glupi jebeni bot

i see this all the time with software designed by americans. on an old job we used a tool called "officevibe" where you'd enter your current impression of your role and workplace once a month. you got some random questions to answer on a 10-degree scale.
when we were presented with the result the stats were terrible because the scale was weighted so that everything below 7 was counted as negative. we were all just answering 5 for "it's okay", 3-4 for "could use improvement", and 6-7 for "better than expected". there had never been a 10 in the stats, and the software took that as "this place sucks".
like, of course you downvote a bad response. you're supposed to help the model get better, right?
- This makes me think of those retail surveys where any scores less than perfect get the employee a talking to by management.
  
  That happens in Google maps too. A 4 star restaurant is not good. But in Japan, 3 star is the norm, and 5 means exceptionally good.
  
  I once got a call from a telecom marketing department because I rated a customer service agent with a 9 out of 10. When I told them it was not for anything the agent did, just that the store the support was in was extremely difficult to find, the caller got a bit aggressive. Like they expected me to shit talk this poor lady who had been so nice to me, just hard to find, and it was all corpo's fault. The store wasn't properly branded and signaled. She couldn't take any comment that was negative on the company, just on the employee. So I told her how ridiculously stupid that system was. That I wanted to change my score to a perfect ten, comment, the best employee this company has, even better than the CEO. The caller got obviously upset. Told her to write down that if they ever call me again I will immediately cancel my contract. She went with, is there anything else I could help you with? Which is call center code for "I want to hang up".
- Recently, saw some survey that explicitly said 1-7 is "poor", 7-8 is "OK", and 9-10 is "great". Wild, not sure what the point of the scale is then.
  Same with book ratings. Looking at StoryGraph, the average ratings I see is somewhere between 3.5 and 4.5. While I would rate a decent book a 3.
  Born in Eastern Europe, live in the US, maybe that's why.
  
  I wonder if it's like the grading system we use in school? <60% is F for fail, 60% to <70% is D which depending on the class can be barely passing or barely failing. >=70% would be A, B, and C grades which are all usually passing, and A in particular means doing extremely well or perfect (>=90%). I just noticed that that rating scale kind of lines up with the typical American grading scale, maybe that's just a coincidence
- From the looks of it, what they're calculating is a net promoter score. The idea is that, in some context, what you actually want to know is whether your target audience would be willing to actually promote your business to their friends and family or not.
  It's very common in retail and other competitive markets, because a customer that had an "okay" experience could still go to a competitor, so only customers who had a great experience (7+ out of ten) are actually loyal, returning clients.
  Don't know if that's the best method to gather impressions on workplace environment though, I don't think many people would consider their workplace "amazing"

“Optimizing for things people love” aka talking to you like an hr team building seminar
It’s frustrating, or maybe it’s a good thing given the tendency for some people to form weird pseudo social relationships with LLMs, to see the evolution of chatgpts language processing
Public chatgpt only had the 3.5, 4, and 4o model but you can play with earlier models like 2 and 3 on huggingface. These were far weirder, often robotic and stilted but sometimes mirroring more natural colloquial English more based on the input
Rather than make something that is authentic and more natural to interact with they instead go for the ultra sanitized HR corporate speak bullshit. Completely bland and inoffensive with constant encouragement and reinforcement to drive engagement that feels so inauthentic (unless you are desperate for connection with anything, I guess). It’s mirrored in other models to some degree, deepseek, llama, etc (I don’t know about grok, fuck going on twitter).
3-5 years until it’s ruined by advertising, tops. If that

Self defense against the IA: tell them they suck until they stop talking to you :D

"objective distribution" yeah right
- and the main problem with gpt are the em-dashes
  
  Yeah, nothing about the blatant hallucinations to even basic questions, the dashes are the major problem
  
  The dashes are a problem because it’s an easy tell that the text is AI. They don’t want you to be able to tell so easily.

It's hard to imagine how horrible 'early gpt' versions were at Croatian if they constantly invented words and grammar for much more popular languages, at the time.
- Gpt aside even google translate invents words regularly especially for augmentative languages.
  Put google translate on Turkish to English and try something like "teakmezliyorlacaklarasacisinimislaslarin Charlie"
  
  👌

There is nothing objective about that 'objective distribution' why would the output automatically center on good?
- Because it's the center duh...
  I feel like for real this will be the reasoning, since they divided the [0,1] interval into 5 equidistant intervals I think they believe that is what the regular distribution of ratings should look like and then compare that to how much different regions deviate from this norm.