')">6</button>
        <button onclick="tecla('*')">×</button>
    
        <button onclick="tecla('1')">1</button>
        <button onclick="tecla('2')">2</button>
        <button onclick="tecla('3')">3</button>
        <button onclick="tecla('-')">−</button>
    
        <button onclick="tecla('0')">0</button>
        <button onclick="tecla('.')">.</button>
        <button onclick="tecla('+')">+</button>
        <button onclick="limpiar()">C</button>
    
        <button class="wide" onclick="calcular()">=</button>
        <button class="wide" onclick="borrar()">⌫</button>
        </div>
    </div>
    
    <script>
        const pantalla = document.getElementById('pantalla');
    
        function tecla(val) {
        pantalla.value += val;
        }
    
        function calcular() {
        try {
            pantalla.value = eval(pantalla.value);
        } catch {
            pantalla.value = 'Error';
        }
        }
    
        function limpiar() { pantalla.value = ''; }
        function borrar() { pantalla.value = pantalla.value.slice(0, -1); }
    </script>
    </body>
    </html>
    
    Guarda esto como calculadora.html y ábrelo en cualquier navegador — no necesita servidor ni dependencias. Es un buen punto de partida para aprender HTML,
    CSS y JavaScript.

> it’s clearly not the case that these things only make use of Spanish training data when you prompt them in Spanish.

It’s not! And I’ve never said that.

Anyways, I’m not even sure what we are arguing about, as it’s 100% fact that SOTA models perform better in English, the only interesting question here how much better, is it negligible or actually makes a difference in real world use-cases.

It’s negligible as far as I can tell. If the LLM can “speak” the language well then you can prompt it in that language and get more or less the same results as in English.

I saw a curious post recently that explored this idea, and showed that it isn’t really the case. The internal layers of the model aren’t really reasoning in English, or in any human language.

Translation in/out of human languages only happens at the edges of the model.

Internal layer activations for the same concept are similar regardless of language, while activations at the top/bottom layers diverge. Meanwhile the pattern is reversed for same language different content.

So we do at least agree on the fact that quality of human language <-> embeddings transition depends on how good target language is represented in the training dataset?

Even if it happens at the edge, on every conversation turn, I may assume non captured small subtleties of meaning over time can accumulate into significant error.

In my experience agents tend to (counterintuitively) perform better when the business language is not English / does not match the code's language. I'm assuming the increased attention mitigates the higher "cognitive" load.

Claude handles human languages other than English just fine.

Supporting 10 different languages in regex is a drop in the ocean. The regex can be generated programmatically and you can compress regexes easily. We used to have a compressed regex that could match any placename or street name in the UK in a few MB of RAM. It was silly quick.

I think it will depend on the language. There are a few non-latin languages where a simple word search likely won't be enough for a regex to properly apply.

Exactly this. Unicode is a big beast to consider in regex concats.

woah. This is a regex use I've never heard of. I'd absolutely love to see a writeup on this approach - how its done and when it's useful.

You can literally | together every street address or other string you want to match in a giant disjunction, and then run a DFA/NFA minimization over that to get it down to a reasonable size. Maybe there are some fast regex simplification algorithms as well, but working directly with the finite automata has decades of research and probably can be more fully optimized.

This was many moons ago, written in perl. From memory we used Regexp::Trie - https://metacpan.org/release/DANKOGAI/Regexp-Trie-0.02/view/...

We used it to tokenize search input and combined it with a solr backend. Worked really remarkably well.

They only need to look at one language to get a statistically meaningful picture into common flaws with their model(s) or application.

If they want to drill down to flaws that only affect a particular language, then they could add a regex for that as well/instead.

More likely their LLM Agent just produced that regex and they didn't even notice.

Why do you need to do it at the client side? You are leaking so much information on the client side. And considering the speed of Claude code, if you really want to do on the client side, a few seconds won't be a big deal.

> a few seconds won't be a big deal

it is not that slow

Depends what its used by, if I recall theres an `/insights` command/skill built in whatever you want to call it that generates a HTML file. I believe it gives you stats on when you're frustrated with it and (useless) suggestions on how to "use claude better".

Additionally after looking at the source it looks like a lot of Anthropics own internal test tooling/debug (ie. stuff stripped out at build time) is in this source mapping. Theres one part that prompts their own users (or whatever) to use a report issue command whenever frustration is detected. It's possible its using it for this.

It looks like it's just for logging, why does it need to block?

Better question - why would you call an LLM (expensive in compute terms) for something that a regex can do (cheap in compute terms)

Regex is going to be something like 10,000 times quicker than the quickest LLM call, multiply that by billions of prompts

This is assuming the regex is doing a good job. It is not. Also you can embed a very tiny model if you really want to flag as many negatives as possible (I don't know anthropic's goal with this) - it would be quick and free.

I think it's a very reasonable tradeoff, getting 99% of true positives at the fraction of cost (both runtime and engineering).

Besides, they probably do a separate analysis on server side either way, so they can check a true positive to false positive ratio.

Reply on news.ycombinator.com