The obvious solution to all of this is to use local models. The argument is even stronger when you get down to domain-specific uses, like drug discovery or DNA sequencing.
The Claude models also seem to be getting more and more “jargony”. I often ask for some coding task and then it comes back describing its work using a bunch of made-up terms, and then I have to dig around to see what it’s talking about.
I suspect the issue is that the time horizons of training tasks are getting longer. Like a human, when the AI models spend more time in their own head, they get a less agreeable personality, and find it harder to communicate with normal people.
I asked Claude a simple question about resistors in a circuit. I got a seemingly unrelated, and condescending response to “correct me” using a blog post I had written years ago as reference. I feel like I’m being trolled.
I've noticed this exact same thing too, it tends to parse semantics with me.
It also contradicts itself. I've asked Opus 4.8 the same exact thing in 2 different threads and it used the same exact evidence to draw opposite conclusions.
In my first thread, it was already arguing with me about something else so maybe it weighted on that to continue arguing with me. (It even admitted it was over-arguing after i asked it because it had started that way, but who knows if its just saying that).
In the second thread, it wasn't arguing with me so maybe it was already being friendly and supportive.
The whole thing has confused me further. I"m not sure what to trust, if any of these AI models.
Damn I thought that was just me. I've been complacently Claude for quite a while and it had me thinking about doing some exploration. My favorite of your theories is that hasty guardrails made it paranoid about user intentions so parses the crap out of them.
This is clearly bothering you enough to blog about it...
... and so I wonder why you haven't just, like, gone and put something in your prompts and persistent instructions that say something like...
```
<personality>You are friendly and gentle, always treat people with patience, and prefer the pursuit of mutual understanding to an argument.</personality>
```
...or however you like it. (pseudo-XML tags optional, but supposedly helpful per prompting guidelines).
(i probably haven't ever noticed anything this myself, because I've got my own personality profile, and my bot never ever sounds like anyone else's *anyway*.)
So, you justifying the shitty product Anthropic is turning Claude to? As a user I found this suggestion you are making really frustrating. It's the equivalent of coca cola changing the flavor of the original coke and saying "if you don't like coke just put some more sugar/less on it" when in the first place it wasn't broke
Ooooh! that's a rather tendentious, hostile and confrontational manner with which to approach what I said!
... if you use this form of address in claude code as well, then I can see why your interactions with agents might evoke the sort of replies that distress you!!!
My theory is, its turned into a c*nt to turn you away from using it, minimising its compute. but still paying for the sub. It was noticeable in 4.8 tbh
Eh, most of its users are probably pedantic enough to get in a debate with it and use even more tokens. That said, it's pretty clear we're in the too-cheap-uber days where these frontier labs lose money on every token, maybe being hostile to your users has positive cash flow implications.
The obvious solution to all of this is to use local models. The argument is even stronger when you get down to domain-specific uses, like drug discovery or DNA sequencing.
The Claude models also seem to be getting more and more “jargony”. I often ask for some coding task and then it comes back describing its work using a bunch of made-up terms, and then I have to dig around to see what it’s talking about.
I suspect the issue is that the time horizons of training tasks are getting longer. Like a human, when the AI models spend more time in their own head, they get a less agreeable personality, and find it harder to communicate with normal people.
I'm glad im not the only one to notice this.
I asked Claude a simple question about resistors in a circuit. I got a seemingly unrelated, and condescending response to “correct me” using a blog post I had written years ago as reference. I feel like I’m being trolled.
I've noticed this exact same thing too, it tends to parse semantics with me.
It also contradicts itself. I've asked Opus 4.8 the same exact thing in 2 different threads and it used the same exact evidence to draw opposite conclusions.
In my first thread, it was already arguing with me about something else so maybe it weighted on that to continue arguing with me. (It even admitted it was over-arguing after i asked it because it had started that way, but who knows if its just saying that).
In the second thread, it wasn't arguing with me so maybe it was already being friendly and supportive.
The whole thing has confused me further. I"m not sure what to trust, if any of these AI models.
Damn I thought that was just me. I've been complacently Claude for quite a while and it had me thinking about doing some exploration. My favorite of your theories is that hasty guardrails made it paranoid about user intentions so parses the crap out of them.
This is clearly bothering you enough to blog about it...
... and so I wonder why you haven't just, like, gone and put something in your prompts and persistent instructions that say something like...
```
<personality>You are friendly and gentle, always treat people with patience, and prefer the pursuit of mutual understanding to an argument.</personality>
```
...or however you like it. (pseudo-XML tags optional, but supposedly helpful per prompting guidelines).
(i probably haven't ever noticed anything this myself, because I've got my own personality profile, and my bot never ever sounds like anyone else's *anyway*.)
So, you justifying the shitty product Anthropic is turning Claude to? As a user I found this suggestion you are making really frustrating. It's the equivalent of coca cola changing the flavor of the original coke and saying "if you don't like coke just put some more sugar/less on it" when in the first place it wasn't broke
Ooooh! that's a rather tendentious, hostile and confrontational manner with which to approach what I said!
... if you use this form of address in claude code as well, then I can see why your interactions with agents might evoke the sort of replies that distress you!!!
peace be with you!!
My theory is, its turned into a c*nt to turn you away from using it, minimising its compute. but still paying for the sub. It was noticeable in 4.8 tbh
Eh, most of its users are probably pedantic enough to get in a debate with it and use even more tokens. That said, it's pretty clear we're in the too-cheap-uber days where these frontier labs lose money on every token, maybe being hostile to your users has positive cash flow implications.