There are many different ways to make an LLM sound more human-like (the author has actually explicitly mentioned ChatGPT sounds more natural). For their public announcements etc. they might as well used specially trained small 24-32B creative writing model or put a LoRA on top of their Haiku. One could also use "antislop" samplers, maybe some encoder-decoder unslopping postprocessing small models etc. Or they simply may have been written by humans.
Anthropic's public writing style is very different from Claude's because it's written by humans. Even in their job ads they ask candidates to not use AI for any writing (or at least they used to).
I find the language of "collaborating with Claude" off-putting. I don't collaborate with Claude, I use Claude. It's a tool in my hands, not a colleague or a friend.
AI companies seem eager to perpetuate the fantasy that all users carefully review all LLM output, and supervise all tool calls.
Anthropic certainly isn't planning to take responsibility for Claude's mistakes - that's the user's responsibility. Describing everything as 'collaborating' is I think part of their efforts to emphasise the user's role in the process.
I think the opposite. AI Companies use the term collaborate to increase the trust in the LLM output.
I trust the output of colleagues I am collaborating with and have an assumption of some shared responsibility. But if I use a tool like numpy/matplotlib then I am accountable for the conclusions I come up with. I can't make an excuse that "matplotlib" created a plot so my decisions were incorrect and not due to my own negligence.
I’ve read the link, but I don’t get how it’s relevant to the OP’s point that Anthropic public communications don’t sound like Claude. Can you enlighten me?
Ive given up trying to fight claudes language, and Im afraid Im getting used to it and can even understand what claude is saying faster or should I say I can parse it faster now.
whats troubling here is that one of the anthropic devs on twitter or github don't seem to acknowledge this issue, If they don't acknowledge there's nothing for them to fix
There are many different ways to make an LLM sound more human-like (the author has actually explicitly mentioned ChatGPT sounds more natural). For their public announcements etc. they might as well used specially trained small 24-32B creative writing model or put a LoRA on top of their Haiku. One could also use "antislop" samplers, maybe some encoder-decoder unslopping postprocessing small models etc. Or they simply may have been written by humans.
Same reason the tech execs kids aren't using the tools created by their parents
Anthropic's public writing style is very different from Claude's because it's written by humans. Even in their job ads they ask candidates to not use AI for any writing (or at least they used to).
Anthropic's guide for job applicants on how you should use Claude when applying for a job there is relevant here: https://www.anthropic.com/candidate-ai-guidance
I find the language of "collaborating with Claude" off-putting. I don't collaborate with Claude, I use Claude. It's a tool in my hands, not a colleague or a friend.
AI companies seem eager to perpetuate the fantasy that all users carefully review all LLM output, and supervise all tool calls.
Anthropic certainly isn't planning to take responsibility for Claude's mistakes - that's the user's responsibility. Describing everything as 'collaborating' is I think part of their efforts to emphasise the user's role in the process.
I think the opposite. AI Companies use the term collaborate to increase the trust in the LLM output.
I trust the output of colleagues I am collaborating with and have an assumption of some shared responsibility. But if I use a tool like numpy/matplotlib then I am accountable for the conclusions I come up with. I can't make an excuse that "matplotlib" created a plot so my decisions were incorrect and not due to my own negligence.
If that were true, they'd cut their token gen speed and reduce all prices to th level their users could actually read what it produces.
I’ve read the link, but I don’t get how it’s relevant to the OP’s point that Anthropic public communications don’t sound like Claude. Can you enlighten me?
why do you care?
Because Simon usually posts interesting comments.
Do I need a different reason?
Ive given up trying to fight claudes language, and Im afraid Im getting used to it and can even understand what claude is saying faster or should I say I can parse it faster now.
gate= some sort of a guard or test
name= a variable name or fields
load bearing= coupled dependencies that can break other things if you change it
Other than the specific phrases like load-bearing etc, I find the biggest tell of all just to be repetition.
Every damn Claude article does the “tell ‘em what you’re going to tell ‘em, tell ‘em, tell ‘em what you told ‘em” routine.
I do really enjoy the style of their blogposts, they remind me of the Cloudflare postmortems. Wish their models could produce it.
Not sure I agree about e.g. GPT not sounding this formulaic though, imo it definitely does.
I don't actually have a problem with the common terms Claude emits, I think they're appropriate for a coding agent. But I do find them overused.
whats troubling here is that one of the anthropic devs on twitter or github don't seem to acknowledge this issue, If they don't acknowledge there's nothing for them to fix
On the bright side, I'll know immediately a post is Claude-generated. If the author didn't bother writing it, I don't bother reading it.
/s?
This is where the concern-trolls barge in with "what about non native English speakers using AI to blogslop everything is actually a good tool!"