104 comments
Lately, like following ChatGPT 3.5 release era, I had found Grammarly now replacing entire sentences with LLM replacement which are random and inject bias to being near useless. It did this some before. The problem is I can't even tell if it's defensible from a grammatical point when it uses an LLM to do this. At some point Grammarly reached too far in the cookie jar and I no longer trusted that someone who read the Grammar rule book verified any part of the elaborate suggestions.
I haven't logged in for months. I guess I was no longer learning from it. I need to take a closer look but sys admin outrage from pushy sales people doesn't bode well in my mind.
Deepl however - 7-8 or more years ago was bright star - i do not say this lightly but their translation (i can only vouch for german<->english) was out of this world in comparison to everything else like google translate, yahoo or whatever.
It was basically so good - i just kept writing in english even though its not my first language but its easier for technical documentation and i use it to translate back to german instead of doing it the other way around.
But then, when the first mass hysteria of chatgpt 1 etc hits - it good worse fast - nowadays - it can not do any good translation service at all - it is far worse than what the competition offered 7-8 years ago! It makes no sense a lot of times.
Benchmark is very AI-pilled: https://www.entrepreneur.com/business-news/ai-is-replacing-j... and is very bullish on it, too: https://gtmnow.com/benchmark-ai-startup-investing-chetan-put...
it also looks like DeepL went from using a CNN engine to an LLM in 2024 (coinciding with its $2 billion valuation and Series C) which is likely why it seems to hallucinate a lot more now: https://en.wikipedia.org/wiki/DeepL_Translator#Translation_m...
I need to find an old school grammar checker that works on the command line with markdown.
You might be interested in Val - https://docs.vale.sh/
Even before all that, I switched to Quillbot because Grammarly would always mess up the formatting in my emails in gmail (but only when you viewed it from mobile). Quillbot isn't totally better than that, but at least its cheaper.
If Gmail suggests the whole email (and whatever model it's using has baked in whatever suggestions Grammarly would've had), nobody needs feedback on an email they didn't even write?
Some people will argue that Grammarly still has a better suggestion system, but models have become increasingly good at writing. I think this product and its revenue model will become obsolete pretty soon.
That's not the impression I have. HN has almost daily articles bemoaning the use of AI to enhance or pad-out human-to-human communication and the comments almost universally agree that the output is bad, using AI to communicate is rude, and they'd rather read the human-written prompt than the AI-"enhanced" version.
I'm not aware of any model that has solved the AI writing problem. Fable 5.1 and Opus 5.5 improve on Fable 5 and Opus 5, but neither are good. They all seem to overuse jargon, misuse literary devices and write ill-fitting metaphor. They all produce prose that is dense and hard to read, but sparse on actual information. Even non-techy people can identify AI writing.
It was a good tool, that worked, paid for it for a couple of months and then they decided to go with AI and using my writing for training.
That is a very bad shortening for "pre-laymen-take-of-AI". Let us try to keep language clean at least here.
If the product in question replaces a tutor, it is AI ("simulation of intelligence"). We cannot restrict the term "AI" to "giving a broader appearance of intelligence with less trust in the results". Either this is still the "AI era" (post cybernetics, post Dartmouth), or we are not there yet.
Models are not good at writing. Like, they are just not producing good writings. They produce differently bad writing.
Or maybe I'm just in a bad European country who pulls this kind of demotivating stints every time I have this kind of problem.
"A contractor of my employer asked me to contact another member of my organisation to persuade them to do something commercially advantageous for them" - at no point is anyone here a consumer.
Business emails are subject to fairly different rules by the looks of it: https://tietosuoja.fi/en/faq-direct-marketing ("Can a person working in a company be sent direct electronic marketing") - it appears there's a fair bit of an exception that doesn't require consent.
Or having their names listed in industry reports, which their marketing departments usually don't like.
[1] almost every cloud solution that exfiltrates your data to the cloud is, but that is besides the point
It's a keylogger in the same way Vim is a keylogger.
Read the full thread on Hacker News →
Related stories
- Ars Technica · 0 points · 7 days ago
- Hacker News · 1 points · 10 days ago
- Hacker News · 797 points · 9 days ago
- Hacker News · 4 points · 10 days ago
- I made ByeBlocked: hide blocked usersgithub.comHacker News · 1 points · 9 days ago
- The Verge · 0 points · 11 days ago