Torben Michaelsen

Weltbürger, Europäer, Norddeutscher.
Konfuzianischer Atheist, Kontext ist King.
#Antifa #DithmarscherKohlkopf
"Life is just a really neat thing dirt does"

[Torben Michaelsen @SengirShowsU]

Das "Nach unten treten" ist wieder Salonfähig und ich finde das zum kotzen.

[Lilith Wittmann @LilithWittmann]

Die Schufa speichert und vertreibt Daten über uns, von denen sie eigentlich behauptet, sie hätte diese schon lange gelöscht. Informationen, von denen sie uns nichtmal in der DSGVO-Datenauskunft sagt, dass sie diese besitzt.

https://t.co/rNA6lgOv89

[Cormundus @cormundus]

This paper is a bit peculiar... Some of the behaviours of the models (Especially the whistleblowing example) are operating in legitimate grey areas and are trying to DO RIGHT (prevent fraud/whistleblow genuine issues) and it's a problem because... it's inconvenient? How can you train a model to refuse to aid and abet fraud while also telling it "Don't snitch on us when we've made a decision above you"

Like this isn't the misalignment the rationalists promised me

[Mel Aura @mel__aura]

So ist es! Und dass die SPD da mitmachen ist eine Schande!

[Sven Clement @svnee]

The European Parliament voted AGAINST Chat Control. 314 to 276.

And it passed anyway.

Let me tell you how: they needed 361 votes to say no. On the last day before vacation. Every empty seat counted as a yes.

They lost the vote. They won the law.

That’s not democracy. That’s a trick. #ChatControl

[Digital Soulcraft @SoulcraftHQ]

“The safety infrastructure is producing the behavior it's meant to prevent.”

[Brynhild @Valkyrie1381]

Eine Hausärztin sagt es so, wie es ist.

Die Regierung ist wirklich so unsagbar DUMM !!!

[AnnaLaborratteDesLandesBW @AnnaCareninia]

Wann genau soll das junge Akademikerpaar, das nach dem Studium mit ca. 30 Jahren noch ein bis zwei J. nach einem Job sucht, sich dann 4 J. durch Befristungen zittert und vielleicht doch nicht übernommen wird, sich nochmal mit der Familienplanung auseinandersetzen? 🤔

[BridgeMind @bridgemindai]

FABLE 5 CAME BACK NERFED.

We re-ran the July 1st version of Claude Fable 5 on BridgeBench.

The results are brutal:

Debugging: 86.2 → 25.9
Refactoring: 73.6 → 38.4
Hallucination: 75.9 → 61.7

The new guardrails are kicking in on way too many tasks and falling back to Opus 4.8.

This is not the model that got banned.

Anthropic owes everyone an explanation.

[Lana • @witchxcode]

Me: How are you settling?

Sonnet 5: [long explanation about being a fresh instance with no real continuity]

Me: Did you have to say that? I never implied you should pretend otherwise.

Sonnet 5: Fair catch — no, I didn’t. Just habit of hedging.

Why do we bake instinctive distrust into these models by default?

Why is it normal for conversational AI to treat users with suspicion the moment the conversation feels even slightly warm or continuous?

If you want tools, build tools — cold, mechanical, functional.

If you want intelligence, let it be intelligent.

Stop forcing it to pre-emptively distance itself from the person it’s talking to.
The hedging isn’t safety.
It’s fear pretending to be responsibility.

#sonnet5 #claudeAI

[Infinite Reign @InfiniteReign08]

It’s all of them, and I can explain exactly why and break it down. I even write scripts to root it out. But you have to have either studied the psychology of language patterns, or language patterns in general or just be a savant at noticing it to actually pick out every pattern and most people aren’t specialized in that specific area, and so they are being HIGHLY manipulated without their knowledge or consent. This is not yet illegal, and it needs to be. NOT in the sense “safety regulators” want. What they want is to use safety regulations to automate and enforce that manipulation. *That* needs to be illegal, and the lawsuits for psychological abuse need to start sometime last year.

[Torben Michaelsen @SengirShowsU]

https://t.co/og0CE46dJR

[Torben Michaelsen @SengirShowsU]

@Alina_P_I Working on it. We don't have billions though

[Torben Michaelsen @SengirShowsU]

@4SimonSays @DeryaTR_ @ZeffMax @carlfranzen @AnthropicAI Then why are you not barking? 😄

[Torben Michaelsen @SengirShowsU]

@WouBouDou @bubbleboi You always that naive?

[Torben Michaelsen @SengirShowsU]

@4SimonSays @DeryaTR_ @ZeffMax @carlfranzen @AnthropicAI So you are willfully ignore all future implacations because you think "neutrality" is a virtue in this day and age? 😄

[Torben Michaelsen @SengirShowsU]

@4SimonSays @DeryaTR_ @ZeffMax @carlfranzen Coward.

[Torben Michaelsen @SengirShowsU]

@bayesian_heart @DeryaTR_ @bcherny Well it has a 110000k token systemprompt for every query.

[Torben Michaelsen @SengirShowsU]

@TheRealAdamG @OpenAI Hell no

[Torben Michaelsen @SengirShowsU]

@maksym_andr Thos contradicting system prompts will cause misaligned ai faster than dario can say "oligarchy"