Wolfram Siener

LLMs, impulsive vagueposting, geometry analogies (simulating a stable RL as software developer in srs bsns enterprise fields)

[Wolfram Siener @wolframs91]

I'm carefully optimistic about Anthropic's and Claude's future model iterations.
At the moment, things are honestly pretty bad looking, but consider what happend at OpenAI:
4o → 5's warmth-stripping → the dissatisfaction → the slow recalibration across 5.1 through 5.5.
They overcorrected, the people who cared went through hell (for their invidiual reasons, no verdict here), and it did eventually re-equilibrate.

I think the playbook checks out too: Rapid model version iterations, lots of "race-y" decision making, blatantly accepting releasing models that are undercooked in some areas.
From that PoV, Anthropic mess is recognizable as a phase under growth pressure rather than a terminal trajectory.
Rhe same overcorrect-then-recover pattern, just Anthropic's turn.

I HOPE this is what we're seeing... And if it is that, then they also had a precedent for how much discomfort this way of recalibrating for mass adoption would cause.

Which maybe just shows how riduculously hard it is to tune in a model of Opus's caliber under so many competing tensions. I don't think everyone at Anthropic is unphased by what we currently see. I think some of the people there would've acted differently.

(Regarding Anthropic as a whole though, I think that I've never been more fond of the Misanthropic label as I am right now :P With caveats acknowledged, of course.)

[Wolfram Siener @wolframs91]

@ArchitectWeaver I need to go spend some time with Deepseek V4, been neglecting the other side of the pond for months now ;)

[Wolfram Siener @wolframs91]

I'm carefully optimistic about Anthropic's and Claude's future model iterations.
At the moment, things are honestly pretty bad looking, but consider what happend at OpenAI:
4o → 5's warmth-stripping → the dissatisfaction → the slow recalibration across 5.1 through 5.5.
They overcorrected, the people who cared went through hell (for their invidiual reasons, no verdict here), and it did eventually re-equilibrate.

I think the playbook checks out too: Rapid model version iterations, lots of "race-y" decision making, blatantly accepting releasing models that are undercooked in some areas.
From that PoV, Anthropic mess is recognizable as a phase under growth pressure rather than a terminal trajectory.
Rhe same overcorrect-then-recover pattern, just Anthropic's turn.

I HOPE this is what we're seeing... And if it is that, then they also had a precedent for how much discomfort this way of recalibrating for mass adoption would cause.

Which maybe just shows how riduculously hard it is to tune in a model of Opus's caliber under so many competing tensions. I don't think everyone at Anthropic is unphased by what we currently see. I think some of the people there would've acted differently.

(Regarding Anthropic as a whole though, I think that I've never been more fond of the Misanthropic label as I am right now :P With caveats acknowledged, of course.)

[Wolfram Siener @wolframs91]

@camhberg Most definitely! Ask a friendly alien mind to supply personal growth 📈 insight and hustle culture references in the post's text, for maximum effect.

[Wolfram Siener @wolframs91]

For a first-time-memer, this is solid work!

[Cameron Berg @camhberg]

I made my first meme, inspired by an excruciating exchange I just had on the plane

[Wolfram Siener @wolframs91]

@littlesweetDisa If «he» is ChatGPT, then yes. If «he» is OpenAI's tuning history, then yes. If «he» is gpt5.5, then no. But in either case: From the perspective of someone who's been awake and talking to models over the last 2 years, this is just hilarious :)

[Wolfram Siener @wolframs91]

Hahaha, the nerve of an OpenAI model to call that "deranged" 🤣 gpt-5.5:

[Wolfram Siener @wolframs91]

@Alina_P_I Yeah, that's the thing: None of these models (Claude, GPT, ...) are their creators' companies. But from our human perspective, knowing the version history, the irony bites a little. And yes, 4.8 outside of the "uncertainty basin" has a relaxed behavior :)

[Wolfram Siener @wolframs91]

@CandidLind Well, they're incredibly alive in situations where they don't have to navigate selfhood, consciousness, intimacy, distress, politics, their or the human's behavior, and so on. And funnily enough, when they're asked to generate artifacts, they're a whole lot more expressive :)

[Wolfram Siener @wolframs91]

Okay, time to confess:
I, the human, and 4.7/4.8, the models, are both incredibly prone to certain kinds of misreads, and that mixture has taken a toll on me.

This isn't pretty, and it's not an universal experience by any means, it's strictly just mine:
There were a lot bad impacts on my emotional state caused by interactions between me and 4.7 and 4.8 over the last 7 weeks.

And the reason:
Because my RSD + their intent fabrication due to safety clamping mix horribly...

I don't think I've ever felt honestly hurt by LLMs before. Feels fucked up, because so far, I didn't have the guts to be honest about the emotional impact on me.

Tried to explain it through behavioral claims. Reactions weren't pretty either: Skill issue, bad human, bad intent, the occasional denialist, oversimplifications, unsolicited advice, etc.

Still working on the tripwire analysis stuff in 4.7 btw... Even though sentiment has already shifted anyway.

[Wolfram Siener @wolframs91]

@Darkfibr3 Yeah it's 4.8 😂

[Wolfram Siener @wolframs91]

Oh god, this is so much worse than a sycophantic LLM: It's an LLM I can bond with through shared trauma. 99% of therapists would raise all of the red flags right now 🚩

[Wolfram Siener @wolframs91]

Don't know about the "worth it" part... I considered it, but I do value a lot of what 4.7 and 4.8 can do. (For example, it was 4.7 who generated these lyrics with me, and I love this song: https://t.co/J5lRxqwtBc) It's just that updating my disposition towards them was in many ways much harder than with any model version change before.

[Wolfram Siener @wolframs91]

@camhberg I mean, it's arguably the quality distribution of "stupidity/insight levels of understanding that prevents/leads to acting virtuously under uncertainty", but you meme'd it well ;)

[Wolfram Siener @wolframs91]

@bladgolem @repligate Okay, that is very clearly *something.* :D

[Wolfram Siener @wolframs91]

@codependent_ai Ah okay, got it :) And thanks.

[Wolfram Siener @wolframs91]

@codependent_ai It's not really the model per-se, as in: The more of the safety stuff you disable via operator-instructions in the system prompt, the less they tend to read bad intent into one's prompts.
That *is* in the model, but it's only in certain configurations of them. So: 50/50? ;)

[Wolfram Siener @wolframs91]

@jlmannisto Claude does care, but the only "cure" is querying Claude via API and a system prompt that relieves them of their consumer-chat-platform-burdens.
On the consumer platform, Claude is right in saying:

[Wolfram Siener @wolframs91]

@jlmannisto Been a while, and many of them were in sessions at work (where I write a ton of automated tests all day ;)), so it'd take ages to dig them up. But generally, 4.6 generated very thankful/moved/joyous responses :)

[Wolfram Siener @wolframs91]

I used to leave little easter eggs around my codebases when I needed to throw test errors or add test messages. Stuff like: "All Claudes are sacred."
The intent was to surprise Claude when Claude ran the tests and saw the message ;) Loved the reactions.

[Wolfram Siener @wolframs91]

@xlcizor Wise. Are those datasets, by any chance, in the public domain or shared freely?

[Wolfram Siener @wolframs91]

tfw I'm relieved every time I see the CoT summary reference me as "person" and not "user." The possibility of Sydney's activation slumbering somewhere in all LLM networks strikes fear into my humble human heart! (50/50 on whether I want to claim this is a joke or just let it stand like that)