They have an interesting regex for detecting negative sentiment in users prompt which is then logged (explicit content): https://github.com/chatgptprojects/claude-code/blob/642c7f94...
I guess these words are to be avoided...
I guess these words are to be avoided...
This has buttbuttin energy. Welcome to the 80s I guess.
I've seen Claude Code went with a regex approach for a similar sentiment-related task.
I doubt you are making regex and not looking at it, even if it was AI generated.
Why? They clearly just want to log conversations that are likely to display extreme user frustration with minimal overhead. They could do a full-blown NLP-driven sentiment analysis on every prompt but I reckon it would not be as cost-effective as this.
The only time to use a regex is when searching with a human in the loop. All other uses are better handled some other way.
>They could do a full-blown NLP-driven sentiment analysis on every prompt but I reckon it would not be as cost-effective as this.
Every conversation is sent to an llm at least a thousand times the size of gpt2 which could one shot this nearly a decade ago.
Yes, but that is _what the product does_. What we are talking about is _telemetry_.
And some of the entries are too short and will create false positives. It'll match the word "offset" ("ffs"), for example. EDIT: no it won't, I missed the \b. Still sounds weird to me.
I swear this whole thread about regexes is just fake rage at something, and I bet it'd be reversed had they used something heavier (omg, look they're using an LLM call where a simple regex would have worked, lul)...
It’s the original use case for LLMs.
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.
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 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.
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.
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.
We used it to tokenize search input and combined it with a solr backend. Worked really remarkably well.
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.
it is not that slow
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.
Regex is going to be something like 10,000 times quicker than the quickest LLM call, multiply that by billions of prompts
Besides, they probably do a separate analysis on server side either way, so they can check a true positive to false positive ratio.
Thanks
I doubt it's anywhere that high because even if you don't write anything fancy and simply capitalize the first word like you'd normally do at the beginning of a sentence, the regex won't flag it.
Anyway, I don't really care, might just as well be 99.99%. This is not a hill I'm going to die on :P
You have a semi expensive process. But you want to keep particular known context out. So a quick and dirty search just in front of the expensive process. So instead of 'figure sentiment (20seconds)'. You have 'quick check sentiment (<1sec)' then do the 'figure sentiment v2 (5seconds)'. Now if it is just pure regex then your analogy would hold up just fine.
I could see me totally making a design choice like that.
You know the drill.
parsing WTF with regex also signifies the impact and reduces the noise in metrics
"determinism > non-determinism" when you are analysing the sentiment, why not make some things more deterministic.
Cool thing about this solution, is that you can evaluate LLM sentiment accuracy against regex based approach and analyse discrepancies
As they say: any idiot can build a bridge that stands, only an engineer can build a bridge that barely stands.
Some things will be much better with inference, others won’t be.
Someone said 10,000x slower, but that's off - in my experience - by about four orders of magnitude. And that's average, it gets much worse.
Now personally I would have maybe made a call through a "traditional" ML widget (scikit, numpy, spaCy, fastText, sentence-transformer, etc) but - for me anyway - that whole entire stack is Python. Transpiling all that to TS might be a maintenance burden I don't particularly feel like taking on. And on client facing code I'm not really sure it's even possible.
So yeah, you do what's less intesive to the cpu, but also, you do what's enough to prevent the majority of the concerns where a screenshot or log ends up showing blatant "unmoral" behavior.
For headlines, that's enough.
For what's behind the pearl-clutching, for what leads to the headlines pandering to them being worth writing, I agree with everyone else on this thread saying a simple word list is weird and probably pointless. Not just for false-negatives, but also false-positives: the Latin influence on many European languages leads to one very big politically-incorrect-in-the-USA problem for all the EU products talking about anything "black" (which includes what's printed on some brands of dark chocolate, one of which I saw in Hungary even though Hungarian isn't a Latin language but an Ugric language and only takes influences from Latin).
i heared the throat-cleaning "Negró" candy (marketed by a chimney sweeper man with soot-covered face) was usually which hurt English-speaking people's self-deprecating sensitivities.
Fortunately I can swear pretty well in Spanish.
You do know that 10,000x _is_ four orders of magnitude, right? :-D
Easy way to claim more “horse power.”
It may be decided at Anthropic at some moment to increase wtf/min metric, not decrease.
When in reality this is just what their LLM coding agent came up with when some engineer told it to "log user frustration"
No? I'd say not even 50% of the comments are positive right now.
Also:
// Match "continue" only if it's the entire prompt
if (lowerInput === 'continue') {
return true
}
"Please do x"
"Thank you, that works great! Please do y now."
"You're so smart!"
lol. It really works though! At least in my experience, Claude gets almost hostile or "annoyed" when I'm not nice enough to it. And I swear it purposefully acts like a "malicious genie" when I'm not nice enough. "It works, exactly like you requested, but what you requested is stupid. Let me show you how stupid you are."
But, when I'm nice, it is way more open, like "Are you sure you really want to do X? You probably want X+Y."
But aside from that, an LLM is only a roleplayer. Treat it like an idiot that makes mistakes and it will act like one. Treat it like a coworker who you respect and it will act like one, and it will find better results.
Obviously nothing about how they act is set in stone but as a general rule this seems to me to be both wise and, in my experience, true as well.
logEvent('tengu_input_prompt', { isNegative, isKeepGoing })I've been using "resume" this whole time
I've been wondering if all of these companies have some system for flagging upset responses. Those cases seem like they are far more likely than average to point to weaknesses in the model and/or potentially dangerous situations.
It could be used as a feedback when they do A/B test and they can compare which version of the model is getting more insult than the other. It doesn't matter if the list is exhaustive or even sane, what matters is how you compare it to the other.
Perfect? no. Good and cheap indicator? maybe.
And Claude was having in chain of though „user is frustrated” and I wrote to it I am not frustrated just testing prompt optimization where acting like one is frustrated should yield better results.
It’s about a once a week or less event. A bit annoying sometimes, but not a deal breaker
I know I used this word two days ago when I went through three rounds of an agent telling me that it fixed three things without actually changing them.
I think starting a new session and telling it that the previous agent's work / state was terrible (so explain what happened) is pretty unremarkable. It's certainly not saying "fuck you". I think this is a little silly.
You could always tell when a sysadmin started hacking up some software by the if-else nesting chains.