this post was submitted on 18 Apr 2025
329 points (94.6% liked)
Technology
70248 readers
3944 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related news or articles.
- Be excellent to each other!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, this includes using AI responses and summaries. To ask if your bot can be added please contact a mod.
- Check for duplicates before posting, duplicates may be removed
- Accounts 7 days and younger will have their posts automatically removed.
Approved Bots
founded 2 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
So, not a single developer thought about filtering useless words locally before triggering the request ?
How can they be so dumb ?
The company I worked for tried that as an experiment on how much money it saves.
Absolutely awful, even removing connectives causes problems.
Dr GPT is smarter when you are polite and spell better in the prompt. I believe u can find some benchmarks proving it.
They talk about separate messages though, if you just send "thanks" it changes nothing to the answer
How would you filter it?
if msg == "thanks" return "you're welcome my dude"thanks, you're clearly a genius, these LLM providers should pay you a lot of money to implement this, you'd save them millions 🙄
That's pretty much how they cenzor stuff right now.
noo the joke was he was supposed to reply
"you're welcome my dude"I wrote "== thanks", not "contains thanks". All this conversation is about messages containing ONLY a SINGLE useless word.
Obviously if it's just at the beginning or the end of a legit message, it's not the same thing...
U use a smaller cheaper LLM that will inject a 20% hallucination.
The writer of this article doesn't consider these words useless though. They are suggesting that these words may improve response quality.
I would argue that being polite also does good to the person writing that line.
The author and the writer they quoted are fucking morons.
You're being downvoted, this is a perfect example of:
*they hated Jesus because he spoke the truth 😂🤣
And your qualifications in computer science are....?
Hi, I have a degree in computer science and work with AI every day.
Feelings aren't a good way to measure things scientifically, they are right about that.
But saying that words can just be filtered is easier said than done. You're back at needing to do a lot of processing to identify and purge these words. This is still going to cost a lot of money and potentially lead to less meaningful inputs. Now you also have to maintain the software that does the word identification, keep it well tested, maintain monitoring and analytics for it, and so on.
So, in short, everyone here is wrong and I'm considering packing it all in and buying a small potato farm with no internet connection.
The big thing here is that 'polite' words are being singled out as extraneous when there are tons of extraneous words being used. The focus is on words that make it seem like AI has feelings or intent.
There is no reason to filter any words, because the entire point of LLMs is to take inefficient human communication and do stuff with it. 'Please' isn't any more of a waste that 'the' or including a period at the end of a sentence.
Not to mention the fact that the whole thing is so horribly inefficient that 'extra' words cost millions of dollars to process. Holy shit that is terrible design.
I'm smart enough to know that an article peppered with assumption and zero facts is dogshit.
Doesn't matter how educated someone is when they write a bunch of words about possibilities with no actual evidence. They are morons because they are spouting a bunch of useless speculation about a shitty and unreliable technology and naval gazing about whether 'being polite' to a bullshit generator is beneficial. I feel dumber for having read both the article and the linked article.
Your never supposed to show certainty unless its like 99.95% I thought
Maybe don't write an article speculating about something possible being true based on another article that is also speculating about something being possible when it being able to confirm it is possible. Like speculating about dinosaurs makes sense as we don't have a way to verify their soft tissues. But when it comes to AI, there are ways to actually confirm the reliability of responses.
Please may be useless. Thank you isn't useless. That tells you that the prior response gave them the answer they were looking for. No response at all could mean that, or that they gave up, or any number of other things.
What if it's a sarcastic thanks ?
Also, the public models are fixed right ? Not perpetually training AFAIK ? So it should really change nothing unless it's linked to those "thumb up/down" buttons
Both authors state that the phrasing from the AI is what is improved based on how they felt about the answers, not the accuracy.
Anecdotally, I use it a lot and I feel like my responses are better when I'm polite. I have a couple of theories as to why.
More tokens in the context window of your question, and a clear separator between ideas in a conversation make it easier for the inference tokenizer to recognize disparate ideas.
Higher quality datasets contain american boomer/millennial notions of "politeness" and when responses are structured in kind, they're more likely to contain tokens from those higher quality datasets.
I haven't mathematically proven any of this within the llama.cpp tokenizer, but I strongly suspect that I could at least prove a correlation between polite token input and dataset representation output tokens
Honestly they were better until recently. GPT (at least) has gotten really good at de-escalation and providing (mostly) factual responses when you get irate
It FEEEEEEEEEEEELS better is what the authors said too. Both articles were completely worthless dreck about how they felt about the responses.
And you FEEEEEEEEL like it doesn't matter. What's the difference?
Yes they were, so I'm offering you an actual theory as to why this may actually be true, yet difficult to "prove".
Smoking was bad for your health long before anyone sat down and took the time to prove it. Autoregressive LLM tokenizer are a very new field of computer science and it's going to take a while for the community to collectively understand everything we're currently doing by trial and error.
And yet doctors saw the tar in the lungs and knew immediately.
Smoking was known to be bad for your health long before anyone did studies because it was easily correlated with coughing and other breathing issues and early death. The evidence was obvious and apparent.
🤦♂️